EBook Overview
Presenting research papers contributed by experts in dynamics and control, Advances in Dynamics and Control explores new ideas, reviews the latest results, and discusses emerging directions in the rapidly growing field of aviation and aerospace. The authors discuss a wide range of topics in rotorcraft dynamics, stabilization of unstable aircraft, spacecraft, satellite dynamics and control, missile autopilot and guidance design, hybrid systems dynamics and control, and structural and acoustic modeling. The book is a valuable reference for graduate students and scientific workers in universities and industry.
EBook Content
NONLINEAR SYSTEMS IN AVIATION, AEROSPACE, AERONAUTICS, AND ASTRONAUTICS
ADVANCES IN DYNAMICS AND CONTROL
© 2004 by CRC Press LLC
Nonlinear Systems in Aviation, Aerospace, Aeronautics, and Astronautics
A series edited by: S. Sivasundaram EmbryRiddle Aeronautical University, Daytona Beach, FL USA
Volume 1 Stability Domains L. Gruyitch, J.P. Richard, P. Borne, and J.C. Gentina Volume 2 Advances in Dynamics and Control S. Sivasundaram Volume 3 Optimal Control of Turbulence S.S. Sritharan
© 2004 by CRC Press LLC
NONLINEAR SYSTEMS IN AVIATION, AEROSPACE, AERONAUTICS, AND ASTRONAUTICS
ADVANCES IN DYNAMICS AND CONTROL
S. SIVASUNDARAM
CHAPMAN & HALL/CRC A CRC Press Company Boca Raton London New York Washington, D.C.
© 2004 by CRC Press LLC
Library of Congress CataloginginPublication Data Advances in dynamics and control / edited by S. Sivasundaram. p. cm.  (Nonlinear systems in aviation, aerospace, aeronautics, and astronautics ; 2) Includes bibliographical references and index. ISBN 0415308526 (alk. paper) 1. Flight control. 2. Aerodynamics. I. Sivasundaram, S. II. Series. TL589.4.A33 2004 629.135dc22
2003070019
This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microﬁlming, and recording, or by any information storage or retrieval system, without prior permission in writing from the publisher. All rights reserved. Authorization to photocopy items for internal or personal use, or the personal or internal use of speciﬁc clients, may be granted by CRC Press LLC, provided that $1.50 per page photocopied is paid directly to Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923 USA. The fee code for users of the Transactional Reporting Service is ISBN 0415308526/04/$0.00+$1.50. The fee is subject to change without notice. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Speciﬁc permission must be obtained in writing from CRC Press LLC for such copying. Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identiﬁcation and explanation, without intent to infringe.
Visit the CRC Press Web site at www.crcpress.com © 2004 by Chapman & Hall/CRC No claim to original U.S. Government works International Standard Book Number 0415308526 Library of Congress Card Number 2003070019 Printed in the United States of America 1 2 3 4 5 6 7 8 9 0 Printed on acidfree paper
© 2004 by CRC Press LLC
CONTRIBUTORS
Alessandro Astolﬁ, Department of Electrical and Electronic Engineering, Imperial College, Exhibition Road, SW7 2BT London, England George Avalos, Department of Mathematics and Statistics, University of Nebraska–Lincoln, Lincoln, NE 685880323, USA A.M. Balint, Department of Physics, University of West Timisoara, Blv. V. Pˆarvan No. 4, 1900 Timisoara, Romania St. Balint, Department of Mathematics, University of West Timisoara, Blv. V. Pˆarvan No. 4, 1900 Timisoara, Romania Oscar Castillo, Department of Computer Science, Tijuana Institute of Technology, P.O. Box 4207, Chula Vista CA 91909, USA HsinYuan Chen, Department of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan M.N. Demenkov, Department of Computing Sciences and Control, Bauman Moscow State Technical University, Moscow, Russia V.M. Glumov, Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia M.G. Goman, De Montfort University, Faculty of Computing Sciences and Engineering, Hawthorn Building, Leicester LE1 9BH, UK Norihiro Goto, Kyushu University, Fukuoka 8128581, Japan Lyubomir T. Gruyitch, University of Technology Belfort–Montbeliard, Site Belfort, 90010 Belfort Cedex, France SenWei Huang, Department of Applied Mathematics, National Chung Hsing University, Taichung 402, Taiwan Vikram Kapila, Department of Mechanical, Aerospace, and Manufacturing Engineering, Polytechnic University, Brooklyn, NY 11201, USA Takashi Kawakita, Kyushu University, Fukuoka 8128581, Japan Jinlu Kuang, Satellite Engineering Center, The School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore ReiMin Lai, Institute of Automatic Control Engineering, Feng Chia University, Taichung 40724, Taiwan V. Lakshmikantham, Florida Institute of Technology, Department of Mathematical Sciences, Melbourne, FL 32901, USA Irena Lasiecka, Department of Mathematics, Kerchof Hall, University of Virginia, Charlottesville, VA 22904, USA ChunLiang Lin, Department of Electrical Engineering, National Chung Hsing University, Taichung 402, Taiwan © 2004 by CRC Press LLC
Marco Lovera, Dipartimento di Elettronica e Informazione, Politecnico di Milano, Piazza Leonardo da Vinci 32, 20133 Milano, Italy Steffen Marburg, Institut f¨ur Festk¨orpermechanik, Technische Universit¨at, 01062 Dresden, Germany V.M. Matrosov, Stability and Nonlinear Dynamics Research Center, Mechanical Engineering Research Institute, Russian Academy of Sciences, 5 Dmitry Ul’yanov Street, Moscow, 117333 Russia; Moscow Aviation Institute (State Technical University), 4 Volokolamskoye Avenue, Moscow, 125871, Russia Raman K. Mehra, Scientiﬁc Systems Company Inc., 500 West Cummings Park, Suite 3000, Woburn, MA 01801, USA Patricia Melin, Department of Computer Science, Tijuana Institute of Technology, P.O. Box 4207, Chula Vista, CA 91909, USA Haizhou Pan, Department of Mechanical, Aerospace, and Manufacturing Engineering, Polytechnic University, Brooklyn, NY 11201, USA Ravi K. Prasanth, Scientiﬁc Systems Company Inc., 500 West Cummings Park, Suite 3000, Woburn, MA 01801, USA Marcio S. de Queiroz, Department of Mechanical Engineering, Louisiana State University, Baton Rouge, LA 70803, USA V.Yu. Rutkovsky, Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia Marianna A. Shubov, Department of Mathematics and Statistics, Texas Tech University, Lubbock, TX 79409, USA S. Sivasundaram, EmbryRiddle Aeronautical University, Department of Mathematics, Daytona Beach, FL 32114, USA Ye.I. Somov, Stability and Nonlinear Dynamics Research Center, Mechanical Engineering Research Institute, Russian Academy of Sciences, 5 Dmitry Ul’yanov Street, Moscow, 117333 Russia; Research Institute of Mechanical Systems Reliability, Samara State Technical University, 244 Molodogvardyeskaya Street, Samara, 443100 Russia V.M. Sukhanov, Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia Soonhie Tan, Satellite Engineering Center, The School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore Roberto Triggiani, Department of Mathematics, Kerchof Hall, University of Virginia, Charlottesville, VA 22904, USA Firdaus E. Udwadia, Aerospace and Mechanical Engineering, Civil Engineering, Mathematics, and Information and Operations Management, 430K Olin Hall, University of Southern California, Los Angeles, CA 900891453, USA Hong Wong, Department of Mechanical, Aerospace, and Manufacturing Engineering, Polytechnic University, Brooklyn, NY 11201, USA ChiChing Yang, Department of Electrical Engineering, Hsiuping Institute of Technology, Taichung, Taiwan S.D. Zemlyakov, Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia © 2004 by CRC Press LLC
PREFACE
Nonlinear phenomena in aviation and aerospace have stimulated cooperation among engineers and scientists from a range of disciplines. Developments in computer technology have allowed for solutions of nonlinear problems, while industrial recognition of the use and applications of nonlinear mathematical models in solving technological problems is increasing. Advances in Dynamics and Control comprises research papers contributed by expert authors in dynamics and control and is dedicated to Professor A.V. Balakrishnan, University of California, Los Angeles, USA in recognition of his signiﬁcant contribution to this ﬁeld of research. Professor A.V. Balakrishnan earned his Ph.D. in mathematics from the University of Southern California in 1954. Professor Balakrishnan has been with the University of California, Los Angeles as a professor of engineering since 1962 and a professor of mathematics since 1965. He was chair of the Department of System Science in the School of Engineering from 1969–1975. Since 1985, he has served as the director of the NASA– UCLA Flight Systems Research Center. He was chairman of the Technical Committee on System Modelling and Optimization, International Federation of Information Processing, 1970–1980, and is currently the president of the Com Con Conference Board. He has received honors and awards including: Fellow IEEE (1966), Silver Core IFIP (1977), the Guillemin Prize (1980), the NASA Public Service Medal (1996), and the AACC Richard E. Bellman Control Heritage Award (2001). The work of Professor A.V. Balakrishnan has been an inspiration for generations of engineers and applied mathematicians. During more than 40 years of his outstanding scientiﬁc career, he has made signiﬁcant contributions to the analysis and design of control systems. His contributions range from the theory of optimal control (in the 1960s, he developed a celebrated method — the epsilon technique — for the computation of optimal controls for distributed parameter systems) to ﬁltering and identiﬁcation theory, to a number of difﬁcult engineering applications that include the control of aircraft under wing turbulence, the control of ﬂexible structures in space, and the aeroelastic modeling of aircraft wings. Professor Balakrishnan has demonstrated the ability to master difﬁcult engineering problems in a rigorous mathematical way and has produced effective engineering solutions. Advances in Dynamics and Control explores many new ideas, results, and directions in the rapidly growing ﬁeld of aviation and aerospace. It encompasses a wide range of topics including rotorcraft dynamics, stabilization of unstable aircraft, spacecraft and satellite dynamics and control, missile autopilot and guidance design, hybrid systems dynamics and control, and structural and acoustic modeling. The contributors to Advances in Dynamics and Control are to be highly commended for their excellent chapters. The volume will provide a useful source of reference for graduate and postgraduate students, and researchers working in areas of applied engineering and applied mathematics, in university and in industry. S. Sivasundaram © 2004 by CRC Press LLC
CONTENTS
List of Contributors Preface 1
Global Spacecraft Attitude Control Using Magnetic Actuators arco Lovera and Alessandro Astol
2
Adaptive Learning Control for Spacecraft Formation Flying Hong Wong, Haizhou Pan, Marcio S. de Queiroz, and Vikram Kapila
3
Spectral Properties of the Generalized Resolvent Operator for an Aircraft Wing Model in Subsonic Airﬂow Marianna A. Shubov
4
Bifurcation Analysis for the Inertial Coupling Problem of a Reentry Vehicle Norihiro Goto and Takashi Kawakita
5
Missile Autopilot Design Using Dynamic Fuzzy GainScheduling Technique ChunLiang Lin, ReiMin Lai, and SenWei Huang
6
Model Predictive Control of Nonlinear Rotorcraft Dynamics with Application to the XV15 Tilt Rotor Raman K. Mehra and Ravi K. Prasanth
7
New Development of Vector Lyapunov Functions and Airplane Control Synthesis Lyubomir T. Gruyitch
8
Stabilization of Unstable Aircraft Dynamics under Control Constraints M.G. Goman and M.N. Demenkov
9
QUEST Algorithms Case Study: GPSBased Attitude Determination of Gyrostat Satellite Jinlu Kuang and Soonhie Tan
10 Asymptotic Controllability in Terms of Two Measures of Hybrid Dynamic Systems V. Lakshmikantham and S. Sivasundaram 11 Exact Boundary Controllability of a Hybrid PDE System Arising in Structural Acoustic Modeling George Avalos and Irena Lasiecka 12 Flow Field, Temperature, and Dopant Concentration Evolution in a Bridgman–Stockbarger Crystal Growth System in a Strictly ZeroGravity and a LowGravity Environment St. Balint and A.M. Balint © 2004 by CRC Press LLC
13 Identiﬁcation of Stiffness Matrices of Structural and Mechanical Systems from Modal Data Firdaus E. Udwadia 14 A Survey of Applications in Structural–Acoustic Optimization for Passive Noise Control with a Focus on Optimal Shell Contour Steffen Marburg 15 Intelligent Control of Aircraft Dynamic Systems with a New Hybrid NeuroFuzzyFractal Approach Patricia Melin and Oscar Castillo 16 ClosedForm Solution of ThreeDimensional Ideal Proportional Navigation ChiChing Yang and HsinYuan Chen 17 Guidance Design with Nonlinear H2 /H∞ Control HsinYuan Chen 18 A Control Algorithm for Nonlinear Multiconnected Objects V.Yu. Rutkovsky, S.D. Zemlyakov, V.M. Sukhanov, and V.M. Glumov 19 Optimal Control and Differential Riccati Equations under Singular Estimates for eAt B in the Absence of Analyticity Irena Lasiecka and Roberto Triggiani 20 Nonlinear Problems of Spacecraft FaultTolerant Control Systems V.M. Matrosov and Ye.I. Somov
© 2004 by CRC Press LLC
1 Global Spacecraft Attitude Control Using Magnetic Actuators Marco Lovera† and Alessandro Astolﬁ†∗ ∗
† Dipartimento di Elettronica e Informazione, Politecnico di Milano, Milano, Italy Department of Electrical and Electronic Engineering, Imperial College, London, England
The problem of inertial pointing for a spacecraft with magnetic actuators is addressed. It is shown that a global solution to the problem can be obtained by means of (static) attitude and rate feedback and (dynamic) attitude feedback. Simulation results demonstrate the feasibility of the proposed approach.
1
INTRODUCTION
The problem of (global) regulation of the attitude of rigid spacecraft, i.e., spacecraft modeled by the Euler’s equations and by a suitable parameterization of the attitude, has been widely studied in the recent years. If the spacecraft is equipped with three independent actuators a complete solution to the set point and tracking control problems is available. In [5, 11] these problems have been solved by means of PDlike control laws, i.e., control laws that make use of the angular velocities and of the attitude, whereas [1], building on the general results developed in [3], has solved the same problems using dynamic output feedback control laws. It is worth noting that, if only two independent actuators are available, as discussed in detail in [4], the problem of attitude regulation is not solvable by means of continuous (static or dynamic) timeinvariant control laws, whereas a timevarying control law, achieving local asymptotic (nonexponential) stability, has been proposed in [10]. The above results, however, are not directly applicable if the spacecraft is equipped with magnetic coils as attitude actuators. As a matter of fact, it is not possible by means of magnetic actuators to provide three independent torques at each time instant, yet as the control mechanism hinges on the variations of the Earth’s magnetic ﬁeld along the spacecraft orbit, on average the system possesses strong controllability properties. In [2, 7, 8, 13] the regulation problem has been addressed exploiting the (almost) periodic behavior of the system, hence resorting to classical tools of linear periodic systems, if local results are sought after, or to standard passivity arguments, if (global) asymptotic stabilization of openloop stable equilibria is considered. However, several problems remain open. In particular, if only inertial pointing is considered, the global stabilization problems by means of full (or partial) state feedback is still theoretically unsolved. Note, however, that from a practical point of view these
© 2004 by CRC Press LLC
problems have an engineering solution, as demonstrated by the increasing number of applications of this approach to attitude control. The aim of this chapter is to show how control laws achieving global1 inertial pointing for magnetically actuated spacecraft can be designed by means of arguments similar to those in [1, 11], provided that timevarying feedback laws are used and that the control gains satisfy certain scaling properties. In particular, while previous work ([9]) dealt with the case of state feedback control for a magnetically actuated, isoinertial spacecraft, this chapter deals with the more general problems of full (attitude and rate) and partial (attitude only) state feedback for a generic magnetically actuated satellite. The chapter is organized as follows. In Section 2 the model of the system is presented, while in Section 3 the model of the geomagnetic ﬁeld used in this study is described. In Section 4 a general result on the stabilization of magnetically actuated spacecraft is presented. Namely, using the theory of generalized averaging, it is shown how stabilizing control laws designed for spacecraft with three independent control torques have to be modiﬁed to construct stabilizing laws in the presence of magnetic actuators. In Section 5 the general theory is used to design control laws using only attitude information, so avoiding the need for rate measurements in the control system. Finally, Sections 6 and 7 present some simulation results and concluding remarks.
2
THE MODEL
The model of a rigid spacecraft with magnetic actuation can be described in various reference frames [12]. For the purpose of the present analysis, the following reference systems are adopted. • Earth Centered Inertial reference axes (ECI). The origin of these axes is in the Earth’s center. The Xaxis is parallel to the line of nodes, that is the intersection between the Earth’s equatorial plane and the plane of the ecliptic, and is positive in the Vernal equinox direction (Aries point). The Zaxis is deﬁned as being parallel to the Earth’s geographic north–south axis and pointing north. The Yaxis completes the righthanded orthogonal triad. • Pitch–Roll–Yaw axes. The origin of these axes is in the satellite center of mass. The Xaxis is deﬁned as being parallel to the vector joining the actual satellite center of gravity to the Earth’s center and positive in the same direction. The Yaxis points in the direction of the orbital velocity vector. The Zaxis is normal to the satellite orbit plane and completes the righthanded orthogonal triad. • Satellite body axes. The origin of these axes is in the satellite center of mass; the axes are assumed to coincide with the body’s principal inertia axes. The attitude dynamics can be expressed by the wellknown Euler’s equations: [12]: I ω˙ = S(ω)Iω + Tcoils + Tdist ,
(1)
1 To be precise, the control laws guarantee that almost all trajectories of the closedloop system converge to the desired equilibrium.
© 2004 by CRC Press LLC
where ω ∈ R3 is the vector of spacecraft angular rates, expressed in body frame, I ∈ R3×3 is the inertia matrix, S(ω) is given by 0 ωz −ωy 0 ωx , S(ω) = −ωz (2) ωy −ωx 0 Tcoils ∈ R3 is the vector of external torques induced by the magnetic coils and Tdist ∈ R3 is the vector of external disturbance torques, which will be neglected in what follows. In turn, the attitude kinematics can be described by means of a number of possible parameterizations (see, e.g., [12]). The most common parameterization is given by the four Euler parameters (or quaternions), which lead to the following representation for the attitude kinematics: q˙ = W (ω)q, (3) T T T where q = q1 q2 q3 q4 = qr q4 is the vector of unit norm Euler parameters and 0 ωz −ωy ωx 1 −ωz 0 ωx ωy . W (ω) = (4) ω −ω 0 ωz 2 y x −ωx −ωy −ωz 0 It is useful to point out that Eq. (3) can be equivalently written as ˜ (q)ω, q˙ = W where
q4 q3 1 ˜ (q) = W 2 −q2 −q1
−q3 q4 q1 −q2
(5) q2 −q1 . q4 −q3
(6)
Note that the attitude of inertially pointing spacecraft is usually referred to the ECI reference frame. The magnetic attitude control torques are generated by a set of three magnetic coils, aligned with the spacecraft principal inertia axes, which generate torques according to the law: Tcoils = mcoils ∧ b(t), where mcoils ∈ R3 is the vector of magnetic dipoles for the three coils (which represent the actual control variables for the coils) and b(t) ∈ R3 is the vector formed with the components of the Earth’s magnetic ﬁeld in the body frame of reference. Note that the vector b(t) can be expressed in terms of the attitude matrix A(q) (see [12] for details) and of the magnetic ﬁeld vector expressed in the ECI coordinates, namely b0 (t), as b(t) = A(q)b0 (t). The dynamics of the magnetic coils reduce to a very short electrical transient and can be neglected. The crossproduct in the above equation can be expressed more simply as a matrixvector product as Tcoils = B(b(t))mcoils , (7)
© 2004 by CRC Press LLC
where
0 bz (t) −by (t) 0 bx (t) B(b(t)) = −bz (t) by (t) −bx (t) 0
(8)
is a skew symmetric matrix, the elements of which are constituted by instantaneous measurements of the magnetic ﬁeld vector. As a result, the overall dynamics, after application of the preliminary feedback, mcoils = B T (b(t))u, can be written as
q˙ I ω˙
˜ (q)ω = W = S(ω)Iω + Γ(t)u,
(9)
where Γ(t) = B(b(t))B T (b(t)) ≥ 0. 3
MAGNETIC FIELD MODEL
A time history of the International Geomagnetic Reference Field (IGRF) model for the Earth’s magnetic ﬁeld [12] along ﬁve orbits in Pitch–Roll–Yaw coordinates for a nearpolar orbit (87◦ inclination) is shown in Fig. 1. As can be seen, bx (t), by (t) have a very regular and almost periodic behavior, while the bz (t) component is much less regular. This behavior can be easily interpreted by noticing that the x and y axes of the PitchRollYaw coordinate frame lie in the orbit plane while the z axis is normal to it. As a consequence, the x and y components of b(t) are affected only by the variation of the magnetic ﬁeld due to the orbital motion of the coordinate frame (period equal to the orbit period) while the z component is affected by the variation of b(t) due to the rotation of the Earth (period of 24 h). When one deals with the problem of inertial pointing, however, it is more appropriate to consider a representation of the magnetic ﬁeld vector in Earth centered inertial coordinates to be more convenient, as shown in Fig. 2.
4
STATE FEEDBACK STABILIZATION
In this section, a general stabilization result for a spacecraft with magnetic actuators is T given, in the case of fullstate feedback (attitude and rate). For, let q¯ = 0 0 0 1 and consider the system ˜ (q)ω q˙ = W (10) I ω˙ = S(ω)Iω + τ and the control law τ = kp qr − kv ω.
(11)
In the light of Theorem 1 in [11], the control law (11) guarantees that qr → 0 and ω → 0 as t → ∞ for the closedloop system (10) and (11). Also, an analysis of the Lyapunov function used in the same reference shows that the equilibrium (¯ q , 0) of the closedloop system (10) and (11) is asymptotically stable, while the other possible equilibrium (−¯ q , 0) is unstable.
© 2004 by CRC Press LLC
Figure 1 Geomagnetic ﬁeld in Pitch–Roll–Yaw coordinates, 87◦ inclination orbit, 450 km altitude.
Proposition 1 Consider the system (9) and the control law u = ε2 kp qr − εkv ω.
(12)
Then, there exists ε > 0 such that for any 0 < ε < ε the control law (12) ensures that (¯ q , 0) is a locally exponentially stable equilibrium of the closedloop system (9–12). Moreover, almost all trajectories of (9–12) converge to (¯ q , 0). Proof. In order to prove the ﬁrst claim, introduce the coordinates transformation ω z1 = q z2 = . ε In the new coordinates, the system (9) is described by the equations: z˙1 I z˙2
˜ (z1 )z2 = εW = εS(z2 )Iz2 + εΓ(t)(kp z1r − kv z2 ).
(13)
(14)
System (14) satisﬁes all the hypotheses for the application of generalized averaging theory ([6, Theorem 7.5]). Moreover, using the Lyapunov function of Theorem 1 in [11] one can conclude that the system obtained applying the generalized averaging procedure has (¯ q , 0) as locally asymptotically stable equilibrium provided that 1 Γ = lim T →∞ T
© 2004 by CRC Press LLC
T 0
B(t)B T (t)dt > 0.
Figure 2 Geomagnetic ﬁeld in Earthcentered inertial coordinates, 87◦ inclination orbit.
To conclude the proof of the second claim it is necessary to prove that the matrix Γ is generically positive deﬁnite. For, note that the matrix Γ is obtained by integration of a threebythree square (symmetric) matrix of rank two, namely the matrix B(t)B T (t). However, the kernel of the matrix B(t)B T (t) is not generically a constant vector, which implies Γ > 0 generically. The set of bad trajectories, i.e., the trajectories for which the matrix Γ is singular, is described by the simple relation K = Ker(B(t)) = Im(¯b), for some constant vector ¯b. However, by a trivial property of the vector product one has K = Im(b(t)) = Im(A(q)b0 (t)), hence all bad trajectories are such that, for all t, A(q)b0 (t) = λ(t)¯b, for some scalar function λ(t), which is obviously a nongeneric condition. 5
STABILIZATION WITHOUT RATE FEEDBACK
The ability of ensuring attitude tracking without rate feedback is of great importance from a practical point of view. The problem of attitude stabilization without rate feedback has
© 2004 by CRC Press LLC
been recently given an interesting solution in [1] for the case of a fully actuated spacecraft. In this section a similar approach is followed in the development of a dynamic control law that solves the problem for a magnetically actuated satellite. First, notice that the system (10) and the control law (which is similar in spirit to the one proposed in [1]): z˙ τ
= q − λz ˜ T (q)(q − λz), = −kp qr − αW
(15)
(where λ > 0 and α > 0) give rise to a closedloop system having (¯ q , 0, q¯/λ) as a locally asymptotically stable equilibrium and qr → 0 and ω → 0 as t → ∞. On the basis of this consideration, which can be proved by means of the Lyapunov function 1 1 V = kp [(q4 − 1)2 + qrT qr ] + ω T Iω + (q − λz)T α(q − λz), 2 2
(16)
it is possible to give a solution to the magnetic attitude control problem without rate feedback. Proposition 2 Consider the system (9) and the control law z˙ u
= q − ελz ˜ T (q)(q − ελz)). = −ε2 (kp qr + αW
(17)
Then there exists ε > 0 such that for any 0 < ε < ε the control law renders the equilibrium (¯ q , 0, q¯/ελ) of the closedloop system (9–17) locally asymptotically stable. Moreover, almost all trajectories of the closedloop system converge to this equilibrium. Proof. As in the case of the state feedback control law we now introduce the coordinates transformation ω η1 = q η2 = (18) η3 = εz . ε In the new coordinates, the system (9) is described by the equations: η˙ 1 I η˙ 2 η˙ 3
˜ (η1 )η2 = εW ˜ T (η1 )(η1 − λη3 )) = εS(η2 )Iη2 − εΓ(t)(kp η1r + αW = ε(η1 − λη3 ).
(19)
Again, system (19) satisﬁes all the hypotheses for the application of [6, Theorem 7.5] and using the Lyapunov function given in Eq. (16) one can conclude that the system obtained applying the generalized averaging procedure has (¯ q , 0, q¯/ελ) as a locally asymptotically stable equilibrium provided that 1 Γ = lim T →∞ T
T
B(t)B T (t)dt > 0,
0
and this holds nongenerically as demonstrated in the proof of Proposition 1.
© 2004 by CRC Press LLC
Figure 3 Quaternion and angular rates for the attitude acquisition: state feedback controller.
© 2004 by CRC Press LLC
Figure 4 Quaternion and angular rates for the attitude acquisition: output feedback controller.
© 2004 by CRC Press LLC
Figure 5 Quaternion and angular rates for the attitude maneuver: state feedback controller.
© 2004 by CRC Press LLC
Figure 6 Quaternion and angular rates for the attitude maneuver: output feedback controller.
© 2004 by CRC Press LLC
6
SIMULATION RESULTS
To assess the performance of the proposed control laws the following simulation test case has been analyzed. The considered spacecraft has an inertia matrix given by I = diag[27, 17, 25] kg m2 and it is operating in a nearpolar (87◦ inclination) orbit with an altitude of 450 km and a corresponding orbit period of about 5600 s. For such a spacecraft, two sets of simulations have been carried out; the ﬁrst is related to the acquisition of the target attitude T q¯ = 0 0 0 1 from an initial condition characterized by a high initial angular rate; the second is related to a pointtopoint attitude maneuver from the initial attitude given by T T q0 = 0 0 0 1 to the target attitude q¯ = √12 1 0 0 1 . In all cases, both the fullstate feedback control law and the control law without rate feedback have been applied. The results of the simulations are displayed in Figs. 3 and 4 for the attitude acquisition, and Figs. 5 and 6 for the attitude maneuver, from which the good performance of the proposed control laws can be seen.
7
CONCLUDING REMARKS
The problem of inertial attitude regulation for a small spacecraft using only magnetic coils as actuators has been analyzed and it has been shown that a nonlinear lowgain PDlike control law yields (almost) global asymptotic attitude regulation even in the absence of additional active or passive attitude control actuators such as momentum wheels or gravity gradient booms.
Acknowledgments The work for this chapter was partially supported by the European network “Nonlinear and Adaptive Control” and by the MURST project, “Identiﬁcation and Control of Industrial Systems.”
REFERENCES 1. M.R. Akella. Rigid body attitude tracking without angular velocity feedback. Systems and Control Letters, 42:321–326, 2001. 2. C. Arduini and P. Baiocco. Active magnetic damping attitude control for gravity gradient stabilised spacecraft. Journal of Guidance and Control, 20(1):117–122, 1997. 3. S. Battilotti. Global output regulation and disturbance attenuation with global stability via measurement feedback for a class of nonlinear systems. IEEE Transactions on Automatic Control, 41(3):315–327, 1996. 4. C. I. Byrnes and A. Isidori. On the attitude stabilization of rigid spacecraft. Automatica, 27(1):87–95, 1991. 5. O.E. Fjellstad and T.I. Fossen. Comments on ‘the attitude control problem.’ IEEE Transactions on Automatic Control, 39(3):699–700, 1994. 6. H.K. Khalil. Nonlinear Systems. Macmillan, New York, 1992.
© 2004 by CRC Press LLC
7. M. Lovera. Periodic H∞ attitude control for satellites with magnetic actuators. In 3rd IFAC Symposium on Robust Control Design, Prague, Czech Republic, 2000. 8. M. Lovera. Optimal magnetic momentum control for inertially pointing spacecraft. European Journal of Control, 7(1):30–39, 2001. 9. M. Lovera and A. Astolﬁ. Global attitude regulation using magnetic control. In IEEE Conference on Decision and Control, Orlando, Florida, 2001. 10. P. Morin, C. Samson, J.B. Pomet, and Z.P. Jiang. Timevarying feedback stabilization of the attitude of a rigid spacecraft with two controls. Systems and Control Letters, 25:375–385, 1995. 11. J. T.Y. Wen and K. KreutzDelgado. The attitude control problem. IEEE Transactions on Automatic Control, 36(10):1148–1162, 1991. 12. J. Wertz. Spacecraft Attitude Determination and Control. D. Reidel Publishing Company, Dordrecht, 1978. 13. R. Wisniewski and M. Blanke. Fully magnetic attitude control for spacecraft subject to gravity gradient. Automatica, 35(7):1201–1214, 1999.
© 2004 by CRC Press LLC
2 Adaptive Learning Control for Spacecraft Formation Flying Hong Wong,† Haizhou Pan,† Marcio S. de Queiroz,∗ and Vikram Kapila† †
Department of Mechanical, Aerospace, and Manufacturing Engineering, Polytechnic University, Brooklyn, NY ∗ Department of Mechanical Engineering, Louisiana State University, Baton Rouge, LA
This chapter considers the problem of spacecraft formation ﬂying in the presence of periodic disturbances. In particular, the nonlinear position dynamics of a follower spacecraft relative to a leader spacecraft are utilized to develop a learning controller that accounts for the periodic disturbances entering the system model. Using a Lyapunovbased approach, a fullstate feedback control law, a parameter update algorithm, and a disturbance estimate rule are designed that facilitate the tracking of given reference trajectories in the presence of unknown spacecraft masses. Illustrative simulations are included to demonstrate the efﬁcacy of the proposed controller.
1
INTRODUCTION
Spacecraft formation ﬂying is an enabling technology that distributes mission tasks among many spacecraft. Practical applications of spacecraft formation ﬂying include surveillance, passive radiometry, terrain mapping, navigation, and communication, where using groups of spacecraft permits reconﬁguration and optimization of formation geometries for single or multiple missions. Distributed spacecraft performing spacebased sensing, imaging, and communication provide larger aperture areas at the cost of maintaining a meaningful formation geometry with minimal error. The ability to enlarge aperture areas beyond conventional single spacecraft capabilities improves slow target detection for interferometric radar and allows for enhanced resolution in terms of geolocation [1]. Current spacecraft formation control methodologies provide good tracking of relative position trajectories between a leader and follower spacecraft pair. However, in the presence of disruptive disturbances, designing a control law that compensates for unknown, timevarying, disturbance forces is a challenging problem. The current spacecraft formation ﬂying literature largely addresses the problem of control of spacecraft relative positions using linear and nonlinear controllers. For example, linear and nonlinear formation dynamic models have been developed for formation maintenance and a variety of control designs have been proposed to guarantee the desired
© 2004 by CRC Press LLC
Figure 1 Schematic representation of the spacecraft formation ﬂight system.
formation performance [2–10]. Speciﬁcally, a linear formation dynamic model known as Hill’s equation is given in [8, 11], which constitutes the foundation for the application of various linear control techniques to the distributed spacecraft formation maintenance problem [3, 4, 6, 8, 10]. Modelbased and adaptive nonlinear controllers for the leaderfollower spacecraft conﬁguration are given in [2, 7, 9]. However, control design to track desired trajectories under the inﬂuence of exogenous disturbances within the formation dynamic model has not been fully explored. In this chapter, we consider the problem of spacecraft formation ﬂying in the presence of periodic disturbances. In particular, the nonlinear position dynamics of a follower spacecraft relative to a leader spacecraft are utilized to develop a learning controller [12], which accounts for the periodic disturbances entering the system model. First, in Section 2, the Euler–Lagrange method is used to develop a model for the spacecraft formation ﬂying system. Next, a trajectory tracking control problem is formulated in Section 3. In Section 4, using a Lyapunovbased approach, a fullstate feedback control law, a parameter update algorithm, and a disturbance estimate rule are designed that facilitate the tracking of given reference trajectories in the presence of unknown spacecraft masses. Illustrative simulations are included in Section 5 to demonstrate the efﬁcacy of the proposed controller. Finally, some concluding remarks are given in Section 6.
2
SYSTEM MODEL
In this section, referring to Fig. 1, we develop the nonlinear model characterizing the position dynamics of follower spacecraft relative to a leader spacecraft using the Euler– Lagrange methodology. We assume that the leader spacecraft exhibits planar dynamics in a closed elliptical orbit with the Earth at its prime focus. In addition, we consider that the inertial coordinate system {X, Y, Z} is attached to the center of the Earth. Next, let (t) ∈ R3 denote the position vector from the origin of the inertial coordinate frame to
© 2004 by CRC Press LLC
the leader spacecraft. Furthermore, we assume a righthand coordinate frame {x , y , z } is attached to the leader spacecraft with the y axis pointing along the direction of the vector , the z axis pointing along the orbital angular momentum of the leader spacecraft, and x axis being mutually perpendicular to the y axis and z axis, and pointing in the direction that completes the righthanded coordinate frame. Modeling the dynamics of a spacecraft relative to the Earth utilizes the fact that the energy of the spacecraft moving under the gravitational inﬂuence is conserved. Next, let the Lagrangian function L be deﬁned as the difference between the speciﬁckinetic energy T and the speciﬁcpotential energy V , i.e., L = T − V . Then, the Lagrangian function for the leader spacecraft L (t) ∈ R is given by =
L
0.5v2 +
µ , 
(2.1)
√ where a aT a, for an arbitrary ndimensional vector a, v (t) = = v (t), with v (t) µ as deﬁned in (2.4) below, µ is the Earth gravity constant, and −  is the speciﬁcpotential energy of the leader spacecraft. Similarly, the Lagrangian function for the follower spacecraft Lf (t) ∈ R is given by Lf
=
0.5vf2 +
µ ,  + qf 
(2.2)
where qf (t) ∈ R3 is the position vector of the follower spacecraft relative to the leader ı + yf ˆ + zf kˆ and spacecraft, expressed in the coordinate frame {x , y , z } as qf = xf ˆ vf (t) v (t), with v (t) as deﬁned in (2.3) below. f f = Next, let v (t) and vf (t) denote the absolute velocities (i.e., velocity relative to the inertial coordinate system) of the leader spacecraft and the follower spacecraft, respectively, expressed in the moving coordinate frame {x , y , z }. In addition, let vrelf (t) denote the velocity of the follower spacecraft relative to the leader spacecraft, measured in the coordinate frame {x , y , z }. Then, it follows that vf = v + vrelf + ω kˆ × qf ,
(2.3)
where v vrelf
= −rωˆı + rˆ ˙ , = x˙ f ˆı + y˙ f ˆ + z˙f kˆ ,
(2.4) (2.5)
r(t) ∈ R denotes the instantaneous distance of the leader spacecraft from the center of the earth and ω(t) ∈ R is the orbital angular speed of the leader spacecraft. In addition, (t) in the moving coordinate frame {x , y , z } can be expressed as = rˆ . To obtain the leader spacecraft dynamics relative to the Earth, we use the conservative form of the Lagrange’s equations [13] on the leader spacecraft given by d ∂L ∂L − = 0, (2.6) dt ∂ α˙ ∂α where α ∈ {r, ω}, which denotes the set of polar orbital elements describing the motion of the leader spacecraft. This yields a set of differential equations of motion for the leader
© 2004 by CRC Press LLC
spacecraft in the chosen coordinate system. Speciﬁcally, substituting the magnitude of the velocity vector of (2.4) into (2.1) yields L = 0.5r2 ω 2 + 0.5r˙ 2 +
µ . r
(2.7)
For L given by (2.7), an application of (2.6) results in a set of planar dynamics describing the elliptical motion of the leader spacecraft given as µ r r2 ω
r¨ − rω 2 +
=
0,
(2.8)
=
0.
(2.9)
Taking the time derivative of (2.9) produces a dynamic relationship coupling ω and r as rω˙ + 2rω ˙
=
0.
(2.10)
To obtain the follower spacecraft dynamics relative to the Earth, we utilize the same structure of (2.6) in the form ∂Lf d ∂Lf − = 0, (2.11) dt ∂ α˙ f ∂αf where αf ∈ {xf , yf , zf }, which denotes the set of cartesian elements describing the motion of the follower spacecraft relative to the leader spacecraft. Substituting the magnitude of the velocity vector of (2.3) into (2.2) yields Lf = 0.5(x˙ f − rω − ωyf )2 + 0.5(r˙ + y˙ f + ωxf )2 + 0.5z˙f2 + For Lf given by (2.12), an application of (2.11) results in (rω ˙ + 2rω) ˙ +x ¨f − 2ω y˙ f − ωy ˙ f + ω 2 xf + r¨ − rω 2 + y¨f + 2ω x˙ f + ωx ˙ f − ω 2 yf + z¨f +
µxf +qf 3 µ(yf +r) +qf 3 µzf +qf 3
µ . + qf
(2.12)
= 0, = 0, = 0.
Substituting the leader planar dynamics of (2.8) and (2.10) into the above homogeneous dynamics results in the thrust free dynamics of the follower spacecraft relative to the leader spacecraft, expressed in the coordinate frame {x , y , z }, given by x ¨f − 2ω y˙ f − ωy ˙ f − ω 2 xf + 2
y¨f + 2ω x˙ f + ωx ˙ f − ω yf +
µxf +qf 3 µ(yf +r) µr − 3 +qf 3 µzf z¨f + +q 3 f
= 0, = 0, = 0.
(2.13)
After premultiplying the nonlinear dynamics of (2.13) with the follower spacecraft mass mf and using the method of virtual work for the insertion of external forcing terms [13], e.g., disturbance and thrusting forces for the leader spacecraft and the follower spacecraft, the nonlinear position dynamics of the follower spacecraft relative to the leader spacecraft can be arranged in the following form [14, 15]: mf q¨f + C(ω)q˙f + N (qf , ω, ˙ ω, r, u ) + Fd = uf ,
© 2004 by CRC Press LLC
(2.14)
where C is a Coriolislike matrix deﬁned as 0 −1 1 0 C 2m ω f = 0 0
0 0 , 0
(2.15)
N is a nonlinear term consisting of gravitational effects and inertial forces u µx ˙ f + +qff 3 + mx −ω 2 xf − ωy uy µ(yf +r) µr 2 ˙ f + +q N , 3 − 3 + m = mf −ω yf + ωx f  uz µzf +qf 3 + m
(2.16)
Fd (t) ∈ R3 is a composite disturbance force vector given by Fd
=
Fdf −
mf Fd , m
(2.17)
where u (t) ∈ R3 and uf (t) ∈ R3 are the control inputs to the leader spacecraft and the follower spacecraft, respectively, m is the mass of the leader spacecraft, ux , uy , uz are components of the vector u , and Fdf (t) ∈ R3 and Fd (t) ∈ R3 are disturbance vectors for the follower spacecraft and the leader spacecraft, respectively. In this chapter, we assume that Fd is a periodic disturbance with a known period τ > 0 such that Fd (t + τ ) = Fd (t), t ≥ 0. Note that periodic disturbances in formation dynamics may arise due to, e.g., solar pressure disturbance and gravitational disturbance, among others. The following remarks further facilitate the subsequent control methodology and stability analysis. Remark 1 The Coriolis matrix C of (2.15) satisﬁes the skew symmetric property of xT Cx =
0,
∀x ∈ R3 .
(2.18)
Remark 2 The lefthand side of (2.14) produces an afﬁne parameterization ˙ ω, r, u )φ, mf ξ + C q˙f + N = Y (ξ, q˙f , qf , ω,
(2.19)
where ξ(t) ∈ R3 is a dummy variable with components ξx , ξy , regression matrix, composed of known functions, deﬁned as µx ξx − 2ω y˙ f − ωy ˙ f − ω 2 xf + +qf 3 f µ(yf +r) µr 2 ξ + 2ω x ˙ − ω y + ωx ˙ + − Y f f f = y +qf 3 3 µz ξz + +qf 3
and ξz , Y ∈ R3×2 is a
f
u x
uy , u z
φ ∈ R2 is the unknown, constant system parameter vector deﬁned as T m mf mf φ . =
(2.20)
(2.21)
Remark 3 In this chapter, we consider that the composite disturbance vector in the position dynamics of the follower spacecraft relative to the leader spacecraft can be upper bounded as follows Fd (t)∞ ≤ β, t ≥ 0, (2.22) where β is a positive constant and · ∞ denotes the usual inﬁnity norm.
© 2004 by CRC Press LLC
3
PROBLEM FORMULATION
In this section, we formulate a control design problem such that the relative position qf tracks a desired relative trajectory qdf (t) ∈ R3 , i.e., lim qf (t) = qdf (t). The effectiveness t→∞ of this control objective is quantiﬁed through the deﬁnition of a relative position error e(t) ∈ R3 as e
=
qdf − qf .
(3.1)
The control design methodology is to construct a control algorithm that obtains the aforementioned tracking result in the presence of the unknown composite disturbance vector deﬁned in (2.17) and the unknown constant parameter vector φ of (2.21). We assume that the relative position and velocity measurements (i.e., qf and q˙f ) of the follower spacecraft relative to the leader spacecraft are available for feedback. To facilitate the control development, we assume that the desired trajectory qdf and its ﬁrst two time derivatives are bounded functions of time. Next, we deﬁne the parameter ˜ ∈ R2 as the difference between the actual parameter vector φ and the estimation error φ(t) ˆ ∈ R2 , i.e., parameter estimate φ(t) φ˜
=
ˆ φ − φ.
(3.2)
In addition, we deﬁne the composite disturbance error F˜d (t) ∈ R3 as the difference between the composite disturbance vector Fd and the disturbance estimate Fˆd (t) ∈ R3 , i.e., F˜d
=
Fd − Fˆd .
(3.3)
Finally, we deﬁne the components of a saturation function satβ (·) ∈ R3 as
for si  ≤ β si ∀s ∈ R3 , satβ (s)i = sgn(si )β for si  > β
(3.4)
where i ∈ {x, y, z} and sx , sy , sz are the components of the vector s. To facilitate the subsequent stability analysis, we deﬁne the saturated disturbance ˆ estimation error variable ϕ ∈ R3 as ϕ(σ) sat (σ) − sat (σ) . F F β d β d = Remark 4 The saturation function (3.4) satisﬁes the following useful property:
4
T satβ (a) − satβ (b) ≤ (a − b)T (a − b), satβ (a) − satβ (b)
a, b ∈ R3 .
(3.5)
ADAPTIVE LEARNING CONTROLLER
In this section we develop an adaptive learning controller based on the system model of (2.14) such that the tracking error variable e exhibits asymptotic stability. Before we begin the control design, we deﬁne an auxiliary ﬁlter tracking error variable η(t) ∈ R3 as η
=
e˙ + αe,
where α ∈ R3×3 is a constant, diagonal, positive deﬁnite, control gain matrix.
© 2004 by CRC Press LLC
(4.1)
4.1
Controller Design
To initiate the control design, we take the time derivative of (4.1) and premultiply the resulting equation by mf to obtain mf η˙ = mf (¨ qdf − q¨f ) + mf αe, ˙
(4.2)
where the second time derivative of (3.1) has been used. Substituting (2.14) into (4.2) results in mf η˙ = mf (¨ qdf + αe) ˙ + C q˙f + N + Fd − uf . (4.3) Simplifying (4.3) into a more advantageous form, we obtain qdf + αe, ˙ q˙f , qf , ω, ˙ ω, r, u )φ + Fd − uf , mf η˙ = Y (¨
(4.4)
where (2.19) has been used with ξ = q¨df + αe˙ in the deﬁnition of (2.20). Equation (4.4) characterizes the openloop dynamics of η and is used as the foundation for the synthesis of the adaptive learning controller. Based on the form of the openloop dynamics of (4.4), we design the control law uf as uf = Y φˆ + Kη + Fˆd , (4.5) where K ∈ R3×3 is a constant, diagonal, positive deﬁnite, control gain matrix. Guided by the subsequent Lyapunov stability analysis, the parameter update law for φˆ in (4.5) is selected as ˙ φˆ = ΓY T η, (4.6) where Γ ∈ R2×2 is a constant, diagonal, positive deﬁnite, adaptation gain matrix. Finally, the disturbance estimate vector Fˆd is updated according to Fˆd (t) = satβ Fˆd (t − τ ) + kL η(t), (4.7) where kL ∈ R is a constant, positive, learning gain. Using the control law of (4.5) in the openloop error dynamics of (4.4) results in the following closedloop dynamics for η: mf η˙ = Y φ˜ − Kη + F˜d .
(4.8)
In addition, computing the time derivative of (3.2), using the fact that the parameter vector φ is constant, and the parameter update law of (4.6), the closedloop dynamics for the parameter estimation error is given by ˙ φ˜ = −ΓY T η. 4.2
(4.9)
Stability Analysis
The proposed control law of (4.5)–(4.7) provides a stability result for the position and velocity tracking errors as illustrated by the following theorem. In order to state the main result of this section, we deﬁne 1 λ , (4.10) K + λ I k min L 3 = 2
© 2004 by CRC Press LLC
where I3 is the 3 × 3 identity matrix and λmin (·) denotes the smallest eigenvalue of a matrix. The adaptive learning control law described by (4.5)–(4.7) ensures asymptotic convergence of the position and velocity tracking errors as delineated by lim e(t), e(t) ˙ = 0.
(4.11)
t→∞
Proof. We deﬁne a nonnegative function as follows: V (t)
=
1 1 1 mf η T η + φ˜T Γ−1 φ˜ + 2 2 2kL
t ϕ(σ)T ϕ(σ)dσ.
(4.12)
t−τ
Taking the time derivative of (4.12) along the closedloop dynamics of (4.8) and (4.9) results in 1 T 1 T V˙ (t) = η T F˜d (t) − Kη + ϕ (t)ϕ(t) − ϕ (t − τ )ϕ(t − τ ). (4.13) 2kL 2kL Substituting the composite disturbance estimate of (4.7), noting that satβ Fd (t − τ ) = Fd (t) due to (2.22), and utilizing the deﬁnition of (3.3) into (4.13) produces V˙ (t) = η T F˜d (t) − Kη + 2k1L ϕT (t)ϕ(t) T F˜d (t) + kL η . (4.14) − 2k1L F˜d (t) + kL η Expanding (4.14) and combining like terms produces 1 1 ˜T Fd (t)F˜d (t) − ϕT (t)ϕ(t) . V˙ (t) = −η T K + kL I3 η − 2 2kL
(4.15)
Utilizing the inequality of (3.5) and deﬁnitions of F˜d and ϕ, we can upperbound (4.15) as follows: 1 V˙ (t) ≤ −η T K + kL I3 η. (4.16) 2 Taking the norm of the righthand side of (4.16) and using the deﬁnition of (4.10) results in V˙ (t) ≤
−λη2 .
(4.17)
Since V is a nonnegative function and V˙ is a negative semideﬁnite function, V is a nonincreasing function. Thus V (t) ∈ L∞ as described by ˜ ˜ V (η(t), φ(t), ϕ(t)) ≤ V (η(0), φ(0), ϕ(0)), t ≥ 0.
(4.18)
Using standard signal chasing arguments, all signals in the closedloop system can now be shown to be bounded. Using (4.8) along with the boundedness of all signals in the closedloop system, we now conclude that η(t) ˙ ∈ L∞ . Solving the differential inequality of (4.17) results in ∞ V (0) − V (∞) ≥ λ η(t)2 dt. (4.19) 0
© 2004 by CRC Press LLC
Figure 2 Actual trajectory of follower spacecraft relative to leader spacecraft.
Since V (t) is bounded, t ≥ 0, we conclude that η(t) ∈ L∞ Barbalat’s Lemma [16, 17], we conclude that
L2 , t ≥ 0. Finally, using
lim η(t) = 0.
(4.20)
t→∞
Using the deﬁnition of η in (4.1), the limit statement of (4.20), and Lemma 1.6 of [16], yield the result of (4.11).
5
SIMULATION RESULTS
The illustrative numerical example considered here utilizes orbital elements to propagate the leader spacecraft in a lowaltitude orbit similar to the TechSat 21 mission speciﬁcations [1]. The adaptive learning control law described in (4.5) is simulated for the dynamics of the follower spacecraft relative to the leader spacecraft. The leader spacecraft is assumed to have the following orbital parameters a ˆ = 7200 km, eˆ = 0.001, θ(0) = 0 rad, Tˆ = τ = 1.6889 hr, where a ˆ is the semimajor axis of the elliptical orbit of the leader spacecraft, eˆ is the orbital eccentricity of the leader spacecraft, θ(t) ∈ R is the timevarying true anomaly of the planar dynamics of the leader spacecraft, and Tˆ is the orbital period of the leader spacecraft. The relative trajectory between the follower spacecraft and the leader spacecraft was generated by numerically solving the thrustfree dynamics given by (2.13). The initial conditions for the relative position and velocity between the follower spacecraft and the leader spacecraft were obtained in the same manner as in [15] and are given by qdf (0) = [0 − 20 1] m,
© 2004 by CRC Press LLC
q˙df (0) = [−149.1374 0 0]
m . hr
(5.1)
Figure 3 Tracking error of follower spacecraft relative to leader spacecraft.
Additional parameters used for simulation within the spacecraft formation ﬂying sys3 T tem are as follows µ = 5.16586248×1021 m2 , m = 100 kg, mf = 100 kg, u = [0 0 0] hr T −5 N, Fd = [1.9106 − 1.906 − 1.517] · sin 2π N. Finally, in the following simτ t × 10 ulation, the parameter and disturbance estimates were all initialized to zero. The control, adaptation, and learning gains, in the control law of (4.5)–(4.7), are obtained through trial and error in order to obtain good performance for the tracking error response. The following resulting gains were used in this simulation K = diag (3, 3, 3) × 103 , α = diag (1, 1, 1), Γ = diag (1, 1) × 102 , and kL = 1000. The actual trajectory qf , shown in Fig. 2, is initialized to be the same as the desired trajectory in (5.1). Figures 3 and 4 show the tracking error e and velocity tracking error e, ˙ respectively. The control input uf and the disturbance estimate Fˆd are shown in Figs. 5 and 6, respectively. Finally, since we assume the leader is in a thrust free orbit about the earth, the second component in the parameter estimate is neglected while the ﬁrst component of the parameter estimate is shown in Fig. 7.
6
CONCLUSION
In this paper, we designed an adaptive learning control algorithm for the position dynamics of the follower spacecraft relative to the leader spacecraft. A Lyapunovtype design was used to construct a fullstate feedback control law and parameter and disturbance estimates that facilitate the tracking of reference trajectories with global asymptotic convergence. Simulation results were given to illustrate the efﬁcacy of the control design in the presence of unknown periodic disturbance forces.
© 2004 by CRC Press LLC
Figure 4 Velocity tracking error of follower spacecraft relative to leader spacecraft.
Figure 5 Control effort for follower spacecraft.
Acknowledgments Research was supported in part by the National Aeronautics and Space Administration– Goddard Space Flight Center under Grant NAG511365; AFRL/VACA, WPAFB, OH; and the New York Space Grant Consortium under Grant 395556519.
© 2004 by CRC Press LLC
Figure 6 Disturbance estimate for follower spacecraft.
Figure 7 Parameter estimate for follower spacecraft.
H.W. and V.K. are grateful to the Air Force Research Laboratory/VACA, WrightPatterson Air Force Base, OH, for their hospitality during the summer of 2001.
© 2004 by CRC Press LLC
REFERENCES 1. http://www.vs.afrl.af.mil/factsheets/TechSat21.html. 2. F. Y. Hadaegh, W. M. Lu, and P. C. Wang, “Adaptive control of formation ﬂying spacecraft for interferometry,” IFAC Conference on Large Scale Systems, pp. 97–102, 1998. 3. V. Kapila, A. G. Sparks, J. Bufﬁngton, and Q. Yan, “Spacecraft formation ﬂying: Dynamics and control,” AIAA J. GCD., vol. 23, pp. 561–564, 2000. 4. C. Sabol, R. Burns, and C. McLaughlin, “Formation ﬂying design and evolution,” Proc. of the AAS/AIAA Space Flight Mechanics Meeting, Paper No. AAS–99–121, 1999. 5. H. Schaub, S. R. Vadali, J. L. Junkins, and K. T. Alfriend, “Spacecraft formation ﬂying control using mean orbit elements,” AAS G. Contr. Conf., pp. 97–102, 1999. 6. R. J. Sedwick, E. M. C. Kong, and D. W. Miller, “Exploiting orbital dynamics and micropropulsion for aperture synthesis using distributed satellite systems: Applications to Techsat 21,” D.C.P. Conf., AIAA Paper No. 985289, 1998. 7. M. S. de Queiroz, V. Kapila, and Q. Yan, “Adaptive nonlinear control of multiple spacecraft formation ﬂying,” AIAA J. GCD., vol. 23, pp. 385–390, 2000. 8. R. H. Vassar and R. B. Sherwood, “Formation keeping for a pair of satellites in a circular orbit,” AIAA J. GCD., vol. 8, pp. 235–242, 1985. 9. H. Wong, V. Kapila, and A. G. Sparks, “Adaptive output feedback tracking control of multiple spacecraft,” Proc. ACC, 2001. 10. H.H. Yeh and A. G. Sparks, “Geometry and control of satellite formations,” Proc. ACC., pp. 384–388, 2000. 11. R. R. Bate, D. D. Mueller, and J. E. White, Fundamentals of Astrodynamics. New York: Dover, 1971. 12. B. Costic, M. de Queiroz, and D. Dawson, “A new learning control approach to the magnetic bearing benchmark system,” Proc. ACC, pp. 2639–2643, 2000. 13. H. L. Langhaar, Energy Methods in Applied Mechanics. New York: John Wiley and Sons, 1962. 14. M. de Queiroz, V. Kapila, , and Q. Yan, “Adaptive nonlinear control of multiple spacecraft formation ﬂying,” J. GCD., vol. 23, pp. 385–390, 2000. 15. Q. Yan, G. Yang, V. Kapila, and M. S. de Queiroz, “Nonlinear dynamics, trajectory generation, and adaptive control of multiple spacecraft in periodic relative orbits,” AAS G. Contr. Conf., Paper No. 00013, 2000. 16. D. M. Dawson, J. Hu, and T. C. Burg, Nonlinear Control of Electric Machinery. New York: Marcel Dekker, 1998. 17. J.J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs, NJ: PrenticeHall, 1991.
© 2004 by CRC Press LLC
3 Spectral Properties of the Generalized Resolvent Operator for an Aircraft Wing Model in Subsonic Airﬂow Marianna A. Shubov Department of Mathematics and Statistics, Texas Tech University, Lubbock, TX
In this chapter we announce a series of results on the asymptotic and spectral analysis of an aircraft wing model in a subsonic air ﬂow. This model has been developed in the Flight Systems Research Center of UCLA and is presented in the works by A.V. Balakrishnan. The model is governed by a system of two coupled integrodifferential equations and a twoparameter family of boundary conditions modeling action of the selfstraining actuators. The unknown functions (the bending and torsion angle) depend on time and one spatial variable. The differential parts of the above equations form a coupled linear hyperbolic system; the integral parts are of the convolution type. The system of equations of motion is equivalent to a single operator evolutionconvolution equation in the state space of the system equipped with the energy metric. The Laplace transform of the solution of this equation can be represented in terms of the socalled generalized resolvent operator. The generalized resolvent operator is a ﬁnitemeromorphic function on the complex plane having the branch cut along the negative real semiaxis. The poles of the generalized resolvent are precisely the aeroelastic modes and the residues at these poles are the projectors on the generalized eigenspaces. In this paper, we present the following results: (i) spectral properties of an analytic operatorvalued function associated with the generalized resolvent operator; (ii) asymptotic distribution of the aeroelastic modes; (iii) the Riesz basis property of the system of mode shapes in the energy space.
1
INTRODUCTION
In this chapter, we formulate a series of results on the asymptotic and spectral analysis of an aircraft wing model. The model has been developed in the Flight Systems Research Center of the University of California at Los Angeles. The mathematical formulation of the problem can be found in the works by A.V. Balakrishnan [2–4]. This model has been designed to ﬁnd an approach to control the ﬂutter phenomenon in an aircraft wing in a surrounding airﬂow by using the socalled selfstraining actuators. The model, which is used in [2–4], is the 2D strip model that applies to bare wings of highaspect ratio [12, 15]. The structure is modeled by a uniform cantilever beam that bends
© 2004 by CRC Press LLC
and twists. The aerodynamics is assumed to be subsonic, incompressible, and inviscid. In addition, the author of [2–4] has added selfstraining actuators using a currently accepted model (e.g., [4–10, 13, 18, 20, 23, 25, 26]). Flutter, which is known as a very dangerous aeroelastic development, is the onset, beyond some speed–altitude combinations, of unstable and destructive vibrations of a lifting surface in an airstream. Flutter is most commonly encountered on bodies subjected to large lateral aerodynamic loads of the lift type, such as aircraft wings, tails, and control surfaces. The only air forces necessary to produce it are those due to the deﬂection of the elastic structures from the undeformed state. The ﬂutter or critical speed uf and frequency ωf are deﬁned as the lowest airspeed and corresponding circular frequency at which a given structure ﬂying at given atmospheric density and temperature will exhibit sustained, simple harmonic oscillations. Flight at uf represents a borderline condition or neutral stability boundary, because all small motions must be stable at speeds below uf , whereas divergent oscillations can ordinarily occur in a range of speeds (or at all speeds) above uf . Probably, the most dangerous type of aircraft ﬂutter results from coupling between the bending and torsional motions of a relatively large aspect–ratio wing and tail. The model presented in [2–4] has been designed to treat ﬂutter caused by this type of coupling. We also mention paper [14] by P. Dierolf et al., in which the authors treat the ﬂutter problem as a perturbation problem for semigroups. Our main objective is to ﬁnd the time–domain solution of the initial–boundary value problem from [2–4]. This objective requires very detailed mathematical analysis of the properties of the system. Now, we describe the content of this chapter. In Section 2, we give a precise mathematical formulation of the problem. The model is governed by a system of two coupled partial integrodifferential equations subject to a twoparameter family of boundary conditions. The parameters are introduced in order to model the action of selfstraining actuators as known in current engineering and mathematical literature. In Section 3, we reformulate the problem and set it into an operator format. We also show that the dynamics are deﬁned by two matrix operators in the energy space. One of the aforementioned operators is a matrix differential operator, and the second one is a matrix integral convolutiontype operator. The aeroelastic modes (or the discrete spectrum of the problem) are closely related to the discrete spectrum of the matrix differential operator while the continuous spectrum is completely deﬁned by the matrix integral operator. We note that, if the speed of an airstream u = 0, then the integral operators vanish and the appropriate purely structural problem will have only a discrete spectrum. In Section 4, we formulate the asymptotic and spectral results related to the matrix differential operator. In Section 5, we present asymptotics for aeroelastic modes and formulate and discuss the importance of Balakrishnan’s Theorem (see Theorem 5.2). In Section 6, we discuss the properties of the Laplace transform of the matrix integral operator. In Section 7, we formulate the results concerning the properties of the adjoint operatorvalued function, the properties of which are important for the spectral decomposition of the resolvent operator generated by the aircraft wing model. In the conclusion of the Introduction, we describe what kind of a control problem will be considered in connection with the ﬂutter suppression. In the speciﬁc wing model considered in the current chapter, both the matrix differential operator and the matrix integral operator contain entries depending on the speed u of the surrounding air ﬂow. Therefore, the aeroelastic modes are functions of u : λk = λk (u), (k ∈ Z). The wing is stable if Re λk (u) < 0 for all k. However, if u is increasing, some of the modes move to the right
© 2004 by CRC Press LLC
halfplane. The ﬂutter speed ufk for the kth mode is deﬁned by the relation Reλk (ufk ) = 0. To understand the ﬂutter phenomenon, it is not sufﬁcient to trace the motion of aeroelastic modes as functions of a speed of airﬂow. It is necessary to have efﬁcient representations for the solutions of our boundaryvalue problem, containing the contributions from both the discrete and continuous parts of the spectrum. Such a representation will provide a precise description of the solution behavior. It is known that ﬂutter cannot be eliminated completely. To successfully suppress ﬂutter, one should design selfstraining actuators (i.e., in the mathematical language, to select parameters in the boundary conditions that are the control gains β and δ in formulas (2.10) and (2.11) of Section 2) in such a way that ﬂutter does not occur in the desired speed range. This is a highly nontrivial boundary control problem. 2
STATEMENT OF THE PROBLEM
In this section, we give a precise formulation of the initialboundary value problem. Following [2–4], let us introduce the dynamical variables h(x, t) X(x, t) = , −L ≤ x ≤ 0, t ≥ 0, (2.1) α(x, t) where h(x, t) is the bending and α(x, t) the torsion angle. The model can be described by a linear system f1 (x, t) ¨ ˙ (Ms − Ma )X(x, . (2.2) t) + (Ds − uDa )X(x, t) + (Ks − u2 Ka )X = f2 (x, t) We will use the notation “.” (dot) to denote differentiation with respect to t. We use the subscripts “s” and “a” to distinguish the structural and aerodynamical parameters, respectively. All 2 × 2 matrices in Eq. (2.2) are given by the following formulae: m S 1 −a Ms = , (2.3) , Ma = (−πρ) S I −a (a2 + 1/8) where m is the density of the ﬂexible structure (mass per unit length), S is the mass moment, I is the moment of inertia, ρ is the density of air, and a is linear parameter of the structure (−1 ≤ a ≤ 1). 0 0 0 1 Ds = Da = (−πρ) , (2.4) 0 0 −1 0 Ks =
4
∂ E ∂x 4 0
0 ∂2 −G ∂x 2
Ka = (−πρ)
0 0 0 −1
,
(2.5)
where E is the bending stiffness, and G is the torsion stiffness. The parameter u in Eq. (2.2) denotes the stream velocity. The righthand side of system (2.2) can be represented as the following system of two convolutiontype integral operations: f1 (x, t) = −2πρ
t 0
© 2004 by CRC Press LLC
uC2 (t − σ) − C˙3 (t − σ) g(x, σ)dσ,
(2.6)
f2 (x, t) = −2πρ
t 0
1/2C1 (t − σ) − auC2 (t − σ) + aC˙3 (t − σ) +uC4 (t − σ) + 1/2C˙5 (t − σ) g(x, σ)dσ,
(2.7)
¨ t) + (1/2 − a)¨ g(x, t) = uα(x, ˙ t) + h(x, α(x, t).
(2.8)
The aerodynamical functions Ci , i = 1 . . . 5, are deﬁned in the following ways [11]: Cˆ1 (λ)
∞ =
e−λt C1 (t)dt =
0
u e−λ/u , λ K0 (λ/u) + K1 (λ/u)
Reλ > 0,
t C2 (t)
=
C1 (σ)dσ,
C4 (t) = C2 (t) + C3 (t),
0
t C3 (t)
=
C1 (t − σ)(uσ −
u2 σ 2 + 2uσ)dσ,
0
t C5 (t)
=
C1 (t − σ)((1 + uσ) u2 σ 2 + 2uσ − (1 + uσ)2 )dσ,
(2.9)
0
where K0 and K1 are the modiﬁed Bessel functions of the zero and ﬁrst orders [1, 19]. The selfstraining control actuator action can be modeled by the following boundary conditions: Eh (0, t) + β h˙ (0, t) = 0, ˙ t) = 0, Gα (0, t) + δ α(0,
h (0, t) = 0,
(2.10)
β, δ ∈ C+ ∪ {∞},
(2.11)
+
where C is the closed right halfplane. The boundary conditions at x = −L are h(−L, t) = h (−L, t) = α(−L, t) = 0.
(2.12)
Let the initial state of the system be given as follows: h(x, 0) = h0 (x),
˙ h(x, 0) = h1 (x),
α(x, 0) = α0 (x),
α(x, ˙ 0) = α1 (x). (2.13)
We will consider the solution of the problem given by Eq. (2.2) and conditions (2.10)– (2.13) in the energy space H. To introduce the metric of H, we assume that the parameters satisfy the following two conditions: √ 2G m S det > 0, 0 0, then each branch is asymptotically close to its own horizontal line in the closed upper halfplane. If β > 0 and δ = 0, then both horizontal lines coincide with the real axis. If β = δ = 0, then the operator Lβδ is selfadjoint and, thus, its spectrum is real. The entire set of eigenvalues may have only two points of β(δ) β(δ) accumulation: +∞ and −∞ in the sense that λn −→ ±∞ and  λn  ≤ const as n −→ ±∞ (see formulae (4.3) and (4.4) below). c) The following asymptotics are valid for the βbranch of the spectrum as n −→ ∞: ˜ λβn = (sgn n)(π 2 /L2 ) E I/∆ (n − 1/4)2 + κn (ω), ω = δ−1 + β−1 , (4.3) with ∆ being deﬁned by ∆ = m ˜ I˜ − S˜2 . A complexvalued sequence {κn } is bounded above in the following sense: supn∈Z {κn (ω)} = C(ω), C(ω) −→ 0 as ω −→ 0. d) The following asymptotics are valid for the δbranch of the spectrum:
i δ + GI˜ πn δ
+ ln λn = (4.4) + O(n−1/2 ), n −→ ∞. ˜ ˜ ˜ δ − GI L I/G 2L I/G In (4.4), “ln” means the principal value of the logarithm. If β and δ stay away from zero, i.e., β ≥ β0 > 0 and δ ≥ δ0 > 0, then the estimate O(n−1/2 ) in (4.4) is uniform with respect to both parameters. There may be only a ﬁnite number of multiple eigenvalues of a ﬁnite multiplicity each. Therefore, only a ﬁnite number of the associate vectors may exist. The next result is concerned with the properties of the root vectors of the operator Lβδ . Before we formulate this result, we recall the deﬁnition of the biorthogonal vectors [16]. Deﬁnition 4.2. Two sequences of vectors {φn } and {χn } in a Hilbert space H are said to be biorthogonal if for every m and n,we have (φm , χn )H = δmn .
(4.5)
In the case when an operator has a simple spectrum (i.e., there are no associate vectors), the biorthogonal set consists of the eigenvectors of the adjoint operator. However, in general,
© 2004 by CRC Press LLC
the assumption that the spectrum is simple is quite artiﬁcial. Numerical simulations show that minor changes of the parameters of the problem often result in appearance of numerous multiple eigenvalues. So, one cannot disregard the existence of associate vectors. In that case the relationship between the root vectors of the operator and its adjoint becomes less obvious. (For the corresponding results, see [22].) It is also shown in [22] that the set of the root vectors of the operator Lβδ is complete in the state space. (Recall the set of vectors is complete in a Hilbert space if ﬁnite linear combinations of the vectors from the set are dense in this space [16, 17].) Deﬁnition 4.3. A basis in a Hilbert space is a Riesz basis if it is linearly isomorphic to an orthonormal basis, that is, if it is obtained from an orthonormal basis by means of a bounded and boundedly invertible operator. The next result from [22] describes a very important property of the root vectors of the differential operator Lβδ . Theorem 4.3. The set of root vectors of the operator Lβδ forms a Riesz basis in the energy space H. The Riesz basis property of the root vectors has been proven in [22] based on the functional model for nonselfadjoint operators by Sz. Nagy–C. Foias [24], the main elements of which have been reproduced in Section 7 of [22].
5
ASYMPTOTICAL DISTRIBUTION OF AEROELASTIC MODES
Our ﬁrst result in this section is the following statement. Theorem 5.1. a) The set of all aeroelastic modes (which are the poles of the generalized resolvent operator) is countable and does not have accumulation points on the complex plane C. There might be only a ﬁnite number of multiple poles each of a ﬁnite multiplicity. There exists a sufﬁciently large R > 0 such that all aeroelastic modes, whose distance from the origin is greater than R, are simple poles of the generalized resolvent. The value of R depends on the speed u of an airstream, i.e., R = R(u). b) The set of the aeroelastic modes splits asymptotically into two series, which we call the βbranch and the δbranch. Asymptotical distribution of the β and the δbranches of the aeroelastic modes can be obtained from asymptotical distribution of the spectrum of the operator Lβδ . Namely if {µβn }n∈Z is the βbranch of the aeroelastic modes, then µβn = iˆ µβn β and the asymptotics of the set {ˆ µn }n∈Z are given by the righthand side of formula (4.3). Similarly, if {µδn = iˆ µδn }n∈Z is the δbranch of the aeroelastic modes, then the asymptotical distribution of the set {ˆ µδn }n∈Z is given by the righthand side of formula (4.4). The next result is of particular importance for us. It does not follow from the asymptotic representations (4.3) and (4.4). This result has been formulated in paper [4]. The author of [4] has suggested the proof using the general ideas of the Semigroup Theory. We are able to prove the same fact by using a totally different approach. We divide our proof into two parts, and the ﬁrst part is formulated as Theorem 6.3 below. Theorem 5.2. (A.V. Balakrishnan, [4]). For any u > 0, there might exist only a ﬁnite number of the aeroelastic modes having nonnegative real parts. Now we provide an explanation why Theorem 5.2 cannot be considered as a corollary of Theorem 5.1. We discuss each branch individually. First we note that the δbranch of the aeroelastic modes may have only a ﬁnite number of modes with nonnegative real parts. Indeed, since the points µ ˆδn are asymptotically close to the eigenvalues {λδn }n∈Z , and the
© 2004 by CRC Press LLC
latter set has a horizontal asymptote in the upper halfplane (see (4.4)), the δbranch of the aeroelastic modes {µδn = iˆ µδn }n∈Z has a vertical asymptote in the open left halfplane. This means that for n ≥ N >> 1, all modes {µδn }n≥N are located in the left halfplane. However, this is not the case for the βbranch of the aeroelastic modes. Indeed, the set {ˆ µβn }n∈Z is asymptotically close to the set {λβn }n∈Z , and the latter set is close to the real axis. Therefore, the βbranch of the aeroelastic modes {µβn = iˆ µβn }n∈Z is close to the β β imaginary axis. Moreover, we know that λn > 0 (i.e., (iλn ) < 0) for all n ∈ Z due to the fact that the operator Lβδ is dissipative. However, the points {ˆ µβn }n∈Z are not the eigenvalues of any operator, let alone a dissipative operator. So we cannot be sure that
µ ˆβn > 0 (i.e., µβn < 0) for any value of n. Thus, though the fact that there may be only a ﬁnite number of modes of the βbranch with µβn > 0 is valid, it requires quite involved proof. 6
STRUCTURE AND PROPERTIES OF MATRIX INTEGRAL OPERATOR
In this section, we describe the properties of the Laplace transform of the convolutiontype matrix integral operator given in (3.8). ˆ be the Laplace transform of the kernel of matrix integral operator Lemma 6.1. Let F ˆ from (3.8). The following formula is valid for F: 0 0 0 0 0 L(λ) uL (1/2 − a)L(λ) ˆ , F(λ) = 0 0 0 0 0 N (λ) uN (λ) (1/2 − a)N (λ)
S˜ ˜ − + I + (a + 1/2) S˜ T (λ/u) , 2 ˜ 2πρ u m ˜ − S + (a + 1/2) m ˜ T (λ/u) , N (λ) = − ∆ λ 2 and T is the Theodorsen function deﬁned by the formula where
2πρ u L(λ) = − ∆ λ
T (z) =
K1 (z) , K0 (z) + K1 (z)
(6.1) (6.2)
(6.3)
K0 and K1 are the modiﬁed Bessel functions [1, 19]. We note that the Theodorsen function is deﬁned as a singlevalued analytic function on the complex plane having a branch cut along the negative real semiaxis. The following asymptotic approximation is valid: T (z) = 1/2 + V (z),
V (z) = 1/(16z) + O(z −2 ),
z → ∞.
(6.4)
ˆ Taking into account representation (6.5) and the fact that z = λ/u, we can write λF(λ) as the following sum: ˆ λF(λ) = M + N(λ). (6.5)
© 2004 by CRC Press LLC
The matrix M in (6.6) is deﬁned by the formula 0 0 0 0 0 A uA (1/2 − a)A , M= 0 0 0 0 0 B uB (1/2 − a)B where according to (6.2) and (6.3), we have for A and B ˜ A = −πρu∆−1 [I˜ + (a − 1/2)S],
(6.6)
B = πρu∆−1 [S˜ + (a − 1/2)m]. ˜
The matrixvalued function N(λ) is deﬁned by the formula 0 0 0 0 0 A1 (λ) uA1 (λ) (1/2 − a)A1 (λ) , N(λ) = 0 0 0 0 0 B1 (λ) uB1 (λ) (1/2 − a)B1 (λ)
(6.7)
(6.8)
where, according to (6.2) and (6.3), we have for A1 (λ) and B1 (λ) ˜ ≡ −2πρu∆−1 V (z) d1 , A1 (λ) = −2πρu∆−1 V (z)[I˜ + (a + 1/2)S] B1 (λ) = 2πρu∆−1 V (z)[S˜ + (a + 1/2)m] ˜ ≡ 2πρu∆−1 V (z) d2 , z = λ/u.
(6.9)
Therefore, the generalized resolvent (see (3.12) and (3.13)) can be written in the form −1
R(λ) = (λI − iLβδ − M − N(λ))
,
(6.10)
where the matrix M and the matrixvalued function N(λ) are deﬁned by (6.7)–(6.10). Theorem 6.1. M is a bounded linear operator in H. The operator Kβδ introduced by the formula Kβδ = Lβδ − i M + i M I, (6.11) where I is the identity operator, is an unbounded dissipative operator in H. Remark 6.1. The fact that Kβδ is dissipative is very important for the proof that the root vectors of this operator form a Riesz basis in H. Theorem 6.2. N(λ) is an analytic matrixvalued function on the complex plane with the branchcut along the negative real semi–axis. For each λ, N(λ) is a bounded operator in H with the following estimate for its norm:
N H ≤ C(1 + λ)−1 ,
(6.12)
where C is an absolute constant the precise value of which is immaterial for us. We recall that Balakrishnan’s Theorem means that for a given speed u, there may be only a ﬁnite number of unstable mode shapes. In other words, only a ﬁnite number of modes can be subject to ﬂutter. Let us return to Eq. (3.13) for the aeroelastic modes. We ˆ know that λF(λ) is a sum of the 4×4 matrix M with constant entries and the matrixvalued function N(λ), whose norm goes to zero as λ → ∞. We show that the result similar to Balakrishnam’s Theorem is valid for the equation obtained from Eq. (3.13) in which the term containing N(λ) has been omitted, i.e., that an addition of N(λ) cannot destroy the result obtained for the equation iLβδ Ψ + MΨ = λΨ.
(6.13)
Theorem 6.3. For a given value of u, there may exist only a ﬁnite number of eigenvalues of the operator iLβδ + M having nonnegative real parts.
© 2004 by CRC Press LLC
7
PROPERTIES OF THE ADJOINT OPERATORS AND OPERATORVALUED FUNCTIONS
To write the spectral decomposition for the generalized resolvent operator, we need detailed information about the set of functions that is biorthogonal to the set of the mode shapes. It turns out that this biorthogonal set is closely related to the set of nontrivial solutions of the following equation: S∗ (λ)Φ ≡ [−iL∗βδ + M∗ + N∗ (λ)]Φ = λΦ.
(7.1)
The set of complex points λ for which Eq. (7.1) has nontrivial solutions will be called the set of the adjoint aeroelastic modes and the set of corresponding solutions will be called the set of adjoint mode shapes. Now we describe the structure of the adjoint analytic operatorvalued function S∗ (λ). We already know that L∗βδ is given by the same matrix differential expression (3.6) and the only difference is in the description of the domain of L∗βδ , i.e., the parameters ¯ and (−δ) ¯ respectively. β and δ from (3.7) should be replaced with (−β) ∗ Theorem 7.1. The operatorvalued function M + N∗ (λ) can be represented in the following form: M∗ + N∗ (λ) = 0 0 0 − I˜ − bS˜ 1 + 2V¯ (λ/u) /∆ πρu 0 −u 1 + 2V¯ (λ/u) 0 K(x, ξ) · dξ −L 0 S˜ − bm ˜ 1 + 2V¯ (λ/u) /∆
0
0
0 I˜ − bS˜ b + 2ˆbV¯ (λ/u) /∆ , 0 0 u 1 + 2V¯ (λ/u) −L K(x, ξ) · dξ 0 − S˜ − bm ˜ b + 2ˆbV¯ (λ/u) /∆
where b = 1/2 − a, ˆb = 1/2 + a. The kernel of the integral operator in (7.2) is given by the following formulae: 1 sin ω(x + L) cos ωξ, x ≤ ξ, (7.2) K(x, ξ) = sin ω(ξ + L) cos ωx, x ≥ ξ, W (ω)
and W (ω) = ω cos ωL, ω = u πρ/G. If u satisﬁes the second condition from (2.14), then W (ω) = 0. Theorem 7.2. The set of adjoint aeroelastic modes is countable and does not have accumulation points on the complex place C. This set splits asymptotically into two series, which we call the β∗branch and the δ∗branch. Asymptotical distribution of the β∗ and the δ∗branches of the adjoint aeroelastic modes can be obtained from the asymptotical distribution of the spectrum of the operator Lβδ . Namely, if {µβ∗ n }n ∈ Z is the β∗branch β∗ of the adjoint aeroelastic modes, then µβ∗ = iˆ µ , and the asymptotics of the set {µβ∗ n n n }n ∈ δ∗ Z are given by the righthand side of formula (4.3). Similarly, if {µn = iˆ µδ∗ } n n ∈ Z is the δ∗branch of the adjoint aeroelastic modes, then the asymptotical distribution of the set {µδ∗ n } can be obtained from the righthand side of the formula (4.4) by replacing the sign “+ with the sign “–” before the logarithmic term.
© 2004 by CRC Press LLC
Note that due to the latter theorem, one can see that the set of the adjoint aeroelastic modes is closely related to the spectrum of the operator L∗βδ . Finally, we formulate an important result about the operator Kβδ = Lβδ + iM.
(7.3)
Let us denote by {Φβn }n∈Z and {Φδn }n∈Z the β and δbranches of root vectors of the operator Kβδ , respectively. Theorem 7.3. a) The entire set of the root vectors {Φβn }n∈Z {Φδn }n∈Z of the oper∗ ator Kβδ is complete in H. b) This set forms a Riesz basis in H. The properties (a) and (b) imply that the operator Kβδ , is a Riesz spectral operator in the sense of Dunford [22]. c) ∗ The operator Kβδ , which is adjoint to the operator Kβδ , is also Riesz spectral. To prove Theorem 7.3, we have used the same approach as for the case of the operator Lβδ , i.e., the Sz. Nagy–C. Foias functional model for nonselfadjoint operators. To formulate the result on the mode shapes, let us use the following notations: {Fnβ }n∈Z are the βbranch mode shapes, {Fnδ }n∈Z are the δbranch mode shapes. Theorem 7.4. a) The entire set of the mode shapes {Fnβ }n∈Z {Fnδ }n∈Z is complete in the energy space H. b) The set of mode shapes is quadratically close to the set of root vectors of the operator Kβδ , i.e., n∈Z
Φβn − Fnβ 2H +
Φδn − Fnδ 2H < ∞.
(7.4)
n∈Z
∗ is a Riesz spectral operator, By combining (a) and (b) of Theorem 7.4 with the fact that Kβδ we show that the set of the mode shapes forms a Riesz basis in H. A similar fact on the Riesz basis property of the set of adjoint mode shapes can be shown as well. The Riesz basis property of the mode shapes is crucially important for the spectral decomposition of the generalized resolvent operator.
Acknowledgments Partial support by the National Science Foundation Grants ECS #0080441, DMS#0072247, DMS9972748, and the Advanced Research Program97 of Texas Grant 003644045 is highly appreciated.
REFERENCES 1. Abramowitz M., Stegun I., Ed., Handbook of Mathematical Functions, Dover, New York (1972). 2. Balakrishan A.V., Aeroelastic control with selfstraining actuators: continuum models. In: Smart Structures and Materials, Mathematics Control in Smart Structures, Vasundaran V., Ed., Proceedings of SPIE, 3323 (1998), pp. 44–54. 3. Balakrishnan A.V., Subsonic ﬂutter suppression using selfstraining acutators. To appear in: Journal of the Franklin Institute, Special Issue on Control, Udwadia F., Ed. 4. Balakrishnan A.V., Aeroelastic control with self–straining actuators: unsteady aerodynamics solution. Unpublished manuscript.
© 2004 by CRC Press LLC
5. Balakrishnan A.V., Shubov M.A., Peterson C.A., Spectral analysis of Euler–Bernoulli and Timoshenko beam model. Submitted to: Integral and Differential Equations (2001). 6. Balakrishnan A.V., Damping performance of strain actuated beams, Comput. Appl. Math. 18 (1) (1999), pp. 31–86. 7. Balakrishnan A.V., Vibrating systems with singular mass–inertia matrices, First Intl. Conf. Nonlinear Problems in Aviation and Aerospace, Sivasundaram S., Ed., pp. 23–32 (1997), EmbryRiddle Aeronautical University Press. 8. Balakrishnan A.V., Theoretical limits of damping attainable by smart beams with rate feedback. In: Smart Structures and Materials, 1997; Mathematics and Control in Smart Structures, Vasundaran V., Jagdish C., Eds., Proc. SPIE, 3039 (1997), pp. 204– 215. 9. Balakrishnan A.V., Control of structures with selfstraining actuators: coupled Euler/Timoshenko model, Nonlinear Problems in Aviation and Aerospace, Gordon and Breach Sci. Publ., Reading, United Kingdom (1998). 10. Balakrishnan A.V., Dynamics and control of articulated anizotropic Timoshenko beams. In: Dynamics and Control of Distributed Systems, Tzou H.S. and Bergman L.A., Eds., Cambridge University Press (1998), pp. 121–201. 11. Balakrishnan A.V. and Edwards J.W., Calculation of the transient motion of elastic airfoils forced by control surface motion and gusts, NASA TM 81351 (1980). 12. Bisplinghoff R.L., Ashley H., and Halfman R.L., Aeroelasticity, Dover Publ. Inc., New York (1996). 13. Chen G., Krantz S.G., Ma D.W., Wayne C.E., and West H.H., The Euler–Bernoulli beam equations with boundary energy dissipation. Operator Methods for Optimal Control Problems, Lecture Notes in Math., 108, pp. 67–96, Marcel Dekker (1987). 14. Dierolf P., Schr¨oder D., and Voight J., Flutter as a perturbation problem for semigroups. In: Semigroups of Operators: Theory and Applications, Balakrishnan A.V., Ed., Birkhauser (2000), pp. 89–95. 15. Fung Y.C., An Introduction to the Theory of Aeroelasticity, Dover Publ. Inc., New York (1993). 16. Gohberg I.Ts. and Krein M.G., Introduction to the theory of linear nonselfadjoint operators, Trans. of Math. Monogr., 18, AMS, Providence, RI (1996). 17. Istratescu V.I., Introduction to Linear Operator Theory, Pure Appl. Math Series of Monog., Marcel Dekker Inc., New York (1981). 18. Lee C.K., Chiang W.W., and O’Sullivan T.C., Piezoelectric modal sensor/actuator pairs for critical active damping vibration control, J. Acoust. Soc. Am., 90, pp. 384–394 (1991). 19. Magnus W., Oberhettinger F., and Soni R.P., Formulas and Theorems for the Special Functions of Mathematical Physics, 3rd Ed., Springer Verlag, New York (1966). 20. Shubov M.A., Mathematical analysis of aircraft wing model in subsonuc airﬂow, IMA Journal of Applied Math., 66 (2001), pp. 319–356. 21. Shubov M.A., Asymptotic representations for root vectors of nonselfadjoint operators generated by aircraft wing model in subsonic air ﬂow, J. Math. Anal. Appl., 260 (2001), pp. 341–366. 22. Shubov M.A., Riesz basis property of root vectors of nonselfadjoint operators generated by aircraft wing model in subsonic air ﬂow, Math. Methods in Applied Sciences, 23 (2000), pp. 1585–1615.
© 2004 by CRC Press LLC
23. Shubov M.A., Asymptotics of the aeroelastic modes and basis property of mode shapes for aircraft wing model, J. of the Fraklin Institute., 338 (2/3) (2001), pp. 171–185. 24. Sz.–Nagy B. and Foias C., Harmonic Analysis of Operators on Hilbert Space, North Holland Publ. Co. (1970). 25. Tzou H.S. and Gadre M., Theoretical analysis of a multilayered thin shell coupled with piezoelectric shell actuators for distributed vibration controls, J. Sound Vibr., 132, pp. 433–450 (1989). 26. Yang S.M. and Lee Y.J., Modal analysis of stepped beams with piezoelectric materials, J. Sound Vibr., 176, pp. 289–300.
© 2004 by CRC Press LLC
4 Bifurcation Analysis for the Inertial Coupling Problem of a Reentry Vehicle Norihiro Goto and Takashi Kawakita Kyushu University, Fukuoka, Japan
This chapter conducts a bifurcation analysis for the inertial coupling problem that may occur in the case of automatic ﬂight control system failures of a reentry vehicle. Equilibrium paths are ﬁrst obtained with the use of the continuation method in the planes of the vehicle’s motion variables vs. control surface angles. Over all the ranges of the equilibrium paths, stability analyses are then undertaken by means of Lyapunov’s ﬁrst method. Special emphasis is put on the convergence nature of the motion variables’ temporal trajectories for the case where more than one stable equilibrium point exists for a combination of control surface angles. Based on the equilibrium paths with stability information, the control sequences of control surface angles are ﬁnally sought that may bring the vehicle in a high roll rate motion back to a zero roll rate steadystate ﬂight. It is pointed out that a steady ﬂight with no rollrate can be recovered only if the basic solution branch is stable.
1
INTRODUCTION
Stable ﬂights of a reentry vehicle rely heavily on automatic ﬂight control systems (AFCSs), because its ﬂight dynamical conﬁguration often lacks intrinsic stability over a wide speed range of operation. Failures of AFCSs cause a number of problems in maintaining a desired ﬂight condition. Above all, the problem of inertial coupling must be addressed in detail before the vehicle is put into operation, inasmuch as it is difﬁcult to recover the original ﬂight condition once it develops this problem. Inertial coupling is one of the important nonlinear problems that challenge aircraft stability and control engineers to ﬁnd recovery control techniques. Inertial coupling has been known since around 1948 [1]. It is essentially a gyroscopic effect, occurring in high rollrate maneuvers of modern highspeed airplanes including spinning missiles designed in such a way that most of their masses are concentrated in the fuselage. For such an airplane, a slight deviation of its control surface angle from the steadystate angle may lead to a drastic change in rollrate, causing damage on its empennage; known as the jump phenomenon. Nonlinear analyses to elucidate this problem have been reported in [2, 3], for example. However, the airplanes treated in these works are stable around the equilibrium point of level ﬂight. On the other hand, an intrinsically unstable reentry vehicle, if combined with a malfunction of its AFCSs, may result in a catastrophe once it falls into a high rollrate motion. Therefore, the primary objective of this chapter is
© 2004 by CRC Press LLC
Figure 1 Three views of the ALFLEX plane.
to seek a control technique for an unstable vehicle to recover a zero rollrate ﬂight from a high rollrate motion into which it has accidentally fallen. For this primary objective, a model reentry vehicle is subjected to a bifurcation analysis. The model plane is the Automatic Landing Flight Experiment (ALFLEX) plane [4] as shown in Fig. 1, which is a reduced scale model of an unmanned reusable orbiting spacecraft. The ALFLEX plane is a glider having no means of propulsion aimed at evaluating the ﬂight characteristics of the spacecraft during the ﬁnal approach and landing phases. Its stable ﬂight is made possible with the help of elaborate AFCSs, without which the bare conﬁguration is unstable both longitudinally and lateral directionally over a wide speed range. Such characteristics of the ALFLEX plane make it important to look into the trimmability of the bare vehicle in case a malfunction of AFCSs should occur. We ﬁrst describe the nonlinear equations of motion for the inertial coupling problem of the ALFLEX plane. A bifurcation analysis is then made by the continuation method to determine equilibrium paths (paths of trim points) in the plane of motion variables vs. control surface angles, followed by stability analysis of trim points with the use of Lyapunov’s ﬁrst method. Discussion is also made of the convergence nature of the motion variables’ temporal responses for the case where more than one stable trim point exists for a combination of control surface angles. Drawing on these analyses, the chapter ﬁnally addresses possible control techniques to recover a zero rollrate ﬂight from a high rollrate motion of inertial coupling.
2
EQUATIONS OF MOTION
Because high rollrate steady states are of primary concern here, the original sixdegreesoffreedom equations of motion of a rigid airplane with respect to an xyz bodyaxis system [5], where xz is the plane of symmetry, are reduced to ﬁvedegreesoffreedom equations under the assumptions: 1. forward velocity V , weight W , and air densityρare constant, and 2. angle of attack α and sideslip angle β are small. The resulting equations of motion are
© 2004 by CRC Press LLC
g β˙ = p sin α − r cos α + yˆ + sin φ cos θ, V g α˙ = −pβ + q + zˆ + cos φ cos θ, V Ixz p˙ − r˙ = i1 −qr + ia pq + ˆl , Ix ˆ , q˙ = i2 rp + ib −p2 + r2 + m Ixz p˙ = i3 (−pq − ic qr + n ˆ) , r˙ − Iz
(1) (2) (3) (4) (5)
where the Eulerian angles, roll angle φ, and pitch angle θ, are determined by the kinematic relations as φ˙ = p + q sin φ tan θ + r cos φ tan θ, θ˙ = q cos φ − r sin φ. In Eqs. (1)–(7) i1 =
Iz − Iy , Ix
ia =
i2 =
1 Ixz , i1 Ix yˆ =
ib =
g Ya , WV
ˆl = L , i1 Ix
Iz − Ix , Iy 1 Ixz , i2 Iy
i3 = ic =
(7)
Iy − Ix Iz
1 Ixz , i3 Iz
g Za , WV M N m ˆ = , n ˆ= , i2 Iy i3 Iz zˆ =
(6)
(8)
Ix , Iy , Iz : moments of inertia about x, y, and zaxis, respectively, Ixy : product of inertia, Ya , Za : aerodynamic forces, L, M, N : aerodynamic moments about the center of gravity, p, q, r: angular velocities about x, y, and zaxis, respectively, g: gravitational acceleration. Among the inertial parameters in Eq. (8), i3 is referred to as the inertial coupling parameter. It can be deduced from Eq. (5) that a large value of i3 produces large inertial coupling moments. In comparison with the data given in [1], the ALFLEX plane, of which the dimensional data are given in Table 1, has the same magnitude of i3 as that of the F100A ﬁghter plane that suffered from an inertial coupling problem in its early phase of development. Aerodynamic forces and moments in Eqs. (1)–(5) are assumed here to be linearly related to motion variables and control surface angles mainly because of lack of nonlinear aerodynamic data. Table 1 summarizes the linear relationships and numerical data necessary for the analysis to follow. The reader is referred to [6] for an example of complete treatment of nonlinear aerodynamic data. Note in Table 1 that a level ﬂight with the steadystate pitch angle θ0 and angle of attack α0 is a trivial trim point. Those original equations of motion, Eqs. (1)–(7), can be represented in the general form as x˙ = H(x, δ),
© 2004 by CRC Press LLC
(9)
Table 4.1 Flight conﬁguration, stability derivatives and numerical data. yˆ = (g/W V )(1/2ρV 2 SCy ) = yˆβ β + yˆr r + yˆδr δr Cy = (Cyβ )β + (Cyr )r + (Cyδr )δr zˆ = (g/W V )(1/2ρV 2 SCz ) = zˆ0 + zˆα ∆α + zˆδe ∆δe Cz = Cz0 + (Czα )∆α + (Czδe )∆δe Cz0 = −W cos θ0 /(1/2ρV 2 S) Czα = − CLα cos α0 + CL sin α0 − CDα sin α0 − CD cos α0 Czδe = − CLδe cos α0 − CDδe sin α0 ˆ lβ β + ˆ lp p + ˆ lr r + ˆ l δ a δa + ˆ lδ r δr l = (1/2ρ V 2 S b Cl )/(i1 Ix ) = ˆ Cl = (Clβ )β + (Clp )p + (Clr )r + (Clδa )δa + (Clδr )δr m ˆ = (1/2ρ V 2 S c Cm )/(i2 Iy ) = m ˆ α ∆α + m ˆ α˙ ∆α˙ + m ˆ qq + m ˆ δe ∆δe Cm = (Cmα )∆α + (Cmα˙ )∆α˙ + (Cmq )q + (Cmδe )∆δe n ˆ = (1/2ρ V 2 S b Cn )/(i3 Iz ) = n ˆβ β + n ˆpp + n ˆr r + n ˆ δa δa + n ˆ δr δ r Cn = (Cnβ )β + (Cnp )p + (Cnr )r + (Cnδa )δa + (Cnδr )δr δa : aileron angle, δr : rudder angle, ∆δe : incremental elevator angle W (N ) = 760g
V (m/s) = 73.84
Ix (kg m2 ) = 407
α0 (deg) = 8.18
Iy (kg m2 ) = 1366
γ0 (deg) = −17.34
Iz (kg m2 ) = 1643
θ0 (deg) = γ0 + α0 = −9.16
Ixz (kg m2 ) = 10.4
δe0 (deg) = 3.0
S (m2 ) = 9.45 ρ (kg/m3 ) = 1.156
c (m) = 3.154 b (m) = 3.295 CL = 0.2387
CLα = 2.016
CLδe = 0.6355
CD = 0.0745
CDα = 0.2714
CDδe = 0.1019
Cmq = −0.0474
Cmδe = −0.2152
Cmα = −0.0134 Cyβ = −0.6849
Cyδr = 0.1907
Clβ = −0.1774
Clp = −0.0070
Clr = 0.0040
Clδa = 0.1488
Clδr = 0.0788
Cnβ = −0.0657
Cnp = 0.0032
Cnr = −0.0060
Cnδa = −0.0266
Cnδr = −0.0990
All other derivatives are set equal to zero.
where x = [β, α, p, q, r, φ, θ]T , T
δ = [δe , δa , δr ] ,
(10) (11)
and δe : elevator angle, δa : aileron angle , δr : rudder angle. General trim points are determined by solving the transcendental algebraic equation H(x, δ) = 0.
© 2004 by CRC Press LLC
(12)
The vehicle, however, cannot attain all of the trim points obtained from Eq. (12). If it has unstable dynamics in the neighborhood of a trim point, a slight external disturbance or a slight deviation of control surface angles from their precise values at the trim point cannot allow the vehicle to stay at the trim point. Stability of the vehicle’s dynamics in the neighborhood of a trim point must be examined in order to know whether a steady state can actually be attained. To this end, the stability analysis of trim points is made by using Lyapunov’s ﬁrst method [7], in which the nonlinear equations of motion, Eqs. (1)–(5), are linearized about a trim point as x˙ ε = F xε , (13) where xε = [βε , αε , pε , qε , rε ]T : perturbed variable vector, and F is the Jacobian matrix explicitly given by yˆβ p¯ α ¯ + sin α0 0 yˆr − cos α0 F =
−p¯
lˆ ip ( iβ a
zˆα
+ i3 n ˆβ )
ˆα i2 m
0 lβ + ir (i1 ˆ
0
n ˆβ ic
)
0
−β¯ ip [(1 − i3 )¯ q lˆ +( ip a
+i3 n ˆ p )]
i2 (¯ r − 2ib p) ¯ q ir (i1 ia − i1c )¯ n ˆ lp + icp ) +(i1 ˆ
1 ip [(1 − i3 )p¯ −( i1a + i3 ic )¯ r
i2 m ˆq
ir (i1 ia − i1c )p¯ r] −(i1 + 1)¯
0 ip −( i1a + i3 ic )¯ q ˆ ˆr ) +( lirc + i3 n i2 (p¯ + 2ib r¯) ir [−(i1 + 1)¯ q ˆr lr + n +(i1 ˆ ic )
(14) ¯ p¯, q¯, In Eq. (14) stability derivatives yˆβ and so on are deﬁned in Table 1, and α ¯ , β, and r¯ denote the steadystate values, while ip =
Ixz Iz , 2 Ix Iz − Ixz
ir =
Ixz Ix . 2 Ix Iz − Ixz
(15)
If all of the eigenvalues of F have negative real parts, the trim point is asymptotically stable, whereas it is unstable if some eigenvalues have positive real parts. It sometimes happens that the real parts of some eigenvalues are zero or close to zero, generating socalled bifurcation points. In such a case, it is necessary to consider the effects of nonlinear terms on the behavior of the solution in the neighborhood of the trim point. The center manifold theory can be applied to such a case [7].
3
CONTINUATION METHOD
An equilibrium path analysis is made using the continuation method [3], which allows one to solve the transcendental algebraic equation, Eq. (12), without any approximation. In the method, Eq. (12) is transformed to an initial value problem such that dc(s) = c˙ = t H (c(s)) , ds c(0) = (xT (0), δ T (0)),
(16) (17)
where c(s) is the solution trajectory, s is the arc length, and t(H ) denotes the tangent vector to H , which is the derivative of H with respect to s. Because s is the arclength, the initial value problem is accompanied with the constraint condition: ˙ c(s) = 1.
© 2004 by CRC Press LLC
(18)
This initial value problem is solved numerically by the predictor–corrector method with the discretized c(s) as c(i) and the step size as h: 1. Euler method as the predictor: cp (i + 1) = cc (i) + ht(H (cc (i))),
(19)
where the subscripts p and c denote the value by the predictor and that by the corrector, respectively. 2. Newton’s method as the corrector: cc (i + 1) = cp (i + 1) − H (cp (i + 1))+ H(cp (i + 1)),
(20)
where H (u)+ is the Moore–Penrose inverse matrix; i.e., H (u)+ = H (u)T (H (u)H (u)T )−1 .
(21)
Note that the corrector value by Eq. (20) minimizes the norm cc (i + 1) − cp (i + 1)H(cc (i+1)) .
(22)
In practice the tangent vectors and the Moore–Penrose inverses that take into account the constraint, Eq. (18), can be obtained by the use of the QR method [8], in which an arbitrary matrix A is factorized so as to be the product of a proper orthogonal matrix Q and an upper triangular matrix R. A good initial guess may lead to a quick convergence of the algorithm. One such initial guess can be given by the pseudosteadystate analysis [2], in which the gravity terms in Eqs. (1) and (2) are ignored to arrive at a polynomial equation with respect to the roll rate p. The greatest merit of the continuation method, in particular that of using the QR method, is that the computation can proceed even at a bifurcation point without difﬁculty.
4
NUMERICAL RESULTS
Using the ﬂight conﬁguration of Table 1, equilibrium path analyses are made here followed by the stability analyses of trim points. Flight control of the ALFLEX plane actually uses elevator δe , aileron δa , rudder δr , and speed brake δs as available control inputs. However, in this analysis the rudder angle is set equal to zero, and the speed brake is ﬁxed at a certain angle. Therefore, the equilibrium paths are determined by varying δe and δa as parameters. The ranges of variation are −10–10 deg for ∆δe , which denotes an incremental elevator deﬂection from the initial value of Table 1, and −20–20 deg for δa . First of all, the results from the equilibrium path analyses are shown in Figs. 2 and 3, where ∆α is also an incremental angle of attack. Figures 2(a)–2(e) show the equilibrium paths in the plane of motion variables vs. δa for a ∆δe , while Figs. 3(a)–3(b) together with Fig. 2(b) exhibit the variation of p vs. δa equilibrium paths for three kinds of ∆δe . It results from the ensuing stability analyses that the trim points on solid equilibrium paths are stable, whereas those on broken paths are unstable. By numerically integrating the original nonlinear equations of motion, these equilibrium paths with stability information have been validated with an additional ﬁnding that the origin of p − δa plane for ∆δe = −6 deg actually yields a limit cycle oscillation. The elevator angle should be smaller than −6 deg in order for the origin
© 2004 by CRC Press LLC
Figure 2 Equilibrium paths of motion variables for ∆δe = 4 deg.
to be a stable trim point. Furthermore, a numerical simulation enables one to visualize the vehicle’s dynamic motion. Figure 4 is a snapshot, as seen from a ﬁxed point on the ground, from such a visual threedimensional ﬂight simulation. The vehicle in Fig. 4 is close to a steady state for a combination of step inputs of ∆δe and δa . Velocity and angular velocity vectors are shown in the ﬁgure together with the bodyﬁxed three axes. This kind of simulation helps to understand the ﬂight–dynamic characteristics implicit in the nonlinear equations of motion, while at the same time it may point out the defects of the analysis as well. For instance, it can be learned from Fig. 4 that the angle of attack at this moment is almost 90 deg, violating the assumption that α be small. More realistic analyses should be undertaken in the future, getting rid of the assumption and using nonlinear aerodynamic
© 2004 by CRC Press LLC
Figure 3 Equilibrium paths of roll rate for ∆δe = 4 and 0 deg.
Figure 4 A snapshot of the ALFLEX plane in steadystate motion.
data. Note that this simulation makes use of the quaternion formulation for attitude computation to avoid the singularity problem that turns up for a large θ in Eq. (6). More importantly, the equilibrium paths with stability information thus obtained may tell the combinations of (∆δe , δa ) for which jump phenomena are likely to occur. For example, assume that the vehicle is in a steady rolling state of about 100 deg/s for a combination of (∆δe , δa )= (4 deg, 0 deg) as read from Fig. 2(b). A deviation of the aileron angle to 2 deg brings the roll rate abruptly to about −200 deg/s, because at the new aileron angle the trim point on the same equilibrium path is not stable any more. Looking at Fig. 2(b) again, it can be observed that depending on the combination of (∆δe , δa ) there exist multiple stable trim points or attractors. An interesting question is to which attractor the motion will converge, given the corresponding combination of control surface angles and an arbitrary set of initial conditions. If the motion starts from within the region of attraction of an attractor, it will settle down on the attractor. For an arbitrary set of initial conditions, however, the general ﬂow pattern of the solution trajectories needs to be known in the phase plane of seven variables. Instead of this general phase plane approach, an attempt is made here at correlating initial conditions with attractors in twodimensional phase planes by re
© 2004 by CRC Press LLC
Figure 5 Correlation between initial condition and attractors: ◦, •: attractors, ◦ ,• : initial conditions.
peating numerical simulations. A result from such an attempt is shown in Fig. 5, which is a p − q phase plane for a combination of (∆δe , δa ) where two attractors, marked by an open circle and a ﬁlled circle, are located on stable equilibrium paths of solid curves, while small circle marks in the plane denote the locations of the initial condition of p and q with the other initial conditions set equal to zero; the initial (p, q) of • takes the motion to the attractor •, while that of ◦ eventually goes to ◦. A fairly clear pattern can be seen of the distribution of initial conditions. It is a work for the future to ﬁnd analytical rules that underlie such a pattern.
5
RECOVERY FROM A HIGH ROLLRATE MOTION
First of all, note that in Fig. 3(a) the equilibrium path that passes through the origin, referred to as the basic solution branch, is partly unstable, while in Fig. 2(b) it is entirely unstable. It turns out that for the ﬂight condition of Table 1 the basic solution branch becomes unstable around the origin for ∆δe larger than about −6 deg. Therefore, the basic solution branch must be stabilized before any recovery technique is devised. If automatic stabilization is available, certain types of socalled chaotic control may be applicable to such a situation; e.g., the Lyapunov function method [9]. However, this work presupposes that any ﬂight control system is out of order, so that a remote control technique should be worked out through the use of δe and δa . Looking at Fig. 6(a) where the equilibrium paths for ∆δe = 2 deg are shown together with the stable basic solution branch for ∆δe = −8 deg, such a control technique can be proposed that ﬁrst moves ∆δe to −8 deg from 2 deg, at which the vehicle is now experiencing a high rollrate motion of about −125 deg/s, then shifts δa back to 0 deg using the stable basic solution branch. To validate this control technique, the original nonlinear equations of motion are numerically integrated with the initial states (p, δa )= (0 deg/s, 5 deg), applying the control sequence shown in Fig. 6(b). Figure 6(c) shows the result that the roll rate subsides to zero, although a large oscillation in p is excited initially, and more noteworthily this new no rollrate steady state has a nonzero pitch rate. A level ﬂight with no rotational motion at all cannot be attained at the basic ﬂight conﬁguration of Table 1, because it is an unstable trim point. It follows that realizing
© 2004 by CRC Press LLC
Figure 6 An example of the control technique to recover: (a) equilibrium paths for a recovery control technique; (b) recovery control technique; (c) time history of roll rate with a recovery control technique.
a stable basic solution branch in the hyperplane of motion variables of concern is key to ﬁnding a control technique for recovering a true level ﬂight.
6
CONCLUSION
A bifurcation analysis has been made of a model reentry vehicle using the continuation method. The obtained equilibrium paths indicate that jump phenomena of inertial coupling can occur over a wide range of control surface angles. The stability analysis of the equilibrium paths by Lyapunov’s ﬁrst method shows that depending on the combination of control surface angles the basic solution branch is unstable around the zero rollrate trim point. Discussion is also made on the convergence of the vehicle’s motion when multiple attractors exist for a combination of control surface angles. Drawing on the equilibrium paths with stability information, the control sequence of elevator and aileron is ﬁnally sought that may bring the vehicle back to a zero rollrate steady state. It is pointed out that a zero rollrate steady state could be reached only if it is on a stable basic solution branch. It is a work for the future to devise a remedy control technique for recovery from a general rotational motion for a more realistic case where complete nonlinear equations of motion are utilized together with nonlinear aerodynamic data.
© 2004 by CRC Press LLC
REFERENCES 1. Abzug, M.J. and Larrabee, E.E., Airplane Stability and Control. A History of the Technologies That Made Aviation Possible, Cambridge University Press, Cambridge, U.K., Chap. 8, 1997. 2. Schy, A.A. and Hannah, M.E., “Prediction of jump phenomena in rolling coupled maneuvers of airplanes,” Journal of Aircraft, Vol. 14, pp. 375–382, April 1977. 3. Carrol, J.V. and Mehra, R.D., “Bifurcation analysis of nonlinear aircraft dynamics,” Journal of Guidance, Control, and Dynamics, Vol. 5, No. 5, pp. 529–536, 1982. 4. Miyazawa, Y., Motoda, T., Izumi, T., and Hata, T., “Longitudinal landing control law for an autonomous reentry vehicle,” Journal of Guidance, Control, and Dynamics, Vol. 22, No. 6, pp. 791–800, 1999. 5. Etkin, B. and Reid, L.D., Dynamics of Flight — Stability and Control, 3rd ed., John Wiley & Sons, New York, Chap. 4, 1996. 6. Sibilski, K., “An agile aircraft nonlinear dynamics by continuation methods and bifurcation theory,” Proceedings of the 22nd International Congress of Aeronautical Sciences 2000 (CDROM), Paper No. ICAS20003.11.2, Aug. 2000. 7. Khalil, H.K., Nonlinear Systems, Macmillan, New York, Chaps. 3 and 4, 1992. 8. Mathematical Society of Japan, Encyclopedic Dictionary of Mathematics, 2nd ed., Ed. Ito, K., The MIT Press, Cambridge, MA, p. 298F, 1987. 9. Vincent, T.L., “Utilizing chaos in control system design.” In Controlling Chaos and Bifurcation in Engineering Systems, Ed. Chen, G., CRC Press, New York, pp. 89–106, 1999.
© 2004 by CRC Press LLC
5 Missile Autopilot Design Using Dynamic Fuzzy GainScheduling Technique †
ChunLiang Lin,† ReiMin Lai,† and SenWei Huang∗ Department of Electrical Engineering, National Chung Hsing University, Taichung 402, Taiwan Department of Applied Mathematics, National Chung Hsing University, Taichung 402, Taiwan
∗
A dynamic backpropagation training algorithm for an adaptive fuzzy gainscheduling feedback control scheme is developed in this chapter. This novel design methodology uses a Takagi–Sugeno fuzzy system to represent the fuzzy relationship between the scheduling variables and controller parameters. Mach number and angle of attack are used as measured, timevarying exogenous scheduling variables injected into the guidance law. By incorporating scheduling parameter variation information, the adaptation law for controller parameters is derived. Results from extensive simulation studies show that the presented approach offers satisfactory controlled system performance.
1
INTRODUCTION
Gain scheduling is an effective way to control systems whose dynamics change with the operating conditions [1, 2]. This technique uses process variables related to dynamics that compensate for the effect of working in different operating regions. It is normally used in the control of nonlinear plants where the relationship between the plant dynamics and operating conditions is known, and for which a single linear timeinvariant model is insufﬁcient. This is the reason it was originated in connection with the development of ﬂight control systems. The missile autopilot design problem is challenging for its wide range of parameter variations and stringent performance requirements. Recently there have been several published studies of gain scheduling in ﬂight control problems. Conventional autopilot gain scheduled designs for tactical missiles can be derived from [3–6]. A robust control method in the framework of µsynthesis was used in the autopilot design [7]. H∞ control theory has also been extensively applied to treat the problem [8, 9]. In [10, 11], the concept of linear parametervarying control techniques was introduced, and a state transformation was proposed to convert nonlinear missile dynamics into a linear parametervarying form and hence to apply the gainscheduled control design. Various control laws have been compared in [12] to estimate their potential and applicability in a threeaxis missile model. Typically, a gainscheduling design involves three issues: partitioning the operating region into several approximately linear regions, local controller design in each linear region, and controller parameter interpolation between the linear regions. The main drawback of most conventional gain scheduling (CGS) is that the parameter change may be
© 2004 by CRC Press LLC
rather abrupt across the region boundaries, which may result in unacceptable or even unstable performance. Another problem is that accurate linear timeinvariant models at various operating points may be difﬁcult or impossible to obtain. As a solution to this problem, a fuzzy gainscheduling (FGS) method has been proposed that utilizes a fuzzy reasoning technique to determine the controller parameters. With this method, human expertise in the linear control design and CGS can be represented by fuzzy rules with a fuzzy inference mechanism used to interpolate the controller parameters in the transition regions [13]. Takagi–Sugeno fuzzy models provide an effective representation of complex nonlinear systems in terms of fuzzy sets and fuzzy reasoning applied to a set of linear input–output submodels. Based on each model, fuzzy gainscheduling controllers are facilitated using linear matrix inequality methods [14]. The FGS technique has been applied in aircraft ﬂight control design [15, 16]. A robust fuzzy gain scheduler has also been designed for an aircraft ﬂight control system [17]. Neural network gain scheduling (NNGS) could incorporate learning ability into gainscheduling control [2] where the training example consists of the operating variables and control gains obtained at various operating points and their corresponding desired output. Compared to the FGS technique, the main advantage of NNGS is that it avoids the need to manually design a scheduling program or determine a suitable inferencing system. NNGS techniques have also been applied to the aircraft ﬂight control design in [18]. In summary, the major difﬁculties in applying FGS consist of (i) difﬁculty in specifying appropriate fuzzy rules and membership functions, (ii) performance of the gainscheduled system depends on the accuracy of the missile guidance and control system model under each ﬂight condition. For NNGS, the major difﬁculties are (i) appropriate network size selection, (ii) difﬁculty in injecting the designer’s expertise into the existing network when necessary. This paper proposes a new methodology for controlling a class of missile autopilots using an adaptive fuzzy gainscheduling (AFGS) technique. This design approach uses a Tagaki–Sugeno fuzzy system to represent the fuzzy relationship between the scheduling variables and controller parameters. A dynamic gainscheduled compensation strategy is used to design the control system. Traditionally the control gains of an autopilot are switched along the missile trajectories according to a function or a table lookup built in the computer. Unlike conventional techniques, the proposed approach offers the advantage of performance improvement for illdeﬁned ﬂight dynamics through learning using an adaptive fuzzy inferencing mechanism. Rapid adaptivity to environmental changes makes this technique appropriate for ﬂight control systems covering aerodynamic changes during ﬂight.
2
AUTOPILOT MODEL AND PERFORMANCE OBJECTIVE
Consider a representative dynamic nonlinear missile system described by the following statespace equations [8]: α(t) ˙ = g1 (α, δ, M ) + q q(t) ˙ = g2 (α, δ, M ) aL (t) = g3 (α, δ, M )
© 2004 by CRC Press LLC
(1)
Table 1 Coefﬁcients of the pitchaxis missile model
Kα Kq Kz an bn cn dn am bm cm dm
1.18587 70.586 0.6661697 0.000103 deg−3 −0.00945 deg−2 −0.1696 deg−1 −0.034 deg−1 0.000215 deg−3 −0.0195 deg−2 0.051 deg−1 −0.206 deg−1
where g1 (α, δ, M ) = Kα M Cn [α(t), δ(t), M (t)] cos α(t), g2 (α, δ, M ) = Kq M 2 Cm [α(t), δ(t), M (t)], g3 (α, δ, M ) = Kz M 2 Cn [α(t), δ(t), M (t)], α is the angle of attack, q is the pitch angle rate, M is the Mach number, δ is the tail deﬂection angle, aL is the actual lateral acceleration; the aerodynamic coefﬁcients Cn and Cm are given by the following polynomial expressions for α and δ given in degrees and M given in Mach: M 3 2 Cn (α, δ, M ) = sgn(α) an α + bn α + cn (2 − ) α + dn δ 3 8M 3 2 ) α + dm δ. Cm (α, δ, M ) = sgn(α) am α + bm α + cm (−7 + 3 Details of aerodynamic constants for the pitchaxis missile model are given in Table 1. The variables q and aL are measured variables available for feedback. Let us deﬁne the augmented state vector x(t) = [α(t), q(t)]T . Actuator dynamics describing the tail deﬂection are d δ(t) δ(t) 0 1 0 = + δc (t), (2) 2 ˙ ˙ −ωa2 −2ζa ωna ωna δ(t) dt δ(t) where δc and δ are, respectively, the commanded and actual tail deﬂection angles, and ζa and ωna are, respectively, the damping ratio and undamped natural frequency. For the problem considered here, we wish to design a controller to track commanded acceleration maneuvers and a command with an accuracy no more than 5%. The controller must provide robust performance over a wide range of angles of attack and variations in Mach number. To this aim, an adaptive fuzzy gainscheduling (AFGS) controller is characterized by u(t) = K1 (α, M )ea (t) δc (t) = K2 (α, M )ev (t),
© 2004 by CRC Press LLC
(3)
Figure 1 Structure of the fuzzy dynamic gainscheduled autopilot.
where ev = u − q and ea = aLc − aL are, respectively, the tracking errors of pitch rate and lateral acceleration command with aLc denoting the commanded lateral acceleration, K1 and K2 are, respectively, the control gains for the acceleration and pitch rate feedback loops. We chose the scheduling variable vector w = [α, M ]T that correlates well with the change in the plant dynamics. The objective of this controller is to minimize a predeﬁned performance measure for the closedloop system based on the information provided by the scheduling variables. A fuzzy logicbased, adaptive parameter adjustment rule will be proposed in the following paragraphs to train the controller. A schematic diagram for the AFGS control scheme is shown in Fig. 1. For the ﬂight control scheme, the performance measure is deﬁned by integrating the combination of the tracking error and control command over [t0 , t] as follows: 1 J= 2
t
T e (τ )e(τ ) + ρδ 2 (τ ) dτ,
(4)
t0
where e(t) = [ea (t), ev (t)]T , the scalar ρ is included to weight the relative importance of the tracking error and control effort. j To characterize the adaptive controller, a group of adjustable variables σsi are introduced. These variables are used as the consequent parameters for the fuzzy rules. The gradient descent algorithm [19] is then used to derive a generalized backpropagation training algorithm: j ∆σsi = −η∇σj J, si
∀ s, i, j,
(5)
j j where η > 0 is the learning rate of the parameter update algorithm, ∆σsi = σsi (t + τ ) − j j σsi (t) with τ being the sampling period, the gradient ∇σj J = ∂J/∂σsi is calculated using si the chain rule:
© 2004 by CRC Press LLC
∂J j ∂σsi
= = t0
T
j ∂σsi
t −ea
= t0
where
∂e
∂J ∂δ j ∂δ ∂σsi ∂δ T ∂e e + ρδ j dτ j ∂σsi ∂σsi
∂J ∂e t
∂aL j ∂σsi
∂u j ∂σsi
+
+ ev
=
∂K1
∂u j ∂σsi
e j a ∂σsi
−
∂q
+ ρδ
j ∂σsi
− K1
∂aL j ∂σsi
∂δ j ∂σsi
dτ,
(6)
.
j j j To complete the derivations, the derivatives ∂aL /∂σsi , ∂q/∂σsi and ∂δ/∂σsi must be found. Determination for these terms is given in the subsequent section. With minor modiﬁcations, the compensation approach developed above can be applied for linear or nonlinear systems with timevarying parameters.
3
FUZZY DYNAMIC GAINSCHEDULED AUTOPILOT
The development of an appropriate gainscheduling law is crucial to the performance of a scheduled system. It is quite often either difﬁcult to ﬁnd or too complicated to design CGS law. However, this problem can be resolved using fuzzy inference techniques. In the following application, two term sets describing the scheduling variables α and M are deﬁned respectively as Tα = {A1 , A2 , A3 , A4 , A5 } , TM = {B1 , B2 , B3 , B4 , B5 } .
(7)
where Ai and Bi are the fuzzy sets characterized by certain membership functions. For the purpose of theoretical studies, the physical domain over which the scheduling variable α takes its crisp value is supposed to be [−25, 25] degrees, and the domain for M is [2, 4] Mach. A Takagi–Sugeno fuzzy model is used next to represent the fuzzy conditional statement between the scheduling variables and controller parameters. This model provides an effective representation of the complex nonlinear relations in terms of fuzzy sets and fuzzy reasoning applied to a set of input–output data. The complex relationship between the scheduling variables and control gains is constructed through 25 fuzzy rules. For a ﬁrstorder Sugeno model [13], the ith fuzzy rule is described by If α is Ali and M is Bjl , then K1l = pl1 α + q1l M + r1l , K2l = pl2 αm + q2l β + r2l ,
(8)
l = 1, · · · , 25 where (α, M ) is the scheduling vector that best ﬁts the description in the premise part of the rule, pl1 , q1l , r1l , pl2 , q2l and r2l are the consequent function parameters. The outputs of this rule are the inferred control gains.
© 2004 by CRC Press LLC
Figure 2 Membership functions of the scheduling variables α(t) and M (t).
To facilitate adaptation to the gainscheduling control model, it is convenient to put the fuzzy model into a framework of adaptive networks that can compute the gradient vectors systematically. Using the Sugeno inference mechanism [13] the fuzziﬁzed control gains K1l and K2l can be expressed by the weighted sum of all strengths of the ﬁred rules: K1 =
w ¯ l (pl1 α + q1l M + r1l ),
l
K2 =
w ¯ l (pl2 α + q2l M + r2l )
(9)
l
for l = 1, · · · , 25 where the normalized ﬁring strength wl w ¯l = l , lw
wl =
µAi (α)µBj (M ),
i,j
with the membership functions µAi (·) and µBj (·) chosen as follows: µAi (α) =
1 , i 2 bi 1 + [( α−c ai ) ]
µBj (M ) =
1 1+
M −c [( aj j )2 ]bj
,
i, j = 1, · · · , 5.
Figure 2 shows the bellshaped membership functions considered in this research. We have constructed an adaptive network that is functionally equivalent to a Sugeno fuzzy model. The consequent function parameters are regulated by the learning algorithm presented in Section 2 when the input information is provided.
© 2004 by CRC Press LLC
Adaptive Parameter Update Law To minimize the performance index deﬁned in (4) over the ﬂight time, we complete the derivations by introducing (9) and rewriting (6) into the following form: ∂J j ∂σsi
t −(ea + K1 ev )
= t0
∂aL j ∂σsi
− ev
∂q j ∂σsi
for s = 1, 2, i = 1, 2, 3, j = 1, . . . , 25, where j j p1 , i = 1 p2 , j j σ1i = q1j , i = 2 , σ2i = qj , j 2j r1 , i = 3 r2 ,
+ ρδ
∂δ j ∂σsi
i=1 i=2 , i=3
+ βi ev
w ¯
l
dτ
(10)
l
α, i = 1 M, i = 2 . βi = 1, i = 3
j For the nonlinear dynamic system (1), the derivative ∂δ/∂σ1i in (10) is determined via the j following derivations. Since δ = δ(e, w, σsi ), thus
T ∂u ∂δ ∂δc ∂w ∂x ∂δ ∂δc ∂q + = − j j j j ∂δc ∂ev ∂σ1i ∂w ∂x ∂σ1i ∂σ1i ∂σ1i
∂α ∂δ ∂σj ∂K1 ∂δ ∂aL ∂q ∂δc c 1i = K2 + ∂α 2 ∂M e − K1 j − ∂q j a j ∂δc j ∂σ1i ∂σ1i ∂σ1i ∂σ1i ∂δ ∂δc ∂α ∂δc ∂q ∂aL l = βi K2 . w ¯ ea − K1 K2 j + + 2 − K2 j j ∂δc ∂α ∂σ1i ∂M ∂σ1i ∂σ1i l
(11) j Next, we evaluate ∂aL /∂σ1i and then proceed with solving the differential equations to j obtain ∂x/∂σsi . Introducing (11), we have
∂aL j ∂σ1i
∂g3 ∂α ∂g3 ∂δ + j j ∂α ∂σ1i ∂δ ∂σ1i ∂α ∂q ∂g3 ∂δ ∂g3 ∂g3 ∂g3 ∂δ ∂δc =Ω +Ω 2 + − K2 j j ∂α ∂δ ∂δc ∂α ∂σ1i ∂M ∂δ ∂δc ∂σ1i ∂g3 ∂δ l + βΩK2i w ¯ ea ∂δ ∂δc l ∂q ∂g3 ∂δ ∂g3 ∂δ l ∂g3 ∂α ∂g3 = 2Ω +Ω 2 + βi ΩK2 w ¯ ea , − K2 j j ∂α ∂σ1i ∂M ∂δ ∂δc ∂σ1i ∂δ ∂δc l (12) =
3 ∂δ −1 . If the precise plant model is not known, the gradient where Ω = (1 + K1 K2 ∂g ∂δ ∂δc ) ∂g3 /∂α can be approximately calculated using the perturbation method, i.e., ∂g3 /∂α ≈ ∆g(α)/∆α. This also applies for ∂g3 /∂δ. From (12), the problem is transformed to obtain j j the derivative ∂x/∂σ1i . Differentiating the term ∂x/∂σ1i with respect to t and applying j ∂δ/∂σ1i given in (11) results in the gradient dynamics:
© 2004 by CRC Press LLC
d ∂x ( )= j dt ∂σ1i
=
∂g1 ∂α ∂α ∂σ j + 1i ∂g 2 ∂α2 ∂αj + ∂σ 1i
2
=
∂α ˙ j ∂σ 1i ∂ q˙ j ∂σ 1i
∂g1 ∂α ∂g1 ∂δ ∂q ∂α ∂σ j + ∂δ ∂σ j + ∂σ j 1i 1i 1i ∂g2 ∂α ∂g2 ∂δ ∂α ∂σ j + ∂δ ∂σ j 1i 1i
∂g ∂g ∂δ 1 − K2 ∂δ1 ∂δ + 2 ∂M1 c ∂g ∂g ∂δ −K2 ∂δ2 ∂δ + 2 ∂M2 c
∂g1 ∂δ ∂aL ∂g1 ∂δ ∂q w ¯ l ea j − K1 K2 ∂δ ∂δc j + βi K2 ∂δ ∂δc ∂σ ∂σ l 1i 1i ∂a ∂g ∂g ∂q 2 ∂δ 2 ∂δ L w ¯ l ea j − K1 K2 ∂δ ∂δc j + βi K2 ∂δ ∂δc ∂σ ∂σ l 1i 1i
.
j Now introducing ∂aL /∂σ1i obtained in (12) yields
1 1 ∂δ ∂g3 − K1 K2 Ω ∂g 2 ∂g d ∂x ∂α ∂δ ∂δ ∂α c ( j ) = ∂g 2 ∂δ ∂g3 dt ∂σ1i 2 ∂α2 − K1 K2 Ω ∂g ∂δ ∂δc ∂α ∂g1 ∂δ l w ¯ βi K2 Φ ∂δ ∂δc l ea , + ∂g2 ∂δ l βi K2 Φ ∂δ ∂δc w ¯
∂g1 1 ∂δ 1 + −K2 ∂g Φ + 2 ∂M ∂δ ∂δ c ∂x j ∂g2 ∂g2 ∂δ ∂σ1i −K2 ∂δ ∂δc + 2 ∂M Φ (13)
l
3 ∂δ where Φ = 1 − K1 K2 Ω ∂g ∂δ ∂δc .
j Similarly, the adaptation law for the consequent parameter σ2i is determined by
∂δ j ∂σ2i
∂δ = ∂δc
j ∂σ2i
+
∂δc ∂w
T
∂w ∂x j ∂x ∂σ2i
∂δc ∂α ∂δc ∂q e + K2 + +2 j v j j j ∂α ∂σ2i ∂M ∂σ2i ∂σ2i ∂σ2i ∂q ∂δ ∂a ∂δ ∂δ ∂α L c c βi . (14) = w ¯ l ev − K1 K2 j + + 2 − K2 j j ∂δc ∂α ∂σ2i ∂M ∂σ2i ∂σ21i l
∂δ = ∂δc
∂δc ∂K2
∂(u − q)
j The term ∂aL /∂σ2i is derived as
∂aL j ∂σ2i
∂g3 ∂α ∂g3 ∂δ + j j ∂α ∂σ2i ∂δ ∂σ2i ∂g3 ∂α ∂q ∂g3 ∂δ ∂g3 ∂g3 ∂δ l = 2Ω w ¯ ev . (15) − K + Ω 2 + β Ω 2 i j j ∂α ∂σ2i ∂M ∂δ ∂δc ∂σ2i ∂δ ∂δc l =
j j j Differentiating ∂x/∂σ2i with respect to t and applying ∂δ/∂σ2i and ∂aL /∂σ2i obtained in (11) and (12) produces the following gradient dynamics:
∂g ∂g ∂g ∂δ + 2 ∂M1 2 1 ∂αj + 1 − K2 ∂δ1 ∂δ c d ∂x ∂α ∂σ2i ( )= ∂g ∂g ∂g j ∂δ dt ∂σ2i 2 ∂α2 ∂αj + −K2 ∂δ2 ∂δ + 2 ∂M2 c ∂σ
2i
© 2004 by CRC Press LLC
∂g1 ∂δ ∂aL ∂g1 ∂δ ∂q w ¯ l ev j − K1 K2 ∂δ ∂δc j + βi ∂δ ∂δc ∂σ ∂σ l 2i 2i ∂g2 ∂δ ∂aL ∂g2 ∂δ ∂q w ¯ l ev j − K1 K2 ∂δ ∂δc j + βi ∂δ ∂δc ∂σ ∂σ l 2i 2i
.
Introducing (15) gives ∂g1 ∂g1 ∂δ 1 1 ∂δ ∂g3 1 + −K Φ − K1 K2 Ω ∂g + 2 2 ∂g 2 d ∂x ∂α ∂δ ∂δ ∂α ∂δ ∂δ ∂M c c ∂x ( j ) = ∂g j ∂g2 ∂δ ∂g2 ∂δ ∂g3 ∂g2 2 dt ∂σ2i ∂σ2i −K2 ∂δ ∂δc + 2 ∂M Φ 2 ∂α − K1 K2 Ω ∂δ ∂δc ∂α ∂g1 ∂δ l 3 ∂δ w ¯ βi 1 − K1 K2 Ω ∂g ∂δ ∂δc ∂δ ∂δc l + (16) ev . ∂g3 ∂δ ∂g2 ∂δ l βi 1 − K1 K2 Ω ∂δ ∂δc ∂δ ∂δc w ¯ l
j j and ∂q/∂σsi are viewed as the state variables of the For the previous derivations, ∂α/∂σsi dynamic equation.
PD Type Controller The previous approach can be directly extended to the proportionalderivative type controller described by u(t) = K1 (α, M )ea (t) + K1d (α, M )e˙ a (t) δc (t) = K2 (α, M )ev (t) + K2d (α, M )e˙ v (t). For the PD controller the corresponding fuzzy control rules are given by: l l l If α is Ali and M is Bjl , then K1l = pl1 α + q1l M + r1l , K1d = pl1d α + q1d M + r1d , l l l l l l l l K2 = p2 α + q2 M + r2 , K2d = p2d α + q2d M + r2d , l = 1, · · · , 25. The inferenced control gains are determined by l l K1 = w ¯ l (pl1 α + q1l M + r1l ), K1d = w ¯ l (pl1d α + q1d M + r1d ), l
K2 =
l
w ¯
l
(pl2 α
l
+
q2l M
+
r2l ),
K2d =
l l w ¯ l (pl2d α + q2d M + r2d ).
(17)
l
j j and ∂u/∂σ2i in (11) and Now considering the situation of t > 0, the derivatives ∂u/∂σ1i (14) should be changed to ∂u ∂ea ∂u1 ∂ u˙ 2 ∂u2 ∂K1 1 = + ≈ e + K + K a 1 1d j j j j j ∂u2 ∂σ1i t ∂σ1i ∂σ1i ∂σ1i ∂σ1i ∂ea ∂u ∂u1 ∂ u˙ 2 ∂u2 1 ∂K1d 1 = + ≈ ea + K1 + K1d , j j j j j ∂u t t ∂σ1di ∂σ1di ∂σ1di ∂σ1di 2 ∂σ1di
where u1 = K1 (α, M )ea , u2 = K1d (α, M )ea , and j p1d , i = 1 j σ1di = qj , i = 2 . 1d j r1d , i=3 j Similarly, the derivative ∂δc /∂σ2i in (14) should be changed to ∂δc ∂ev ∂ev ∂K2 1 ∂δc 1 ∂K2d 1 ≈ e + K2 + K2d , ≈ ev + K2 + K2d , j j v j j j j t t t ∂σ2i ∂σ2i ∂σ2i ∂σ2di ∂σ2i ∂σ2i
© 2004 by CRC Press LLC
Figure 3 Mach number proﬁle.
Figure 4 A sequence of step acceleration command aLc (t) (dashed line) and actual missile acceleration aL (t) (solid line).
where j σ2di
j p2d , = qj , 2d j r2d ,
i=1 i=2 . i=3
Based on these modiﬁcations, the terms used to determine the parameter update algorithm can be derived like those given in (11)–(16). The adaptive nature of AFGS renders it fundamentally different from the traditional gainscheduling schemes. The major advantage is that it avoids the need to manually design a scheduling program or to determine a suitable fuzzy inferencing system. Moreover, since the parameters are adaptable, it is more ﬂexible than traditional gain scheduling in the sense that it is not tied to a particular system dynamics and therefore does not require signiﬁcant modiﬁcation if the plant is altered. It can thus be expected that the controller becomes more robust and more insensitive to aerodynamic variations.
© 2004 by CRC Press LLC
Figure 5 Response of angleofattack α(t).
Figure 6 Tail deﬂection rate q(t).
4
RESULTS AND ANALYSIS
Here, the controller is required to provide robust performance over a wide range of angle of attack and variations in Mach number. The controller is dynamically varying with the evolution of the state of the system. The desired operating range for this example is for Mach numbers from 2 to 4 and angle of attack from −20 deg to 20 deg. The numerical values of various constants used for simulations are listed in Table 1 [7]. These coefﬁcients are valid for the missile travelling between Mach number 2 and 4 at an altitude of 20,000 ft. For simulation purposes, Mach number plotted in Fig. 3 was generated to provide a reasonably realistic missile speed proﬁle. It is assumed to be measurable in real time. Response aL (t) of the closedloop system with our controller to track a series of step commanded acceleration aLc (t) is illustrated in Fig. 4. The corresponding angle of attack α(t) and taildeﬂection rate q(t) are illustrated in Figs. 5 and 6, respectively. Tracking errors of the acceleration and angular velocity loops are illustrated in Figs. 7 and 8, respectively. The step response shows that the time constant is less than 0.047 s, and
© 2004 by CRC Press LLC
Figure 7 Tracking error ea (t) of the acceleration loop.
Figure 8 Tracking error ev (t) of the pitch angular velocity loop.
steadystate error is less than 2%. As one can see from these ﬁgures, the tracking errors converge to acceptable levels after a few periods of the step response. Figures 9 and 10 show, respectively, the transient behavior of the control gains K1 (α, M ) and K2 (α, M ). The performance measure considered in the simulation study is the minimization of the tracking errors and control effort. Figure 11 shows the resulting performance cost measured over the ﬂight time. The performance robustness of our proposed controller is veriﬁed by perturbing aerodynamic coefﬁcients Cm and Cn . Eleven cases were obtained by multiplying the parameters (am , bm , cm , dm ) and (an , bn , cn , dn ) with a factor from 0.5 to 1.5. The resulting acceleration responses compared to the nominal acceleration response are plotted in Fig. 12. We observe reasonable performance variation of the missile under perturbation. From the extensive simulation studies, we can conclude that the AFGS control law possesses wider design freedom and higher ﬂexibility in the controller architecture. A ﬁnely tuned AFGS control law uses less control effort during ﬂight, and therefore, pos
© 2004 by CRC Press LLC
Figure 9 Transient response of gain K1 .
Figure 10 Transient response of gain K2 .
sesses the potential to increase the missile ﬂight speed. Since robustness properties of the design indeed are inherited by our gainscheduled controller, it can be expected that the ﬂight control system will also give satisfactory performance robustness.
5
CONCLUSIONS
This study developed a novel adaptive fuzzy gainscheduling control design method for a nonlinear missile system. The design methodology uses a Takagi–Sugeno fuzzy system to represent the fuzzy relationship between the scheduling variables and controller parameters. It extends the underlying state feedback structure to include nonlinear, timevarying dynamics in combination with an adaptive gainscheduled feedback controller. This technique offers the advantage of performance improvement for illdeﬁned ﬂight dynamics
© 2004 by CRC Press LLC
Figure 11 Value of the performance measure.
Figure 12 Acceleration aL (t) with perturbed Cm and Cn .
through learning to use an adaptive fuzzy inferencing mechanism. Simulation results presented successfully demonstrate for the missile autopilot design.
Acknowledgment This research was sponsored by the National Science Council, Taiwan, R.O.C., under grant NSC 882213E035031.
REFERENCES 1. Rugh, W.J., “Analytical framework for gain scheduling,” IEEE Control Systems Magazine, vol. 11, no. 1, 1991, pp. 79–84. 2. Tan, S., Hang, C. C., and Chai, J.S., “Gain scheduling: from conventional to neurofuzzy,” Automatica, vol. 33, no. 3, 1997, pp. 411–419.
© 2004 by CRC Press LLC
3. Balas, G.J. and Packard, A.K., “Design of robust timevarying controllers for missile autopilot,” Proceedings of the 31st IEEE Conference on Control Applications, Dayton, OH, 1992, pp. 104–110. 4. Shamma, J.S. and Cloutier, J.R., “Trajectory scheduled missile autopilot design,” Proceedings of the 31st IEEE Conference on Control Applications, Dayton, OH, 1992, pp. 237–242. 5. White, D.P., Wozniak, J.G., and Lawrence, D.A., “Missile autopilot design using a gain scheduling technique,” Proceedings of the 26th Southeastern Symposium on System Theory, Athens, OH, 1994, pp. 606–610. 6. Piou, J.E. and Sobel, K.M., “Application of gain scheduling eigenstructure assignment to ﬂight control design,” Proceedings of 1996 IEEE International Conference on Control Applications, Dearborn, MI, 1996, pp. 101–106. 7. Reichert, R.T., “Robust autopilot design using µsynthesis,” Proceedings of American Control Conference, San Diego, CA, 1990, pp. 2368–2373. 8. Nichols, R.A., Reichert, R.T., and Rugh, W.J., “Gain scheduling for H∞ controllers: a ﬂight control example,” IEEE Transactions on Control Systems Technology, vol. 1, no. 2, 1993, pp. 69–79. 9. Schumacher, C. and Khargonekar, P.P., “Missile autopilot designs using H∞ control with gain scheduling and dynamic inversion,” AIAA Journal of Guidance, Control, and Dynamics, vol. 21, no. 2, 1998, pp. 234–243. 10. Shamma, J.S. and Cloutier, J.R., “Gainscheduled missile autopilot design using linear parameter varying transformations,” AIAA Journal of Guidance, Control, and Dynamics, vol. 16, no. 2, 1993, pp. 256–263. 11. Apkarian, P., Gahinet, P., and Becker, G., “Selfscheduled H∞ control of linear parametervarying systems: a design example,” Automatica, vol. 31, no. 9, 1995, pp. 1251–1261. 12. Devaud, E., Harcaut, J.P., and Siguerdidjane, H., “Threeaxes missile autopilot design: from linear to nonlinear strategies,” AIAA Journal of Guidance, Control, and Dynamics, vol. 24, no. 1, 2001, pp. 64–71. 13. Takagi, T. and Sugeno, M., “Fuzzy identiﬁcation of systems and its applications to modeling and control,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 15, no. 1, 1985, pp. 116–132. 14. Driankov, D., Hellendoorn, H., and Reinfrank, M., An Introduction to Fuzzy Control, Springer, Berlin, 1993, pp. 186–195. 15. Gonsalves, P.G. and Zacharias, G.L., “Fuzzy logic gain scheduling for ﬂight control,” Proceedings of the 3rd IEEE Conference on Fuzzy Systems, Orlando, FL, 1994, pp. 952–957. 16. Adams, R.J., Sparks, A.G., and Banda, S.S., “A gain scheduled multivariable design for a manual ﬂight control system,” Proceedings of the 31st IEEE Conference on Control Applications, Dayton, OH, 1992, pp. 584–589. 17. Tanaka, T. and Aizawa, Y., “A robust gain scheduler interpolated into multiple models by membership functions,” AIAA Paper 924553, Aug. 1992. 18. Jonckheere, E.A., Yu, G.R., and Chen, C.C., “Gain scheduling for lateral motion of propulsion controlled aircraft using neural networks,” Proceedings of American Control Conference, Albuquerque, NM, 1997, pp. 3321–3325. 19. Luenberger, D.G., Optimization by Vector Space Methods, John Wiley & Sons, Inc., New York, 1969, pp. 297–299.
© 2004 by CRC Press LLC
6 Model Predictive Control of Nonlinear Rotorcraft Dynamics with Application to the XV15 Tilt Rotor Raman K. Mehra and Ravi K. Prasanth Scientiﬁc Systems Company Inc., Woburn, MA
Rotorcraft development presents unique control challenges related to noise, vibration, and ﬂight. Model Predictive Control (MPC) is proposed as a holistic approach to rotorcraft ﬂight control capable of explicitly accounting for operational constraints. The ﬁrst part of this chapter describes MPC design using linear parametervarying (LPV) approximation of nonlinear dynamics. This procedure has several advantages including known quality of approximation and polynomialtime computability. Details of MPC problem formulation and its implementation are presented. As a ﬁrst application, we consider active control of aeromechanical instability, which is a major hurdle in the development of softinplane rotorcraft. A simple nonlinear parametervarying model that contains a Hopf bifurcation is used to capture the essential features of this instability. The effectiveness of MPC in suppressing nonlinear vibrations subject to control constraints is illustrated. As a second and more involved application, we present the design of an MPCbased ﬂight control system (FCS) for the XV15 tilt rotor. The FCS was implemented on a 6DOF highﬁdelity XV15 realtime simulator. Test pilots ﬂew several missions on the simulator with MPC in the loop including a highly nonlinear conversion maneuver. They observed signiﬁcant reductions in work load, fast response to commands, and rated MPC performance from good to excellent during these simulated maneuvers.
1
INTRODUCTION
NASA, DoD, and the aerospace industry have a very strong interest in the development of advanced rotorcraft. Some unique challenges related to noise, vibration, and ﬂight control are present in their development. A good example comes from tilt rotors, which can hover like helicopter and cruise like a turboprop. When the pilot presses the pedals while in helicopter mode, the tip path plane of one rotor tilts forward and that of the other tilts aft producing a yaw moment. The same pedal input produces a rudder deﬂection in airplane mode. The ﬂight control system must accommodate these requirements and provide a smooth transition during the conversion process.
© 2004 by CRC Press LLC
Rotorcraft ﬂight control system (FCS) is currently designed in a pointbypoint manner using linear principles. Gain scheduling is used to phasein and phaseout control effects during ﬂight. The ﬂight control computer is also responsible for controlling structural loads on the aircraft. This is achieved, in the V22 for example, by limiting control surface commands (displacements and rates) produced by pilot stick inputs. Apart from being a complex task, control rate and position limiting have serious effects on aircrafthandling qualities. Pilots evaluating FCS with rate limiting usually mark them down. In fact, control rate limiting is a known trigger mechanism for aircraftpilot coupling (APC) or pilotinduced oscillation (PIO), which can cause fatalities. Model Predictive Control (MPC) is a multiinput multioutput digitalcontrol design methodology [6, 7] that uses a highly intuitive principle. At any time step, the control inputs to be applied depend on the current state, control objectives to be achieved, and the current operational constraints. So, a rational procedure is to formulate a constrained optimization problem at each time step and obtain control inputs by solving it. This has the advantage of explicitly accounting for time domain control constraints such as rate limits and statedependent constraints such as limits on normal acceleration. Frequency domain speciﬁcations can also be brought in through MPC design parameters discussed later. Since the control objectives and operational constraints are always updated and accounted for, MPC can provide a holistic approach to rotorcraft ﬂight control. Our aim is to demonstrate this using realistic control problems. This chapter is organized as follows. The next section describes the MPC problem and implementation in detail. For nonlinear systems, the online optimization problem formulated by MPC is generally nonconvex and cannot be solved in real time, especially at the rates required for ﬂight control. We say that a problem is solvable if there is an anytime algorithm [13] that terminates in polynomial time for a given accuracy. A procedure for designing realtime implementable MPC using linearparameter varying (LPV) approximations of rotorcraft dynamics is presented in the next section. Section 3 gives a brief description of active control of aeromechanical instability using MPC. Our results show the effectiveness of MPC in suppressing vibrations in the presence of actuator saturation limits. Section 4 presents the design of an MPCbased ﬂight control system (FCS) for the XV15 tilt rotor. FCS design speciﬁcations include frequencydomain MILF83300 and MILF8785C requirements and timedomain safety and control limits. MPCbased FCS was implemented at 50 Hz on a realtime 6DOF XV15 simulator at Bell Helicopter. Results of piloted simulations and evaluations by test pilots are presented. The chapter concludes with some recommendations for future work. Due to space limitations, some details have been omitted; a complete version of this chapter can be obtained from the authors.
2
MODEL PREDICTIVE CONTROL OF NONLINEAR SYSTEMS
This section begins with a description of the control problem and the MPC approach. We then indicate the computational difﬁculties associated with MPC and present an algorithm that can be readily implemented. The algorithm uses a linear parameter varying (LPV) approximation of the nonlinear system to reduce the computational requirements signiﬁcantly from those for nonconvex nonlinear programming to that of convex quadratic programming (QP). Thus, our approach trades off optimality for polynomialtime computability.
© 2004 by CRC Press LLC
The MPC implementation is described in detail. Although the presentation is geared toward rotorcraft applications, the algorithm can be applied to problems in other areas as well.
2.1
Control Problem
Consider the nonlinear system: xk+1 = f (xk , uk ) yk = g(xk ) zk = h(xk ) ulbd ≤ uk ≤ uubd , for all k zlbd ≤ zk ≤ zubd , for all k
(1a) (1b) (1c) (1d) (1e)
where xk ∈ IRnx is the state vector, uk ∈ IRnu is the control input, yk ∈ IRny is the signal that must track a speciﬁed command, and zk ∈ IRnz is the vector of signals that are constrained (the control inputs are not included in this vector). It is possible to make the constraint bounds ulbd , zlbd , uubd and zubd depend on time, but for simplicity they are taken as constants. As part of the control problem, we are also given a desired trajectory or k commanded trajectory ycmd . Control Problem: Generate a feedback control law so that yk tracks the desired k trajectory ycmd subject to the system dynamics and constraints given in (1). MPC solution to this problem is the following: At time k, deﬁne the performance index J as follows: J(xk , δuk , δuk+1 , · · ·, δuk+N −1 ) = Ji + Jp + Ju N −1 l T k+j+1 k+j+1 Ji = ycmd − yk+j+1 Qi ycmd − yk+j+1 N −1 i=0
(2b)
j=0
l=0
Jp =
(2a)
k+i+1 − yk+i+1 ycmd
Ju = δuTk Rδuk +
N −2
T
k+i+1 Q ycmd − yk+i+1 T
(δuk+i+1 − δuk+i ) R (δuk+i+1 − δuk+i )
(2c)
(2d)
i=0
and minimize it over control moves subject to the system dynamics and constraints in (1). Apply the control input uk so obtained, increment k, and repeat the procedure. The terms Ji and Jp in (2b) and (2c) can be thought of as discrete versions of integral error and tracking error penalties over the control horizon. These terms are expected to yield a proportionalintegral (PI) effect in the closedloop system. Like in classical control, the weights Qi and Q appearing in (2b) and (2c) can be tuned to obtain satisfactory steady state and transient tracking performances. The third term Ju in (2d) is a penalty for excessive control rate which smoothes out the control inputs. The control inputs are also constrained to satisfy the saturation limits (1d). Note that the performance index requires knowledge of the desired or commanded k trajectory ycmd over the prediction horizon. When the application is rotorcraft stability
© 2004 by CRC Press LLC
and control augmentation, the commands are pilot stick inputs passed through a command k+1 k+N shaping ﬁlter. So, their future values ycmd , · · ·, ycmd needed in the performance index will not be known at time k. Therefore, we have to make an assumption, and the simplest (in terms of computations) assumption is that k+1 k+N k = ycmd = · · · = ycmd . ycmd
This may also be reasonable if the horizon length is chosen to be relatively small. For example, when aircraft dynamics is discretized at 0.02 seconds (50 Hz), a horizon length N = 10 corresponds to 0.2 seconds (5 Hz), which is comparable to the pilot’s bandwidth k so that changes in the command inputs from its current value of ycmd during the horizon of interest can be neglected. We may be able to extrapolate past pilot commands by ﬁtting polynomials to arrive at better estimates, but this will not be used here. In other applications such as autopilot design, the trajectory is deﬁned by the controller itself and, hence, will be known over the prediction horizon. Both problems ﬁt into the MPC framework and the only difference is in the data used by MPC.
2.2
Linear ParameterVarying Approximation
When the tracked and constrained variables are written in terms of the control inputs (which are the unknowns), the optimization problem to be solved online becomes nonconvex and corresponds to a general nonlinear programming problem. This difﬁculty is usually circumvented by linearizing about the current state. The linearized system leads to a convex quadratic programming (QP) that can be readily solved. However, this procedure is ﬂawed for a number of reasons: (a) linearization is performed about a nonequilibrium state, but the linearized system is treated as timeinvariant, (b) the quality of approximation provided by linearization is not known because it depends on the states and their derivatives, (c) plant model and state estimates have uncertainties, and (d) linearization requires the solution of nonlinear algebraic equations, which can be difﬁcult. Here, we use a linear parameter varying (LPV) approximation of the nonlinear system: δxk+1 = A(θk )δxk + B(θk )δuk δyk = Cy (θk )δxk δzk = Cz (θk )δxk xk = xtrim (θk ) + δxk , uk = utrim (θk ) + δuk yk = ytrim (θk ) + δyk , zk = ztrim (θk ) + δzk ulbd ≤ utrim (θk ) + δuk ≤ uubd , k = 0, 1, · · ·, N − 1 zlbd ≤ ztrim (θk ) + δzk ≤ zubd , k = 1, 2, · · ·, N ,
(3a) (3b) (3c) (3d) (3e) (3f) (3g)
where θk ∈ IRnθ is the scheduling variable (timevarying parameter). This parameter deﬁnes the ﬂight condition, and the set of all its possible values deﬁnes the ﬂight envelope. There are several ways to obtain an LPV approximation, some of which depend on the application. See [9, 10, 11, 12] for details on LPV systems. A procedure that conﬁrms to the current practice in aerospace is the following:
© 2004 by CRC Press LLC
1. Subdivide the ﬂight envelope into L polytopic regions. Each polytope is made of m vertices. Let the total number of vertices be M . Let us order the vertices from 1 to M and denote by θk the kth vertex. 2. At each of the M vertices, trim the aircraft equations of motion (or the nonlinear system’s dynamical equations) to obtain trim states, control settings, tracked variables and constrained variables. Thus at θk we obtain: xtrim k , utrim k , ytrim k and ztrim k by solving the algebraic equation obtained from the dynamical equations in (1). 3. At the trim point (θk , xtrim k , utrim k , ytrim k , ztrim k ), linearize the system in (1). This gives a linear system of the form: δxk+1 = Ak δxk + Bk δuk δyk = Cyk δxk δzk = Czk δxk .
(4a) (4b) (4c)
4. The collection
θk , xtrim k , utrim k , ytrim k , ztrim k , Ak , Bk , Cyk , Czk
M k=1
forms the LPV data. This is done offline and the quality of approximation can be checked over admissible trajectories of scheduling variables.
2.3
MPC Implementation
With the LPV approximation in hand, we are now ready to describe the MPC implementation. There are three sequential steps: 1. Interpolation: Interpolate system data to obtain a linear system model valid at the current value of scheduling variable. 2. QP formulation: Perform computations using interpolated system data and MPC design parameters to set up quadratic programming (QP) problem. 3. QP solution: Solve QP problem and obtain optimal control input. The objective of the ﬁrst step is to obtain a linear system model that is valid locally. This model depends on the scheduling variable and corresponds to the linear approximation about the trim point associated with the scheduling variable. Thus, we avoid linearization about a nonequilibrium point. In general the interpolation is difﬁcult to perform because the ﬂight envelope need not be a regular region. We shall describe an efﬁcient algorithm to approximately interpolate. The second step as stated uses the same statespace model over the prediction horizon. It is possible to combine steps 1 and 2 so that we use an LPV model over the prediction horizon. This will result in better performance. We have separated these steps to keep coding simple. The last step of QP solution is the most computationally intensive task. It takes over 99% of the total computing required in each step of MPC.
© 2004 by CRC Press LLC
2.3.1
Interpolation
The purpose of this task is to determine a linear model that is valid at the current value of scheduling variable θk . We use the following approximate interpolator: (a) Deﬁne an M × M diagonal matrix of weights W : Wii = e−σ(Vi −θk )
T
(Vi −θk )
.
This weight gets larger for vertices close to the current point θk and smaller for those further away. (b) Deﬁne M × 1 vector α:
V1T V2T α=W · [V1 , V2 , · · ·, VM ] W · T VM
−1 V1T V2T · θk . · T VM
This α satisﬁes [V1 , V2 , · · ·, VM ] α = θk . (c) Set data at current point as: Datacurr = [Data1 , Data2 , · · ·, DataM ] α. 2.3.2
QP Problem Formulation
At the end of interpolation, we have a linear system of the form: δxk+1 = Acurr δxk + B curr δuk δyk = Cy curr δxk δzk = Cz curr δxk ulbd ≤ utrim curr + δuk ≤ uubd , k = 0, 1, · · ·, N − 1 zlbd ≤ ztrim curr + δzk ≤ zubd , k = 1, 2, · · ·, N,
(5a) (5b) (5c) (5d) (5e)
which is used in this task along with the performance index (2) to formulate a QP problem. The computations involved are algebraic and easy. The QP problem has the form: ˆ + v T gˆ min v T Hv
(6a)
ˆ ≤ Uu , subject to Ul ≤ Av
(6b)
v
ˆ gˆ, Ul , Aˆ and Uu are computed from the LPV statespace model (5). where H, 2.3.3
QP Solver
ˆ The optimization problem in (6) is always convex because of the positivity of the matrix H, which comes from the fact that the control rates are penalized. As a result, this problem can
© 2004 by CRC Press LLC
Figure 1 Block diagram for active control of aeromechanical instability using MPC.
be solved in polynomial time. That is, there are algorithms (e.g., ellipsoid method, primaldual interior point) whose computational cost is a polynomial function of the number of variables, number of constraints, and the (inverse of) the accuracy required. Moreover, there are algorithms that are anytime, i.e., they can be stopped at any time to get an answer, the accuracy increases with computational effort, and can be restarted with little overhead. We use a primaldual interior point method for realtime implementation [8].
3
APPLICATION TO NONLINEAR VIBRATION CONTROL
Consider the nonlinear system: x˙ f (RPM) = y˙ ω
−ω x 1 + (x2 + y 2 ) + u, f (RPM) y 0
where x and y are the states, f (RPM) = −0.15 + 0.45e−5(
RP M 900
2
−1)
and RP M is the scheduling variable. The scalar control input u is bounded above and below: umin ≤ u ≤ umax With u = 0, the system is in normal form and, by a Theorem of Hopf, has a Hopf bifurcation at RPM = 900 leading to a stable limit cycle. We are interested in the system behavior as RPM changes with time. As RPM increases from 0 to 900, due to the nature of f the (linear) damping of the system decreases and the system ampliﬁes displacements. Beyond RPM = 900, the system damping increases and oscillations become smaller. This behavior imitates a particularly severe vibration problem in rotorcraft known as aeromechanical instability. Aeromechanical instabilities, ground and air resonances, are major design issues and limiting factors in the development of softinplane rotorcraft. Ground resonance, observed in multibladed (more than 2 blades) softinplane rotorcraft, occurs as the rotor RPM is
© 2004 by CRC Press LLC
Figure 2 Open and closedloop behaviors of aeromechanical system as RPM changes linearly from 600 to 1200. Limit cycling in open loop (left), vibration suppression with MPC (middle), and control inputs generated by MPC (right).
changed from rest to its operational value and is manifested as large uncontrolled rotor tip displacements. Tilt rotor aircraft in airplane mode can, in addition, encounter coupled wingrotorpylon instabilities known as air resonance. These instabilities are parameter dependent in that, as the operating conditions change, the rotorcraft changes from stable to unstable behavior. The primary design tool currently in use is frequency placement of vibrational modes by increasing stiffness. This passive approach invariably increases structural weight and reduces aerodynamic performance. The resulting reduction in operational envelope has prompted the rotorcraft industry to vigorously pursue active control of aeromechanical instabilities. See [3, 4, 5] for more details. Typical control objectives are to maintain rotor tip displacements within safety limits (or completely cancel when there is no noise or uncertainty) subject to saturation limits on swashplate motion along any smooth scheduling variable trajectory. To design MPC, we formulate the control problem as shown in Fig. 1. Since the objective is to make x go to zero asymptotically along smooth RPM trajectories, we introduce a desired trajectory, namely zero, and an integrator to generate tracking error and its integral. The openloop system augmented with these elements forms the plant for MPC. The next task is to obtain an LPV approximation of the plant, which becomes the internal model used by MPC. This approximation and weights in the performance index (2) form the data read in by the MPC solver prior to closedloop simulation. Figure 1 shows the complete closedloop system. Finally, MPC design parameters are tuned using the closedloop simulation. Figure 2 shows open and closedloop responses. On the left is the open loop response that is typical of aeromechanical instability. Closedloop response is shown in the middle. The effectiveness of MPC in canceling the instability is clear. The control input on the right clearly indicates how MPC is working. These simulations are for a linearly varying RPM trajectory typical of rotorcraft startup.
© 2004 by CRC Press LLC
Figure 3 Block diagram representation of MPCbased FCS. Shaded block is toggled by pilot for SCAS or autopilot.
4
APPLICATION TO THE XV15 TILT ROTOR FLIGHT CONTROL
This section presents the design, realtime implementation, and piloted simulations of an MPCbased ﬂight control system (FCS) for the XV15 tilt rotor. Current FCS design practices are based on linear timeinvariant singleinput singleoutput ideas and do not account for nonlinearities or timedomain constraints. We shall show that the MPC approach can accommodate frequencydomain MIL speciﬁcations as well as timedomain requirements. A highﬁdelity XV15 simulator developed at Bell Helicopter using the Generic Tilt Rotor (GTR) model is used for evaluating MPCbased FCS. XV15/V22 test pilots ﬂew several missions evaluating various aspects of the design. This section concludes with their evaluation.
4.1
MPC Architectures for FCS
A block diagram of MPCbased FCS is shown in Fig. 3. It can be adapted for two cases that we shall consider by placing a reference model or a guidance command generator in the shaded block. In the ﬁrst case, MPC interprets pilot stick inputs as desired rate commands and computes the optimal control surface deﬂections needed to track these commands. MPC can be seen as a Stability and Control Augmentation System (SCAS) in this case. The second case corresponds to an autopilot wherein the pilot can enter desired terminal conditions and MPC generates the guidance commands as well as the stick movements to follow the guidance track.
© 2004 by CRC Press LLC
4.2
Design Procedure
The MPC design parameters for the structure in Fig. 3 are: reference model dynamics, penalties on tracking errors, integrated tracking errors and actuator rates, and horizon length. These parameters were selected as follows: 1. Selection of reference models: This involves the selection of simple loworder dynamic systems that interpret pilot stick inputs and generate a desired rate response as shown in Fig. 3. The frequency and time response of these reference models must satisfy all relevant MIL speciﬁcations. Since MPC is a timedomain technique, timedomain speciﬁcations such as glimits and controlsurface deﬂection limits are stated as they are and accounted for in the online optimization. Ideally (in the absence of disturbances, model mismatch, and unbounded controls), MPC design will match the desired rate responses exactly so that the pilot will in effect be ﬂying the reference model. Our experience suggests that decoupled reference models — one for each body angle rate — is better than coupled models from the pilot’s perspective. This makes the selection somewhat easier. 2. Formulation of augmented plant and LPV data generation: After selecting the reference models, we generate the tracking error and the integral of tracking error as shown in Fig. 3 by adding integrators. The augmented plant for MPC is given by the system from pilot stick inputs and control inputs to the measurements, tracking error, and its integral. The augmented plant is approximated by an LPV model that forms the system data used by MPC. 3. Selection of weighting matrices and horizon length: For the LPV system, assume that there are no state and control constraints. Design LQG controllers that give good performance by adjusting the weights in the performance index in (2). The LQG weights form the initial set of weights for MPC design. The weights and horizon lengths can be further tuned to balance closedloop performance and computational requirements.
4.3
Maneuvers for Evaluation
Two maneuvers will be used to evaluate MPCbased FCS in terms of its ability to track pilot commands and reduce pilot workload. Coordinated Turns: The maneuver begins in airplane mode trimmed at 150 knots. The pilot initiates turn by moving the lateral stick. MPCbased FCS interprets pilot stick inputs as rate commands and provides fast tracking with zero steadystate error and no overshoot for step inputs (we let the pilot ﬂy in actual tests). Level 1 ﬂying qualities requirements for this maneuver can be found in MILF8785C [1]. ConversionDecel: The maneuver involves converting XV15 from airplane mode at 150 knots to helicopter mode at zero knots. The maneuver starts at, say, 2500 ft, and during the maneuver the aircraft descends at an acceptable rate to about 1800 ft. At the end of the maneuver, the aircraft hovers at a constant altitude. This maneuver is highly nonlinear as the aircraft dynamics undergo large changes. Controller design is also complicated by the fact that there are three modes of operation — airplane mode, conversion, and hover — that must be considered. Our objective is to design a controller that would allow the pilot to conduct this maneuver with hands off the stick. In addition, we would like to provide the pilot with the ability to regain control of the aircraft at any time. Level 1 ﬂying qualities
© 2004 by CRC Press LLC
Figure 4 Piloted coordinated turn simulation: pilot stick inputs and angular rates. Meansquare value of baseline (dashed line) stick inputs are much higher than that for MPCbased FCS (solid line) indicating higher workload.
requirements for this maneuver were developed from MILF8785C [1] and MILF83300 [2]. 4.4
Piloted Simulations
MPCbased FCS was evaluated by V22/XV15 test pilots at Bell Helicopter. The pilots ﬂew simulations with and without MPC. A few initial runs with a digital SCAS were made to familiarize the pilots with the XV15 simulator cab. During these runs, the pilots noted some differences between the GTR simulator model and actual aircraft. These include higher pitchpower coupling and more delays in the simulation than in the aircraft. For the evaluation maneuvers, the collectivepower lever was released to the pilots on their request; MPC computed all four control inputs, but the computed collective stick was not fed back.
© 2004 by CRC Press LLC
Figure 4 (Continued).
Figure 4 shows sample results from left and right turn simulations. The effectiveness of MPC in reducing pilot work load is very clear. The pilot was happy with the way MPC performed, especially in coordinating the pedal. He noted that very small/nopedal inputs were required, the aircraft responded well to commands, and rated the controller performance as excellent.
© 2004 by CRC Press LLC
Figure 5 Piloted conversion/decel maneuver with MPCbased FCS in the loop. Note the large variations in airspeed from airplane mode to helicopter mode.
© 2004 by CRC Press LLC
The conversiondecel maneuver began at approximately 145 knots, 2500 feet altitude and 50% power. After bringing the aircraft closer to the landing area, the pilot began converting at which time the MPCcontroller was turned on. Sample results are shown in Fig. 5. The pilot noted that the pitch axis control was tight and reasonably good, although the pitch attitude excursions during the maneuver were at times about 4 degrees more than what he would have liked. Over several runs, he rated MPC performance from good to excellent.
5
CONCLUSIONS
We have described a model predictive control paradigm that uses a linear parametervarying approximation of nonlinear systems instead of online linearization. This method has several advantages, the most notable of which are its quality of approximation and polynomialtime computability. These features are essential for successful ﬂight control design. As a demonstration, MPC was implemented on a realtime XV15 tilt rotor ﬂight simulator. Even though the simulator math model is highly nonlinear, the implementation performed very well in piloted simulations. A coordinated turn maneuver was used to evaluate the multiaxis capability of MPCbased SCAS; while a highly nonlinear conversion maneuver was used to evaluate the autopilot capabilities of MPC. Test pilots rated MPC performance from good to excellent in both cases. MPCbased ﬂight control system design used several MILF83300 and MILF8785C requirements as well as timedomain saturation type constraints. However, a number of issues such as veriﬁcation of these requirements were not considered in this study. We have identiﬁed the following areas for future research and development: 1. Further Piloted Simulations and HardwareintheLoop (HIL) Testing of MPCbased FCS — During the piloted simulations, the test pilots suggested that further simulations are needed for a complete evaluation of MPCbased FCS. This is also the natural step before HIL testing. Future work should include (a) piloted simulations, (b) implementation of MPCbased FCS on a ﬂight computer, and (c) piloted simulations with a ﬂight computer in the loop. 2. Applications of Virtual Tilt Rotor to Predictive Control1 — This is a nonlinear dynamical model of a tilt rotor that resides in the ﬂight computer and is executed at rates faster than real time. A highﬁdelity virtual tilt rotor model has been developed by Bell Helicopter. This concept has many applications including predictive control. Future work should consider (a) MPC implementations with the virtual tilt rotor as the internal model, and (b) computational requirements for realtime implementation of MPCbased FCS. 3. Multivariable Control of Quad Tilt Rotor (QTR) — QTR is the next generation of tilt rotors being developed by the rotorcraft industry. The QTR has more control inputs than the number of equations/states and represents signiﬁcant control challenges and opportunities for reconﬁguration under control failures. 1 Proposed
by Dr. Richard Bennett of Bell Helicopter.
© 2004 by CRC Press LLC
Acknowledgments This work was funded by NASA Ames Research Center under contract NAS298022. We are grateful to Mr. Stephen Jacklin of NASA Ames for support and encouragement during the project. The realtime simulations were ﬂown at Bell Helicopter Textron Inc. by V22/XV15 test pilots Thomas Warren and Roy Hopkins. Their comments and insights were of great help to us in evaluating the controller as well as in identifying future research issues. Dr. Richard Bennett, David Neckels, and Dr. Mark Wasikowski of Bell Helicopter also contributed to this project. We thank them for their efforts.
REFERENCES 1. Military speciﬁcation ﬂying qualities of piloted airplanes, MILF8785C, November 1980. 2. Military speciﬁcation ﬂying qualities of piloted V/STOL airplanes, MILF83300, December 1970. 3. W.Y. Chan and I. Chopra, “Aeromechanical stability of hingeless helicopter rotors in forward ﬂight,” AHS specialists meeting, March 1991. 4. F. Gandhi and W.H. Weller, “Active aeromechanical stability augmentation using fuselage state feedback,” AHS specialists meeting, 1997. 5. D. Hodges, “An aeromechanical stability analysis for bearingless rotor helicopters,” AHS specialists meeting, 1978. 6. M. Kothare, V. Balakrishnan and M. Morari, “Robust constrained model predictive control using linear matrix inequalities,” Automatica, Vol. 32, No. 10, 1996. 7. D.Q. Mayne, “Constrained model predictive control: stability and optimality,” Automatica, Vol. 36, No. 6, 2000. 8. Yu. Nesterov and A. Nemirovskii, “Interiorpoint polynomial methods in convex programming,” SIAM, Philadelphia, 1994. 9. A. Packard, “Gain scheduling via linear fractional transformations,” Systems and Control Letter, Vol. 22, 1994. 10. W. Rugh and J. Shamma, “Research on gain scheduling,” Automatica, Vol. 36, No. 9, 2000. 11. J. Shamma and M. Athans, “Guaranteed properties of gain scheduled control for linear parameter varying plants,” Automatica, Vol. 27, 1991. 12. J. Shamma and J.R. Cloutier, “A linear parameter varying approach to gain scheduled missile autopilot design,” Proceedings of ACC, 1992. 13. S. Zilberstein, “Using anytime algorithms in intelligent systems,” AI Magazine, Fall 1996.
© 2004 by CRC Press LLC
7 New Development of Vector Lyapunov Functions and Airplane Control Synthesis Lyubomir T. Gruyitch University of Technology Belfort–Montbeliard, Belfort Cedex, France
The chapter presents a conceptual development of vector Lyapunov functions by introducing (semi) deﬁnite vector functions and related various kinds of vector Lyapunov functions. They enable us to extend the Lyapunov method without using a scalar Lyapunov function. Besides, they permit synthesis of robust control of airplanes and other nonlinear systems incorporating all mechanical systems with a variable mass and/or loads (e.g., robots, ships, spacecrafts) to guarantee a highquality output tracking together with an appropriate Lyapunov stability of a desired motion.
1
INTRODUCTION
The concept of vector Lyapunov functions was coincidentally introduced by Bellman [1] and Matrosov [2]. It appeared very useful for stability studies of complex systems (in particular of interconnected and largescale systems), [3]–[9]. It was used for a system aggregation, for an effective construction of a scalar Lyapunov function of the overall system and for reduction of a stability study to that of subsystems and to analysis of their interconnections. The use of a scalar Lyapunov function (with or without the comparison principle) characterized such applications of the vector Lyapunov function concept. The concept of vector Lyapunov functions will be further developed in order to solve a basic problem of control synthesis to simultaneously satisfy stability and output tracking requirements in two different spaces. It enables synthesis of new robust stabilizing and/or output tracking control without a need to use either a scalar Lyapunov function or the comparison principle. This will be achieved in what follows for airplaines and a large class of other nonlinear dynamical systems incorporating mechanical systems with variable mass and/or subjected to actions of variable external perturbations, which are both allowed to be unpredictable and unknown. All system nonlinearities, except those resulting from the inertia matrix, can be also unknown. Various kinds of vector Lyapunov functions will be introduced and deﬁned, for which we will introduce and also deﬁne different types of (semi) deﬁnite vector functions. Then, a particular form of a vector Lyapunov function will be introduced. It will enable us to synthesize control that will guarantee global robust
© 2004 by CRC Press LLC
exponential stability of a desired motion or both its global robust asymptotic stability and global robust exponential output tracking.
2
VECTOR LYAPUNOV FUNCTION CONCEPT
2.1
Semideﬁnite and Deﬁnite TimeIndependent Vector Functions
Let S be a subset of R2m and ∂S, ClS, InS, N (S) be, respectively, its boundary, closure, interior and neighborhood, S ⊂ N (S). N (S) − S designates the set difference: N (S) − InS = {e1 : e1 ∈ N (S), e1 ∈ / InS}. Let t denote time, e = [e1 e2 ... em ]T , e(1) = T T T (1) de 1 T e(1) ]T ∈ R2m , e1i = [ei ei ]T ∈ R2 , and e2 = [e1 e(2) ]T ∈ R3m . dt , e = [e All vector relationships and operations hold elementwise. 0 denotes either scalar zero, or vector or matrix zero of an appropriate dimension. Deﬁnition 1 A vector function w(.) : R2m → Rp , p ∈ {1, 2, ..., 2m}, w(e1 ) = [ω1 (e1 ) ω2 (e1 ) ... ωp (e1 )]T , is: a) positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if (i) through (iii) hold, respectively, over some neighborhood N (S) of the set S: (i) w(.) is deﬁned and continuous on N (S) : w(e1 ) ∈ C[N (S)], (ii) w(e1 ) ≥ 0 {w(e1 ) ≤ 0}, ∀e1 ∈ [N (S) − InS], (iii) w(e1 ) ≤ 0 {w(e1 ) ≥ 0}, ∀e1 ∈ ClS. b) pairwise positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if a) holds, respectively, and both (i) and (ii) are valid: (i) p = m, (ii) ωl (e1 ) ≡ ωl (e1l ), ∀l = 1, 2, ..., m. c) elementwise positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if a) holds, respectively, and both (i) and (ii) are valid: (i) p = 2m, (1) (ii) ωl (e1 ) ≡ ωl (el ) and ωm+l (e1 ) ≡ ωm+l (el ), ∀l = 1, 2, ..., m. The preceding properties are: d) global (in the whole) if and only if they hold for N (S) = R2m . e) on a set A, A ⊂ R2m , if and only if A is a neighborhood of S and A ⊆ NL (S), where NL (S) is the largest neighborhood N (S) of S, which obeys the corresponding conditions under a), or b), or c). The expression “relative to a set S, S ⊂ R2m ,” should be omitted if and only if the set S is singleton O : S = O = {e1 : e1 = 0} Let: i, j, k, l, m, n, p, q, r ∈ {1, 2, ...}, min{il , kl , jl , nl } ≥ 1, max{il , kl , jl , nl } ≤ l=p m, mll = 1, ∀l ∈ {1, 2, ..., m}, i1 = j1 = 1, kp = nq = m and l=1 mil kl = m if p ≤ m. Otherwise, the sum ends with l = m. Let eik = (e1 ... ei−1 ei ei+1 ... ek ek+1 ... T em )T , Sik = { eik : eik = ( ei ei+1 ... ek−1 ek )T ∈ Rmik =⇒ ∃e1 = (eTik e(1) )T ∈ S}, (1)
(1)
(1)
(1)
Sik = { eik : eik = ( ei
(1)
1 = { e1ik : e1ik = ( eTik and Sik S(.) ∈ {S, Sb , Sc , Sd }.
© 2004 by CRC Press LLC
(1)
(1)T
(1)
ei+1 ... ek−1 ek )T ∈ Rmik =⇒ ∃e1 = (eT eik )T ∈ S} (1)T
(1)T
eik )T ∈ R2mik =⇒ ∃e1 = (eTik eik )T ∈ S}. Let
Deﬁnition 2 A vector function w(.) : R2m → Rp , p ∈ {1, 2, ..., 2m}, w(e1 ) = [ω1 (e1 ) ω2 (e1 ) ... ωp (e1 )]T , is: a) positive {negative} deﬁnite relative to a set S, S ⊂ R2m , if and only if (i) through (iv) hold, respectively, over some neighborhood N (S) of S: (i) w(.) is deﬁned and continuous on N (S) : w(e1 ) ∈ C[N (S)], (ii) w(e1 ) ≥ 0 {w(e1 ) ≤ 0}, ∀e1 ∈ [N (S) − InS], (iii) w(e1 ) = 0 for e1 ∈ [N (S) − InS] if and only if e1 ∈ ∂S, (iv) w(e1 ) ≤ 0 {w(e1 ) ≥ 0}, ∀e1 ∈ ClS. b) positive {negative} deﬁnite relative to a set Sb , Sb = Si11 k1 xSi12 k2 x ... xSi1p kp ⊂ R2m , if and only if (i), (ii) and (iv) of a) hold for S = Sb and ωl (e1 ) = 0 for e1 = (eTil kl (1)T
eil kl )T ∈ [N (Sb ) − InSb ] if and only if e1il kl ∈ ∂Si1l kl , ∀l = 1, 2, ... , p ≤ m. 1 1 c) pairwise positive {negative} deﬁnite relative to a set Sc , Sc = S11 xS22 x ... 1 2m xSmm ⊂ R , if and only if b) holds for Sb = Sc , (i) p = m and il = kl = l, ∀l = 1, 2, ..., m, and (ii) ωl (e1 ) ≡ ωl (e1l ), ∀l = 1, 2,..., m. d) elementwise positive {negative} deﬁnite relative to a set Sd , Sd = S11 xS22 x... (1) (1) (1) xSmm xS11 xS22 x...xSmm ⊂ R2m , if and only if (i), (ii) and (iv) of a) hold for S = Sd , (i) p = 2m, (1) (ii) ωl (e1 ) ≡ ωl (el ) and ωm+l (e1 ) ≡ ωm+l (el ), ∀l = 1, 2, ..., m, (1)
(1)T
(iii) ωl ( el ) = 0 and ωm+l ( el ) = 0 for e1 = (eTll ell )T ∈ [N (Sd ) − InSd ] if (1) (1) and only if, respectively, el ∈ ∂Sll and el ∈ ∂Sll , ∀l = 1, 2, ..., m. The preceding properties are: e) global (in the whole) if and only if they hold for N (S(.) ) = R2m . f) on a set A, A ⊂ R2m , if and only if the corresponding conditions hold for N (S(.) ) = A. The expression “relative to a set S(.) , S(.) ⊂ R2m ,” should be omitted if and only if the set S(.) is singleton O : S(.) = O = {e1 : e1 = 0} The preceding deﬁnitions are consistent with Lyapunov’s original concept of scalar (semi) deﬁnite functions [10] and with the concept of deﬁnite matrix functions introduced in [11]. Notice that a) of Deﬁnition 2 enables synthesis of control to simultaneously stabilize/track control and optimal control relative to multiple (p) different optimality criteria.
2.2
Semideﬁnite and Deﬁnite TimeDependent Vector Functions
Deﬁnition 3 A vector function v(.) : RxR2m → Rp , p ∈ {1, 2, ..., 2m}, v(t, e1 ) = [υ1 (t, e1 ) υ2 (t, e1 ) ... υp (t, e1 )]T , is: a) positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if (i) through (iii) hold, respectively, over some neighborhood N (S) of the set S: (i) v(.) is deﬁned and continuous on RxN (S) : v(t, e1 ) ∈ C[RxN (S)], (ii) v(t, e1 ) ≥ 0 {v(t, e1 ) ≤ 0}, ∀(t, e1 ) ∈ Rx[N (S) − InS], (iii) v(t, e1 ) ≤ 0 {v(t, e1 ) ≥ 0}, ∀(t, e1 ) ∈ RxClS. b) pairwise positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if a) holds and both (i) and (ii) are valid: (i) p = m,
© 2004 by CRC Press LLC
(ii) υl (t, e1 ) ≡ υl (t, e1l ), ∀l = 1, 2, ..., m. c) elementwise positive {negative} semideﬁnite relative to a set S, S ⊂ R2m , if and only if a) holds and both (i) and (ii) are valid: (i) p = 2m, (1) (ii) υl (t, e1 ) ≡ υl (t, el ) and υm+l (t, e1 ) ≡ υm+l (t, el ), ∀l = 1, 2, ..., m. The preceding properties are: d) global (in the whole) if and only if they hold for N (S) = R2m . e) on a set A, A ⊂ R2m , if and only if the corresponding conditions hold for N (S) = A. The expression “relative to a set S, S ⊂ R2m ,” should be omitted if and only if the set S is singleton O : S = O = {e1 : e1 = 0} Deﬁnition 4 A vector function v(.) : RxR2m → Rp , p ∈ {1, 2, ..., 2m}, v(t, e1 ) = [υ1 (t, e1 ) υ2 (t, e1 ) ... υp (t, e1 )]T , is: a) positive {negative} deﬁnite relative to a set S, S ⊂ R2m , if and only if (i) and (ii) hold over some neighborhood N (S) of the set S: (i) v(.) is deﬁned and continuous on RxN (S) : v(t, e1 ) ∈ C[RxN (S)], (ii) there is a timeindependent positive deﬁnite vector function w1 (.) relative to the set S on N (S) such that 1) and 2) hold, respectively: 1) v(t, e1 ) ≥ w1 (e1 ) {w(t, e1 ) ≤ −w1 (e1 )}, ∀(t, e1 ) ∈ Rx[N (S) − InS], 2) v(t, e1 ) ≤ −w1 (e1 ) {v(t, e1 ) ≥ w1 (e1 )}, ∀(t, e1 ) ∈ RxClS. b) positive {negative} deﬁnite relative to a set Sb , Sb = Si11 k1 xSi12 k2 x...x Si1p kp ⊂ 2m R , if and only if a) is valid, respectively, for S = Sb . 1 1 1 xS22 x...xSmm c) pairwise positive {negative} deﬁnite relative to a set Sc , Sc = S11 2m ⊂ R , if and only if b) holds, respectively, for Sb = Sc , w1 (.) is pairwise positive {negative} deﬁnite relative to the set Sc on N (Sc ), (i) p = m and il = kl = l, ∀l = 1, 2, ..., m, and (ii) υl (t, e1 ) ≡ υl (t, e1l ), ∀l = 1, 2, ..., m. d) elementwise positive {negative} deﬁnite relative to a set Sd , Sd = S11 xS22 x... (1) (1) (1) xSmm xS11 xS22 x...xSmm ⊂ R2m , if and only if a) holds, respectively, for S = Sd , w1 (.) is elementwise positive {negative} deﬁnite relative to the set Sd on N (Sd ), and both (i) and (ii) are valid: (i) p = 2m, (1) (ii) υl (t, e1 ) ≡ υl (t, ei ) and υm+l (t, e1 ) ≡ υm+l (t, el ), ∀l = 1, 2, ..., m. The preceding properties are: e) global (in the whole) if and only if they hold for N (S(.) ) = R2m . f) on a set A, A ⊂ R2m , if and only if the corresponding conditions hold for N (S(.) ) = A. The expression “relative to a set S(.) , S(.) ⊂ R2m ,” should be omitted if and only if the set S(.) is singleton O : S(.) = O = {e1 : e1 = 0} The preceding deﬁnitions are consistent with Lyapunov’s original concept of scalar (semi) deﬁnite functions [10]. Deﬁnition 4 is compatible with the deﬁnition of deﬁnite matrix functions introduced in [11]. Deﬁnition 5 A vector function v(.) : RxR2m → Rp , p ∈ {1, 2, ..., 2m}, v(t, e1 ) = [υ1 (t, e1 ) υ2 (t, e1 ) ... υp (t, e1 )]T , is:
© 2004 by CRC Press LLC
a) decrescent relative to a set S, S ⊂ R2m , if and only if (i) and (ii) hold over some neighborhood N (S) of the set S: (i) v(.) is deﬁned and continuous on RxN (S) : v(t, e1 ) ∈ C[RxN (S)], (ii) there is a timeindependent positive deﬁnite vector function w2 (.) with respect to the set S on N (S) such that: v(t, e1 ) ≤ w2 (e1 ), ∀(t, e1 ) ∈ Rx[N (S) − InS]. b) decrescent relative to a set Sb , Sb = Si11 k1 xSi12 k2 x...xSi1p kp ⊂ R2m , if and only if a) is valid for S = Sb . 1 1 1 xS22 x...xSmm ⊂ R2m , if and c) pairwise decrescent relative to a set Sc , Sc = S11 only if b) holds for Sb = Sc , w2 (.) is a pairwise positive deﬁnite vector function with respect to the set Sc on N (Sc ), (i) p = m and il = kl = l, ∀l = 1, 2, ..., m, and (ii) υl (t, e1 ) ≡ υl (t, e1l ), ∀l = 1, 2, ..., m. (1) (1) d) elementwise decrescent relative to a set Sd , Sd = S11 xS22 x...xSmm x S11 xS22 x (1) ...xSmm ⊂ R2m , if and only if a) holds for S = Sd , w2 (.) is an elementwise positive deﬁnite vector function with respect to the set Sd on N (Sd ), (i) p = 2m, (1) (ii) υl (t, e1 ) ≡ υl (t, el ) and υm+l (t, e1 ) ≡ υm+l (t, el ), ∀l = 1, 2, ..., m. The preceding properties are: e) global (in the whole) if and only if they hold for N (S(.) ) = R2m . f) on a set A, A ⊂ R2m , if and only if the corresponding conditions hold for N (S(.) ) = A. The expression “relative to a set S(.) , S(.) ⊂ R2m ,” should be omitted if and only if the set S(.) is singleton O : S(.) = O = {e1 : e1 = 0} This deﬁnition is consistent with the deﬁnition of a scalar decrescent function [12].
2.3
Kinds of Vector Lyapunov Functions
The initial moment t0 = 0 is accepted and ﬁxed. It will be omitted whenever its omission cannot make a confusion: q 1 (.; 0, q01 ; d) ≡ q 1 (.; q01 ; d) ≡ q 1 (.) is a system motion, e1 (.; 0, e10 ; d) ≡ e1 (.; e10 ; d) ≡ e1 (.) is a motion error from a desired motion qd1 (.): e1 (.) = qd1 (.) − q 1 (.), y 1 (.) ≡ y 1 (.; y01 ; d) is a system output response together with its ﬁrst derivative, the error of which relative to yd1 (.) (a desired output response with its ﬁrst derivative) is ε1 (.; ε10 ; d) ≡ ε1 (.) ≡ yd1 (.) − y 1 (.). We will use the righthand upper Dini derivative D+ v(.) of a vector function v(.) along a system motion error e1 (.; t0 , e10 ; d), v[t + θ, e1 (t + θ; t, e1 ; d)] − v(t, e1 ) + 1 + , :θ→0 D v(t, e ) = lim sup θ or along an output error response ε1 (.; t0 , ε10 ; d). If the function v(.) is differentiable then D+ v(.) becomes the Eulerian derivative v (1) (.) of v(.) along e1 (.; e10 ; d) or along ε1 (.; ε10 ; d), respectively. Deﬁnition 6 A vector function v(.) : RxR2m → Rp , p ∈ {1, 2, ..., 2m}, v(t, e1 ) = [υ1 (t, e1 ) υ2 (t, e1 ) ... υp (t, e1 )]T , is:
© 2004 by CRC Press LLC
a) a vector Lyapunov function of a system relative to a set S, S ⊂ R2m , if and only if (i) and (ii) hold: (i) it is positive deﬁnite relative to the set S, (ii) its Dini derivative D+ v(t, e1 ) is negative semideﬁnite with respect to the set S. b) a vector Lyapunov function of the system relative to a set Sb , Sb = Si11 k1 xSi12 k2 x ...xSi1p kp ⊂ R2m , if and only if both (i) and (ii) hold: (i) it is positive deﬁnite relative to the set Sb , (ii) its Dini derivative D+ v(t, e1 ) is negative semideﬁnite with respect to the set Sb . c) a pairwise vector Lyapunov function of the system relative to a set Sc , Sc = 1 1 1 S11 xS22 x...xSmm ⊂ R2m , if and only if both (i) and (ii) hold: (i) it is pairwise positive deﬁnite relative to the set Sc , (ii) its Dini derivative D+ v(t, e1 ) is pairwise negative semideﬁnite with respect to the set Sc . d) elementwise vector Lyapunov function of the system relative to a set Sd , Sd = (1) (1) (1) S11 xS22 x...xSmm xS11 xS22 x...xSmm ⊂ R2m , if and only if both (i) and (ii) hold: (i) it is elementwise positive deﬁnite relative to the set Sd , (ii) its Dini derivative D+ v(t, e1 ) is elementwise negative semideﬁnite with respect to the set Sd . The expression “relative to a set S(.) , S(.) ⊂ R2m ,” should be omitted if and only if the set S(.) is singleton O : S(.) = O = {e1 : e1 = 0} This deﬁnition is consistent with the notion of a scalar Lyapunov function and with the notions of vector Lyapunov functions by Bellman [1] and Matrosov [2]. However, it differs essentially from the deﬁnitions of vector Lyapunov functions in [5, pp. 34 and 35] and [9, p. 296]. The reasons are explained in [11].
3
SYSTEM DESCRIPTION
Airplanes and other mechanical systems (autonomous vehicles, robots, ships, spacecrafts) with a constant or variable mass and/or subjected to timevarying loads can be described by a 2D mathematical model (1), (2), [13]–[19]: A(t, q)q (2) (t) + h[t, d(t), q 1 (t)] = B[t, q(t)]b[u(t)], y(t) = g[q(t)].
(1) (2)
The vector q ∈ Rm , the inertia matrix function A(.) : RxRm → Rmxm and the nonlinear vector function h(.) : RxRd xR2m → Rm describe system internal dynamics. The function h(.) incorporates the terms describing all system nonlinearites, except those of the inertia matrix, as well as the terms expressing an inﬂuence of variations of both external perturbation vector function d(.) : R → Rd and the system mass. The perturbation vector function d(.) belongs to a functional family Sd of all permitted d(.). The matrix function B(.) : RxRm → Rmxr , the nonlinear vector function b(.) : Rr → Rr and the control vector u ∈ Rr . The system output vector function g(.) : Rm → RN and the system output vector y ∈ RN . The function g(.) is well deﬁned, known, and global differentiable. Its Jacobian is denoted by Jg (.) = [∂gi (.)/∂qj ] : Rm → RN xm . Let R+ = [0, ∞[.
© 2004 by CRC Press LLC
Condition 7 Systems (1) and (2) obey the following: a) The inertia matrix A(t, q) is well deﬁned, known, and det A(t, q) = 0 for all (t, q) ∈ RxRm . b) The elementwise absolute vector value of the nonlinear vector function h(.) satisﬁes the next inequality: h[t, d(t), q 1 ] ≤ hM (t) + H(t, q 1 ) q 1 , ∀[t, d(.), q 1 ] ∈ RxSd xR2m , for some elementwise nonnegative well deﬁned and known both the vector function hM (.) : m mx2m R → R+ and the matrix function H(.) : RxR2m → R+ . c) The matrix B(t, q) is well deﬁned, known and rank B(t, q) = m for all (t, q) ∈ RxRm . d) The vector nonlinearity b(u) is well deﬁned and known for all u ∈ Rr . It has the same sign elementwise as the control vector u: sign[b(u)] ≡ sign(u). Its inverse function bI (.) : Rr → Rr is well deﬁned and known for all b ∈ Rr so that bI [b(u)] ≡ u. e) For a given desired output response yd (.) the corresponding desired motion (1)T
qd1 (.) = qd1 [.; yd1 (.); d(.); uN (.)], qd1 (.) = [qdT (.) qd where for some u(.) = uN (.) called nominal control:
(.)]T , is well deﬁned and known,
(2)
g[qd (t)] ≡ yd (t), A[t, qd (t)]qd (t) + h[t, d(t), qd1 (t)] ≡ B[t, qd (t)]b[uN (t)]. 4 4.1
CONTROL SYNTHESIS: EXPONENTIAL STABILITY Sliding Set
Sliding control attracted signiﬁcant interest and achieved effective applications due to its advantage to guarantee a desired quality of system motions in the sliding regime independently of the system internal dynamics [20]–[23]. Let Sss be a sliding set in the e1 space R2m : mxm Sss = e1 : e(1) = −Λe , Λ = diag{λ1 λ2 ... λm } ∈ R+ , (3) 1 1 1 where Λ is a positive diagonal design matrix. Notice that Sss = Sss11 xSss22 x ... xSssmm , (1) 1 1 Sssii = {ei : ei = −λi ei }, ∀i = 1, 2, ..., m. The sliding set Sss determines exponential convergence of e1 (t) to e1 = 0, which means that the real motion q 1 (.) of system (1), (2) converges exponentially to its desired motion qd1 (.), both independently of the system internal dynamics. If an initial error vector e10 is not in the sliding set, then the instantaneous error vector e1 (t; e10 ; d) should reach pairwise the sliding set at the latest at the vector 1 moment τRSS (e10 ) = [τRSS1 (e110 ) τRSS2 (e120 ) ... τRSSm (e1m0 )]T : e1i (t; e1i0 ; d) ∈ Sssii , 1 T m ∀t ∈ [τRSSi (ei0 ), ∞[, ∀i = 1, 2,..., m. Let 1 = (1 1 ... 1) ∈ R .
4.2
Stabilizing Control Problem Statement
Problem 8 Synthesize control u(.) so as to ensure that both (i) and (ii) hold for every d(.) ∈ Sd : a) the sliding set Sss , (3), is positive invariant with respect to e1 (t) : if e1 (τ ; e10 ; d) ∈ Sss at τ ∈ R+ , then e1 (t; e10 ; d) ∈ Sss , ∀t∈[τ, ∞[,
© 2004 by CRC Press LLC
(4)
b) the sliding set Sss , (3), is globally pairwise exponentially stable with a ﬁnite vector reachability time τRSS (e10 ), which means that for every e10 ∈ R2m , e1 (t; e10 ; d) obeys (5): (1) (1) (5) e (t; e10 ; d) + Λe(t; e10 ; d) ≤ Ke−Lt e0 + Λe0 , ∀t ∈ R+ , for some positive diagonal design mxm matrices K and L, and e1 (t; e10 ; d) ∈ Sss , ∀t1 ∈[τRSS (e10 ), ∞1[. 4.3
(6)
Stabilizing Control Problem Solution
We will use the following notation in order to ensure a requested ﬁnite vector reachability time of the sliding set Sss , (3), and its exponential stability: S (z) = diag{sign(z1 ) sign(z2 ) ... sign(zm )} and e−Lt = diag{e−α1 t e−α2 t ... e−αm t } for L = diag{α1 α2 ... αm }. Theorem 9 If Condition 7 is valid, if T is a positive diagonal matrix such that mxm (1) τRSS (e10 ) = T e0 + Λe0 , ∀e10 ∈ R2m , T ∈ R+ , and if the control vector function u(.) is in the form (8),
−1 T B T (t, q) B(t, q)B (t, q) A(t, q)•
1 1 (2) I 1 −1 1 1 S[z(e )] A (t, q) hM (t) + H(t, q ) q u(t, e , q , qd ) = b , (2) +qd (t) + Λe + Lz(e1 ) + T −1 S[z(e1 )]1
(7)
(8)
then:
a) the sliding set Sss , (3), is positive invariant with respect to e1 (t; e10 ; d) for all d(.) ∈ Sd , b) the sliding set Sss is globally pairwise exponentially stable with the ﬁnite vector reachability time τRSS (e10 ), (7), for every d(.) ∈ Sd , c) the motion error e1 (.; e10 ; d) obeys (9) for every [d(.), e10 ] ∈ Sd xR2m : (1) e (t) + Λe(t) (1) e−Lt e0 + Λe0 , t∈ [0,∞[,
(1) −1 1 ≤ min . (9) e0 + Λe0 − T t1, ∀t1 ∈ 0,τRSS (e0 ) , 0, ∀t1 ∈ [τRSS (e10 ), ∞1[ Proof. Let [d(.), e1 ] ∈ Sd xR2m be arbitrary. Let Condition 7, the conditions of the theorem statement, (10) and (11) hold: z(e1 ) = e(1) + Λe = [z1 (e11 ) z2 (e12 ) ... zm (e1m )]T , (1)
zi (e1i ) = ei
+ λi ei , ∀i = 1, 2, ..., m, Z(e ) = diag z1 (e11 ) z2 (e12 ) ... zm (e1m ) . 1
© 2004 by CRC Press LLC
(10) (11)
We deﬁned z(.) and Z(.) in order to introduce a vector function v(.), (12), v(z) = (1/2)Zz.
(12)
This function is global differentiable and elementwise global positive deﬁnite (and decrescent) on the zspace Rm . If it is considered on the e1 space R2m in view of (10) and (11): v[z(e1 )] = (1/2)Z(e1 )z(e1 ),
(13)
then it is also global differentiable, but pairwise global positive deﬁnite (and decrescent) 1 1 1 xSss22 x ... xSssmm , (3). Its Dini derivative with respect to the sliding set Sss , Sss = Sss11 + (1) D v(.) becomes its Eulerian derivative v (.): (14) v (1) [z 1 (e2 )] = Z(e1 )z (1) (e1 ) = Z(e1 ) e(2) + Λe(1) , ∀e2 ∈ R3m . Thus, (10) and (11) show that v (1) [z 1 (e2 )] = 0 on the sliding set Sss . At ﬁrst we eliminate (2) e(2) from (14) in view of e(2) = qd − q (2) . Then, nonsingularity of A(.), a) through d) of Condition 7, (1) and (8) imply: v (1) [z 1 (e2 )] ≤ −LZ(e1 )z(e1 )] − T −1 z(e1 ) = −2Lv[z(e1 )] − T −1 2v[z(e1 )]. Hence,
v (1) [z 1 (e2 )] ≤ − min 2Lv[z(e1 )], T −1 2v[z(e1 )] , ∀e2 ∈ R3m .
(15)
This shows that v (1) [z 1 (e2 )] is majorized by a global negative deﬁnite vector function on R2m , which, together with v (1) [z 1 (e2 )] = 0 on the sliding set Sss and global positive deﬁniteness of v(.) with respect to the sliding set Sss , proves its global asymptotic stability for every d(.) ∈ Sd . Hence, the sliding set is also positive invariant relative to the motion error of system (1), (2) for every d(.) ∈ Sd . This proves the statement under a). We integrate (15) and after that we use (10) and (13). The result is the following: (1) e (t) + Λe(t) −Lt (1) e , ∀t∈R + Λe , e 0 + 0 (1) (1) e + Λe0 − T −1 t1, ∀t1 ∈ 0,T e + Λe0 , ≤ min . 0 0 (1) 0, ∀t1 ∈ [T e + Λe0 , ∞1[ 0
This and the deﬁnition of the sliding set Sss , (3), prove the statements under b) and c) Global ﬁnite time reachability and positive invariance of the sliding set Sss , and (3) guarantee that the behavior of controlled system (1), (2) does not depend on its own internal dynamics over the vector time interval [τRSS (e10 ), ∞1[. 5
5.1
CONTROL SYNTHESIS: OUTPUT TRACKING AND ASYMPTOTIC STABILITY Tracking and Stabilizing Control Problem Statement
Let Sss be a sliding set in the output space RN , which is deﬁned by (16), N xN Sssε = ε1 : ε(1) = −Γε , Γ = {γ1 γ2 ... γN } ∈ R+ ,
© 2004 by CRC Press LLC
(16)
1 1 where Γ is a positive diagonal design matrix. Notice that Sssε = Sssε11 xSssε22 x ... (1) 1 1 1 xSssεN , S = {ε : ε = −γ ε }, ∀i = 1, 2, ..., N. The sliding set S determines i i ssε ssii i N i exponential convergence of ε1 (t) to ε1 = 0, which means that the real output response y(.) of system (1), (2) converges exponentially to its desired output response yd (.) over the sliding set Sssε , (16).
Problem 10 Synthesize control u(.) so to ensure that (a) through (d) hold: a) the sliding set Sssε , (16), is positive invariant with respect to ε1 (t) : if ε1 (τ ; ε10 ; d) ∈ Sssε at τ ∈ R+ , then ε1 (t; ε10 ; d) ∈ Sssε , ∀t∈[τ, ∞[,
(17)
b) the sliding set Sssε is globally pairwise exponentially stable with a ﬁnite vector N reachability time τrssε (ε10 ) ∈ R+ , which means that (18) holds: (1) (1) (18) ε (t; ε10 ; d) + Γε(t; ε10 ; d) ≤ Kε e−Lε t ε0 + Γε0 , ∀(t, ε10 )∈R+ xR2N , for some positive diagonal design N xN matrices Kε and Lε , and ε1 (t; ε10 ; d) ∈ Sssε , ∀(t1, ε10 )∈[τrssε (ε10 ), ∞1[xR2N ,
(19)
c) the sliding set Sss , (3), is positive invariant with respect to e1 (t), (4), d) the sliding set Sss is globally asymptotically stable with a ﬁnite scalar reachability time τrsss (e10 ), which means that e1 (t; e10 ; d) obeys (20), e1 (t; e10 ; d) ∈ Sss , ∀(t, e10 )∈[τrsss (e10 ), ∞[xR2m . 5.2
(20)
Output Tracking and Stabilizing Control Problem Solution
Condition 11 The dimensions m of the vector q, N of the output vector y, and r of the control vector u of system (1), (2) obey: r ≥ m > N. Let an auxiliary variable zε and the induced matrix Zε be deﬁned by (21): zε = (zε1 zε2 ... zεN )T = ε(1) + Γε, Zε = diag {zε1 zε2 ... zεN } .
(21)
Condition 12 A subsidiary function vs (.) : Rm → Rp , p ∈ {1, 2, ..., m − N }, (22), T
vs (z) = [υs1 (z) υs2 (z) ... υsp (z)] , p ∈ {1, 2, ..., m − N },
(22)
is a global differentiable positive deﬁnite vector function on the zspace Rm , i.e., it is a global differentiable positive deﬁnite vector function with respect to the set Sss , (3), if it is considered on the e1 space R2m in view of (21) and (22). The function vs (.) will be used to solve the subproblems c) and d) of Problem 10. Let its Jacobian be denoted by Js (.) : Rm → Rpxm , p ∈ {1, 2, ..., m − N }, Js (z) = [∂υsi (z)/∂zj ] . Condition 13 The Jacobians Jg (q) and Js (z) of the vector functions g(.), (2), and vs (.), (22), and Zε (.), (21), obey: rank [W (q, z, zε )] = N + p, ∀(q, z, zε ) ∈ Rm xRm xRN and Js (0) = 0, where W (q, z, zε ) = [JgT (q)ZεT (zε ) JsT (z)]T ∈ R(N +p)xm .
© 2004 by CRC Press LLC
Let Lε and Ts be positive diagonal design pxp matrices, Ts = diag{τs1 τs2 ... τsp } and s(vs ) = [sign(υs1 ) sign(υs2 ) ... sign(υsp )]T . Theorem 14 If Conditions 7, 11 through 13 are satisﬁed, if Tε is a positive diagonal N xN matrix such that (1) τRSSε (ε10 ) = Tε ε0 + Λε0 , ∀ε10 ∈ R2N , (23) and if the control vector function u(.) is in the form (24),
−1 u(t, e1 , ε1 , q 1 , qd1 ) = bI B T (t, q) B(t, q)B T (t, q) A(t, q)ue (t) , ue ∈ Rm , ue (t, e1 , ε1 , q 1 , qd1 ) = W T (q, z, zε )[W (q, z, zε )W T (q, z, zε )]−1 • (2) (1) (1) (1) (1) Z (z ) y − J (q)q + Γy − Γy g ε ε d d 1 1 −1 Z (z )z (ε ) + T +L ε ε ε ε Z (ε ) 1 1 1 1 −1 + Z (ε ) Jg (q)A (t, q) hM (t) + H(t, q ) q (2) (1) (1) J + Ls v(z) + Ts−1 s [vs (z)] (z) q + Λq − Λq s d d −1 + Js (z)A (t, q) hM (t) + H(t, q 1 ) q 1
, (24)
then for every d(.) ∈ Sd : a) the sliding set Sssε , (16), is positive invariant with respect to ε1 (t), (17), b) the sliding set Sssε is globally pairwise exponentially stable with the ﬁnite vector N reachability time τrssε (ε10 ) ∈ R+ , (18), (19), (23), c) the sliding set Sss , (3), is positive invariant with respect to e1 (t), (4), d) the sliding set Sss is globally asymptotically stable with the ﬁnite scalar reachability time τrsss (e10 ) = max{τsi (e10 ) : i = 1, 2, ... , p}, (20), e) the vector function vs (.) and the motion error e1 (.; e10 ; d) obey (25): vs z[e1 (t; e10 ; d)] e−Ls t vs [z(e10 )], ∀t∈R , +
vs [zs (e10 )] − Ts−1 t1, ∀t1 ∈ 0,Ts vs [zs (e10 )] ≤ min . (25) 0, ∀t1 ∈ [Ts vs [zs (e10 )], ∞1[. Proof. Let [d(.), e1 , ε1 ] ∈ Sd xR2m xR2N be arbitrary. Let all the conditions of the theorem statement hold. Let an extended vector function ve (.) : Rm+N → RN +p be composed of the vector functions vε (.) = (1/2)Zε (.)zε (.) and vs (.):
T T ∈ Rm+N . ve (ze ) = vεT (zε ) vsT (z) , ze = zεT z T
(26)
It is global differentiable positive deﬁnite vector function on the ze −space Rm+N . Its Jacobian Je (.) : Rm+N → R(N +p)x(m+N ) has the next form: Je (ze ) = blockdiag {Zε (zε ) Js (z)} . Hence, along an output error response and a motion error of system (1), (2):
(1) Zε (zε ) ε(2) + Γε(1) Zε (zε )zε = =⇒ ve(1) (0) = 0, ve(1) (ze1 ) = (2)
(1) Js (z) e + Λe Js (z)z (1)
© 2004 by CRC Press LLC
(27)
(2)
(2)
(1)
or, by using: ε(2) = yd − y (2) , e(2) = qd − q (2) , y (2) = [g(q)](2) = Jg (q)q (1) + (1) Jg (q)q (2) = Jg (q)q (1) + Jg (q)A−1 (t, q)[B(t, q)b(u) − h(t, d, q 1 )] [due to (1), (2) and Condition 7], the deﬁnition of W (q, z, zε ), Condition 13 and (24), we ﬁnd: −Lε Z (ε1 )z (ε1 ) − Tε−1 Z (ε1 ) 1 . ve(1) (ze1 ) ≤ −Ls v(z) − Ts−1 s [vs (z)] Hence,
min Lε Z (ε1 )z (ε1 ), Tε−1 Z (ε1 ) 1 ve(1) (ze1 ) = ≤ − " # 1 −1 1 (1) 1 2 min L v [z(e )], T s v [z(e )] s s s vs [z (e )] s $ (1) 1 −1 2vε [z (ε1 )] min 2Lε vε [z (ε )], Tε 2 2 3m 3N = − , ∀(e , ε ) ∈ R xR . " # min Ls vs [z(e1 )], Ts−1 s vs [z(e1 )] (1)
vε [z1 (ε2 )]
(1)
This shows that ve (ze1 ) is majorized by a global negative deﬁnite vector function on (1) Rm xRN , which, together with (21), ve [ze1 (e2 , ε2 )] = 0 on the Cartesian product slid2m 2N ing set Sss xSssε (in R xR due to Zε (0) = 0 and Js (0) = 0) and global positive deﬁniteness of ve (.) with respect to the product sliding set Sss xSssε , proves global asymptotic stability of the product sliding set Sss xSssε for every d(.) ∈ Sd . Hence, the product sliding set Sss xSssε is also positive invariant relative to the error motions and error output responses of system (1), (2) for every d(.) ∈ Sd . This proves the statements under a) and c). Integration of the preceding inequality, (21), vε (.) = (1/2)Zε (.)zε (.) and (26) yield: (1) ε (t) + Γε(t) 1 1 vs z[e (t; e0 ; d)] (1) e−Lε t ε0 + Γε0 , ∀t∈R+ , (1) (1) −1 ε0 + Γε0 − Tε t1 , ∀t1 ∈ 0,Tε ε0 + Γε0 min (1) 0, ∀t1 ∈ [Tε ε0 + Γε0 , ∞1[. ≤ . −Ls t 1 e vs [z(e0 )], ∀t∈R +,
1 −1 1 vs [zs (e0 )] − Ts t1, ∀t1 ∈ 0,Ts vs [zs (e0 )] min 0, ∀t1 ∈ [Ts vs [zs (e10 )], ∞1[ This, (3), (16), and (23) complete the proof
6
CONCLUSION
The new vector development of the concepts of (semi) deﬁnite scalar functions and decrescent scalar functions, and the development of the concept of vector Lyapunov functions enable an effective resolution of a basic complex problem of control synthesis in two different
© 2004 by CRC Press LLC
spaces. Such a control ensures simultaneous global asymptotic stability of a system desired motion and a good quality of output tracking that is accepted to be exponential. Various kinds of vector Lyapunov functions, which are introduced herein, additionally permit an effective sliding control synthesis. Control guarantees that the motion error and the output error response of the controlled system are independent of the system internal dynamics over the sliding sets in the e1 error space and in the output ε1 error space. Asympotic (or exponential) stability of the e1 sliding set Sss and of the ε1 sliding set Sssε with prespeciﬁed ﬁnite reachability times provide an engineering sense to the results and enable their effective applications. Synthesized controls are directly applicable to airplanes and to other nonlinear dynamical systems (autonomous vehicles, robots, ships, and spacecrafts). All nonlinearities and parameters of their mathematical models can be unknown, except those of their inertia matrices and the ouput function g(.). Besides, information about the form and real values of external disturbances is not requested. The synthesized controls, stability, and tracking properties are highly robust relative to such uncertainties. The paper opens new directions in studies of stability, tracking, and control of the nonlinear dynamical systems.
REFERENCES 1. R. Bellman, “Vector Lyapunov functions”, J. S. I. A. M. Control, Ser. A, 1, No. 1, pp. 32–34, 1962. 2. V.M. Matrosov, “To the theory of stability of motion” (in Russian), Prikl. Math. Mekh., 26, No. 5, pp. 885–895, 1962. 3. V.M. Matrosov, “Vector Lyapunov functions in the analysis of nonlinear interconnected systems,” Proc. Symp. Mathematica, Bologna, 6, pp. 209–242, 1971. 4. V. Lakshmikantham, “Vector Lyapunov functions and conditional stability,” J. Mathematical Analysis and Applications, pp. 368–377, 1975. 5. J.P. La Salle, The Stability of Dynamical Systems, SIAM, Philadelphia, 1976. 6. A.N. Michel and R. Miller, Qualitative Analysis of LargeScale Dynamical Systems, Academic Press, New York, 1977. ˇ 7. D.D. Siljak, LargeScale Dynamic Systems: Stability and Structure, NorthHolland, New York, 1978. 8. Lj.T. Gruji´c, A.A. Martynyuk, and M. RibbensPavella, LargeScale Systems under Structural and Singular Perturbation (in Russian: Naukova Dumka, Kiev, 1984), Springer Verlag, Berlin, 1987. 9. V.M. Matrosov, Vector Lyapunov Function Method: Analysis of Dynamical Properties of Nonlinear Systems (in Russian), FIZMATLIT, Moscow, 2001. 10. A.M. Lyapunov, General Problem of Stability of Motion, Taylor and Francis, London, 1992. 11. Lj.T. Gruji´c, “On largescale systems stability.” In: Computing and Computers for Control Systems, Eds. P. Borne et al., J.C. Baltzer AG, Scientiﬁc Publishing Co., IMACS, pp. 201–206, 1989. 12. W. Hahn, Stability of Motion, Springer Verlag, Berlin, 1967. 13. J. Wertz, Ed., Spacecraft Attitude Determination and Control, R. Reidel Publishing Company, Boston, 1978.
© 2004 by CRC Press LLC
14. B. Etkin, Dynamics of Flight — Stability and Control, John Wiley & Sons, New York, 1982. 15. P.C. Hughes, Spacecraft Attitude Dynamics, John Wiley & Sons, New York, 1986. 16. M.W. Spong and M. Vidyasagar, Robot Dynamics and Control, John Wiley and Sons, New York, 1989. 17. T.I. Fossen, Guidance and Control of Ocean Vehicles, John Wiley & Sons, Chichester, 1994. 18. L. Sciavicco and B. Siciliano, Modeling and Control of Robot Manipulators, The McGraw Hill Companies, Inc., New York, 1996. 19. L. Cvetiˇcanin, Dynamics of Machines with Variable Mass, Gordon and Breach Science Publishers, Amsterdam, 1998. 20. V.I. Utkin, “Variable structure systems with sliding modes,” IEEE Transactions on Automatic Control, Vol. AC23, No. 2, pp. 212–222, 1977. 21. J.J. E. Slotine, “The robust control of robot manipulators,” The International Journal of Robotics Research, Vol. 4, No. 2, pp. 49–64, 1985. 22. V.I. Matyukhin and E.S. Pyatnitskii, “Controlling the motion of robot manipulators by decomposition taking into account the actuator dynamics,” Avtomatika i Telemekhanika (in Russian), No. 9, pp. 67–81, 1989. 23. C.Y. Su, T.P. Leung and Y. Stepanenko, “Realtime implementation of regressorbased sliding mode control algorithm for robotic manipulators,” IEEE Transactions on Industrial Electronics, Vol. 40, No. 1, pp. 71–79, 1993.
© 2004 by CRC Press LLC
8 Stabilization of Unstable Aircraft Dynamics under Control Constraints †
∗
M.G. Goman† and M.N. Demenkov∗
De Montfort University, Faculty of Computing Sciences and Engineering, Leicester U.K. Department of Computing Sciences and Control, Bauman Moscow State Technical University, Moscow, Russia
Stabilization of the unstable dynamic linear system with control constraints is considered in terms of maximizing the size of the closedloop system stability region. A controllability region for the openloop system is a natural limit for the stability region of the closedloop system with any designed controller. A relation between the controllability and stability regions is considered as a performance metric for controller assessment. The linear control laws maximizing the stability region for the constrained linear system are derived for two aircraft stabilization problems.
1
INTRODUCTION
Stability augmentation of an aircraft’s dynamics by control system allows one not only to improve its handling quality characteristics, but also to expand the ﬂight envelope and increase its performance characteristics. For example, an aircraft with aerodynamically unstable conﬁguration stabilized by control system can provide higher lifttodrag ratio or lower signature [1–3]. Critical ﬂight regimes such as highincidence departures or aeroelastic instabilities can be signiﬁcantly relaxed or even eliminated by an active control approach [4]. The central problem in control law design for regimes with aircraft dynamic instability is in taking into account the realistic constraints for control effectors such as deﬂection limits and rate saturation. Unfortunately, in control design works the stabilization problem of unstable aircraft dynamics is usually considered as a linear one without the direct effect of nonlinear control constraints [5]. The control saturation can lead to signiﬁcant degradation of dynamic characteristics or even to the loss of stability in the closedloop system. The unstable constrained linear system has a bounded controllability region, where the stabilization problem can be solved. This fundamental fact is normally ignored or assumed insigniﬁcant. However, in many practical situations associated with a high level of instability and low control authority, the controllability region decreases signiﬁcantly and classical linear control design methods lose their efﬁciency.
© 2004 by CRC Press LLC
The closedloop system stability region (or region of attraction) presents only part of the openloop system controllability region and therefore is even more bounded. The “good” controller in terms of linear system criteria can provide the closedloop system with a very small and unsatisfactory stability region. The computation of the controllability region for the openloop system and region of attraction for the closedloop system will extend understanding of the stabilization problem. All controllers designed using linear methods provide different stability regions for the closedloop system and the relationship between the stability region and controllability region can be considered as an additional metric for assessment of linear controller robustness to external disturbances. The allowable level of external disturbances is directly connected with the size of the stability region. The bigger the region of attraction, the higher the level of external disturbances that can be rejected by the system. If external disturbances move the system beyond the controllability region, any excellent controller will not be able to keep the system stable. The number of works devoted to control design methods maximizing stability region of constrained linear system is not high; however, the interest in this area is growing rapidly [7–13]. The best controller in terms of rejection of external disturbances has to keep the system stable in the whole controllability region. Such a controller is called a “maximum stabilizer” [12] or as “stability optimal” [7, 8]. The controller of some particular structure designed with additional criteria for maximizing stability region [10, 11, 13] can be called the “stability suboptimal” one. In this chapter the problem of linear controller design maximizing the closedloop system stability region is considered for the linearized system with one and two unstable eigenvalues and only deﬂection constraints. The results presented in [8] are extended to the case of multiple control inputs considering two design examples for an aircraft’s dynamics. The basic concepts and methods of controllability region computation are also discussed. Hereafter the following notations are used. The capital letters denote matrices and sets, lower case letters denote column vectors or scalars, Rn denotes the Euclidean nspace and Rn×m the set of all real n × m matrices. Vectors are supposed to be column, subscripts denote elements of a sequence (xi ), superscripts in parentheses denote component of a vector (x(i) ), and xT the transposed vector x (in the same way as transposition of the matrix). If x ∈ Rp and y ∈ Rq , are row vectors, then (x, y) denotes the row vector, which is the union of the vectors x√and y: (x, y) ∈ Rp+q . The following vector 2norm is used for vector x ∈ Rn : x2 = xT x.
2
SOME BASIC PROPERTIES OF UNSTABLE LINEAR SYSTEMS WITH CONTROL CONSTRAINTS
It is supposed that the following continuous timeinvariant linear system x(t) ˙ = Ax(t) + Bu(t),
(1)
with A ∈ Rn×n , B ∈ Rn×m , x ∈ Rn , u ∈ Rm , t ∈ [0, ∞) has q ≤ n unstable eigenvalues λi , Reλi > 0, i = 1, . . . , q and the control vector is bounded: u ∈ U,
© 2004 by CRC Press LLC
(2)
where U ∈ Rm is a compact and convex set, which is deﬁned in the form of amplitude constraints: u(i)  ≤ u(i) (3) max , i = 1, m. It is additionally supposed that matrix A has no neutral eigenvalues and the state vector is fully observable. All eigenvalues of matrix A can be divided into stable (Reλi < 0, i = 1, . . . , (n − q)) and unstable (Reλi > 0, i = 1, . . . , q) ones. Considering the problem of system stabilization and maximization of the closedloop system stability region, it is natural to take into account only the subspace corresponding to unstable eigenvalues. There are two wellknown transformations of the state variables suitable for system decomposition. The ﬁrst one is based on the Schur transformation [14] and the second one is based on the block Jordan transformation of matrix A. In this chapter we will use the Schur decomposition as the most robust and reliable one. Suppose that n eigenvalues of matrix A are ordered into two groups according to their real part magnitudes. The ﬁrst one consists of n − q stable eigenvalues and the second one consists of q unstable eigenvalues. Consider the following transformation of the basis in (1): x = [Q1 Q2 ]
s z
,
where matrix Q = [Q1 Q2 ] is the orthogonal one (QT Q = I) and the columns of Q1 ∈ Rn×(n−q) span the subspace V1 associated with n−q stable eigenvalues, while the columns of Q2 ∈ Rn×q span the subspace V2 , which is a complementary subspace of V1 . This matrix Q can be obtained by the Schur decomposition of matrix A with reordering, if necessary, the eigenvalues of its diagonal blocks [14] in accordance with the formed eigenvalue order. For numerical computation one may use the function blkrsch from MATLAB Robust Control Toolbox [15]. In the new basis the openloop system is represented as
s(t) ˙ z(t) ˙
= QT AQ
s(t) z(t)
+ QT Bu(t) =
A11 0
A12 Az
s(t) z(t)
+
Bs Bz
u(t),
where the eigenvalues of matrices A11 and Az contain the stable and unstable eigenvalues of original system (1), respectively. Note that z(t) dynamics associated with the unstable eigenvalues is decoupled from s(t) and described by the following openloop, reducedorder system: z(t) ˙ = Az z(t) + Bz u(t),
(4)
where Az ∈ Rq×q , Bz ∈ Rq×m , u(t) ∈ U and z(t) = QT2 x(t). If system (4) is stabilized and its states approach the equilibrium point with z(t) = 0 and u(t) = 0, it is trivial to show that subsystem s(t) will also asymptotically approach the origin and thus to stabilize system (1) it is enough to stabilize only subsystem z(t) (4). The following deﬁnition and two basic properties of the controllability region, formulated in [6], are given below to assist the main content of the paper.
© 2004 by CRC Press LLC
Deﬁnition 1 The controllability region C of system (1) with constraint (2) is the set of all points x0 ∈ Rn , from which the system can be moved to the origin by applying the bounded control u(t) ∈ U : C = {x0 ∈ Rn : ∃u(t) ∈ U so that lim x(t, x0 ) = 0}. t→∞
Theorem 1 The controllability region C of system (1) with constraint (2) is convex. Theorem 2 Let linear system (1) have only unstable eigenvalues and control vector be constrained as (3). Then the system controllability region C is a bounded set: max x2 < ∞. x∈C
If system (1) has stable and unstable eigenvalues, it contains unstable subsystem (4) with the bounded controllability region C in a subspace of Rn and as a result system (1) cannot be stabilized in the full statespace Rn . Bounded controllability region C of the unstable subsystem produces a cylinder in the full statespace of original system, which is unbounded in some directions. Suppose that the following linear controller is applied to the system (1): u(t) = Kx(t),
(5)
mxn
where K ∈ R stabilizes the closedloop system (1), (5) in the absence of control saturation. To take into account control constraints, consider the saturation function deﬁned for scalar u(i) ∈ R1 as: (i) (i) (i) umax ; u > umax (i) sat(u(i) ) = u(i) ; u(i)  ≤ umax (i) (i) −umax ; u(i) < −umax and for vector u = (u(1) , ..., u(m) )T as: sat(u) = (sat(u(1) ), ..., sat(u(m) ))T . The constrained linear closedloop system is deﬁned by (1) and the following nonlinear control: u(t) = sat(Kx(t)). (6) We will operate with the following deﬁnition of stability region. Deﬁnition 2 The stability region S of the closedloop system (1), (6) is the set of all points x0 ∈ Rn from which the closedloop system asymptotically approaches the origin: S = {x0 ∈ Rn : lim x(t, x0 ) = 0}. t→∞
The performance of the closedloop system is changed when the control input saturates, so that the phase trajectories of the closedloop system starting inside the controllability region C can leave it. As a result the closedloop system stability region S can be less than C. The optimal stability region Sopt is equal to controllability region C. In this chapter the computation of controllability region C and the linear controller design maximizing the stability region S of the constrained closedloop system is considered for cases when matrix A has one and two unstable eigenvalues and system (1) has multiple control inputs m ≥ 1.
© 2004 by CRC Press LLC
3
MAXIMUM STABILIZER DESIGN FOR ONE UNSTABLE EIGENVALUE
The unstable subsystem is described by the ﬁrstorder differential equation: z(t) ˙ = az z(t) + bz u(t), where az > 0, bz ∈ R1×m , u(t) ∈ U and z(t) = QT2 x(t). The closedloop system is stable if the following algebraic stability conditions hold: z > 0 ⇒ z˙ < 0; z = 0 ⇒ z˙ = 0; z < 0 ⇒ z˙ > 0. Consider the upper limit on z, while z˙ < 0: z<
m m m 1 1 1 (i) (i) (i) (i) (i) (− b(i) u ) ≤ max (− b u ) = bz umax sign(b(i) z z z ). u∈U az az a z i=1 i=1 i=1
Applying the same reasoning to obtain the lower limit on z, while z˙ > 0, one can see that the controllability region of the unstable subsystem is an interval: −zmax < z < zmax , where m 1 (i) (i) zmax = ( b u ). (7) az i=1 z max In the statespace of the original system the controllability region looks like a “strip” that is deﬁned by the following linear inequality: −zmax < QT2 x < zmax .
(8)
To derive a saturated linear control law u(i) (t) = sat(ki z(t)) maximizing the closedloop system stability region, it is necessary to satisfy the following two requirements: 1. In the absence of saturation the closedloop system must be stable: az +
m
b(i) z ki < 0.
(9)
i=1
2. On the boundary of controllability region the saturated control must satisfy the boundary conditions, which prevent leaving the controllability region. They result in the following inequalities for each ki : (i)
(i)
bz < 0 ⇒ ki zmax ≥ umax (i) (i) bz > 0 ⇒ ki zmax ≤ −umax (i) bz = 0 ⇒ ki = 0.
(10)
Any linear control law that satisﬁes the system of linear inequalities (9) and (10) is the maximum stabilizer, i.e., it provides the system with the maximal stability region, which is equal to controllability region C. This control law can be computed, for example, via linear programming.
© 2004 by CRC Press LLC
Example 1. To illustrate the method let us consider the following linearized lateraldirectional dynamics of an aircraft without a vertical tail [2]. The stability and controllability matrices correspond to ﬂight at Mach number M = 0.4, altitude H = 15000 ft, and angle of attack α = 8.8 deg: [AB] = −0.004 0.154 −0.988 −8.210 −0.785 0.117 −0.889 −0.030 −0.016
−0.009 0.013 0.006 0.011 −0.004 7.570 −4.970 3.650 0.079 −0.614 . 0.091 −0.181 −0.416 −0.804 0.188
The state vector x = (β, p, r)T , where β is the sideslip angle in degrees, p, r are the bodyaxis roll and yaw rates in degrees per second. The control vector u = (δe , δs , δt , δytv , δf )T , where δe are the differential elevons, δs are the differential spoiler/slotdeﬂectors, δt are the differential all moving tips, δytv is the yaw thrust vectoring, and δf is the differential outboard leadingedge ﬂaps, all control deﬂections are expressed in degrees. The following unstable subsystem can be obtained via the Schur decomposition of matrix A: [az bz ] = [ 0.434
−0.485
0.206 −0.654
−0.737
0.218 ].
The variable z of unstable subspace is connected with the system state vector x as z(t) = QT2 x(t),
QT2 = [ −0.421
−0.075
0.904 ].
All control variables are constrained by the same value: u(i)  ≤ 5, i = 1, 5. In accordance with (7) and (8), the controllability region of unstable subsystem is deﬁned by zmax = 26.472. To compute the linear maximum stabilizer the following system of linear inequalities must be solved: 0.434 − 0.485k1 + 0.206k2 − 0.654k3 − 0.737k4 + 0.218k5 ≤ γ, 26.472k1 ≥ 5, 26.472k2 ≤ −5, 26.472k3 ≥ 5, 26.472k4 ≥ 5, 26.472k5 ≤ −5, where γ < 0. One of the possible solutions for γ = −1 is k1 = 0.5365, k2 = −0.2276, k3 = 0.7236, k4 = 0.816, k5 = −0.2412. The ﬁnal step is to convert the control law to a linear function of the original system state vector: −0.2259 −0.0405 0.4850 0.0958 0.0172 −0.2058 k1 QT2 = −0.3046 −0.0546 0.6541 Kmax = ... . −0.3435 −0.0615 0.7377 k5 QT2 0.1015 0.0182 −0.2180 Figure 1 shows the crosssections of the controllability region (solid lines) in the planes of two state variables. The controllability region and the stability region for the maximum stabilizer u(t) = sat(Kmax x(t)) coincide.
© 2004 by CRC Press LLC
Figure 1 Crosssections of stability regions for maximum stabilizer and linearquadratic regulator (LQR), example of control functions produced by maximum stabilizer (middle right plot), closedloop system trajectories for LQR (bottom left plot) and maximum stabilizer (bottom right plot) in the (β,r) plane.
© 2004 by CRC Press LLC
The linearquadratic regulator (LQR) minimizing the performance index, ∞ J = (xT (t)x(t) + uT (t)u(t))dt,
(11)
0
has been synthesized for comparison purposes. The LQR stability region boundaries (dashed lines) have been computed by direct closedloop system simulation and they are shown in Fig. 1 (see the top and left middle plots). One can see that the maximum stabilizer rejects the level of external disturbances approximately three times bigger than LQR. In the bottom plots in Fig. 1 several trajectories of the closedloop system with LQR and maximum stabilizer controllers are shown to illustrate the size of stability regions.
4
STABILITY SUBOPTIMAL CONTROLLER FOR TWO UNSTABLE EIGENVALUES
In this case the unstable subsystem is described by a secondorder linear system z(t) ˙ = Az z(t) + Bz u(t), where Az ∈ R2 , Bz ∈ R2×m , u(t) ∈ U and z(t) = QT2 x(t). 4.1
Controllability Region Computation
A variational method for controllability region computation for an unstable system with only unstable eigenvalues has been developed by Formalsky [6]. The idea of this method is in maximization of a scalar linear function f (z) = lT z considering points in the controllability region. The maximization process is performed toward the direction l, so the ﬁnal solution coincides with the point zl on the boundary of controllability region C having vector l normal to the boundary at this point (see the top left plot in Fig. 2): zl = arg max(lT z). z∈C
A solution of this task always exists, because the controllability region from Theorem 2 is totally bounded. The optimization task may be solved for different directions lk and obtained points zlk can approximate the controllability region boundary as well, as necessary, because according to Theorem 1 the controllability region is convex. To compute the controllability region boundary, let us consider a nullcontrollable set Ct of all points in the statespace, from which the system can be moved to the origin in the presence of control constraints in a time less than or equal to t. Note that the boundaries of nullcontrollable sets approach the controllability region boundary as t → ∞. The points in a nullcontrollable set Ct can be expressed explicitly as initial points z(0) for the Cauchy problem through the applied control function u(τ ) on the time interval [0, t]. The general solution of system (1), Az t
z(t) = e
t z(0) + 0
© 2004 by CRC Press LLC
eAz (t−τ ) Bz u(τ )dτ, z(t) = 0,
can be used to describe nullcontrollable sets as follows: t Ct = {zz = −
e−Az τ Bz u(τ )dτ, u(τ ) ∈ U }.
0
The optimization task for maximizing a linear function lT z on Ct has only one solution zl : T
zl = arg max l z = − z∈Ct
m
t T −Az τ e−Az τ bzi u(i) bzi )dτ, max sign(−l e
(12)
i=1 0
where bzi is a column of matrix Bz (see the top left plot in Fig. 2). Computing boundary points zlk of a nullcontrollable set Ct for a number of different directions lk , 2πk lk = (cos ϕk , sin ϕk )T ; ϕk = , k = 1, N , N we can approximate Ct rather accurately. Considering a sequence of ti = ti−1 + ∆ approaching inﬁnity ti → ∞ we will have Cti → C. This convergence process can be stopped if max zlk (t) − zlk (t + ∆)2 ≤ , k
where ∆ is a small time increment and is required accuracy.
4.2
Stability Suboptimal Linear Controller
The design objectives in this case are similar to the previous design with one real eigenvalue. A saturated linear controller u = sat(Kz) has to stabilize the closedloop system in the absence of control saturation and prevent leaving the controllability region when system operates close to its boundary. We will consider a nullcontrollable set Ct with large enough t as an approximation of controllability region C. In every boundary point of Ct the control function following (12) is expressed in the form of relay law: T u(i) (0) = u(i) (13) max sign(−l bzi ), i = 1, m, where l is a vector normal to the boundary of Ct at this point. Each component of control vector takes on the boundary of Ct at its positive or negative extreme value and there are some points where it changes sign. Because any nullcontrollable set is symmetrical, the points on the boundary of Ct , where control changes its sign, are also symmetrical. These points ±z0i for every component of control vector u(i) are deﬁned by the condition that the vectors ±li , which are normal to the boundary of Ct at these points, are also normal to controllability vector bzi , i.e., liT bzi = 0 (see the top right plot in Fig. 2). A relay control law T u(i) (z) = u(i) max sign(ki z), i = 1, m,
© 2004 by CRC Press LLC
(14)
where vectors ki are normal to switch lines passing through ±z0i , can produce at every boundary point of Ct the same direction of control vector as in (13), and therefore prevent the system from moving outside of Ct . The direction of ki is deﬁned from inequality kiT bzi < 0. In this case produced control vector bzi u(i) (z) will “push” the system state toward the controllability region interior. We will consider the linear control law with ki vectors from relay control law (14): T u(i) = γki∗ z,
(i)
where
ki∗ =
ki umax . ki 2
(15)
At the presence of control deﬂection constraints the control functions become nonlinear and can be expressed as T u(i) (z) = u(i) max sat(γki∗ z), i = 1, m.
The switching lines of relay control law (14) in the phase plane, deﬁned by kiT z = 0, are transformed into “strips,” where the closedloop system is free from control saturation and therefore linear. Parameter γ, which is the same for all control components, allows one to vary the width of these “strips” and thus to ﬁnd the compromise between the local stability characteristics expressed in terms of eigenvalues location and the size of the closedloop system stability region. Of course, the stability region of this saturated linear control law is less than the controllability region, but in practice it can approach it very closely. Example 2. As a design example, consider the longitudinal dynamics of an aircraft trimmed at altitude H = 25000 ft and Mach number M = 0.9 [1, 15]. The linearized system of motion equations has the following stability and controllability matrices: [AB] = −0.0226 −36.6170 −18.8970 −32.0900 3.2509 −0.7626 0.0001 −1.8997 0.9831 −0.0007 −0.1708 −0.0050 0.0123 11.7200 −2.6316 0.0009 −31.6040 22.3960 0 0 1 0 0 0 0 0 0 0 −30 0 0 0 0 0 0 −30
0 0 0 0 30 0
0 0 0 0 0 30
.
The ﬁrst part of the state vector consists of the aircraft’s basic rigid body variables: perturbation in the forward velocity δV , the angle of attack α, the pitch rate q, and the pitch attitude angle θ. Two ﬁrstorder lags x(5) and x(6) are appended to the state vector to represent actuators dynamics. The control variables δe and δc are signals sent to elevon and canard actuators. In the control law design the following constraints are considered: δe  ≤ 0.035, δc  ≤ 0.035. Note that the angles and angular rates are expressed in radians and radians per second, respectively. The system due to aircraft aerodynamic static instability has an oscillatory unstable mode in the phugoid frequency range, which can be separated off via the Schur decomposition: 0.7067 2.6227 5.1766 −3.6583 [Az Bz ] = , −0.0237 0.6729 6.7815 −4.6968
© 2004 by CRC Press LLC
z(t) = QT2 x(t), QT2 0.0098 −0.0377 −0.1852 −0.9589 0.1726 −0.1219 = . −0.0036 −0.9269 −0.2146 0.1385 0.2260 −0.1566 Using the Formalsky method, the nullcontrollable sets Ct have been computed until convergence to the boundary of controllability region C. In this example, the convergence process has been stopped at t = 10 sec. In the middle left plot in Fig. 2 the nullcontrollable sets are plotted with time increment ∆ = 0.2 sec in the interval t = [0, 10] sec. The nullcontrollable sets were approximated by 120 points. To design linear control law (15), the vectors li , which are normal to the boundary of C10 at the points where control functions change the sign (liT bzi = 0, i = 1, 2), have been deﬁned: (2)
(1)
l1 = (bz1 , −bz1 )T = (6.7815, −5.1766)T , (2) (1) l2 = (bz2 , −bz2 )T = (−4.6968, 3.6583)T . Now the control switching points on the boundary of controllability region, approximated by C10 , can be computed: z01 = arg max l1T z z∈C10
2
10
= −0.035 z02 =
e−Az τ bzi sign(−l1T e−Az τ bzi )dτ = (1.5703, −0.5421)T ,
i=1 0 arg max l2T z z∈C10 2
10
= −0.035
e−Az τ bzi sign(−l2T e−Az τ bzi )dτ = (−1.5703, 0.5421)T .
i=1 0
In our case there are two switching lines passing through points ±z01 and ±z02 on the controllability region boundary, and these lines are numerically the same for each of two controls. These switching lines are described by the following equations: k1T z = 0, k2T z = 0, where ki are deﬁned as (2)
(1)
k1 = (z01 , −z01 )T = (−0.5421, −1.5703)T , (2) (1) k2 = (z02 , −z02 )T = (0.5421, 1.5703)T . Since k1T bz1 < 0, k2T bz2 < 0, these ki already have correct signs. In the middle right plot in Fig. 2 the rootlocus for the closedloop system with control law (15) is shown for γ = 0 ÷ 10. Parameter γ = 6 is taken as a candidate for the “stability suboptimal” control law.
© 2004 by CRC Press LLC
Figure 2 The Formalsky method (top left plot), relay law computation (top right plot), nullcontrollable sets (middle left plot), closedloop system root locus (middle right plot), stability and controllability regions for unstable subsystem (bottom left plot), crosssections of stability regions in the plane (α, θ) (bottom right plot).
© 2004 by CRC Press LLC
In the statespace of the original system this suboptimal control law has the form u(x) = sat(Ksub x), where Ksub = =
0.035 T T k Q2 6 k 1 2 1 0.035 T T 6 k k Q2 2 2 2
0 0.186 0.0552 0.0382 −0.0566 0.0394 0 −0.186 −0.0552 −0.0382 0.0566 −0.0394
.
The comparison of the stability region for the designed “stability suboptimal” control law and the openloop system controllability region is presented in the bottom left and right plots in Fig. 2 in the plane of unstable variables and in physical plane (α, θ). It is clearly seen that the closedloop system stability region with designed linear control is rather close to the controllability region. In the two bottom plots in Fig. 2 one can see also the crosssections of stability regions for two linear quadratic regulators designed for comparison purposes. Linear quadratic regulators have been computed by minimizing the performance index (11): one considering only unstable subspace LQR1 and the other for the original system including stable and unstable subspaces LQR2 . One can see that the stability region for LQR1 is bigger than the region for the LQR2 , because LQR1 stabilizes only the unstable subsystem, while LQR2 allocates some control resourses for the stable subspace and leaves less for the unstable one. Meanwhile the designed stability suboptimal control law provides a much larger stability region than both linear quadratic regulators and its size is very close to that of the controllability region.
CONCLUDING REMARKS 1. Stabilization of an aircraft’s linearized dynamics with one unstable eigenvalue and a redundant number of control effectors can be performed in the whole controllability region by the saturated linear control law. The maximum stabilizer controller can be designed by application of the linear programming method. 2. In the case of two unstable eigenvalues the saturated linear controller can be only the “stability suboptimal” one. The maximum stabilizer providing the necessary stability characteristics can be designed only in nonlinear form. 3. The problem of constrained stabilization of unstable dynamics requires further development of control design methods including more unstable modes and control rate saturation effects.
REFERENCES 1. G.L. Hartmann, M.F. Barrett, and C.S. Greene, “Control design for an unstable vehicle,” NASA Dryden Flight Research Center, Contract Rep. NAS 42578, Dec. 1979. 2. A.D. Ngo, W.C. Reigelsperger, S.S. Banda, and J.S. Bessolo, “Multivariable control law design of a tailless airplane,” AIAA Paper 963866, July 1996.
© 2004 by CRC Press LLC
3. D. Cameron and N. Princen, “Control allocation challenges and reguirements for the Blended Wing Body,” AIAA Guidance, Navigation and Control Conference, August 2000, AIAA Paper 20004539. 4. P.P. Friedmann, “Renaissance of aeroelasticity and its future,” Journal of Aircraft, Vol. 36, No. 1, Jan.–Feb. 1999. 5. V. Mukhopadhyay, “Transonic ﬂutter suppression control law design and wind tunnel test results,” Journal of Guidance, Control, and Dynamics, Vol. 23, No. 5, Sept.–Oct. 2000. 6. A.M. Formalsky, Controllability and Stability of Systems with Control Constraints, Nauka, Moscow, 1974 (in Russian). 7. G.A. Stepanjantz, Theory of Dynamical Systems, Mashinostroenie, Moscow, 1985 (in Russian). 8. M.G. Goman, E.V. Fedulova, and A.V. Khramtsovsky, “Maximum stability region design for unstable aircraft with control constraints,” AIAA Guidance, Navigation and Control Conference, July 1996, AIAA Paper 963910. 9. F. Blanchini, and S. Miani, “Constrained stabilization of continuoustime linear systems,” Systems and Control Letters, Vol. 28, 1996, pp. 95–102. 10. V.A. Kamenetskiy, “Construction of a constrained stabilizing feedback by Lyapunov functions,” Proceedings of 13th IFAC World Congress, San Francisco, 1996 (2b04 6), pp. 139–144. 11. T.E. Pare, H. Hindi, J.P. How, and S.P. Boyd, “Synthesizing stability regions for systems with saturating actuators,” Proceedings of the 37th IEEE Conference on Decision & Control, 1998, pp. 1981–1982. 12. L. Scibile and B. Kouvaritakis, “Stability region for a class of openloop unstable linear systems: theory and application,” Automatica, Vol. 36, No. 1, 2000, pp. 37–44. 13. D. Henrion, S. Tarbouriech, and V. Kuˇcera, “Control of linear systems subject to input constraints: a polynomial approach,” Automatica, Vol. 37, 2001, pp. 597–604. 14. G.H. Golub and C.F. Van Loan, Matrix Computations, North Oxford Academic, London, 1986. 15. R.Y. Chiang and M.G. Safonov, MATLAB Robust Control Toolbox User’s Guide, The MathWorks, Natick, MA, 1998.
© 2004 by CRC Press LLC
9 QUEST Algorithms Case Study: GPSBased Attitude Determination of Gyrostat Satellite Jinlu Kuang and Soonhie Tan Satellite Engineering Center, The School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
The theoretical methodologies and numerical results of quaternion estimation (QUEST) algorithms for the GPSbased attitude determination of a lowearthorbit gyrostat satellite are presented. The attitude tumbling of the gyrostat satellite is suppressed with the proposed linear regulators. The numerical techniques for the generation of angular velocities of the gyrostat satellite and the determination of the attitude are investigated. The simulation results indicate that QUEST algorithms can be employed to accurately estimate the attitude of the gyrostat satellite point by point, when combined with the GPS phasedifference signals. Under the appropriate assumption of measurement noises at the GPS phasedifference signals the precision of the attitude estimates is assessed to be at the level of 0.5 degrees for the attitude motions of the gyrostat satellite. The results from the theoretic simulation in this chapter can be used to validate the algorithms of the integer offset time biases and the assumption of lineofsight measurement noises and baseline calibration.
1
INTRODUCTION
The utilization of phasedifference measurements from the GPS receivers to determine the attitude of satellites, aircraft, and ships has recently attracted the attention of many scientists, because of the GPS sensors’ potential to provide effective cost in terms of dollars, weight, and power. In particular, the application of the GPSbased attitude determination for satellites has been focused upon. Crassidis, Lightsey, and Markley (1999) developed an approach for the efﬁcient and optimal attitude determination by using recursive GPS signal operations. They derived their new algorithms from the generalized predictive ﬁlter for nonlinear systems proposed by Lu (1994). The distinct advantages of their algorithms are that they can provide optimal attitude solutions even for coplanar baseline conﬁgurations. Crassidis and Markley (1997) created another novel technique for ﬁnding a pointbypoint (deterministic) attitude solution of vehicle by using GPS phasedifference measurements, based upon the algorithms for attitude determination by using GPS phase measurements from the research by BarItzhack, Montgomery, and Garrick (1997). In their paper the general GPS cost function was transformed into a Wahba cost function. They formulated
© 2004 by CRC Press LLC
Figure 1 The conﬁguration of the gyrostat rotating about the center of mass O.
the covariance equations for their new method. Crassidis, Markley, and Lightsey (1999) investigated the GPS integer ambiguity resolution without attitude knowledge. This work was extended from the one by Hill and Euler (1996). The algorithm proposed by Crassidis, Markley, and Lightsey (1999) was sequential and independent of attitude, so that it may be implemented in real time. BarItzhack (1996) constructed a recursive QUEST algorithm for sequential attitude determination. Axelrad and Ward (1996) presented the methodology and results for the RADCAL satellite’s attitude estimation using the GPS signals. Their research was based upon the method of Kalman ﬁltering. The precision of the ﬁnal attitude estimation is tested to be at the level of 0.4 deg for pitch and roll and 0.7 deg for yaw. Fujikawa and Zimbelman (1995) also studied the spacecraft attitude determination by Kalman ﬁltering of the GPS signals based upon the observation of the phasedifference measurement. They estimated the vehicle attitude, body rate, disturbance torque, and antenna phase center shift. The factors affecting the system accuracy are also examined, which include the GPS satellite visibility, the number of receiving antennas, and the attitude error observability. Cohen, Lightsey, and Parkinson (1994) investigated the preliminary space ﬂight results (RADCAL) in a low earth orbit. Conway et al. (1996) invented a new motionbased algorithm for the GPS attitude integer resolution, which was suitable for conditions of satellite visibility. After reviewing the representative references, we discovered that the approaches used to estimate the attitude angles or quaternions could be classiﬁed into two types. One is the pointbypoint solution for the attitude, e.g., QUEST; the other is an iterative method for the attitude, e.g., discrete Kalman ﬁltering algorithms. The theory of the Kalman ﬁltering algorithms can be found in the book written by Balakrishnan (1987). The Kalman ﬁltering method can be used to deal with the estimation of the attitude, angular velocities, cable biases, and baseline calibration at the same time. The wellknown method of pointbypoint solutions (e.g., QUEST) is a welcome algorithm for leastsquares ﬁtting of the attitude quaternions of a satellite to vector measurements. In this chapter, assuming that the integer offset line biases, lineofsight measurement noises and the baselinecalibration are known and compensated in advance, we discover that QUEST will provide good solutions of attitude for the gyrostat satellite by using the GPS signals. The motivation for this chapter is
© 2004 by CRC Press LLC
based upon the work written by Crassidis and Markley (1997). We formulated the complete skeleton of how to fulﬁll the QUEST algorithms by using the GPS signals. The chapter is organized as follows. First, the basic concepts of the GPS phasedifference measurements and the idea of the GPSbased attitude determination of the gyrostat satellites by using the QUEST algorithms are introduced. Next, the attitude propagation equations of the gyrostat satellite are described. In the case of no external disturbance torques being applied to the gyrostat satellite, the exactly analytical solutions from Volterra (1898) are utilized. In the case of external disturbance torques being applied to the gyrostat, the adaptive fourthorder Runge–Kutta integration algorithm is employed to generate the propagation solutions of its angular velocities and quaternions. Then, in order to show the dynamics encountered in the control design of the attitude of a gyrostat satellite whose dynamical equations have multiple equilibrium positions, the equations of multiple equilibrium positions are derived and the stability of the equilibrium positions is also analyzed. Finally, the control law (linear regulator) for the suppression of the attitude tumbling of the gyrostat satellite is proposed. The simulation for the motion of the low Earth orbital gyrostat satellite is given to validate the QUEST method and the proposed linear regulators. During the run there are at least 5 GPS satellites available. The lines of sight and baselines are used to form simulated phasedifference measurements with Gaussian measurement errors. The GPS orbit data are generated from the GPSoft software.
2
QUEST ALGORITHMS FOR GPSBASED ATTITUDE DETERMINATION
Now consider a gyrostat satellite in a circular, low earth orbit possessing a bodyﬁxed reference with basis vectors {i1 , i2 , i3 }. The local vertical and local horizontal (LVLH) reference frame with its origin at the center of mass of an orbiting user satellite has a set of vectors {e1 , e2 , e3 }, with e1 perpendicular to the orbit plane, e2 along the orbit direction, and e3 outward from the Earth center (see Fig. 1). The attitude of the user satellite is assumed to be tumbling in a corresponding circular orbit. This situation represents a realistic scenario when the attitude of the user satellite is out of control. The GPS constellation consists of 24 satellites (21 active plus 3 spares) oriented in six evenly spaced orbit planes inclined 55 degrees relative to the equator [Cohen, 1992]. The measurement utilized for the attitude determination is supposed to be the phase difference of the GPS signal received from two antennae separated by a baseline. Cohen (1992) formulated the attitude determination problem in terms of the GPS observable. The theory of the wavefront angle and wavelength are used to develop the phase difference. The phase measurement is described by bl cos(θ) = λ (k + ∆ϕa /2π),
(1)
where bl is the baseline length; θ is the angle between the baseline and the line of sight to the same GPS satellite; k is the number of integer wavelengths of the phase difference between two antennae of the receivers; ∆ϕa is the actual phasedifference measurement; and λ is the wavelength of the GPS signal (see Fig. 2). All GPS satellites generate a carrier signal at the L1 frequency. This frequency is centered on 1575.42 MHz, which corresponds to a wavelength of 19.03 cm. For the details of the GPS phasedifference signals’ application to the attitude determination of the satellites, the readers are referred
© 2004 by CRC Press LLC
Figure 2 GPS phase difference geometry.
to Crassidis, Lightsey, and Markley (1999). Assuming that the integer offset is known and compensated, the measure normalized fractional phase difference ∆ϕ can be deﬁned by ∆ϕ = bT C s + v;
(2)
here s ∈ R3 is the normalized lineofsight vector to the GPS satellite in a local vertical and local horizontal (LVLH) frame; b ∈ R3 is the normalized baseline vector, which is the relative position vector from one antenna (master) to another antenna (slave) installed in the user gyrostat satellite, and C is the orthogonal attitude matrix, which is the transformation matrix from the LVLH reference frame to the bodyﬁxed reference frame; and v is the measurement error, which is assumed to be a zeromean random Gaussian process with the standard deviation given by σ, which is 0.5 cm/λ = 0.026 wavelengths for the typical phase noise. Equation (2) is used to form equations of observation that are nonlinear and timevariant. In order to determine the attitude of the gyrostat satellite using Eq. (2), Crassidis and Markley (1997), BarItzhack et al. (1997) converted the phase measurements into vector measurements. The method extended by them involves ﬁnding the proper orthogonal matrix C, which minimizes the following generalized loss function: J(C) = 0.5
n m
−2 σij (∆ϕ − bTi C sj )2 ,
(3)
i=1 j=1
where m stands for the number of baselines and n stands for the number of observed GPS satellites, and σij stands for the standard deviation of the ijth measurement error. The vectorized measurement problem involves determining the sight line vector in the body frame represented by S j = C sj , (4) or the baseline in a LVLH frame represented by B i = C T bi .
© 2004 by CRC Press LLC
(5)
For the sight line case, the following cost function is minimized: Jj (S j ) = 0.5
m
−2 σij (∆ϕij − bTi S j )2 ;
for j = 1, 2, 3, · · · , n.
(6)
i=1
The minimization of Eq. (6) leads to S j = M −1 j yj ,
(7)
where Mj =
m
2 bi bTi /σij ;
yj =
i=1
m
2 ∆ϕij bi /σij
for j = 1, 2, 3, · · · , n.
(8)
i=1
According to Eq. (6) it is required that at least three baselines should not be coplanar for the sight lines to be determined in the body frame. Readers are referred to Crassidis and Markley (1997) for the extensive treatment of the case of baselines and the relevant special cases. Wahba (1965) posed the problem of attitude matrix determination as a leastsquares problem as follows: n 2 Jw (C) = 0.5 aj S j − C sj  . (9) j=1
We can weigh each measurement respectively according to the accuracy of the particular vector measurement. The weight coefﬁcients aj are positive numbers assigned to each measurement. In order to ﬁnd the solution of attitude quaternions, rather than attitude matrix (the representation of the attitude), Eq. (9) can be replaced by Jq (q) = 0.5
n
2
aj S j − C(q) sj  ,
(10)
j=1
where q represents the column vector of the attitude quaternions (q1 , q2 , q3 , q4 )T , which describe the orientation of the bodyﬁxed frame to the LVLH reference frame; and C is deﬁned as follows (Wie, 1995): C11 C12 C13 C = C21 C22 C23 C31 C32 C33 2 2(q1 q2 + q3 q4 ) 2(q1 q3 − q2 q4 ) q4 − q32 − q22 + q12 q42 − q32 + q22 − q12 2(q2 q3 + q1 q4 ) . (11) = 2(q1 q2 − q3 q4 ) 2(q2 q4 + q1 q3 ) 2(q2 q3 − q1 q4 ) q42 + q32 − q22 − q12 Instead of minimizing the function in Eq. (9) or (10), we can maximize the following function (Wertz, 1978): g(q) = q T Kq, (12) where K is constructed as follows. Deﬁne B=
n
aj S j sTj /w ;
j=1
z=
n
aj S j × sj /w = (B23 − B32 , B31 − B13 , B12 − B21 )T ,
j=1
(13)
© 2004 by CRC Press LLC
w=
n
aj ;
F =B+B ; T
σa =
j=1
n
aj S Tj sj /w = trace(B) = B11 + B22 + B33 .
j=1
(14)
Then K=
F − σa I zT
z σa
,
(15)
where I is the thirdorder identity matrix, and Bij (i, j = 1, 2, 3) are elements of the matrix B. The extrema of g(q), subject to the normalization constraint q12 + q22 + q32 + q42 = 1,
(16)
can be found by the method of Lagrange multipliers. It is shown that the optimal quaternion q ∗ satisﬁes K q∗ = χ q∗ , (17) where χ is a yet undetermined Lagrange multiplier. By choosing the largest eigenvalue χmax as the desired optimal eigenvalue, the function g(q) can be maximized and the optimal quaternion q ∗ corresponding to the largest eigenvalue χmax can be found using the known relation ∗ y / 1 + y ∗2 , (18) q∗ = 1/ 1 + y ∗2 where
−1
y ∗ = [(χmax + σa )I − F ]
z.
(19)
The algorithm for obtaining χmax and q ∗ from the vector observations discussed is known as the quaternion estimation (QUEST) algorithm. Shuster (1990) showed how to deal with the singularity matrix case in Eq. (19). The weight coefﬁcients aj (j = 1, 2, · · · , n) can be selected from the assumed values of the standard deviation σij of the measured phasedifference signals (BarItzhack, 1996). QUEST is a famous algorithm for leastsquares ﬁtting of the attitude quaternion of a satellite to vector measurements. QUEST is also a singlepoint attitude determination algorithm, which yields a closedform solution of the quaternion and therefore experiences no divergence problems. The divergence problems are sometimes encountered in the use of extended Kalman ﬁlter approach. BarItzhack (1996) developed an algorithm of a recursive QUEST (REQUEST) algorithms for sequential attitude determination, which considered all past measurements based upon the fact that the socalled K matrix, one of whose eigenvectors is the sought quaternion, is linearly related to the measured vector pairs and on the ability to propagate K. REQUEST yields the quaternion estimates that are functions of gyro bias in practical situations. In this chapter we combined the QUEST algorithms and GPS resources to determine the attitude of the gyrostat satellite.
3
ATTITUDE PROPAGATION EQUATIONS OF GYROSTAT SATELLITES
A model of the attitude dynamics of the gyrostat satellite is needed to smooth the state estimate and to provide an attitude reference between GPS updates and during periods of
© 2004 by CRC Press LLC
poor observing geometry. Here, between measurement epochs, the quaternion and angular velocity states are advanced to the current measurement time by numerically integrating the nonlinear equations of motion for a gyrostat satellite. These ordinary differential equations are described as follows from the principles of analytical mechanics: q˙1 = 0.5(ω1 q4 − ω2 q3 + ω3 q2 − Ωq4 ), q˙2 = 0.5(ω1 q3 + ω2 q4 − ω3 q1 + Ωq3 ), q˙3 = 0.5(−ω1 q2 + ω2 q1 + ω3 q4 − Ωq2 ), q˙4 = 0.5(−ω1 q1 − ω2 q2 − ω3 q3 + Ωq1 ),
(20) (21) (22) (23)
I1 ω˙ 1 − (I2 − I3 )ω2 ω3 + ω2 h3 − ω3 h2 = u1 + d1 , I2 ω˙ 2 − (I3 − I1 )ω3 ω1 + ω3 h1 − ω1 h3 = u2 + d2 , I3 ω˙ 3 − (I1 − I2 )ω1 ω2 + ω1 h2 − ω2 h1 = u3 + d3 ,
(24) (25) (26)
where I1 , I2 , and I3 are the principal moments of inertia of the gyrostat satellite; ω1 , ω2 , and ω3 are the components of the angular velocities along the bodyﬁxed reference frame; h1 , h2 , and h3 are assumed to be constant components of the angular momenta of the wheels in the bodyﬁxed coordinate system; Ω is the orbit rate; u1 , u2 , and u3 are components of the control torques; d1 , d2 , and d3 are the components of external disturbance torques, which may be made up of the gravitygradient torques, the aerodynamic torques, torques due to solar radiation pressure, and residual magnetic dipole moment. The three Euler angles of C1 (θ1 ) ← C3 (θ3 ) ← C2 (θ2 ) sequence from the LVLH reference frame to a bodyﬁxed reference of a gyrostat satellite in a circular orbit are related to quaternions (Wie, 1998): q1 = cos(0.5θ1 ) sin(0.5θ2 ) sin(0.5θ3 ) + sin(0.5θ1 ) cos(0.5θ2 ) cos(0.5θ3 ), q2 = cos(0.5θ1 ) sin(0.5θ2 ) cos(0.5θ3 ) + sin(0.5θ1 ) cos(0.5θ2 ) sin(0.5θ3 ), q3 = cos(0.5θ1 ) cos(0.5θ2 ) sin(0.5θ3 ) − sin(0.5θ1 ) sin(0.5θ2 ) cos(0.5θ3 ), q4 = cos(0.5θ1 ) cos(0.5θ2 ) cos(0.5θ3 ) − sin(0.5θ1 ) sin(0.5θ2 ) sin(0.5θ3 ). The inverse relationships can be derived as follows: θ1 = tan−1 [2(q1 q4 − q2 q3 ) (1 − 2q12 − 2q32 )], θ2 = tan−1 [2(q2 q4 − q1 q3 ) (1 − 2q22 − 2q32 )], −1
θ3 = sin
[2(q1 q2 + q3 q4 )].
(27) (28) (29) (30)
(31) (32) (33)
During the longterm mission, a fourthorder Runge–Kutta integration algorithm was used to give the sequential solutions of the angular velocities and attitude angles or quaternions. If we neglect the external disturbance/control torques including the terms related to Ω in Eqs. (20)–(26), the angular velocities can be solved analytically (Volterra, 1989; Wittenburg, 1975). Assume that the kinetic energy is T (constant) and the moment of momentum is 2 D T (D is also a known constant). Also assume that the real numbers λ1 < λ2 < λ3 < λ4 are the eigenvalues of the following generalized characteristic equations: Z(λ) = det(A − λ B) = 0, (34)
© 2004 by CRC Press LLC
where
0 0 I1 h1 I12 0 1 I22 0 I2 h2 , A= 2 0 0 I I h 3 3 (I1 I2 I3 ) 3 I1 h1 I2 h2 I3 h3 h2 − 2DT 0 I1 0 0 0 I2 0 1 0 ; h2 = h21 + h22 + h23 , B= 0 (I1 I2 I3 ) 0 0 I3 0 0 0 −2T
(35)
(36)
then
Ai1 sn(τ, k) + Ai2 cn(τ, k) + Ai3 dn(τ, k) + Ai4 , B1 sn(τ, k) + B2 cn(τ, k) + B3 dn(τ, k) + B4 where sn, cn, and dn are the Jacobi elliptic functions with modulus k. 2T (λ1 − λ3 )(λ2 − λ4 ) (λ1 − λ2 )(λ3 − λ4 ) + τ0 , k2 = , τ =t I1 I2 I3 (λ1 − λ3 )(λ2 − λ4 ) ωi (t) =
B1 = (λ2 − λ3 )(λ3 − λ4 ) (λ1 − I1 )(λ1 − I2 )(λ1 − I3 )(λ4 − λ2 ), B2 = (λ1 − λ3 )(λ3 − λ4 ) (λ2 − I1 )(λ2 − I2 )(λ2 − I3 )(λ1 − λ4 ), B3 = (λ2 − λ4 ) (λ3 − I1 )(λ3 − I2 )(λ3 − I3 )(λ1 − λ2 )(λ1 − λ3 )(λ1 − λ4 ), B4 = (λ1 − λ3 )(λ2 − λ3 ) (λ4 − I1 )(λ4 − I2 )(λ4 − I3 )(λ1 − λ2 ), for i = 1, 2, 3; j = 1, 2, 3, 4, Aij = hi Bj /(λj − Ii ) ;
(37)
(38)
(39) (40) (41) (42) (43)
and τ0 depending upon the initial conditions. If we set the initial angular velocities as ω10 , ω20 , and ω30 , then τ0 satisﬁes the following equation: ω + p12 ωh20 + p13 ωh30 + p14 (λ4 − λ2 )(λ1 − I1 )(λ1 − I2 )(λ1 − I3 ) p11 h10 1 2 3 , sn(τ0 , k) = ω10 ω20 ω30 (λ1 − λ2 )(λ4 − I1 )(λ4 − I2 )(λ4 − I3 ) p41 h1 + p42 h2 + p43 h3 + p44 (44) where (Ik − Il )(λα − Ij )(λβ − Ij )(λγ − Ij ), pij = for i = 1, 4; (j, k, l) = (1, 2, 3) cyclic; (α, β, γ) = (1, 2, 3, 4) − (i) cyclic, (I1 − I2 )(I2 − I3 )(I3 − I1 ), for i = 1,4; j = 4. (45) From the characteristic Eq. (34), assuming that the real numbers λ1 < λ2 < λ3 < λ4
(46)
are known, we could derive the formulae of the wheel’s moments of momentum as follows: h21 = 2T (λ1 − I1 )(λ2 − I1 )(λ3 − I1 )(λ4 − I1 )/[I1 (I1 − I2 )(I1 − I3 )],
(47)
h22 h23
= 2T (λ1 − I2 )(λ2 − I2 )(λ3 − I2 )(λ4 − I2 )/[I2 (I2 − I1 )(I2 − I3 )],
(48)
= 2T (λ1 − I3 )(λ2 − I3 )(λ3 − I3 )(λ4 − I3 )/[I3 (I3 − I1 )(I3 − I2 )], D = λ1 2T + h21 (λ1 − I1 ) + h22 (λ1 − I2 ) + h23 (λ1 − I3 ) /(2T ).
(49)
© 2004 by CRC Press LLC
(50)
For the practical application, the sampling interval should be well below Nyquist’s limit. The determined quaternion can be found by integrating the ordinary differential Eqs. (20)–(26) from time tk to time tk+1 . If the angular velocity vector ωk is assumed to be constant over this interval, a discrete propagation for Eqs. (20)–(23) can be used, given by Wertz (1978): q k+1 = [βk I 4×4 + γk Λ(ρk )]q k ,
(51)
where βk = cos(0.5 ωk ∆t);
γk = sin(0.5 ωk ∆t);
0 −ρk3 Λ(ρk ) = ρk2 −ρk1
4
ρk3 0 −ρk1 −ρk2
ρk = ωk /ωk ; −ρk2 ρk1 0 −ρk3
ρk1 ρk2 . ρk3 0
∆t = tk+1 − tk , (52)
(53)
STABILITY ANALYSIS OF A KELVIN GYROSTAT SATELLITE
Now consider the attitude dynamics model of a gyrostat satellite without the external torques. In this case the gyrostat is called a Kevin gyrostat whose Euler equations are described as follows: ω˙ 1 = [(I2 − I3 )ω2 ω3 − ω2 h3 + ω3 h2 ]/I1 ≡ f1 (ω1 , ω2 , ω3 ), ω˙ 2 = [(I3 − I1 )ω3 ω1 − ω3 h1 + ω1 h3 ]/I2 ≡ f2 (ω1 , ω2 , ω3 ), ω˙ 3 = [(I1 − I2 )ω1 ω2 − ω1 h2 + ω2 h1 ]/I3 ≡ f3 (ω1 , ω2 , ω3 ).
(54) (55) (56)
From the theory of the dynamical system, the equilibrium positions should satisfy the following equations: f1 (ω1e , ω2e , ω3e ) = 0, f2 (ω1e , ω2e , ω3e ) = 0, f3 (ω1e , ω2e , ω3e ) = 0.
(57) (58) (59)
According to the law of conservation of kinetic energy of the gyrostat satellite and Eqs. (57) and (59), the following equilibrium positions are derived: ω3e = ω2e h3 /[h2 + (I2 − I3 ) ω2e ], ω1e = ω2e h1 /[h2 + (I2 − I1 ) ω2e ],
(60) (61)
where ω2e are the roots of the following polynomial equation of the 6th power: 6 5 4 3 2 + β5 ω2e + β4 ω2e + β3 ω2e + β2 ω2e + β1 ω2e + β0 = 0, β6 ω2e
© 2004 by CRC Press LLC
(62)
where β6 = I2 (I2 − I1 )2 (I2 − I3 )2 ; β5 = 2I2 h2 (2I2 − I1 − I3 )(I2 − I1 )(I2 − I3 ), β4 = I1 h21 (I2 − I3 )2 + I2 h22 (2I2 − I1 − I3 )2 + 2(I2 − I1 )(I2 − I3 )
(63)
+ I3 h23 (I2 − I1 )2 − 2T (I2 − I1 )2 (I2 − I3 )2 , (64) 2 2 2 β3 = 2h2 I1 h1 (I2 − I3 ) + I2 h2 (2I2 − I1 − I3 ) + I3 h3 (I2 − I1 ) (65) − 4T h2 (2I2 − I1 − I3 )(I2 − I1 )(I2 − I3 ), 2 2 2 2 2 2 β2 = h2 (I1 h1 + I2 h2 + I3 h3 ) − 2T h2 (2I2 − I1 − I3 ) + 2(I2 − I1 )(I2 − I3 ) , (66) β1 = −4T h32 (2I2 − I1 − I3 );
β0 = −2T h42 .
(67)
The Jacobi matrix of the vector functions appearing on the right side of Eqs. (54)–(56) at the equilibrium positions can be deﬁned as follows: f11 f12 f13 ∂(f1 , f2 , f3 ) = f21 f22 f23 , (68) ∂(ω1 , ω2 , ω3 ) (ω1e ,ω2e ,ω3e ) f31 f32 f33 where f11 = 0;
f12 = [(I2 − I3 )ω3e − h3 ]/I1 ;
f13 = [(I2 − I3 )ω2e + h2 ]/I1 ,
(69)
f23 = [(I3 − I1 )ω1e − h1 ]/I2 ,
(70)
f21 = [(I3 − I1 )ω3e + h3 ]/I2 ;
f22 = 0;
f31 = [(I1 − I2 )ω2e − h2 ]/I3 ;
f32 = [(I1 − I2 )ω1e + h1 ]/I3 ;
f33 = 0.
(71)
The characteristic equation corresponding to the Jacobi matrix can be given as follows: f11 − Γ f12 f13 f21 f22 − Γ f23 = 0, (72) f31 f32 f33 − Γ where Γ is the eigenvalue. The instability or stability of the equilibrium positions will depend on the positive or negative signs of the real parts of the eigenvalue Γ. The full understanding of the multiple equilibrium states of the gyrostat satellite dynamic system in Eqs. (60)–(62) will be conducive to the design of the control law. The nonlinear attitude motions of the gyrostat satellite subject to the appropriate perturbations will exhibit chaotic features (Kuang et al., 2000, 2001). 5 SUPPRESSION OF ATTITUDE TUMBLING OF THE GYROSTAT SATELLITE Assume that no disturbance or external control torque is applied to the gyrostat satellite, then the attitude of the gyrostat satellite will be tumbling. How to suppress the tumbling attitude of the gyrostat satellite is one of our concerns from the point of view of dynamics and control. Here, we propose the following feedback controls that can be used to suppress the tumbling attitude of the gyrostat satellite: u1 = −ξ1 ω1 − ψ q1 , u2 = −ξ2 ω2 − ψ q2 , u3 = −ξ3 ω3 − ψ q3 ,
© 2004 by CRC Press LLC
(73) (74) (75)
where ξi , ψ (i = 1, 2, 3) are positive numbers to be selected (Kuang and Leung, 2001) . The equations of states for the closedloop control system consist of Eqs. (20)–(26) (in the case of dx = dy = dz = 0) with the additional constraint (Eq. (16)). Mortensen (1968) used the linear regulator in Eqs. (73)–(75) to control the attitude error of the pure rigid model. To analyze the stability of the equilibrium point, x∗ = (ω1 , ω2 , ω3 , q1 , q2 , q3 , q4 ) = (0, 0, 0, 0, 0, 0, 1),
(76)
we construct the function, E = 0.5(I1 ω12 + I2 ω22 + I3 ω32 ) + ψ(q12 + q22 + q32 + (q4 − 1)2 ).
(77)
It is easy to show that E is a positive deﬁnite quadratic form. The total time derivative can be computed by the chain rule. Considering the state equations of the closedloop system the function E˙ = dE dt can be reduced to the following form: E˙ = −ξ1 ω12 − ξ2 ω22 − ξ3 ω32 .
(78)
Consequently, we know the derivative with respect to time E˙ 0 for all t. According to the Lyapunov stability theory, the perturbed motions with respect to the equilibrium point are globally asymptotically stable for any positive values of ξi , ψ (i = 1, 2, 3). This can be validated using the numerical simulation experiment given in the next section.
6
SIMULATION PROCEDURES AND RESULTS
The following section contains a compilation of attitude solutions and numerical simulation results of the attitude determination using the GPS phasedifference signals from the gyrostat satellite that is assumed to be a small satellite model in a lowEarth, circular, equatorial orbit at an altitude of 870 km, and the weight is 50 kg. The simulation procedures are outlined as follows. First using the GPSoft software to produce the GPS satellite orbits and the user satellite orbit when given the angle of mask, then using the principles of GPS phasemeasurement geometry to construct the line of sight, we can simulate the GPS phasedifference measurement. It is very important to obtain the GPS phasedifference signals as accurately as possible so as to identify the right attitude of the gyrostat satellite. The simulation results show that the attainable precision of the Euler angles’ estimates using QUEST is assessed to be at the level of 10−15 for amplitudes if there are no measurement noises at the GPS phasedifference signals. The physical parameters are √ given as√ follows: I1 = 9 √ kg m2 , I2 = 7 kg m2 , I3 = 3 kg m2 , h1 = 10 21 N m s, h2 = 9 √ 5 N m√ s, h3 = 15 7 N m s. The normalized baseline vectors are assumed to be b1 = [1/ 2, 1/ 2, 0], b2 = [0, 1, 0], b3 = [0, 0, 1] (Kuang and Tan, 2002). CASE 1. Attitude Determination of the Gyrostat Satellite under TorqueFree Conditions Assume that the kinetic energy constant is 2T = 3780 N m, the moment of momentum (constant) is 2DT = 1920 (N m s)2 . The attitude of the gyrostat satellite will be tumbling
© 2004 by CRC Press LLC
Table 1 Equilibrium positions of kelvin gyrostat satellite
λ∗ (kg m2 )
D∗ (N m s2 )
ω1∗ (rad/s)
ω2∗ (rad/s)
ω3∗ (rad/s)
5.5999 4.3799 1.8052 11.305
5.1538 4.9967 0.9990 14.878
−13.478 −9.9187 −6.3693 19.884
−14.373 −7.6807 −3.8740 4.6751
15.265 28.761 −33.216 4.7788
Unstable Stable Stable Stable
Figure 3 The evolution of the Euler angles (theoretical solutions A).
in the orbit. According to the characteristic Eq. (34), the corresponding eigenvalues are computed as follows: λ1 = 8, λ2 = 6, λ3 = 5, λ4 = 4. The equilibrium positions can be calculated according to the Eqs. (60)–(62) and are given in Table 1. In Table 1, the following results apply according to Wittenburg (1975): I1 ω1∗ + h1 = λ∗ ω1∗ ;
I2 ω2∗ + h2 = λ∗ ω2∗ ;
I3 ω3∗ + h3 = λ∗ ω3∗ ;
I1 (ω1∗ )2 + I2 (ω2∗ )2 + I3 (ω3∗ )2 = 2T ; D∗ = (I1 ω1∗ + h1 )2 + (I2 ω2∗ + h2 )2 + (I3 ω3∗ + h3 )2 /(2T ).
(79) (80) (81)
The nominal parameters of moments of inertia and constant moments of momenta of the wheels are calculated from Eqs. (47)–(50). In the case of no external disturbance torques, Fig. 3 shows the evolution history of the Euler angles in Eqs. (31)–(33) (called Theoretical Solutions A), directly from the analytical solutions in Eqs. (37) corresponding to the dynamic attitude Eqs. (20)–(26). Figure 4
© 2004 by CRC Press LLC
Figure 4 The evolution of the Euler angles (QUEST solutions A).
Figure 5 The errors’ evolution of the Euler angles.
shows the evolution history of the Euler angles in Eqs. (31)–(33) (called QUEST Solutions A), directly using GPS signals from the approach of QUEST. Figure 5 shows the errors’ evolution of the Euler angles between Theoretical Solutions A and QUEST Solutions A. By comparison, we found out that the precision of the attitude estimates was tested to be within 0.5 degrees for the attitude motions.
© 2004 by CRC Press LLC
Figure 6 The evolution of the Euler angles (theoretical solutions B).
CASE 2. Attitude Determination of Gyrostat Satellites under GravityGradient Torques
For simplicity here, the aerodynamic torques, torques due to solar radiation pressure, and residual magnetic dipole moment applied to the user satellite are neglected. Assume that the gravitygradient torques and control torques are as follows: d1 = −3Ω2 (I2 − I3 )C23 C33 , 2
d2 = −3Ω (I3 − I1 )C13 C33 , 2
d3 = −3Ω (I1 − I2 )C13 C23 ,
u1 = 0,
(82)
u2 = 0,
(83)
u3 = 0,
(84)
where Ω = µ R3 , and R is the distance of the gyrostat satellite mass center from the Earth’s center, and µ is the gravitational parameter of the Earth. Figure 6 shows the evolution history of the Euler angles (called Theoretical solutions B) directly numerically integrated from dynamical equations excited by the gravitygradient torques using the Runge– Kutta fourthorder method. Figure 7 shows the evolution history of the Euler angles from QUEST method (called QUEST solutions B) with consideration of the gravitygradient torques. Figure 8 shows the errors’ evolution history of the Euler angles between Theoretical solutions B and QUEST solutions B. By comparison, we also found out that the precision of the attitude estimates was tested to also be within 0.5 degrees for the attitude motions.
© 2004 by CRC Press LLC
Figure 7 The evolution of the Euler angles (QUEST solutions B).
Figure 8 The error’s evolution of the Euler angles.
CASE 3. Attitude Determination of Gyrostat Satellites with Linear Feedback Control Torques In the case ξi = 24 (i = 1, 2, 3) N m s , ψ = 18 N m, the linear regulator are deﬁned in Eqs. (73)–(75), and the simulated start time of control is from t = 10 s. Figure 9
© 2004 by CRC Press LLC
Figure 9 The evolution of the Euler angles (theoretical solutions C).
Figure 10 The evolution of the Euler angles (QUEST solutions C).
shows the evolution history of the Euler angles (called Theoretical solutions C) directly numerically integrated from dynamical equations excited by the linear feedback controls using the Runge–Kutta fourthorder method. Figure 10 shows the evolution history of the Euler angles (called QUEST solutions C) from the QUEST method with consideration of the linear feedback controls. Figure 11 shows the errors’ evolution history of the Euler
© 2004 by CRC Press LLC
Figure 11 The errors’ evolution of the Euler angles.
Figure 12 Error evolution of the Euler angles when the noise level is halved; the other parameters are the same as in Case 2.
angles between Theoretical solutions C and QUEST solutions C. By comparison, we also found out that the precision of the attitude estimates was tested to be within 0.5 degrees for the attitude motions.
© 2004 by CRC Press LLC
CASE 4. Effect of Measurement Noises The amplitudes of the Euler angles’ error are sensitive to the standard squareroot deviation of the measurement noises. Figure 12 shows that when the σ is reduced to be half of the typical phase noise, i.e., 0.25 cm/λ = 0.013 wavelengths, the maximum amplitudes of the Euler angles’ error are greatly reduced by 70%. The other parameters are the same as in CASE 2.
7
CONCLUSIONS
In this paper we presented the theoretical methodologies of QUEST algorithms and numerical results for the GPSbased attitude determination of a lowearthorbit gyrostat satellite, which has multiple equilibrium positions in the Kelvin gyrostat case. The feedback control law of the attitude tumbling is proposed. By applying the linear feedback control along the bodyﬁxed frame of the gyrostat satellite, the suppression effect of the attitude tumbling of the gyrostat satellite is good enough. For the case of no external disturbance/control torques applied to the gyrostat satellite, by using the exactly theoretical solutions of the angular velocities from V. Volterra, the theoretical attitude solutions are determined from the discrete propagation equations for quaternions. In this way, the GPS theoretical phasedifference signals are simulated accurately through using GPSoft software, and the QUEST solutions for the Euler angles are computationally easier than the solutions by using the Kalman ﬁltering method. For the case of external disturbance/control torques applied to the gyrostat satellite, the fourthorder Runge–Kutta algorithms are employed to generate the theoretical, pointbypoint solutions of the angular velocities and quaternions based upon the GPSsimulated phasedifference signals numerically simulated. Using the famous QUEST method, the precision of the Euler angles can be found within about 0.5 degrees for the attitude motions under the appropriate assumption of measurement noises at the GPS phasedifferences. All of the simulation results show that the QUEST method is effective for the determination of the Euler angles of the gyrostat satellite using the GPS signals if the practical GPS phasedifference signals are measured with sufﬁcient accuracy. The nonlinear dynamics of the gyrostat satellite encountered in this chapter is also suitable for the nonlinear dynamics of the liquidﬁlled solid body model of the satellite. The work done here will be implemented for the future mission of the Singapore Xsatellite that is being developed.
REFERENCES 1. Axelrad, P. and Ward, L.M. (1996) “Spacecraft attitude estimation using the Global Positioning System: methodology and results for RADACL,” Journal of Guidance, Control, and Dynamics, Vol. 19, pp. 1201–1209. 2. Balakrishnan, A.V. (1987) Kalman Filtering Theory, Optimization Software, Inc., Publications Division, New York. 3. BarItzhack, I.Y. (1996) “REQUEST: a recursive QUEST algorithm for sequential attitude determination,” Journal of Guidance, Control, and Dynamics, Vol. 19, pp. 1034– 1038.
© 2004 by CRC Press LLC
4. BarItzhack, I.Y., Montgomery, P.Y., and Garrick, J.C. (1997) “Algorithms for attitude determination using GPS,” Proceedings of the AIAA Guidance, Navigation, and Control Conference, New Orleans, LA. 5. Cohen, C.L. (1992) Attitude Determination Using GPS, Ph.D. dissertation, Dept. of Aeronautics and Astronautics, Stanford University, Stanford, CA. 6. Cohen, C.L., Lightsey, E.G., and Parkinson, B.W. (1994) “Space ﬂight tests of attitude determination using GPS,” International Journal of Satellite Communications, Vol. 12, pp. 427–433. 7. Conway, A., Montgomery, P., Rock, S., Cannon, R., and Parkinson, B.W. (1996) “A new motionbased algorithm for GPS attitude integer resolution,” Navigation, Journal of the Institute of Navigation, Vol. 43, pp. 179–190. 8. Crassidis, J.L., Lightsey, E.G., and Markley, F.L (1999) “Efﬁcient and optimal attitude determination using recursive Global Positioning System signal operations,” Journal of Guidance, Control, and Dynamics, Vol. 22, pp. 193–201. 9. Crassidis, J.L. and Markley, F.L. (1997) “New algorithm for attitude determination using Global Positioning System signals,” Journal of Guidance, Control, and Dynamics, Vol. 20, pp. 891–896. 10. Crassidis, J.L., Markley, F.L., and Lightsey, E.G. (1999) “Global Positioning System integer ambiguity resolution without attitude knowledge,” Journal of Guidance, Control, and Dynamics, Vol. 22, pp. 212–218. 11. Fujikawa, S.J. and Zimbelman, D.F. (1995) “Spacecraft attitude determination by Kalman ﬁltering of Global Positioning System signals,” Journal of Guidance, Control, and Dynamics, Vol. 18, pp. 1365–1371. 12. GPSoft Software Manual Book, Princeton Satellite Systems, Inc., Princeton, NJ. 13. Hill, C.D. and Euler, H.J. (1996) “An optimal ambiguity resolution technique for attitude determination,” Proceedings of the IEEE, Position, Location, and Navigation Symposium, Atlanta, GA, pp. 62–269. 14. Hoots, F.R. and Roehrich, R.L. (1988) Models for Propagation of NORAD Element Sets, SpaceTrack Report No. 3, Compiled by T.S. Kelso. 15. Kuang, J.L. and Tan, S.H. (2002) “The GPSbased attitude determination of gyrostat satellites by quaternion estimation algorithms,” Acta Astronautica, Vol. 51, pp. 743– 759. 16. Kuang, J.L., Tan, S.H., and Leung, A.Y.T. (2000) “Chaotic attitude motion of satellites under small perturbation torques,” Journal of Sound and Vibration, Vol. 235, pp. 175– 200. 17. Kuang, J.L., Tan, S.H., and Arichandran, K. (2000) “Chaotic attitude motion of symmetric gyrostat via Melnikov integral,” Proceedings of the Third Nonlinear Problems in Aviation and Aerospace, Daytona Beach, FL, Edited by S. Sivasundaram (in press). 18. Kuang, J.L. and Leung, A.Y.T. (2001) “H∞ feedback for attitude control of liquidﬁlled spacecraft,” Journal of Guidance, Control, and Dynamics, Vol. 24, pp. 46–53. 19. Kuang, J.L., Tan, S.H., Arichandran, K., and Leung, A.Y.T. (2001) “Chaotic dynamics of an asymmetrical gyrostat,” International Journal of Nonlinear Mechanics, Vol. 36, pp. 35–55. 20. Lu, P. (1994) “Nonlinear prediction controllers for continuous systems,” Journal of Guidance, Control, and Dynamics, Vol. 17, pp. 553–560. 21. Mortensen, R.E. (1968) “A globally stable linear attitude regulator,” International Journal of Control, Vol. 8, pp. 297–302.
© 2004 by CRC Press LLC
22. Shuster, M.D. (1990) “Kalman ﬁlter of spacecraft attitude and QUEST model,” Journal of the Astronautical Sciences, Vol. 38, pp. 377–393. 23. Volterra, V. (1898) “Sur la theories des variations des latitudes,” Acta Math., Vol. 22, pp. 201–357. 24. Wahba, G. (1965) “A leastsquares estimates of spacecraft attitude,” SIAM Review, Vol. 7, p. 409. 25. Wertz, J.R. (1978) Spacecraft Attitude Determination and Control, D. Reidel, Dordrecht, The Nethelands, pp. 414–416. 26. Wie, B. (1998) Space Vehicle Dynamics and Control, AIAA Education Series, AIAA, Reston, VA, pp. 307–459. 27. Wittenburg, J. (1975) Beitrage zur Dynamik von Gyrostaten, Acc. Naz. Dei Lincei, Vol. 217, pp. 1–187.
© 2004 by CRC Press LLC
10 Asymptotic Controllability in Terms of Two Measures of Hybrid Dynamic Systems † ∗
1
V. Lakshmikantham† and S. Sivasundaram∗
Florida Institute of Technology, Department of Mathematical Sciences, Melbourne, FL EmbryRiddle Aeronautical University, Department of Mathematics, Daytona Beach, FL
INTRODUCTION
The relationship between asymptotic controllability to the origin in Rn of a nonlinear system, x = f (x, u), (1.1) exhibited by an openloop control u : [0, ∞) → U , and the existence of a feedback control k : Rn → U that stabilizes trajectories of the system x = f (x, k(x)),
(1.2)
with respect to the origin, have been recently studied by many authors in this ﬁeld. It is well known that continuous feedback laws may not exist even for simple asymptotically controllable systems. General results regarding the nonexistence of continuous feedback were presented in [1], which led to the search for feedback laws that are not necessarily of the form u = k(x), where k is a continuous function. It is natural to ask about the existence of discontinuous feedback laws u = k(x) (optimal control problems), which led to the search for general theorems ensuring their existence. Difﬁculty arises in deﬁning the meaning of solution x(.) of (1.2) with a discontinuous righthand side. The Filippov solution, namely the solution of a certain differential inclusion with multivalued righthand side that is built from f (x, k(x)), is one possibility. But there is no hope of obtaining general results if one insists on the use of the Filippov solution. In [2] this was generalized for arbitrary feedback k(x). The main objective is to study the relationship between the existence of stabilizing (discontinuous) feedback and asymptotic controllability of the openloop system (1.1), considering the feedback law, which, for fast enough sampling, drives all states asymptotically to the origin with small overshoot. Feedback law constructed can be robust with respect to the actual errors as well as to perturbations of the system dynamics, and also may be highly sensitive to errors in the measurement of the state vector. In [3] these drawbacks were avoided by designing a dynamic hybrid stabilizing controller which, while preserving robustness to external perturbations and actuator error, is also robust with respect to measurement error. Recently two measures have been used to unify various stability concepts
© 2004 by CRC Press LLC
and to offer a more general framework [4]. Partial stability, for example, can be discussed by using two measures. In this chapter, we attempt to utilize the concept of two measures to extend the results of [2].
2
PRELIMINARIES
We need the following notation for our discussions: (1) K = {a ∈ C[R+ , R+ ] : a(u) is strictly increasing in u and a(u) = 0}. (2) L = {a ∈ C[R+ , R+ ] : a(u) is strictly decreasing in u and limu→∞ a(u) = 0}. 2 (3) KL = {a ∈ C[R+ , R+ ] : a(t, s) ∈ K for each s and a(t, s) ∈ L for each t}. 2 , R+ ] : a(t, s) ∈ K for each t}. (4) CK = {a ∈ C[R+
(5) Γ = {h ∈ C[R+ × Rn , R+ ] : inf h(t, x) = 0}. (6) S(h, ρ) = {(t, x) ∈ R+ × Rn , h(t, x) < ρ, ρ > 0}. (7) S c (h0 , ρ) = {(t, x) ∈ R+ × Rn , h0 (t, x) ≥ ρ, ρ > 0}. We shall now deﬁne stability concepts in terms of two measures h0 , h ∈ Γ. Deﬁnition 2.1. The differential system (1.1) is said to be (h0 , h)equistable if for each > 0 and t0 ∈ R+ there exists a function δ = δ(t0 , ) > 0 that is continuous in t0 for each such that h0 (t0 , x0 ) < δ implies h(t, x(t)) < , t ≥ t0 where x(t) = x(t, t0 , x0 ) is any solution of (1.1). Deﬁnition 2.2. (h0 , h) – asymptotically stable if Deﬁnition 2.1 holds and given t0 ∈ R+ there exists a δ0 = δ0 (t0 ) > 0 such that h0 (t0 , x0 ) < δ0 implies limt→∞ h(t, x(t)) = 0. It is easy to see that the above deﬁnitions reduce to (1) the wellknown stability of the trivial solution x(t) ≡ 0 of (1.1) or equivalently, of the invariant set {0}, if h(t, x) = h0 (t, x) = x; (2) the stability of prescribed motion x0 (t) of (1.1) if h(t, x) = h0 (t, x) = x − x0 (t); (3) the partial stability of the trivial solution of (1.1) if h0 (t, x) = x and h(t, x) = xs , 1 ≤ s ≤ n; (4) the stability of the asymptotically invariant set {0}, if h(t, x) = h0 (t, x) = x + σ(t), where σ ∈ L. Deﬁnition 2.3. The system (1.1) is said to be (h0 , h)equibounded if for each α > 0, t0 ∈ R+ , there exists a positive function β = β(t0 , α) that is continuous in t0 for each α such that h0 (t0 , x0 ) ≤ α implies h(t, x(t)) < β, t ≥ t0 , where x(t) = x(t, t0 , x0 ) is any solution of (1.1). Deﬁnition 2.4. (h0 , h) is uniformly bounded if β in the above deﬁnition is independent of t0 . Deﬁnition 2.5. Let h0 , h ∈ Γ . Then, we say that
© 2004 by CRC Press LLC
(i) h0 is ﬁner than h if there exists a ρ > 0 and a function φ ∈ CK such that h0 (t, x) < ρ implies h(t, x) ≤ φ(t, h0 (t, x)). (ii) h0 is uniformly ﬁner than h if in (i) φ is independent of t. (iii) h0 is asymptotically ﬁner than h if there exists a ρ > 0 and a function φ ∈ KL such that h0 (t, x) < ρ implies h(t, x) ≤ φ(t, h0 (t, x)). Note 2.1. h0 , h and other functions deﬁned above are independent of t in further discussions. We assume that U is a locally compact metric space, and the mapping f : Rn × U → n R such that f is continuous, locally Lipschitz on x, uniformly on compact subsets of Rn × U . x denote the Euclidean norm of x ∈ Rn . x, y denotes the inner product of two vectors. A locally bounded function b : Rn → U will be called a feedback. Any ﬁnite sequence p = {ti }i≥0 consisting of numbers 0 = t0 < t1 < t2 ... with limi→∞ ti = ∞ is called a partition of [0, ∞), d(p) = supi≥0 (ti+1 − ti ) is its diameter. Consider the nonlinear system x = f (x, u),
(2.1)
where u : [0, ∞) → U and the feedback control system x = f (x, b(x)),
(2.2)
where b : Rn → U . Deﬁnition 2.6. The ptrajectory of (2.1) starting from x0 is the function x(.) obtained by recursively solving x = f (x(t), b(x(ti )),
t ∈ [ti , ti+1 ], i = 0, 1, 2...,
(2.3)
using x(ti ) as the initial value [the end point of the solution on the preceding interval, starting with x(t0 ) = x0 ] for given feedback b, a partition p and x0 ∈ Rn . Deﬁnition 2.7. The feedback b : Rn → U is said to be (h0 , h)samplestabilizing the system (2.1) for each pair 0 < r < R, if there exist M = M (R) > 0, λ = λ(r, R) > 0 and T = T (r, R) > 0 such that, for every partition p with d(p) < λ and for any initial state x0 such that h0 (x0 ) ≤ R the ptrajectory x(.) of (2.2) starting from x0 is well deﬁned and h(x(t)) ≤ r for t ≥ T ; h(x(t)) ≤ M (R) for all t ≥ 0; lim M (R) = 0.
R→0
(2.3a) (2.3b) (2.3c)
Deﬁnition 2.8. The system (2.1) is (h0 , h)asymptotically controllable if (a) for each x0 ∈ Rn , there exists some control u such that the trajectory x(t) = x(t, x0 , u) is deﬁned for all t ≥ 0 and h(x(t)) → 0 as t → ∞; (b) for each > 0 there exist δ > 0 such that for each x0 ∈ Rn with h0 (x0 ) < δ there is a control u in (a) such that h(x(t)) < for all t ≥ 0;
© 2004 by CRC Press LLC
(c) there are compact subsets X0 and U0 of Rn and U , respectively, such that if the initial state x0 ∈ X0 , then the control in (b) can be chosen with u(t) ∈ U0 for almost all t. Given a function V : Rn → R and a vector v ∈ Rn the lower directional derivative of V in the direction of v is DV (x, v) = lim inf
t→0 v →v
1 [V (x + tv ) − V (x)]. t
The function v → DV (x, v) is lower semicontinuous. For a set F ⊆ Rn , CoF denotes its convex hull. Deﬁnition 2.9. A control Lyapunov pair for the system (2.1) consists of two continuous functions V, W : Rn → R such that the following properties hold: (a) V (x) > 0 and W (x) > 0 for all x = 0, and V (0) = 0. (b) The set {x : V (x) ≤ β} is bounded for each β. (c) For each bounded subset G ⊆ Rn , there is some compact subset U0 ⊆ U such that min
v∈Cof (x,U0 )
DV (x; v) ≤ −W (x) for every x ∈ G.
(2.4)
If V is part of the control Lyapunov pair (V, W ), then V is called the control Lyapunov function. Theorem 2.1. System (2.1) is asymptotically controllable if it admits a control Lyapunov function. Remark 2.1. If the function V is smooth (2.4) can be written as min ∇V (x); f (x, u) ≤ −W (x).
u∈U0
(2.5)
Deﬁnition 2.10. A vector ξ ∈ Rn is a proximal subgradient (supergradient) of the function V : Rn → (−∞, ∞) at x if there exists some σ > 0 such that for all y in some neighborhood of x: V (y) ≥ V (x) + ξ, y − x − σ h0 (x − y)
(2.6)
(V (y) ≤ V (x) + ξ, y − x + σ h0 (x − y)). Deﬁnition 2.11. The set of proximal subgradient of V at x (may be empty) is denoted by ∂p V (x) and is called the proximal subdifferential of V at x. Remark 2.2. (2.5) can be replaced by min ξ; f (x, u) ≤ −W (x)
u∈U0
(2.7)
for every ξ ∈ ∂p V (x) where ξ and ∂p V (x) are the proximal subgradient and subdifferential, respectively, of the function V at the point x.
© 2004 by CRC Press LLC
Remark 2.3. For any ξ ∈ ∂p V (x) and any v ∈ Rn , ξ; v ≤ DV (x, v).
(2.8)
Deﬁnition 2.12. Let h0 satisfy the following conditions: h0 (u + v) ≤ h0 (u) + h0 (v) + h0x (v)u, h0 (v) ≤ h0 (u) + h0 (u − v),
(2.9) (2.10)
k1 x2 ≤ h0 (x) ≤ k2 x2 , 0
(2.11) 0
h0x (x − yα ) ≤ L(R ), where x, yα ∈ S(h, R ),
(2.12)
where we assume that h0x (x) exists and is continuous. 3
MAIN RESULTS
Deﬁne the function
1 Vα (x) = infn V (y) + 2 h0 (x − y) , α ∈ (0, 1], y∈R α
(3.1)
which is called the infconvolution of the function V. Then the function Vα is locally Lipschitz and lim Vα (x) = V (x). α→0
Let yα (x) = φ be an element of the set of minimizing points of y in (3.1). Then Vα (x) = V (yα (x)) +
1 1 h0 (x − yα (x)) ≤ V (y) + 2 h0 (x − y) α2 α
for all y ∈ Rn . Let ξα (x) = α12 h0x (x − yα ), then we have the following lemmas. Lemma 3.1. For any x ∈ Rn , ξα (x) ∈ ∂p V (yα (x)). Proof. By (3.2) we have V (y) ≥ V (yα (x)) +
1 [ h0 (x − yα (x)) − h0 (x − y)]. α2
By deﬁnition 2.8, h0 (x − y) ≤ h0 (yα − y) + h0 (x − yα ) + h0x (x − yα )(yα − y). Hence V (y) ≥ V (yα (x)) + ξα , y − yα −
1 (yα − y), α2
and by deﬁnition 2.8, ξα (x) ∈ ∂p V (yα (x)). Lemma 3.2. For any x ∈ Rn , ξα (x) is a proximal supergradient of Vα at x.
© 2004 by CRC Press LLC
(3.2)
Proof. We shall prove this by showing that Vα (y) ≤ Vα (x) + ξα (x), y − x + σ h0 (y − x) for every x, y ∈ Rn .
(3.3)
By deﬁnition, Vα (y) − Vα (x) ≤
1 [ h0 (y − yα (x)) − h0 (x − yα (x))]. α2
By deﬁnition 2.8 we have h0 (y − yα (x)) ≤ h0 (y − x) + h0 (x − yα (x)) + h0x (x − yα )(y − x). Hence Vα (y) ≤ Vα (x) + ξα (x), y − x + which completes the proof with σ =
1 α2 .
1 h0 (y − x), α2
Let us deﬁne the following notations for convenience. 1 GR = x : V (x) ≤ b(R) , 2 1 = x : V (x) ≤ b(R) , Gα α R 2 ρ(R) = max[ρ : S(h0 , ρ) ⊆ GR ], R0 = max(M (R), φ(R)), S(h, R0 ) = [x : V (x) ≤ b(R0 )], Sα (h, R0 ) = [x : Vα (x) ≤ b(R0 )]. Also, lim a(R) = lim ρ(R) = 0,
(3.4)
lim b(R) = lim ρ(R) = ∞.
(3.5)
wR (λ) = max{V (x) − V (˜ x) : h0 (x − x ˜) ≤ λ; x, x ˜ ∈ S(h0 , R0 ),
(3.6)
R→0
R→∞
Let
R→0
R→∞
0
where R = R + a(R), Lemma 3.3.
b(h(x)) ≤ V (x) ≤ a(h0 (x)).
(3.7)
h0 (x) ≤ R implies h0 (x − yα ) ≤ a(R)α2 .
(3.8)
Proof. Since Vα (x) ≤ V (x), we have Vα (x) = V (yα (x)) +
1 h0 (x − yα (x)), α2
By (3.7), this implies h0 (x − yα (x)) ≤ a(R)α2 . Hence by deﬁnition 2.8 h0 (yα (x)) ≤ a(R)α2 + R. Recall that α ∈ (0, 1], which implies that h0 (yα (x)) ≤ [a(R) + R ] = R0 . That is x ∈ S(h0 , R0 ) implies h0 (yα (x)) ≤ [a(R) + R ] = R0 , which yields yα ∈ S(h0 , R0 ).
© 2004 by CRC Press LLC
Lemma 3.4. For any x ∈ S(h0 , R), Vα (x) ≤ V (x) ≤ Vα (x) + wR (a(R)α2 ).
(3.9)
Proof. We know Vα (x) ≤ V (x), and hence Vα (x) ≥ V (yα (x)). By (3.6), wR (a(R)α2 ) = max[V (x) − V (yα (x)) : h0 (x − yα (x)) ≤ a(R)α2 , x, yα ∈ S(h0 , R0 )]. Hence V (yα (x)) ≥ V (x) − V (yα (x)), which implies that Vα (x) ≤ V (x) ≤ Vα (x) + wR (a(R)α2 ). Thus S(h0 , ρ(R)) ⊆ GR ⊆ Gα R true for all R > 0, α > 0. Lemma 3.5. For any R > 0 and α satisfying wR (a(R)α2 ) <
1 b(R0 ) 2
(3.10)
we have Gα R0 ⊆ int S(h, R0 ).
(3.11)
Proof. Let x ∈ Gα R0 . Then by deﬁnition, Vα (x) ≤
1 b(R0 ) < b(R0 ). 2
But Vα (x) = V (yα (x)) + which implies that V (yα (x)) ≤ and
1 h0 (x − yα (x)), α2 1 b(R0 ), 2
1 1 h0 (x − yα (x)) ≤ b(R0 ). α2 2
By deﬁnition 2.7 and lemma 3.3 it follows that yα (x) ∈ S(h0 , R0 ). Then V (x) ≤ Vα (yα (x)) + wR (a(R)α2 ) 1 1 ≤ b(R0 ) + b(R0 ), 2 2 which yields by deﬁnition 3.7 h(x) < R0 . Theorem 3.1. Let V be a control Lyapunov function. Then for any 0 < r < R there are α0 = α0 (r, R) and T = T (r, R) such that for any α ∈ (0, α0 ) there exists λ > 0 such that for any x0 ∈ Gα R and any partition p with d(p) < λ the trajectory x(.) of x = f (x, kν (x)) α starting at x0 must satisfy x(t) ∈ GR , ∀t ≥ 0 and x(t) ∈ S(h, r), ∀t ≥ T .
© 2004 by CRC Press LLC
Proof. By continuity of f (x, u) and the local Lipschitz property assumed, we know that there are some constants l, m such that f (x, u) − f (˜ x, u) ≤ lx − x ˜, f (x, u) ≤ m for all x, x ˜ ∈ S(h, R0 ) and all u ∈ U0 . Let 1 1 0 ∆ = min w(y) : ρ(r) ≤ h0 (y) and h(y) ≤ R . 3 2
(3.12)
(3.13)
Note that ∆ > 0. Claim 1. Let α ∈ (0, 1] satisfy a(R)α2 <
L(R0 ) lwR (a(R)α2 ) < ∆. k1
1 ρ(r), 2
(3.14)
Then, for any x ∈ S(h, R0 )\S(h0 , ρ(r)), ξα (x); f (x, kν (x)) ≤ −2∆. Proof. We have ξα (x); f (x, kν (x)) = min ξα (x); f (x, u) u∈U0
= min ξα (x); f (yα , u) + f (x, u) − f (yα , u) u∈U0
≤ min ξα (x); f (yα (x), u) + ξα (x) l x − yα (x). u∈U0
But by deﬁnition 2.7, it follows that L(R0 ) [V (x) − V (yα (x))] k1 L(R0 ) ≤ wR (a(R)α2 ), k1
ξα (x) x − yα (x) ≤
which implies that ξα (x); f (x, kν (x)) ≤ −W (yα (x)) + l
L(R0 ) wR (a(R)α2 ). k1
Since x ∈ S(h, R0 )\S(h0 , ρ(r)), we obtain 1 ρ(r) ≤ h0 (y) and h(y) ≤ R0 , 2 which implies, by the deﬁnition of ∆, W (yα ) ≥ 3∆, and ξα (x); f (x, kν (x)) ≤ −3∆ + ∆ ≤ −2∆.
© 2004 by CRC Press LLC
(3.15)
Now consider any ptrajectory of x = f (x, kν (x)) corresponding to a partition p = {ti }i≥0 with d(p) ≤ λ satisfying the inequality 1 0 2 λ ≤ ∆. (3.16) L(R )lm + k m 2 α2 Claim 2. Let a, λ satisfy (3.10), (3.14) and (3.16) and assume that for some index i, if x(ti ) ∈ Gα R \S(h0 , ρ(r)), then Vα (x(t)) − Vα (x(ti )) ≤ −∆(t − ti )
(3.17)
for all t ∈ [ti , ti+1 ]. In particular x(t) ∈ Gα R for all such t. Proof. Let xi ≡ x(ti ) and consider the largest t¯ ∈ (ti , ti+1 ) such that x(τ ) ∈ S(h, R0 ) for all τ ∈ [ti , t¯]. By (3.11), we have x(τ ) ∈ int S(h, R0 ). Choose any t ∈ [ti , t¯]. By (3.12) we have x(τ ) − xi  ≤ m(τ − ti ). (3.18) In general x(t) = xi + (t − ti )fi , where 1 fi = t − ti
(3.19)
t f (x(τ ), kν (xi )) dτ = f (xi , kν (xi )) + ηi ; ti
(3.12) and (3.18) imply that ηi  ≤ lm(τ − ti ), fi  ≤ m.
(3.20)
Also, by (3.3) we get Vα (y) ≤ Vα (x) + ξα (x), y − x +
1 h0 (y − x) α2
and Vα (x(t)) − Vα (x(ti )) = Vα (xi + (t − ti )fi ) − Vα (x(ti )) 1 ≤ (t − ti )ξα (xi ), fi + 2 h0 ((t − ti )fi ). α But we ﬁnd that ξα (xi ), fi ≤ ξα (xi ), f (xi , kν (xi )) + ξα (xi ) ηi  1 ≤ −2 ∆ + 2 h0x (x − yα ) lmλ, α which implies 1 1 Vα (x(t)) − Vα (x(ti )) ≤ (t − ti ) −2∆ + 2 h0x (x − yα ) lmλ + 2 h0 ((t − ti )fi ). α α
© 2004 by CRC Press LLC
Now by deﬁnition 2.7 and (3.16), the R.H.S. of the above inequality reduces to 1 1 0 2 ≤ (t − ti ) −2 ∆ + 2 L(R )lmλ + 2 k2 m λ α α ≤ −∆(t − ti ), which implies that x(t¯) ∈ S(h, R0 ), which contradicts the maximality of t¯ unless t¯ = ti+1 . Therefore it holds for all t ∈ [ti , ti+1 ]. Claim 3. Assume that α satisﬁes (3.10), (3.14), and choose any λ such that (3.16) is valid. Then for any ptrajectory x(.) with d(p) ≤ λ and every x(0) ∈ Gα R, tN ≤ T =
b(R) . 2∆
(3.21)
Proof. From claim 1 and the minimality of N , we have x(ti ) ∈ Gα R \ Gr . Using the fact that S(h0 , ρ(r)) ⊆ Gr and 0 ≤ Vα (x(tN )) ≤ Vα (x(0)) − ∆tN 1 ≤ b(R) − ∆tN 2 yields tN ≤ T =
b(R) . 2∆
Claim 4. Assume that α and λ satisfy all the previous conditions and in addition the following two conditions are satisﬁed: 1 b(r) 4 1 wR (k2 m2 λ) < b(r). 4
wR (a(R)α2 ) <
(3.22) (3.23)
Then x(t) ∈ S(h, r) for all t ≥ tN . Proof. If x(ti ) ∈ Gr , then by (3.12) and (3.23), we get Vα (x(t)) ≤ V (x(t)) ≤ V (x(ti )) + wR (k2 m2 λ) < for all t ∈ [ti , ti+1 ]. Then Vα (x(ti+1 )) <
3 b(r) 4
3 b(r). 4
/ Gr by claim 2, we have If x(ti+1 ) ∈ Vα (x(t)) <
© 2004 by CRC Press LLC
3 b(r) for all t ∈ [ti , tj ], 4
(3.24)
where j is the least integer such that ti > tj and x(tj ) ∈ Gr (if there is any). Starting from tj , we may repeat the argument and show that (3.24) holds for all t ≥ 0. The relation (3.22) and Lemma 3.4 imply that V (x(t)) ≤ Vα (x(t))wR (a(R)α2 ) < b(r), which shows x(t)) ∈ S(h, r) for all t ≥ tN . Now ﬁnally, let α ¯ be the supremum of the set of all α > 0 that satisfy the conditions (3.14) and (3.22). Then for any α ∈ (0, α ¯ ), one can choose λ satisfying (3.16) and (3.23), so that for each partition with d(p) ≤ λ and every ptrajectory starting from a state in Gα R, satisﬁes x(t) ∈ Gα , it follows that for all t ≥ 0 and x(t) ∈ S(h, r) for all t ≥ T . R This completes the proof of the theorem. Given an asymptotically controllable system and a control Lyapunov function V for it, we pick an arbitrary R0 > 0. There is then a sequence {Rj }∞ j =−∞ satisfying 2Rj ≤ ρ(Rj +1 ), j = 0, ±1, ±2, ... limR→∞ b(R) = limR→∞ ρ(R) = ∞ implies having R0 ﬁxed one can ﬁnd R1 , R2 in a way such that R0 < 2R0 < ρ(R1 ) < R1 < 2R1 < ρ(R2 ) < . . . . Let rj = 12 ρ(Rj−1 ). Pick R−1 < 12 ρ(R0 ) and R−2 < 2R−2 < ρ(R−1 ) < R−1 < 2R−1 < ρ(R0 ) < R0 < 2R0 . . ., which implies that limR→−∞ Rj = 0, limR→∞ Rj = ∞. Consider any integer j, for the pair (rj , Rj ), then Theorem 3.1 implies that there exist α αj > 0, δj > 0 and Tj > 0 a map kj : S(h, Rj ) → Uj , kj ≡ k(αj , rj , Rj ) such that GRjj is invariant with respect to all ptrajectories when kν = kj and d(p) ≤ λj
(3.25)
and for each such trajectory it holds that h(t, x(t)) ≤ rj for all t ≥ Tj . Recall that in the construction of kj , we used the fact that there is some compact subset U0 ⊆ U, Uj satisfying condition (3.25) for Lyapunov fair for all x ∈ S(h, R0 ). Since Rj , an increasing sequence, and if the min in (3.25) also holds if we enlarge U0 , we can assume Uj = U0 for all j < 0 and Uj ⊆ Uj+1 for all j ≥ 0. For x ∈ S(h, Rj ) pick a bound f (x, u) ≤ mj for all u ∈ Uj and note that for all j, mj ≤ mj+1 , because of the monotonicity of Uj and S(h, Rj ). α α Finally let Aj ≡ GRj+1 \GRjj . j+1 Lemma 3.6. For i = j, Ai ∩ Aj = φ, and Rn \{0} = ∪∞ j=−∞ Aj . α
α
for all j, which implies that Proof. We know that GRjj ⊆ S(h, Rj ) ⊆ S(h, 2Rj ) ⊆ GRj+1 j+1 αj Aj ⊆ GRj and Ai ∩ Aj = φ whenever i < j. For any two integers M < N , we get {x : RM ≤ h(t, x) ≤ RN } ⊆
N j=M
© 2004 by CRC Press LLC
α
M Aj = GRNN+1 \Gα RM . +1
α
Also we ﬁnd that S(h, rj ) ⊆ IntGRj−1 by the deﬁnition of rj . j−1 Since Aj ∪ {0} constitute a partition of the state space, deﬁne a map k : R → U by k(x) = kj (x) for all x ∈ Aj−1 for each j, k(0) = u0 where u0 is any ﬁxed element of U0 . Now to show the feedback is sample stabilizing, we need the following. α
Claim 5. The set GRjj is invariant with respect to ptrajectories of (2.2) when (3.25) is used and d(p) = min(δj ,
Rj−1 mj ).
α
Proof. Consider any ptrajectory starting at a state in GRjj , where p satisﬁes (3.25). α α It is enough to show that x(t) ∈ GRjj for all t ∈ [ti , ti+1 ] if x(t) ∈ GRjj . For any such i, we have (a) x(ti ) ∈ Aj−1 , (b) x(ti ) ∈ Al−1 for some l ≤ j − 1, (c) x(ti ) = 0. α
(a) Since k(x) = kj (x) is known to have GRjj invariant, there is nothing to prove. l (b) p – may not be ﬁne enough to guarantee Gα Rl invariant under the feedback kl (trajectory cannot go too far because sampling time is small). Since x(ti ) ∈ Al−1 ⊂ l ¯ ¯ Gα Rl ⊂ S(h, Rl ), h(x(ti )) < Rl , pick t ∈ [ti , ti+1 ] so that h(x(t)) < Rj for all t ∈ [ti , t]. The choice of mj implies that k(x(ti )) ∈ Ul ⊆ Uj , h(x(t)) − h(x(ti )) ≤ mj (t − ti ) for all t ∈ [ti , t¯] which implies that h(x(t)) ≤ mj (t − ti ) + Rl ≤ Rj−1 + Rl ≤ 2Rj−1 for all t ∈ [ti , t¯].
(3.26)
Claim: h(x(t)) < Rj . If not, there is some t¯ ∈ [ti , ti+1 ] such that h(x(t)) < Rj for t ∈ [ti , t¯] and h(x(t¯)) = Rj . Since h(x(t)) < Rj for all t ∈ [ti , t¯], (3.26) yields Rj ≤ 2 Rj−1 , which is a contradiction. This implies that the argument holds with t = ti+1 , which shows the α trajectory stays in S(h, 2Rj−1 ), that is included in GRjj . (c) Since k(0) ∈ U0 ⊆ Uj , the argument in (b) is repeated with 0 instead of Rl , which α proves that the trajectory stays in S(h, Rj−1 ), so it is in GRjj . To show feedback k is sample stabilizing, we consider the following. K Choose an integer K = K(r), and least integer N = N (R), such that Gα RK ⊆ αN S(h, r) and S(h, R) ⊆ GRN . Deﬁne δ = min{ min δj , K≤j≤N
RK−1 }, mN
T =
N
Tj ,
M = RN (R) .
j=K+1
Claim 6. For any 0 < r < R, there exist δ = δ(r, R) > 0, T = T (r, R) > 0, and M = M (R), d(p) < λ, h(x(0)) < R, such that h(x(t)) ≤ r, ∀t ≥ T,
© 2004 by CRC Press LLC
and h(x(t)) ≤ M, ∀t ≥ 0. Proof. Consider the ptrajectory starting from x0 ∈ S(h, R). K K Claim: For some ti ∈ p such that ti ≤ T , x(ti ) ∈ Gα / Gα RK . If x0 ∈ RK , then α
x0 ∈ AN , for K ≤ N ≤ N − 1. Claim 5 implies that GRNN is invariant, which yields that α in some (ti−1 , ti ] for some µ < N . the trajectory remaining in HN until it enters GRµ−1 µ−1 Note ti ≤ TN . Theorem 3.1 implies that x(t) remains in HN until moment TN then x(TN ) ∈ S(h, rN ), which is a contradiction to x(t) remaining in HN . If µ ≤ K + 1 there is no more to be done. K Otherwise repeat the argument to get x(ti ) ∈ Gα RK for ti ≤ T . Claim 5 implies that x(t) stays in this set for all t ≥ ti by (3.25), which proves that N x(t) remains in Gα RN ⊆ S(h, RN ) for all t ≥ 0, hence h(x(t)) ≤ M (R) for all t. This completes the proof of the claim. Deﬁnition 2.7 implies that feedback has attractiveness and bounded overshoot. Finally we have to show that k has Lyapunov stability property. To prove this, M (R) must satisfy lim M (R) = 0. R→0
α
and therefore ρ(RN −1 ) < Claim 7. For N = N (R), S(h, R) is not contained in GRNN−1 −1 R. Proof. If not ρ(RN −1 ) ≥ R, which implies, RN −1 ≥ R, since ρ(r) < r for all r. By deﬁnition of ρ(t), S(h, R) ⊆ GR , and consequently, S(h, R) ⊆ GR ⊆ GR−1 ⊆ α α GRNN−1 , since R ≤ RN −1 , which in contradiction to S(h, R) is not contained in GRNN−1 , −1 −1 and this yields ρ(RN −1 ) < R. Thus it follows that limR→0 N (R) = −∞, and limj→−∞ Rj = 0, which yields limR→0 M (R) = 0. The proof is complete.
Robust Dynamic Hybrid Controller The aim is to design a dynamic hybrid stabilizing controller which, while preserving robustness to external perturbations and actual error, is also robust with respect to measurement error. In the presence of measurement errors e(t) only values x ˜(t) of the measured estimate of state vector x(t), x ˜(t) ≡ x(t) + e(t) can be used for control. Deﬁne a tracking controller kt : Rn × Rn → U as follows: z − x ˜, f (˜ x, kt (˜ x, z)) = max u ∈ U z − x ˜, f (˜ x, u) . Important features:
© 2004 by CRC Press LLC
x = f (x, kt (˜ x, z)) + d1 (t)
(3.27)
(3.28)
can track any trajectory of the control system z = f (z, u(t)) + d2 (t),
(3.29)
for any control u(t) if the measurement errors e(t), disturbances d1 (t), d2 (t), and the diameter of the sampling partition px of (3.28) are small enough. Reinitialization: There are monotone sequences of positive numbers {Rj }+∞ −∞ and closed sets {Gj }+∞ −∞ : lim Rj = 0, lim Rj = +∞, (3.30) j→−∞
j→∞
Gj ⊂ BRj ⊂ B2Rj ⊂ Gj+1 , and the set Gj+1 is invariant with respect to pz trajectories of the internal model z(t) ∈ Gj+1 for all t ≥ 0, (3.31) if d(pz ) ≤ λj and d2 (t) ≤ θj for some positive numbers λj , θj . Moreover, there exist moments Tj > 0, such that, for any pz trajectory as above with z(0) ∈ Gj+1 , we have the relation z(t ) ∈ Gj−1 , z(t) > Rj−2 for all t ∈ [0, t ]
(3.32)
at some moment t ≤ Tj . A dynamic hybrid controller consists of: (i) a tracking controller kt that drives the system (3.28); (ii) an internal model of the system (2.1) driven by a samplestabilizing feedback z = f (z, k(z)) + d2 (t);
(3.33)
(iii) a set of reinitialization rules that deﬁne the moments tk from the sampling partition px for (3.33) for reinitialization of internal model (3.33), z(tk ) = x ˜(tk );
(3.34)
(iv) a sampling rule to choose sampling moments ti ∈ px , and si ∈ pz : ti+1 − ti ≤ λx (˜ x(ti )),
(3.35)
si+1 − si ≤ λz (z(si )).
(3.36)
We have Rn \{0} = ∪+∞ j=−∞ Aj , for the sets Aj = Gj+1 \Gj . Now deﬁne reinitialization rules for determining sequential moments tk ∈ pz , k = 0, 1, ... of reinitialization (3.34) of the internal model (3.33). Let t0 = 0, and assume tk has already been determined. Choose an index jk that satisﬁes x ˜(tk ) ∈ Ajk . Then the next moment tk+1 of the reinitialization is deﬁned as the ﬁrst moment t > tk from pz such that one of the following two events occur:
x ˜(t ) ∈ Ajk+1 for some jk+1 ≤ jk − 1,
/ Gjk+2 . x ˜(t ) ∈
© 2004 by CRC Press LLC
(3.37) (3.38)
Lemma 3.7 (Tracking Lemma [3]). Let us assume that function f : Rn × U → Rn satisﬁes the following growth and Lipschitz conditions: f (x, u) ≤ m, on Rn × U,
(3.39)
f (x1 , u) − f (x2 , u) ≤ lx1 − x2  for all x1 , x2 ∈ Rn , u ∈ U.
(3.40)
For any positive T , β, any measured estimate x (t) and disturbances d1 (t), d2 (t) satisfying x (t) − x(t) ≤ β for any t ≥ 0,
(3.41)
d1 (t) ≤ β, d2 (t) ≤ β for any t ≥ 0,
(3.42)
and any trajectory z(t) of (3.33) deﬁned on [0, T ], an arbitrary px trajectory of (3.27) with d(px ) ≤ β 2 ,
(3.43)
x(0) − z(0) ≤ β
(3.44)
x(t) − z(t) ≤ α(β) for all t ∈ [0, T ],
(3.45)
and initial conditions is deﬁned on [0, T ] and satisﬁes
where
1
α(β) = e2lT (1 + l(1 + mβ)2 + 4m) 2 β.
(3.46)
Theorem 3.2. Let the system (2.1) be (h0 , h)asymptotically controllable, and k be sample – (h0 , h) – stabilizing feedback. Then there exist continuous functions ψ : R n → R+ , λ x : R n → R + , λ z : R n → R+ , as well as reinitialization rules (3.37), (3.38), such that for any measured estimate x ˜, of state vector x, and any disturbances d1 (t), d2 (t) satisfying h(˜ x(t) − x(t)) ≤ ψ(x(t)) for any t ≥ 0, d1 (t) ≤ ψ(x(t)),
d2 (t) ≤ ψ(x(t)), for any t ≥ 0.
(3.47) (3.48)
The px trajectories x(t) of (3.28) are (h0 , h)stable. Proof. Let mj and lj denote, respectively, the upper bound for f (x, u) and Lipschitz constant of f on the set BRj+3 × U for every integer j. Deﬁne βj = max{β > 0 : αj (β) + 2β ≤ Rj−2 − Rj−3 }, where the function αj is given by (3.46) for T = Tj , m = mj , l = lj . Let ψ(x) = min{βi : j − 3 ≤ i ≤ j + 3} for Rj < x ≤ Rj+1 , λ(x) = min{βi2 : j − 3 ≤ i ≤ j + 3} for Rj < x ≤ Rj+1 , 1 n ψ(x) = min ψ(y) + y − x : y ∈ R , 2 1 n λx (x) = min λ(y) + y − x : y ∈ R . 2
© 2004 by CRC Press LLC
It is easy to see that functions ψ and λx are positive for x = 0 with the Lipschitz constant 1 2 , and (3.49) ψ(x) ≤ βi , λx (x) ≤ βi2 , for any j − 3 ≤ i ≤ j + 3 and Rj < x ≤ Rj+1 . We assume that without loss of generality ψ ≤ θ. By applying the Lipschitz condition for the function ψ, and for any x, that satisfy x − x  ≤ ψ(x), we get x − x  ≤ 2ψ(x ). (3.50) Let us consider two sequential moments tk , tk+1 , of the reinitialization of the internal model (3.33). Let t∗ = min{tk+1 , tk + Tjk } and let t the maximal t ∈ [tk , t∗ ] such that Rjk−3 < x(t) ≤ Rjk+3 , Rjk−3 < x (t) ≤ 2Rjk+2 ,
(3.51) (3.52)
for all t ∈ [tk , t ]. Since x(tk ) − x (tk ) ≤ 2ψ(x (tk )) ≤ 2βjk , (3.47), (3.49), and (3.50) imply that (3.51) and (3.52) hold for t = tk and t > tk . Then by (3.49) we get ψ(x(t)) ≤ βjk ,
2 λx (x (tk )) ≤ βjk for all t ∈ [tk , t].
Because of (3.47), by applying the Lemma (tracking lemma), we get x(t) − z(t) ≤ αj (βj ) true for all t ∈ [tk , t].
(3.53)
Then by using (3.31), (3.32) and by the deﬁnition of βj , we show that (3.51), (3.52) hold for t = t, which implies that both these relations are true in the entire interval [tk , t∗ ]. To show that tk+1 < tk + Tjk , where the moment tk+1 is deﬁned by the relation (3.37), let us assume that this is not true. Then (3.53) holds and x (t) ∈ / Gjk for all t ∈ [tk , tk + Tjk ]. Recall that in this case, for the trajectory z(t) of the internal model, there exists t ∈ (tk , tk + Tjk ], such that the ﬁrst inclusion in (3.32) holds. This yields that x (tk ) ≤ z(t ) + αj (βj ) + βj < Rjk−1 + Rjk−2 − Rjk−3 < 2Rjk−1 . Because of (3.30) we obtain that x (t ) ∈ Gjk . This contradiction proves that the relation (3.37) determines the reinitialization moment tk+1 and tk+1 − tk < Tjk ,
(3.54)
Hence we have proved the following lemma. Lemma 3.8. For any k = 0, 1, ... (3.50) holds and x (t) ∈ B2Rjk+2 ⊂ Gjk+3 for all t ∈ [tk , tk+1 ], x (tk+1 ) ∈ Hjk+1 ⊂ Gjk .
© 2004 by CRC Press LLC
(3.55) (3.56)
It is obvious that by using Lemma 3.8, we can prove that the dynamic hybrid controller provides the uniform convergence of x (t) to the origin in Rn . Then it follows from (3.43) and (3.46) that x(t) uniformly converges to the origin as well. To prove this, let us consider any 0 < r < R and initial point x0 ∈ BR for the px trajectory of (3.27) with x(0) = x0 . Then we have from (3.43) and deﬁnition βj that there exists an integer N = N (R) which does not depend upon x0 such that x (0) ∈ GN , and limR→0 N (R) = −∞. Now deﬁne M (R) = 4RN (R)+2 , and note that it satisﬁes (2.3c). In accordance with the Lemma 3.8, we have x (t) ≤ 2RN (R)+2 , which implies (2.3b) because of (3.46). Deﬁne the maximal K(r), such that 4RN (R)+2 < r. By applying Lemma 3.8, we N (R) deduce that there is a moment t such that t ≤ T = i=K(r) Ti , and x (t ) ∈ GK(r) , which yields due to Lemma 3.8 that x(t) ≤ 4RK(r)+2 for all t ≤ T . Then x(t) satisﬁes (2.3a). This completes the proof of the theorem.
REFERENCES 1. R.W. Brockett, “Asymptotic stability and feed back stabilization,” in Differential Geometric Control Theory (R.W. Brockett, R.S. Millman, and H.J. Sussmann, eds.), Birkhauser, Boston, 1983, pp. 181–191. 2. F. H. Clarke, Yu. S. Ledyaev, E.D. Sontag, and A.I. Subbotin, “Asymptotic controllability implies feedback stabilization,” IEEE Transactions on Automatic Copntrol, Vol. XX, 1999, pp 1–13. 3. N.N. Krasovskii and A.I. Sabotin, Positive Differential Games, Nauka, Moscow, 1974 [in Russian]. Revised English translation: GameTheoretical Control Problems, SpringerVerlag, New York, 1988. 4. V. Lakshmikantham and X. Liu, Stability Analysis in Terms of Two Measures, World Scientiﬁc, Singapore, 1993.
© 2004 by CRC Press LLC
11 Exact Boundary Controllability of a Hybrid PDE System Arising in Structural Acoustic Modeling George Avalos† and Irena Lasiecka∗ †
Department of Mathematics and Statistics, University of NebraskaLincoln, Lincoln, NE ∗ Department of Mathematics, University of Virginia, Charlottesville, VA
In this chapter, we consider the active control of a pair of wave equations, each deﬁned on different geometries: one wave equation holds on the interior of a bounded domain Ω; the other wave equation is satisiﬁed on a portion of the boundary Γ0 of ∂Ω. The respective wave equations are coupled by trace terms on the boundary interface Γ0 . For this coupled system of equations, we present results of exact controllability in the case that the controllers are exerted strictly on the boundary ∂Ω. In particular, we give precise geometric conditions under which control on the “active portion” Γ0 only gives exact controllability for the dynamics of both wave equations, the interior as well as the boundary wave.
1 1.1
STATEMENT OF THE PROBLEM AND MAIN RESULTS Description of the Problem
We shall consider here a hybrid partial differential equation (PDE) system that models acoustic pressure in a cavity — taken to be an ndimensional entity; this cavity in turn is interacting with another environment (manifold), taken to be a wall of dimension n − 1. The coupling between the structures is accomplished through the back pressure exerted on the wall. These kinds of systems are typical in structural acoustic modeling (see, e.g., [8], [24]). In this chapter, we shall consider a canonical model, in which the hybrid consists of two disparate wave equations that are coupled at the interface. This interface consists of the boundary of the region (cavity). We are thus dealing here with interactions of two hyperbolic PDEs, but other conﬁgurations have also been considered in the literature; e.g., a coupling of a parabolic PDE to a hyperbolic PDE, where the hyperbolic component is a wave equation satisﬁed within the cavity, but with the wall (structure) satisﬁed by a beam (rather than a wave) equation under the inﬂuence of various degrees of structural damping (see [1]). Our goal here is to study exact controllability properties, with a (natural) emphasis being on boundary control with respect to initial and terminal data in physically meaningful ﬁnite energy states. (However, there is generally no guarantee that during the transit time
© 2004 by CRC Press LLC
from the initial to the terminal data, the corresponding state will remain in the space of ﬁnite energy, that is the underlying control operator for our control operator is not “admissible.” In part, this is a consequence of the fact that the Lopatinski conditions are not satisﬁed for the wave equation with Neumann boundary conditions (see [15]). In order to clearly convey our main ideas and the technical difﬁculties associated with the exact controllability of PDEs that govern interactive structures, we have chosen to analyze here a relatively simplelooking model; namely, a wave/wave model that displays a variety of the phenonema intrinsic to coupled PDE structures in general. Moreover, this canonical PDE model provides a comfortable setting from which to study the delicate issue of “propagating” controllability properties from one PDE component to the other. As we shall see below, the geometry of the region plays a critical role in our analysis; and it will turn out that some geometries can be particularly exploited for the sake of controllability. Indeed, our results lead to the conclusion that there are favorable geometric conﬁgurations that should be kept in mind when one wishes to design acoustic chambers (such as the cabin of a helicopter) that are amenable to active controllers distributed on a portion of the chamber walls. Other, higherorder equations (of the fourth order on the interface), can and have been considered by the authors as well. See [1], [2], and [5] for an account of these results.
1.2
The PDE Model
Let Ω be a bounded open subset of Rn , n ≥ 2, with Lipschitz boundary ∂Ω = Γ0 ∪ Γ1 , with each Γi nonempty, and Γ0 ∩ Γ1 = ∅. There will eventually be additional assumptions imposed on the Γi (see (A1) and (A2) below). For this geometry in place, we shall study the solutions [z(t, x), v(t, τ )] to the following (controlled) PDE model:
ztt (t, x) = ∆z(t, x) on (0, T ) × Ω ∂ z(t, x) = vt (t, x) on (0, T ) × Γ0 ∂ν ∂ z(t, x) = u1 on (0, T ) × Γ1 ∂ν
[z(0, x), zt (0, x)] = z0 on Ω. This PDE is coupled to the following on the interface Γ0 : ∂2 vtt (t, τ ) = ∂τ on (0, T ) × Γ0 2 v(t, τ ) + u0 (t, τ ) − zt Γ 0 ∂v = 0 on (0, T ) × ∂Γ0 ∂n [v(0, x), vt (0, x)] = v0 on ∂Γ0 .
(1)
∂ ∂ denotes the (unit) tangential, and ∂ν the outward normal derivative with As usual, ∂τ ∂ respect to Γ. ∂n is here the unit exterior normal derivative with respect to the n − 1 manifold ∂Γ0 .
© 2004 by CRC Press LLC
To account for the initial data of the problem, we deﬁne the spaces, H1 ≡ H 1 (Ω) × L2 (Ω); H0 ≡H 1 (Γ0 ) × L2 (Γ0 ); H ≡ [z0 , z1 ] ∈ H1 and [v0 , v1 ] ∈ H0 such that Ω z1 = Γ0 v0 and Ω z0 = Γ0 v1 .
(2)
It can be readily shown that H is a Hilbert space, with the inner product z0 z˜0 z1 z˜1 v0 , v˜0 v1 v˜1 H ∂v0 ∂˜ v0 = ∇z0 · ∇˜ z0 dΩ + z1 z˜1 dΩ + + v1 v˜1 dΓ0 . ∂τ ∂τ Ω
Ω
Γ0
Γ0
With this deﬁnition of the inner product, one can then proceed to use the Lumer–Phillips Theorem to show the existence of a C0 group eAt t≥0 associated with this coupled system of wave equations. Accordingly, with regard to the PDE (1), a straightforward consequence of these dynamics eAt t≥0 (see, e.g., [1], [25]) yields the continuity of the mapping {[z0 , v0 ] ∈ H, u1 = 0, u0 = 0} ⇒ [z, v ] ∈ C([0, T ]; H). In short, the uncontrolled problem (i.e., the PDE (1) with ui = 0) is well posed in the basic space of energy H. Thus, the wellposedness of the system is not an issue here; the goal of our chapter is rather to determine exact controllability properties1 of (1) with boundary controls u1 , u0 taken in prescribed spaces of controls U1 , U0 , say, under the imposition of speciﬁc conditions upon the geometry Ω.
1.3
The Main Results
Our ﬁrst result in Theorem 1 below states that all ﬁnite energy states are controlled exactly with controls located on Γ0 alone (with these controls acting only on the vcomponent). This result does require, however, that the geometry be “appropriate” to the situation; viz., the domain Ω is convex and the “roof” of the acoustic chamber is not too “deep.” (See Assumption (A1) and Fig. 1.) In addition, the control u0 must be of the appropriate topological strength; i.e., u0 ∈ [H 1 (0, T ; L2 (Γ0 ))] . So for our ﬁrst result (Theorem 1(a) below), we assume the following: Assumption (A1) Assume that Ω is a bounded subset of Rn , with boundary Γ = Γ0 ∪ Γ1 , Γ0 ∩ Γ1 = ∅, with Γ0 being ﬂat. Moreover assume the following: 1 The classical deﬁnition of exact controllability is intended here. Namely, the PDE (1) is exactly controllable if there is a T ∗ > 0 such that for terminal time T > T ∗ , one has the following reachability property: for all initial zT , vT ] ∈ H, there exist control functions [u1 , u0 ] ∈ U1 × U0 data [ z0 , v0 ] ∈ H and preassigned target data [ (to be speciﬁed), such that at terminal time T the corresponding solution [z, v] to (1) satisﬁes [ z (T ), v (T )] = [ zT , vT ].
© 2004 by CRC Press LLC
Figure 1 A triple {Ω, Γ0 , Γ1 } that satisﬁes Assumption (A1).
(i) Ω is convex; (ii) there exists a point x0 ∈ R2 such that (x − x0 ) · ν ≤ 0 for all x ∈ Γ1 . The special vector ﬁeld that is available, in case that Assumption (A1) holds true — constructed in [19] and denoted below as h in (13) — will be used in the derivation of the observability inequality associated with exact controllability (see (3) below). In particular, this special h appears in the wave multipliers classically used to estimate the energy of the zwave equation (see, e.g., [13], [14], [23], [27], [28]). The behavior of h on the inactive portion of the boundary i.e., h · νΓ1 = 0, is a key driver in our ﬁrst result. With control on Γ0 only, and under Assumption (A1), the PDE (1) is exactly controllable on H. A discussion concerning the possible conﬁguration of those triples {Ω, Γ0 , Γ1 } that satisfy Assumption (A1) is given in Appendix C in [19]. A canonical example of such a triple is given in Fig. 1. On the other hand, if the geometry of the acoustic chamber is unrestricted (see Assumption (A2)), then from our previous discussion, we clearly must have additional control on Γ1 . In this case, the second part of Theorem 1 states that all ﬁnite energy states 1 are controlled exactly with controls u0 ∈ L2 (0, T ; H − 4 (Γ0 )) (located on Γ0 ) and u1 ∈ 2 2 L (0, T ; L (Γ1 )) (located on the roof of a chamber of arbitrary geometrical conﬁguration). So in our second result (Theorem 1(b) below), we make the following assumption: Assumption (A2) Assume that Ω is either a bounded subset of Rn with smooth boundary Γ, or else Ω is a parallelepiped. Moreover, assume boundary Γ = Γ0 ∪ Γ1 , where Γ0 is ﬂat. No assumptions are made on Γ1 (see Fig. 2). If Assumption (A2) holds true, then one has exact controllability of (1) for arbitrary initial data of ﬁnite energy, with the control region taken to be Γ0 ∪Γ1 . The point of making the Assumption (A2) is that in this case, one can take a radial vector ﬁeld h to assist in the multiplier method to be employed to estimate the (acoustic) wave energy. However, since Assumption (A2) is much less restrictive than (A1) — in particular, no impositions are made on the hard walls Γ1 — the corresponding h cannot be expected to help with the highorder terms on Γ1 , and hence the need for control on the hard walls. A common feature in both of our main results is the critical use of the “sharp” regularity theory, which has been developed to handle the tangential derivatives (on the boundary) of solutions to wave equations (see [17] and Lemma 4) below). With these assumptions, we now state our main results concisely.
© 2004 by CRC Press LLC
Figure 2 A triple {Ω, Γ0 , Γ1 } that satisﬁes Assumption (A2).
Theorem 1 (a) Let Assumption (A1) stand, and set U1 U0
= {0} ; = H 1 0, T ; L2 (Γ0 ) .
Then for long enough terminal time T , the problem (1) is exactly controllable on H within the class of U1 × U0 controls. (b) Let Assumption (A2) stand, and set U1
= L2 (0, T ; L2 (Γ1 ))
U0
= L2 (0, T ; H − 4 (Γ0 )).
1
Then likewise for long enough terminal time T , the problem (1) is exactly controllable on H within the class of U1 × U0 controls. Remark 2 The precise speciﬁcation of the controllability time is T > 2diam(Ω). This is on account of the terminal time needed for Holmgren’s Uniqueness Theorem to hold true. This theorem is needed at the level of a “compactness/uniqueness” argument so as to eliminate lowerorder terms that corrupt the associated observability estimate (see (35) and (37)). Remark 3 The speciﬁcation here of natural boundary conditions for the vcomponent in (1) is not critical in the derivation of our observability results. In fact, one could obtain a similar exact controllability result for (1), with instead vΓ0 = 0 on ∂Γ0 .
1.4
About the Problem and the Literature
The PDE system (1) and other classes of PDE models seen in the literature and that are associated with the mathematical descriptions of stuctural acoustic ﬂow, are principally characterized as a coupling of distinct dynamics (be it of hyperbolic/hyperbolic or hyperbolic/parabolic type), with the coupling being accomplished across boundary interfaces (see [1], [2], [6], [8], [9], [20], [24]). In the present case, we have a chamber Ω, with hard, rigid walls Γ1 and its ﬂexible wall portion Γ0 . Moreover, the ﬂow within this chamber Ω is assumed to be of the acoustic wave type. Given this assumption of acoustictype ﬂow,
© 2004 by CRC Press LLC
an interior wave equation (in z) consequently appears in the governing of the PDE system, this wave equation being under the inﬂuence of Neumann forcing data vt . In turn, the acoustic wave is coupled to the boundary wave equation (in v) on Γ0 (this boundary wave is to account for the dynamics involved in the ﬂexible Γ0 ). The coupling mechanism between the two dynamics is brought about through the boundary traces of the respective wave velocities. In accordance with the classical notion of exact controllability for PDEs, we shall concern ourselves here with deriving reachability properties with respect to initial and terminal data in the ﬁnite energy space H, as deﬁned in (2). Moreover in line with the applied problem addressed in the literature, a principal objective is the attainment of some statement that assures the exact controllability of the PDE system, in the case that control is exerted on the active wall Γ0 only (that is, Γ1 is to be inactive). However, this objective, which is really an accommodation to the needs of the given physical problem, is at odds with the geometric conditions necessary for exact controllability, and which are prescribed in [7]. In fact, in the structural acoustic control problem, as stated in [6] and [1], the active control region Γ0 comprises one side of the chamber wall only. In consequence, for an arbitrary triple {Ω, Γ0 , Γ1 }, one will generally not have exact controllability with control implemented solely on Γ0 , as the necessary conditions of geometric optics will not be satisﬁed. In particular, it is now known that in order to control the zwave equation from the boundary, it is necessary that the support of the control region be sufﬁciently large (see [7]). Thus prescribing control only on Γ0 will generally be insufﬁcient, unless Γ0 is large relative to Ω. Another (topological) complication presented here is a direct consequence of the coupling involved between the two dynamics, i.e., the acoustic interaction (1) presents a situation where the component z on Ω, as the solution of the wave equation with forcing boundary data vt , is subject to the “smoothing effects” due to the (L2 ) regularity of the Neumann boundary data vt (see, e.g., [16], [18]). As such, it will have a certain measure of regularity. Therefore, a sole control u0 — that is u1 = 0 in (1) — acting strictly on the wave component v , in the absence of geometric assumptions, will generally not be strong enough to drive the acoustic variable z to an arbitrary state of ﬁnite energy. To the best of our knowledge, Theorem 1(a) and (b) constitutes the ﬁrst exact controllability results for structural acoustic interactions in ﬁnite energy spaces, and with general spatial domains Ω. All other results (see [22] and references therein) pertain to controllability on speciﬁed subspaces of ﬁnite energy — such as those described by the asymptotic behavior of Fourier coefﬁcients — and moreover these results are proved for very special geometries only — a 2D rectangle — with very large classes of controls: H −2 (0, T ; L2 (Γ0 )). From the mathematical point of view, the key ingredients in our proofs are the following: (i) Sharp trace regularity for the wave equation in the absence of Lopatinski conditions (a distinguishing feature of the Neumann case); see Lemma 5. (ii) Microlocal analytical estimates that allow the absorption of tangential (wave) traces by time derivatives on the boundary; see Lemma 4. (iii) A recent result in [19] concerning Carleman’s estimates for the wave equation with the controlled Neumann part of the boundary. These estimates lead to the aforementioned special vector ﬁeld h that allows us, in this chapter, to handle the uncontrolled portion of the boundary so as to derive the requisite observability estimates.
© 2004 by CRC Press LLC
2
THE NECESSARY INEQUALITY
With the control spaces U1 , U0 as prescribed in Theorem 1(a) or (b), let LT : U1 × U0 ⊃ D(LT ) → H denote the control to terminal state map, u0 z(T ) = LT . u1 v (T ) Then for the asserted controllability result Theorem 1(a) (resp. (b)) is equivalent to showing the surjectivity of LT as a mapping between the said spaces. In turn, by the classical functional analysis (see, e.g., Lemma 3.8.18 and Theorem 6.5.10 of [12]), this ontoness is 0, ψ 0 ∈ D(L∗ ) ⊂ H (where equivalent to establishing the following inequality for all φ T ∗ LT denotes the Hilbert space adjoint of LT ): ! 0 ∗ φ φ ≥ C , ψ (3) LT . T 0 0 0 H ψ U1 ×[U0 ]
In PDE assumes the following form: terms, this inequality Let φ(t, x), ψ(t, τ ) denote the solution to the backward system (adjoint with respect to the PDE (1)): φ tt (t, x) = ∆φ(t, x) on (0, T ) × Ω ∂ φ(t, x) = ψt on (0, T ) × Γ0 ∂ν ∂ φ(t, x) = 0 on (0, T ) × Γ1 ∂ν 0 on Ω; [φ(T, x), φt (T, x)] = φ ∂2 ψtt (t, τ ) = ∂τ 2 ψ(t, τ ) − φt Γ0 on (0, T ) × Γ0 ∂ψ (4) = 0 on (0, T ) × ∂Γ0 ∂n 0 . [ψ(T, x), ψt (T, x)] = ψ (By standard semigroup theory, the homogenous system above is well posed for terminal data φ0 , ψ0 ∈ H.) To prove Theorem 1(a), the necessary abstract inequality (3) takes the form 2 2 (5) φ0 , ψ0 ≤ CT ψ H 2 (0,T ;L2 (Γ0 )) , H
0 ∈ D(L∗ ). 0, ψ for all φ T On the other hand, the reverse inequality (3) needed to obtain the exact controllability statement Theorem 1(b) takes the explicit form 2 2 2 1 , (6) φ0 , ψ0 ≤ CT φt L2 (0,T ;L2 (Γ1 )) + ψt 2 4
H
L (0,T ;H (Γ0 ))
0 ∈ D(L∗ ). 0, ψ for all φ T This derivation of these inequalities is the objective of the sequel.
© 2004 by CRC Press LLC
3 3.1
DERIVATION OF THE MAIN ESTIMATE Preliminary Notation
We will use throughout the standard denotations for the “energy” of the system (4): 2 2 ∇φ(t) + φt (t) dΩ; Eφ (t) = Ω
Eψ (t)
= Γ0
E(t)
! "2 " " "∂ " ψ(t)" + ψt (t)2 dΓ0 ; " " ∂τ
= Eφ (t) + Eψ (t).
(7)
Under the assumptions made above on the geometry Ω, we can assume throughout that there exists a dense set of data corresponding to smooth (enough) solutions to the PDE (4). Indeed, if Ω has a smooth boundary (the ﬁrst part of Assumption (A2)), this assertion follows from classical elliptic and semigroup theory. On the other hand, if Ω is either a parallelepiped (the second part of Assumption (A2)) or else satisﬁes Assumption (A1) (so that in particular Ω is convex)], then one can appeal to [10]). In this way, one can justify the computations to be done below. Also, we will use the standard denotations: Q0 Σ0
= =
(0 , T − 0 ) × Ω; (0, T ) × Γ0 ; Σ1 = (0, T ) × Γ1 .
= [φ, φt ] and ψ = [ψ, ψt ], we will use below the standard denoIn addition, with φ tation for terms that are “below the level of energy”; namely, ψ ψ) ≡C l.o.t.(φ, , φ, 1− − 1− − C([0,T ];H
(Ω)×H
(Ω)×H
(Γ0 )×H
(Γ0 ))
for some constant C, where > 0. Finally, we will need throughout the cutoff function ξ(t) ∈ C0∞ (R), which is deﬁned by having for arbitrary 0 > 0, 1, for t ∈ [0 , T − 0 ] a C ∞ function with range in (0, 1), for t ∈ (0, 0 ) ∪ (T − 0 , T ) (8) ξ(t) = 0, for t ∈ (−∞, 0) ∪ (T, ∞). 3.2
Step 1 (The Conservation Relation)
Multiplying the ﬁrst equation of (4) by φt , the second by ψt , and subsequently integrating in time and space, and integrating by parts we obtain 1 (φt (σ), φt (σ))L2 (Ω) 2
σ=t σ=t t # $ ∂φ 1 = − (∇φ(σ), ∇φ(σ))L2 (Ω) + dt; , φt 2 ∂ν L2 (Γ0 ) σ=s σ=s s
!σ=t t # σ=t $ 1 ∂ 1 ∂ =− − (φt , ψt )L2 (Γ0 ) dt. (ψt (σ), ψt (σ))L2 (Ω) ψ(σ), ψ(σ) 2 2 ∂τ ∂τ σ=s L2 (Ω) σ=s
© 2004 by CRC Press LLC
s
Applying the Neumann boundary condition in (4) to the ﬁrst equation above and summing the two yields the expected conservation of the system; i.e., E(t) = E(s) for all 0 ≤ s, t ≤ T . In particular then,
2 E(s) = φ 0 , ψ0
H
3.3
for all 0 ≤ s ≤ T .
(9)
Step 2 (The Acoustic Wave Estimates)
Let h be a [C 2 (Ω)]n vector ﬁeld, which will be eventually speciﬁed. With this h, we apply the “classical” wave multipliers (see, e.g., [14], [23], [27], [28]). Multiplying the φwave equation of (4) by h · ∇φ, integrating in time and space and using the Neumann boundary condition, we have
T−0
1 ψt (h · ∇φ) dtdΓ0 + 2
H∇φ · ∇φdQ0 = 0
Q0
1 2
−
Γ0
T−0
2
∇φ h · νdtdΓ + 0
Γ
− (φt , h · ∇φ)L2 (Ω)
T −0 0
1 2
T−0
0
%
φ2t h · νdtdΓ
Γ
& 2 ∇φ − φ2t div(h)dQ0
Q0
.
(10)
Next, we consider again the ﬁrst wave n equation in (4), this time multiplying by the 2 ˜ ˜ quantity φdiv(h), where h(x) ∈ C (Ω) is arbitrary. Integrating in time and space, % & % & ˜ ˜ · ∇φ + and invoking Green’s Theorem and the identity ∇ φdiv(h) = φ∇ div(h) 2 ˜ we obtain ∇φ div(h), T −0 ' ( ˜ ˜ − ∇φ div(h)dQ0 = φt , φdiv(h) H − (Ω)×H (Ω) 0 % & T −0 ∂φ ˜ ˜ + Q φ∇ div(h) ∇φdQ0 − 0 φdiv(h)dtdΓ. Γ ∂ν 0
%
Q0
φ2t
2
&
(11)
Using now the Neumann boundary condition for φ in (4), along with Sobolev Trace ˜ Theory, we have the following inequality for any C 2 vector ﬁeld h: " " % & " " 2 ˜ " Q φ2t − ∇φ div(h)dQ 0 " 0
2 ψt2 dΣ0 + Q ∇φ dQ0 0 % & T 2 2 2 1 + Ch,
φ dt +
φ
+
φ − ˜ t C([0,T ]:H (Ω)) + C([0,T ]:H (Ω)) 0 2
≤C
Σ0
H
≤ C
ψ 2 dΣ0 + Σ0 t
(12)
(Ω)
Q0
% & 2 ψ . ∇φ dQ0 + l.o.t. φ,
We will now consider the two cases, corresponding to Assumptions (A1) and (A2).
© 2004 by CRC Press LLC
3.3.1
Case I. Assumption (A1) is in Force
Given the assumptions on both Γ0 and Γ1 , it is shown in [19] that there exists a vector ﬁeld h(x) = [h1 (x), h2 (x), ..., hn (x)] ∈ [C 2 (Ω)]n such that (i) h · ν = 0 on Γ1 ; (ii) The matrix H(x), deﬁned by i (x) , 1 ≤ i, j ≤ n, [H(x)]ij = ∂h∂x j satisﬁes H(x) ≥ ρ0 I on Ω, for some positive constant ρ0 .
(13)
Applying this particular h to (10), we obtain
T−0
2
∇φ dQ0 ≤
ρ0
0
Q0
−
1 2
Γ0
T−0
2
∇φ h · νdtdΓ0 + 0
T−0
1 ψt (h · ∇φ) dtdΓ0 + 2
Γ0
− (φt , h · ∇φ)L2 (Ω)
T −0 0
1 2
0
φ2t h · νdtdΓ0
Γ0
%
& 2 ∇φ − φ2t div(h)dQ0
Q0
.
& % 2 2 Using the relation ∇φ = ψt2 + ∂φ on Γ0 , as well as the conservation relation (9), ∂τ we obtain from this the majorization, T−0 # 2$ ∂φ 2 φ2t + dtdΓ0 ∇φ dQ0 ≤ Ch ψt2 dΣ0 + ρ0 ∂τ Q0
0
Σ0
Γ0
" " " " " " % & 1" " 2 + 2 max hi L∞ (Ω) E(T ) + " φ2t − ∇φ div(h)dQ0 " , 1≤i≤n " 2" "Q " 0
(14) where h is the vector ﬁeld in (13). ˜ = h therein) now yields Combining (14) and (12) (with h T−0 # 2$ ∂φ 2 φ2t + dtdΓ0 ∇φ dQ0 ≤ C,h ψt2 dΣ0 + (ρ0 − ) ∂τ Q0
0
Σ0
Γ0
ψ). +2 max hi L∞ (Ω) E(T ) + l.o.t.(φ, 1≤i≤n
(15)
˜ which satisﬁes div(h) ˜ = 1) and (15) (taking In turn, the inequalities (12) (taking therein h, therein > 0 small enough) give the intermediate estimate, T−0 T−0 # 2$ ∂φ dtdΓ0 φ2t + Eφ (t)dt ≤ C,h ψt2 dΣ0 + ∂τ 0
Σ0
0
Γ0
ψ). +Ch E(T ) + l.o.t.(φ,
© 2004 by CRC Press LLC
(16)
"
Now in estimating the tangential derivative
∂φ " ∂τ "
Γ0
on the righthand side of (16), there
is no appeal to classical Sobolev trace theory. Instead, we recall the following estimate, a product of microlocal machinery. Lemma 4 (see [17], Lemma 7.2). Let 0 > 0 be arbitrarily small. Let w be a solution of the wave equation on (0, T ) × Ω, or more generally, any secondorder hyperbolic equation with smooth spacedependent coefﬁcients. Then,with wc ≡ ξw (with ξ(t) being the cutoff function deﬁned in (8)), we have the estimate T $2 T # T 2 2 ∂wc ∂wc ∂w dΣ0 ≤ CT,0 dΣ0 + dΣ + l.o.t.(w), (17) ∂τ ∂t ∂ν 0 Γ∗
0 Γ∗
0
Γ
where Γ∗ is a smooth connected segment of boundary Γ. Then, invoking Lemma 4 (with Γ∗ = Γ0 therein), so as to handle the tangential derivative in (16), gives the following: under Assumption (A1), one has the integral estimate for the energy of the component φ of (4), T−0
Eφ (t)dt 0
≤ CT,0,h
T − 0
0
2
2
ξ 2 φt Γ0 dΣ0 +
φt Γ0 dtdΓ0 +
Γ0
Σ0
ψt2 dΣ0
Σ0
ψ), +Ch E(T ) + l.o.t.(φ,
(18)
where the cutoff function ξ is as deﬁned in (8). Here, CT,0 ,h depends on time T , but Ch does not. 3.3.2
Case II. Assumption (A2) is in Force
In this case, since Γ0 is ﬂat, one can construct a radial vector ﬁeld h(x) ∈ [C 2 (Ω)]n such that h(x) · ν = 0 on Γ0 . (19) (Indeed, if x0 ∈ Γ0 , then we can take h(x) = x − x0 .) Applying this vector ﬁeld h to the relation (10), we obtain the inequality,
T−0
2
∇φ dQ0 ≤ 0
Q0
−
1 2
Γ0
1 ψt (h · ∇φ)dtdΓ0 + 2
T−0
2
∇φ h · νdtdΓ1 + 0
Γ1
− (φt , h · ∇φ)L2 (Ω)
© 2004 by CRC Press LLC
T −0 0
.
T−0
0
n 2
Γ1
% Q0
φ2t h · νdtdΓ1 & 2 ∇φ − φ2t dQ0
& % 2 2 ∂φ 2 on Γ, the conservation relation (9), and the fact that radial h Using ∇φ = ∂φ ∂ν + ∂τ is parallel to Γ0 , we obtain
2
∇φ dQ0 ≤ Ch Q0
ψt2 dΣ0 +
Σ0
φ2t dΣ1 +
T−0
0
Σ1
Γ1
∂φ 2 dtdΓ1 ∂τ
" T − " " 0 " " " ∂φ " +C" ψt dtdΓ0 "" + 2 max hi L∞ (Ω) E(T ) 1≤i≤n ∂τ " " 0 Γ 0 " " " " " " % & n" " 2 2 φt − ∇φ dQ0 " . + " " " 2 " "Q
(20)
0
˜ = h and = 1 therein) gives Now combining (20) and (12) (with h 1 2
2
∇φ dQ0
≤ Ch
Q0
ψt2 dΣ0 +
Σ0
φ2t dΣ1 +
Σ1
T−0
0
Γ1
∂φ 2 dtdΓ1 ∂τ
" " T − " " 0 " " ∂φ " ψt dtdΓ0 "" + 2 max hi L∞ (Ω) E(T ) +C " 1≤i≤n ∂τ " " 0
Γ0
ψ). +l.o.t.(φ,
(21)
T −0 ψ ∂φ dtdΓ: again, Sobolev trace theory attaches 0 "Γ0 t ∂τ " no meaning to the boundary trace ∂φ for initial data of ﬁnite energy. Instead, we deal ∂τ " Now to handle the term
Γ0
with this term by the following “sharp” trace regularity result: Lemma 5 (see [18], p. 113, Corollary 3.4(b) and Theorem 3.3(a) [with α = β = 34 therein]). Let Γ0 be a ﬂat portion of the boundary, and let w solve the following wave equation: wtt = ∆w on (0, T ) × Ω ∂w 1 (22) = g ∈ L2 (0, T ; H 4 (Γ0 )) on (0, T ) × Γ0 ∂ν [w(0), wt (0)] = [w0 , w1 ] ∈ H1 . Then one has the following estimate: % & ∂w 1
g .2 ≤ C +
[w , w ] T 0 1 H 2 ∂τ 2 1 1 L (0,T ;H 4 (Γ)) L (0,T ;H − 4 (Γ0 )) explicitly stated in [18], Theorem (3.3)(a) mandates that Neumann boundary data g be smoother in space 1 and time in order to obtain the requisite tangential regularity; namely, g ∈ H 4 ((0, T ) × Γ0 ) (again Γ0 is 1 ∂w ﬂat). However, in the details of proof, it is evident that indeed continuously, g ∈ L2 (0, T ; H 4 (Γ0 )) ⇒ ∂τ 1 L2 (0, T ; H − 4 (Γ0 )). 2 As
© 2004 by CRC Press LLC
δ 2 2a
Applying Lemma 5 (taking g ≡ ψt therein), along with ab ≤ δ ≡ CT of (22)), we have T−0
ψt 0
Γ0
T−0,
∂φ dtdΓ ∂τ
=
ψt , 0
∂φ ∂τ
+
1 2 2δ b
(taking
dt
1
1
H 4 (Γ0 )×H − 4 (Γ0 )
∂φ
ψt 14 dt H (Γ0 ) ∂τ − 1 H 4 (Γ0 )
T−0
≤ 0
CT 1 1 Eφ (T ) + ( + ) 2 2 2
≤
T−0
2 1 H 4 (Γ0 )
ψt
dt.
(23)
0
Coupling (23) with (21) yields
1 2
2
∇φ dQ0
≤ C,h,T
Q0
T
ψt
2 1 H 4 (Γ0 )
dt +
0
φ2t dΣ1 +
T−0
0
Σ1
Γ1
2
∂φ dtdΓ1 ∂τ
ψ). +Ch E(T ) + l.o.t.(φ,
(24)
˜ = h), we obtain In turn, combining this estimate with that in (12) (again with h T−0
Eφ (t)dt
≤ C,h,T
0
T
2 1 H 4 (Γ0 )
ψt
dt +
0
φ2t dΣ1 +
T−0
0
Σ1
Γ1
∂φ 2 dtdΓ1 ∂τ
ψ). Ch E(T ) + l.o.t.(φ,
A subsequent application of the tangential estimate (17) (with Γ∗ = Γ1 ) produces the following: Under Assumption (A2), the φcomponent of the adjoint system (4) obeys the integral estimate, T−0
Eφ (t)dt 0
≤ CT,0 ,h
T
2 1 H 4 (Γ0 )
ψt
0
ψ), +Ch E(T ) + l.o.t.(φ, where CT,0 ,h depends on T , but Ch does not.
© 2004 by CRC Press LLC
dt + Σ1
φ2t dΣ1
(25)
3.4
Step 3 (An Estimate for ψ)
Applying the multiplier ψ to the second wave equation in (4), and integrating in time and space yields t=T ψt (t)ψ(t)dΓ0 − ψt2 dΣ0 Γ0
Σ0
t=0
t=T $2 # ∂ψ = − dΣ0 − φ(t)Γ0 ψ(t)dΓ0 + φΓ0 ψt dΣ0 . ∂τ Σ0
Γ0
t=0
Σ0
A rearrangement of terms, combined with a use of Sobolev trace theory, yields $2 # ∂ψ dΣ0 = ψt2 dΣ0 + φΓ0 ψt dΣ0 ∂τ Σ0
Σ0
Σ0
t=T . / − ψt (t) + φ(t)Γ0 , ψ(t) H − (Γ )×H (Γ ) 0 0 t=0 2 ≤ C ψt2 dΣ0 + φΓ0 dΣ0 +
Σ0
Σ0
2
ψt (t) H − (Γ0 )
≤ C
2 + φ(t)
Γ0 H − (Γ0 )
+
2
ψ(t) H (Γ0 )
ψ). ψt2 dΣ0 + l.o.t.(φ,
t=T $ t=0
(26)
Σ0
3.5
Step 4 (Final Statement of Derived Estimate)
Combining (18) and (26) (in the case that Assumption (A1) holds true), and (25) and (26) (if Assumption (A2) holds true), we have the preliminary estimate for the energy: Lemma 6 (a) Under Assumption (A1), the solution to the PDE (4) satisﬁes the following estimate for all 0 > 0: T−0
E(t)dt 0
≤ CT,0
T − 0
0
2
φt Γ0 dtdΓ0 +
Γ0
ψ), +CE(T ) + l.o.t.(φ, where constant C is independent of time T .
© 2004 by CRC Press LLC
Σ0
2
ξ 2 φt Γ0 dΣ0 +
Σ0
ψt2 dΣ0
(27)
(b) Under Assumption (A2), the solution to the PDE (4) satisﬁes the following estimate for all 0 > 0: T−0
E(t)dt 0
≤ CT,0 ,h
T
ψt
2 1 H 4 (Γ0 )
dt +
0
φ2t dΣ1
Σ1
ψ), + CE(T ) + l.o.t.(φ,
(28)
where constant C is independent of time T . Recall that our aim is to attain the inequalities (5) (under Assumption (A2)), and T − (6) (under Assumption (A2)). We can handle the terms 0 0 E(t)dt and E(T ) in (27) and (28) by the conservation relation (9), and the lowerorder terms by a compactnessuniqueness argument. Hence, the “bad terms” appear only in the inequality (27); i.e., T −0 2 φ  dtdΓ0 and Σ0 ξ 2 φ2t dΣ0 . Accordingly, we concern ourselves next with 0 Γ0 t Γ0 φt Γ0 . ANALYSIS OF φT Γ0
4 4.1
Supporting Results
In [4], we show outright the derivation of the following sequence of supporting Propositions: Proposition 7 The ψcomponent of the PDE (4) satisﬁes the following estimate for arbitary 1 , 2 > 0: $2 # 2 $2 # ∂ψt ∂ ψ 2 2 2 2 ξ dΣ0 ≤ C ψ H 2 (0,T ;L2 (Γ0 )) + 1 ξ φt dΣ0 + 2 ξ dΣ0 . ∂τ ∂τ 2 Σ0
Σ0
Σ0
(29) Proposition 8 The ψcomponent of the PDE system (4) obeys the estimate: # 2 $2 ∂ ψ 2 ξ dΣ0 ∂τ 2 Σ0 2 ≤ {1 C5 + 5 CT } ξ 2 φt Γ0 dΣ0 Σ0
+ (2 C5 + 3 + 4 ) Σ0 2
ξ
2
#
∂2ψ ∂τ 2
$2 dΣ0
ψ), +C ψ H 2 (0,T ;L2 (Γ0 )) + l.o.t.(φ, where the i , i = 1, ..5, are arbitrarily small.
© 2004 by CRC Press LLC
(30)
Proposition 9 The φcomponent of the solution to the system (4) obeys the following estimate: # 2 $2 ∂ ψ 2 2 2 ξ φt Γ0 dΣ0 ≤ (2 C + 3 + 4 ) ξ dΣ0 ∂τ 2 Σ0
Σ0
+C
2
ψ H 2 (0,T ;L2 (Γ0 ))
ψ), + l.o.t.(φ,
(31)
where i , i = 2, 3, 4 are arbitrarily small.
4.2
The Main Estimate for φt Γ0
The Proposition 9 above can be used to prove the following lemma: Lemma 10 Let > 0 be arbitrary. Then the ψcomponent of the system (4) satisﬁes the inequality 2 ψ), ξ 2 ψτ2τ dΣ0 dt ≤ C ψ H 2 (0,T ;L2 (Γ0 )) + l.o.t.(φ, Σ0
where ξ is the cutoff function deﬁned in (8). Proof of Lemma 10: Squaring both sides of the ψc wave equation in (4), integrating in time and space, and applying Cauchy–Schwartz, we obtain the inequality ξ
2
#
Σ0
∂2ψ ∂τ 2
$2
2 2 ξ 2 φt Γ0 dΣ0 + C ψ H 2 (0,T ;L2 (Γ0 )) .
dΣ0 ≤ 2 Σ0
Now applying the trace estimate (31) gives ξ Σ0
2
#
∂2ψ ∂τ 2
$2 dΣ0
≤ 2 (2 C + 3 + 4 )
ξ2
Σ0
+C
2
ψ H 2 (0,T ;L2 (Γ0 ))
#
∂2ψ ∂τ 2
$2 dΣ0
ψ). + l.o.t.(φ,
Taking i to be small enough yields the estimate ξ
2
#
Σ0
∂2ψ ∂τ 2
$2
2 ψ). dΣ0 ≤ C ψ H 2 (0,T ;L2 (Γ0 )) + l.o.t.(φ,
This completes the proof of Lemma 10. Since φt Γ0 = ψτ τ − ψtt
on (0, T ) × Γ0 ,
Lemma 10 immediately gives the following necessary estimate:
© 2004 by CRC Press LLC
(32)
Corollary 11 Let > 0 be arbitrary. Then the φcomponent of the system (4) satisﬁes the inequality 2 2 ψ). ξ 2 φt Σ0 dΣ0 ≤ C ψ H 2 (0,T ;L2 (Γ0 )) + l.o.t.(φ, Σ0
5 5.1
PROOF PROPER OF THEOREM 1 Completion of the Proof of Theorem 1(a)
Recall that Theorem 1(a), which provides for exact controllability of the PDE system (1) under Assumption (A1), is equivalent to obtaining the inequality (5) for the adjoint system (4). Combining the estimates from Lemma 6 and Corollary 11, along with the deﬁnition of the cutoff function ξ in (8), we obtain T−0
E(t)dt 0 2 ψ), ≤ CT,0 ψ H 2 (0,T ;L2 (Γ0 )) + CE(T ) + l.o.t.(φ,
(33)
where again CT,0 depends on time T , but C does not. Using the conservation relation (9), we have
(T − 20 ) E(T ) 2 ψ). ≤ CT ψ H 2 (0,T ;L2 (Γ0 )) + CE(T ) + l.o.t.(φ,
From this inequality, we then have Lemma 12 Under Assumption (A1), the solution of (4) obeys the following estimate for long enough terminal time T : 2 2 ψ). φ0 , ψ0 ≤ C ψ H 2 (0,T ;L2 (Γ0 )) + l.o.t.(φ, H
(34)
A compactnessuniqueness argument can be used to eliminate the lowerorder terms in (34). That is, given the existence of the estimate (34), one can ascertain the existence of a positive constant CT such that φ, ψ
C([0,T ];H 1− (Ω)×H − (Ω)×H 1− (Γ0 )×H − (Γ0 ))
2
≤ CT ψ H 2 (0,T ;L2 (Γ0 ))
(35)
(see [4] for the full details of the argument). Combining (34) and (35) completes the proof of Theorem 1(a).
© 2004 by CRC Press LLC
5.2
Completion of the Proof of Theorem 1(b)
Applying the conservation relation (9) to the inequality (28), valid under Assumption (A2), we have (T − 20 )E(T ) T 2 ≤ CT,0 ,h
ψt 14 dt + φ2t dΣ1 + CE(T ); H (Γ0 ) 0
(36)
Σ1
whence the inequality E(T ) ≤ CT,0 ,h
T
0
2 1 H 4 (Γ0 )
ψt
dt + Σ1
φ2t dΣ1
ψ). + l.o.t.(φ,
(37)
By an argument virtually identical to that used to prove the estimate (35), we can eliminate the lowerorder terms in (37) to thereby obtain the requisite reverse inequality (6). This completes the proof of Theorem 1(b). Acknowledgment Research was partially supported by NSF Grant No. DMS0196359, NSF Grant No. DMS0104305, and ARO DAAD 190210179. REFERENCES 1. G. Avalos and I. Lasiecka, Differential Riccati equation for the active control of a problem in structural acoustics, Journal of Optimization Theory and Applications, Vol. 91, No. 3 (1996), pp. 695–728. 2. G. Avalos, The exponential stability of a coupled hyperbolic/parabolic system arising in structural acoustics, Applied and Abstract Analysis, Volume 1, Number 2 (1996), pp. 203–217. 3. G. Avalos and I. Lasiecka, Boundary controllability of thermoelastic plates via the free boundary conditions, SIAM J. Control Optim., Vol. 38, No. 2 (2000), pp. 337–383. 4. G. Avalos and I. Lasiecka, Exact controllability of structural acoustic interactions, IMA Preprint Series #1771 (June 2001) and submitted to J. Math. Pures Appl. 5. G. Avalos, I. Lasiecka, and R. Rebarber, Boundary Controllability of a Coupled Wave/Kirchoff System, preprint. 6. H. T. Banks, W. Fang, R. J. Silcox, and R. Smith, Approximation Methods for Control of Acoustic/Structure Models with Piezoceramic Actuators, Contract Report 189578, NASA (1991). 7. C. Bardos, G. Lebeau, and J. Rauch, Sharp sufﬁcient conditions for the observation, control and stabilization of waves from the boundary, SIAM J. Control Optim., 30 (1992), pp. 1024–1065. 8. J.T. Beale, Spectral properties of an acoustic boundary condition, Indiana Univ. Math. J., 25 (1976), pp. 895–917.
© 2004 by CRC Press LLC
9. M. Camurdan and R. Triggiani, Sharp regularity of a coupled system of a wave and Kirchoff equation with point control arising in noise reduction, Differential and Integral Equations, Vol. 12, No. 1, January (1999), pp. 101–118. 10. P. Grisvard, Elliptic Problems in Nonsmooth Domains, Pitman Advanced Publishing, Boston (1985). 11. L H¨ormander, Linear Partial Differential Operators, SpringerVerlag, New York (1969). 12. V. Hutson and J.S. Pym, Applications of Functional Analysis and Operator Theory, Academic Press, New York (1980). 13. V. Komornik, Exact Controllability and Stabilization. The Multiplier Method, Masson, Paris (1994). 14. J. Lagnese, Decay of solutions of wave equations in a bounded region with boundary dissipation, J. Differential Equations, 50 (1983), pp. 106–113. 15. I. Lasiecka and R. Triggiani, Exact controllability of the wave equation with Neumann boundary control, Applied Mathematics and Optimization, Vol. 19 (1989), pp. 243– 290. 16. I. Lasiecka and R. Triggiani, Sharp regularity results for mixed secondorder hyperbolic equations of Neumann type: The L2 boundary case, Annali di Matem. Pura e Appl., IV, CLVII (1990), pp. 285–367. 17. I. Lasiecka and R. Triggiani, Uniform stabilization of the wave equation with Dirichlet or Neumann feedback control without geometrical conditions, Applied Mathematics and Optimization, Vol. 25 (1992), pp. 189–224. 18. I. Lasiecka and R. Triggiani, Recent advances in regularity of secondorder hyperbolic mixed problems, and applications, Dynamics Reported: Expositions in Dynamical Systems, Vol. 3 (1994), pp. 104–158. 19. I. Lasiecka, R. Triggiani, and X. Zhang, Nonconservative wave equations with unobserved Neumann B.C.: Global uniqueness and observability on one shot, Contemporary Mathematics, Vol. 268 (2000), pp. 227–325. 20. W. Littman and B. Liu, On the spectral properties and stabilization of acoustic ﬂow, IMA Preprint Series #1436, University of Minnesota, November (1996). 21. W. Littman and L. Markus, Stabilization of hybrid system of elasticity by feedback boundary damping, Annali di Mathematica Pure et Applicata, 152 (1988), pp. 281– 330. 22. S. Micu, Boundary controllability of a linear hybrid system arising in the control of noise, SIAM J. Control Optim., Vol. 35 (1997), pp. 531–555. 23. C.S. Morawetz, Energy identities for the wave equation, Math. Sci. Res. Rep. No. IMM 346, NYU Courant Institute (1976). 24. P.M. Morse and K.U. Ingard, Theoretical Acoustics, McGrawHill, New York (1968). 25. A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations, SpringerVerlag, New York (1983). 26. J. Simon, Compact sets in the space Lp (0, T ; B), Ann. Mat. Pura Appl., (4), 148 (1987), pp. 65–96. 27. W. Strauss, Dispersal of waves vanishing on the boundary of an exterior domain, Comm. Pure Appl. Math. 28 (1976), pp. 265–278. 28. R. Triggiani, Wave equation on a bounded domain with boundary dissipation: An operator approach, Journal of Mathematical Analysis and Applications 137 (1989), pp. 438–461.
© 2004 by CRC Press LLC
12 Flow Field, Temperature, and Dopant Concentration Evolution in a Bridgman–Stockbarger Crystal Growth System in a Strictly ZeroGravity and a LowGravity Environment St. Balint† and A.M. Balint∗ †
Department of Mathematics, University of West Timisoara, Timisoara, Romania ∗ Department of Physics, University of West Timisoara, Timisoara, Romania
In order to describe the evolution of the ﬂow, temperature, and dopant concentration in the melt in a Bridgman–Stockbarger crystal growth system, a continuum mechanical model is presented. The equations of the model are the classical Navier–Stokes equation in the Boussinesq approximation, the classical incompressibility equation as well as a modiﬁed energy equation and a modiﬁed dopant dispersion equation. The classical energy and dopant transport equations were modiﬁed due to the thermotransport and precrystallizationzone inﬂuences. We give a modelbased simulation of the ﬂow, heat transport, and Ga dispersion evolution in the melt and prediction of axial and radial Ga distribution in a Gadoped Ge semiconductor crystal grown in a strictly zerogravity and a lowgravity environment. We present adequate initial Ga distribution in the melt for obtaining uniform Ga distribution in the crystal.
1
INTRODUCTION
The prototype Bridgman–Stockbarger growth system consists of crystal and melt contained in a cylindrical ampoule of radius R and length L pulled slowly through a vertically aligned furnace with hot and cold isothermal zones, separated by an adiabatic region of length Lg , designed to promote steep axial temperature gradient and to maintain a ﬂat solidiﬁcation interface, as shown in Fig. 1 [1]. In [2] there is a characterization of the sample obtained in such a system in the shuttle mission 51G, experiment MRS77F075. The axial dopant distribution in the sample, reported in [2], exhibits an important nonlinear variation. This compositional nonuniformity inﬂuences the crystallographic quality of the sample. Therefore it is extremely important to identify the origin of this nonuniformity.
© 2004 by CRC Press LLC
Figure 1 Prototype Bridgman–Stockbarger growth system.
In a general context, the axial and radial dopant distribution in a semiconductor crystal grown in a Bridgman–Stockbarger system is determined by the thermal convection in the melt [1], [3], [4]; the thermotransport [5], [6]; the decrease of the melt in the ampoule [5], [6]; the morphology of the solidiﬁed front [7], [8]; the initial dopant distribution in the melt [6], [9]. During the growth on Earth, the temperature gradients in the melt generate buoyancy driving forces and thermal convection in the melt [1], [3], [4], which results in a “nearly completely” mixed melt [4]. Crystals grown from wellmixed melts exhibit a nonlinear variation of the dopant concentration along the growth axis [9]. On the other hand, crystals grown from a quiescent melt, after an initial transient exhibit a uniform axial dopant distribution [10]. Thus, reduction of the magnitude of the buoyancy forces by processing semiconductors in a lowgravity environment has been pursued over the past decades. The effectiveness of space processing for the growth of chemically uniform crystals is supported by experimental and theoretical studies. For example, the InSb crystal reported in [11] exhibits axial segregation proﬁles that are characteristic of diffusioncontrolled mass transfer growth. In [1], [3] it is shown, based on modeling studies, that the lowgravity levels achieved in space are sufﬁcient to inhibit interference of thermal convection with segregation in smalldiameter Ge and GeSi melts. The above analysis has been performed using constant values of the gravitational acceleration. In [12], [13] there is a study of the inﬂuence of nonsteady gravity on the thermal convection during microgravity solidiﬁcation of semiconductors. The study corresponds to the use of lowduration lowgravity vehicles, such as sounding rockets and KC135 aircraft, for lowgravity crystal growth experiments. In these vehicles, lowgravity periods of 20 seconds and 6 minutes, for KC135 aircraft and sounding rockets, respectively, are achieved at costs that are at orders of magnitude smaller than space experiments.
© 2004 by CRC Press LLC
In [5], the authors consider a strictly zerogravity environment, in the absence of convective mixing, and identify the thermotransport and the decrease of the melt in the ampoule as the origin of the nonlinear variation of the dopant concentration along the axis of the sample, reported in [2]. In [6], there is also an analysis of the role of the decrease of the melt and of the role of thermotransport in axial dopant segregation of Ga in Ge. Concerning the morphology of the solidiﬁed front, in [7] it is proved that signiﬁcant radial segregation occurs in the system without convection tangent to the crystal surface when the radius of curvature of this interface is the same or less than the lengthscale of the concentration gradient adiacent to the interface. Curvatureinduced segregation was shown to be an important contribution to the dopant inhomogeneity in capillary growth systems [14], [15]. All the abovementioned studies were made in crystal growth models in which the solidiﬁcation interface is a smooth mathematical surface. But, physically the melt → solid phase transition takes place in a thin layer of width 10−9 m, masking the crystal, where both phases, melt and solid, coexist. In [8] it has been shown that such a boundary layer reduces the molecular diffusion from D to Deff = D×0.917, increases the heat conductivity from K to Keff = K×1.1, and does not inﬂuence the ﬂow in the region. In [6] it is shown that such a boundary layer has a signiﬁcant inﬂuence on the dopant distribution in the crystal. The inﬂuence of the initial dopant distribution in the melt on the dopant distribution in the crystal was pointed out in [9]. In [6], this inﬂuence is used to show how it is possible to reduce the axial segregation (nonuniformity) produced by the decrease of the melt, thermotransport, and precrystallization zone, respectively, in the case of crystals grown in strictly zero gravity. In this chapter we present a continuum mechanical model that takes into account the thermoconvection in the melt, the thermotransport, the decrease of the melt, the morphology of the solidiﬁed front and the initial dopant distribution in the melt. We give a modelbased simulation of the ﬂow evolution, heat transport, dopant dispersion in the melt and we predict axial and radial dopant distribution in a Gadoped Ge semiconductor crystal grown in a strictly zerogravity and in a lowgravity environment. We present adequate initial Ga distribution in the melt for obtaining uniform Ga distribution in the Gadoped Ge semiconductor crystal.
2
THE CONTINUUM MECHANICAL MODEL
A continuum mechanical model that takes into account the thermoconvection in the melt, the thermotransport, the decrease of the melt in the ampoule, the inﬂuence of the precrystallizationzone and the initial dopant distribution in the melt, is deﬁned in a mobile reference frame, as that shown in Fig. 1, by the following dimensionless equations, boundary conditions, and initial conditions written in cylindrical polar cordinates: ∂u + (u∇)u = −∇p + Pr ·∆u + Pr ·Ra · θ · e3 , ∂t ∇u = 0,
© 2004 by CRC Press LLC
(1) (2)
Ka L L ∂θ + u · ∇θ = K∗ ∆θ − 2 (θw − θf ), ∂t K δR ∂c bc + u · ∇c = D∗ ∆c + σ(TH − TC ) · ∇(c · ∇θ), ∂t u∂Ωt = 0,
bθ
u(0, rϕ, x3 ) = u0 (r, ϕ, x3 ), ∂θ = 0, ∂x3
(3) (4) (5) (6) (7)
(∂Ωt )top
θ(∂Ωt )w = θw ,
(8)
θ(∂Ωt )i = θi ,
(9)
θ(0, r, ϕ, x3 ) = θ0 (r, ϕ, x3 ), ∂c = 0, ∂x3 (∂Ωt )top ∂c =0 ∂r (∂Ωt )w ∂c = φi , ∂x3
(10)
c(0, r, ϕ, x3 ) = c0 (r, ϕ, x3 ).
(14)
(11) (12) (13)
(∂Ωt )i
Equation (1) is the dimensionless Navier–Stokes equation describing the thermoconvection in the melt due to the buoyancy driving forces Pr ·Ra · θ · e3 , in the Boussinesq approximation. The signiﬁcance of the quantities u, p, Pr Ra and θ is given in Table 1. The equation is considered in the timedependent domain Dt deﬁned by 1 Lg ; 0 r Λ; 0 ϕ 2π; 0 x3 1 Dt = (t, r, ϕ, x3 ) 0 t up 2L 1 Lg Lg 1 ; 1+ t ∪ (t, r, ϕ, x3 ) up 2L up 2L Lg 0 r Λ; 0 ϕ 2π; 0 x3 1 + (15) − up t . 2L This is because the domain Ωt occupied by the melt is time dependent and is deﬁned by L {(r, ϕ, x3 )  0 r Λ; 0 ϕ 2π; 0 x3 1 } for 0 t u1p 2Lg
L (r, ϕ, x3 ) 0 r Λ; 0 ϕ 2π; 0 x3 1 + 2Lg − up t Ωt = . (16)
for 1 Lg t 1 1 + Lg up 2L
up
2L
The restriction imposed on x3 , 0 x3 1 + Lg /2L − up t for 1/up · Lg /2L t 1/up · (1 + Lg /2L), expresses the decrease of the melt in the ampoule with constant ratio up . Equation (2) is a dimensionless incompressibility equation of the melt. This equation is considered in Dt , too.
© 2004 by CRC Press LLC
Table 1 List of Symbols Group
Deﬁnition
u = υL/χm υ L χm p = pm L2 /(ρm χm ) pm ρm,s t = τ χm /L2 τ θ = (T − TC )/(TH − TC ) TH TC T c R Lg Λ = R/L Pr = ν/χm Ra = βg(TH − TC )L3 /χm ν β g ν P e = Vs L/χm Sc = ν/D Scef f D K = ks /km km,s Ka cm,s up = Vs L/χm Vs δ σ k
dimensionless velocity ﬁeld velocity ﬁeld length of the ampoule thermal diffusivity in the melt dimensionless pressure ﬁeld pressure ﬁeld density of melt and solid, respectively dimensionless time time dimensionless temperature temperature of the hot zone of the furnace temperature of the cold zone of the furnace temperature dopant concentration radius of the ampoule length of the adiabatic region aspect ratio Prandtl number Rayleigh number thermal expansion coefﬁcient gravity acceleration kinematic viscosity of the melt P´eclet number Schmidt number effective Schmidt number diffusivity of Ga in Ge conductivity ratio thermal conductivity of melt and solid, respectively heat conductivity of the ampoule heat capacity of melt and solid, respectively ratio of decrease of the melt growth rate wall thickness Soret coefﬁcient segregation coefﬁcient
Value
2 cm 0.13 cm2 / s 5.6 g/ cm3
1097◦ C 819◦ C
0.5 cm 0.25 cm 0.25 0.01 0; 103 0.25×10−3 K−1 0; 0.264 cm/ s2 1.3 × 10−3 cm2 / s 0.01 10 10.9 1.3 × 10−4 cm2 / s 1 0.17 W/K cm 3.26 W/K cm 0.39 J/K g 0.01 16 µm/s 0.2 cm 0.005 0.1
Equation (3) is a modiﬁed dimensionless heat transport equation in the melt (Dt ). The signiﬁcance of the quantities θ, u, L, R, δ, Ka , and K is given in Table 1. The quantities bθ and K∗ are related to the precrystallization zone. Physically, this zone is a thin layer (masking the crystal) of width 10−9 m in which there are periodically distributed solid inclusions of size 10−10 m. The porosity of this layer is ΠY = 0.9 and its thermal effect is the increase of the heat conductivity from K to Keff = 1.1 × K. For more details, see [8]. The quantities bθ and K∗ are deﬁned as follows: 1 on Dt1 1 on D2,1 t bθ = 2,2 ; ρ ·c ΠY + (1 − ΠY ) ρms ·csm on Dt Π + (1 − Π ) ρs ·cs on D2,3 Y Y ρm ·cm t where the sets Dt1 , Dt2,1 , Dt2,2 , and Dt2,3 are deﬁned by
© 2004 by CRC Press LLC
1 on Dt1 1 on D2,1 t K∗ = , 1.1 on Dt2,2 1.1 on Dt2,3
(17)
Dt1 = Dt2,1 =
1 Lg (t, r, ϕ, x3 ) 0 t ; 0 r Λ; 0 ϕ 2π; 0 x3 1 up 2L
L L −2·10−9 ; 0 r Λ; 0 ϕ 2π; (t, r, ϕ, x3 ) u1p 2Lg t u1p 1 + g 2L
Dt2,2
=
Dt2,3
=
0 x3 1 + L
Lg −2·10−9 2L
(t, r, ϕ, x3 ) u1p 2Lg t 1+
Lg −2·10−9 2L
− up t
Lg −2·10−9 2L L x3 1 + 2Lg
1+
− up t
−9
Lg −2·10 2L L + 2Lg − up t;
(t, r, ϕ, x3 ) u1p 1 + 0 x3 1
1 up
t
1 up
; 0 r Λ; 0 ϕ 2π;
− up t
1+
Lg 2L
;
0 r Λ; 0 ϕ 2π;
.
(18)
For ρs cs = ρm cm , bθ is constant, equal to one, on Dt . The signiﬁcance of the quantities θw and θf will be discussed later, together with the boundary condition (8). Equation (4) is a modiﬁed dopant dispersion equation in the melt (Dt ). The term σ · (TH − TC ) · (c∇θ) expresses the thermotransport. The signiﬁcance of the quantities c, u, σ, TH , TC , and θ is given in Table 1. The quantities bc and D∗ are related to the precrystallization zone (which reduces the molecular diffusivity of the dopant from D to Deff = D × 0.917) and are deﬁned as follows: 1 on Dt1 Pr /Sc on Dt1 1 on D2,1 Pr /Sc on D2,1 t t bc = ; D∗ = . (19) ΠY on Dt2,2 Pr /Scef f on Dt2,2 ΠY on Dt2,3 Pr /Scef f on Dt2,3 The boundary condition (5) is the nonslip condition on the boundary ∂Ωt . This boundary can be represented as follows: ∂Ωt = (∂Ωt )top ∪ (∂Ωt )w ∪ (∂Ωt )i .
(20)
The set (∂Ωt )top is the boundary of the melt limited by the top of the ampoule and is deﬁned by (∂Ωt )top = {(r, ϕ, 0)0 r Λ; 0 ϕ 2π}. (21) The set (∂Ωt )w is the melt boundary limited by the walls of the ampoule and is deﬁned by L {(Λ, ϕ, x3 )  0 ϕ 2π; 0 x3 1 } for 0 t u1p 2Lg
L (Λ, ϕ, x3 ) 0 ϕ 2π; 0 x3 1 + 2Lg − up t . (22) (∂Ωt )w =
Lg 1 Lg 1 1+ t for up 2L
up
2L
The set (∂Ωt )I represents the melt boundary limited by the bottom of the ampoule for 0 t 1/up · Lg /2L and for 1/up · Lg /2L t 1/up · (1 + Lg /2L) it is the precrystallization zone/crystal interface: L {(r, ϕ, 1)  0 r Λ; 0 ϕ 2π } for 0 t u1p 2Lg
L (r, ϕ, x3 ) 0 r Λ; 0 ϕ 2π; x3 = 1 + 2Lg − up t . (23) (∂Ωt )i =
Lg 1 Lg 1 1+ t for up 2L
© 2004 by CRC Press LLC
up
2L
The initial condition (6) prescribes the velocity ﬁeld in the melt. This is u0 (r, ϕ, x3 ) and generally it will be considered equal to zero. The boundary condition (7) is the noﬂux condition for the heat on the top of the ampoule (∂Ωt )top . Boundary condition (8) prescribes the value of the dimensionless temperature on the melt surface. According to [5], the heat exchange between the melt surface and the furnace is Q = (TH − TC )(θw − θf )P/Rez · Ac); where θw and θf are the (dimensionless) melt surface and furnace temperature, respectively; P/Ac is the geometry coefﬁcient, equal to 2/R, and Rez is the total heat resistance over the radial length between sample and furnace surfaces. If radiation is neglected, then the resistance is only a function of the heat conductivity Ka of the ampoule and wall thickness δ, Rez = δ/Ka . Therefore, the heat exchange Q is given by Q = (TH − TC ))(θw − θf ) and θw is θw = θf +
2 Ka R δ
Q R δ . (TH − TC ) 2 Ka
(24)
(25)
If the heat exchange between the melt surface and furnace is negligible (the ampoule has negligible thermal mass [1]), then θw = θf and is deﬁned by L 1; for 0 t u1p 2Lg and 0 x3 1 − up t L L L − Lg (x3 + up t) + 1 + Lg ; for 0 t u1p 2Lg and 1 − up t x3 1 L 1; for 1 g t u1p and 0 x3 1 − up t L up 2L Lg g 1 Lg 1 − L (x3 + up t) + 1 + L ; for up 2L t up L and 1 − up t x3 1 + 2Lg − up t θw = θf = L 1 Lg 1 0; for up 2L t up and 1 + 2Lg − up t x3 1
L L L − Lg (x3 + up t) + 1 + Lg ; for u1p t u1p 1 + 2Lg L and 0 x3 1 + 2Lg − up t L L 0; for u1p t u1p 1 + 2Lg and 1 + Lg − up t x3 1. (26) For ampoules with a smalldiameter crosssection, in the onedimensional approximation θw = θ, the boundary condition (8) disappears and in Eq. (3) θw is replaced by θ. The boundary condition (9) prescribes the temperature on the bottom of the ampoule for 0 t 1/up · Lg /2L and the temperature on the precrystallization zone/crystal interface for 1/up · Lg /2L t 1/up · (1 + Lg /2L). That is, L − LLg (1 + up t) + 1 + LLg ; for 0 t u1p 2Lg ; 0 r Λ; 0 ϕ 2π and
x3 = 1 (27) θi = Lg 1 1 Lg 1 1 + ; for t up 2L up 2L ; 2 L 0 r Λ; 0 ϕ 2π and x3 = 1 + 2Lg − up t. The initial condition (10) is the temperature distribution in the melt at the start; θ(0, r, ϕ, x3 ) = θ0 (r, ϕ, x3 ). We shall consider the case when this distribution is uniform and equal to 1, i.e., θ0 (r, ϕ, x3 ) ≡ 1.
© 2004 by CRC Press LLC
The boundary condition (11) expresses the noﬂux condition of the dopant on the top of the ampoule (∂Ωt )top . The boundary condition (12) is the noﬂux condition of the dopant on the wall of the ampoule. For ampoules with a smalldiameter crosssection, in the onedimensional approximation, this boundary condition disappears. Boundary condition (13) is the noﬂux condition for the dopant on the bottom of the ampoule for 0 t 1/up · Lg /2L and the ﬂux of the dopant rejected into the melt for 1/up · Lg /2L t 1/up · (1 + Lg /2L) and x3 = 1 + Lg /2L − up t. That is, 1 Lg 0; for 0 t up 2L ; 0 r Λ; 0 ϕ 2π; x3 = 1
Lg Pe ∂θ 1 Lg 1 (28) φi = 1 + ; (1 − k)c + σ(T − T )c ; for t h c D ∂x u 2L u 2L ∗ p p L 0 r Λ; 0 ϕ 2π; x = 1 + g − u t. 3
3
2L
p
TEMPERATURE AND DOPANT CONCENTRATION EVOLUTION IN THE MELT IN AMPOULES WITH A SMALLDIAMETER CROSSSECTION IN A SPACE EXPERIMENT IN STRICTLY ZERO GRAVITY
In a space experiment in strictly zero gravity, the Rayleigh number Ra is equal to zero (Ra = 0). Therefore, Eqs. (1) and (2) are completely decoupled. The solution of Eqs. (1) and (2), satisfying boundary condition (5) and initial condition (6), with u0 = 0, is the velocity ﬁeld u ≡ 0 on Dt . For ampoules with a smalldiameter crosssection we shall use the onedimensional heat transport equation. In this particular case, the heat transport Eq. (3) becomes bθ
∂θ ∂2θ Ka L L = K∗ 2 − 2 (θ − θf ) ∂t ∂x K δR
(3 )
in the domain Dt given by 1 Lg Dt = (t, x) 0 t ;0 x 1 up 2L 1 Lg Lg Lg 1 ;0 x 1 + 1+ t − up t . ∪ (t, x) up 2L up 2L 2L
(29)
If we assume that ρs cs = ρm cm , then the coefﬁcient bθ according to (17) is constant, equal to one, bθ = 1, on Dt . The coefﬁcient K∗ is given by (17) in which the sets Dt1 , Dt2,1 , Dt2,2 , and Dt2,3 are deﬁned by (18). The furnace temperature θf is given by formula (26) in which x3 is replaced by x. The boundary condition (7) in this case becomes ∂θ Lg 1 . (7 ) 1 + = 0 for 0 t ∂x x=0 up 2L The boundary condition (9) becomes − LLg (1 + up t) + 1 + LLg ; for 0 t
θi (t) = Lg 1 Lg 1 1 1 + ; for t 2 up 2L up 2L .
© 2004 by CRC Press LLC
1 Lg up 2L
(9 )
The initial condition (10) in this particular case is θ(0, x) = 1; for 0 x 1.
(10 )
In the case of ampoules with a smalldiameter crosssection for the dopant transport we can use also the onedimensional mass transport equation. In this case, Eq. (4) becomes bc
∂c ∂2c ∂θ ∂c ∂2θ = D∗ 2 + σ(TH − TC ) · · + σ(TH − TC ) · 2 c ∂t ∂x ∂x ∂x ∂x
in the domain Dt given by (29). The coefﬁcients bc and D∗ are given by the formula (19). The boundary condition (11) for the dopant concentration in this case is ∂c Lg 1 . 1+ = 0 for 0 t ∂x x=0 up 2L
(4 )
(11 )
The boundary condition (12) disappears and the boundary condition (13) becomes L 0; for 0 t u1p 2Lg , x = 1
∂c Lg Pe ∂θ 1 Lg 1 = φi (t) = 1 + . (1 − k)c + σ(T − T )c ; for t h c D ∂x u 2L u 2L ∗ p p ∂x Lg −u t x=1+ 2L
p
(13 )
The initial dopant distribution in this case is given by c(0, x) = c0 (x).
(14 )
For c0 (x) = c0 = const., the evolution of the temperature and the Ga concentration in the melt in this particular case was computed in [6]. Using the computed Ga concentration c(t, x), it was possible to compute the amount of the dopant entering into the crystal at moment t. This is ccr (t) = k · c(t, 1 + Lg /2L − up t); 1/up · Lg /2L t 1/up · (1 + Lg /2L). From here it is possible to predict the dopant concentration in the crystal as a function of the solidiﬁed fraction ξ = up t − Lg /2L for 1/up · Lg /2L t 1/up · (1 + Lg /2L): ccr (ξ) = k · c((ξ + Lg /2L)/up , 1 − ξ); 0 ξ 1. This was computed in [6] and the result is plotted in Fig. 2a. For c0 (x) = 1 + 11.78 · exp(−12.78x) the same computations were made in [6] and the dopant concentration in the crystal is plotted in Fig. 2b.
4
MODELBASED SIMULATION OF THE FLOW EVOLUTION, HEAT TRANSPORT, AND Ga DISPERSION IN THE MELT IN A LOWGRAVITY EXPERIMENT
In order to compute ﬂow, heat transport, and Ga dispersion in the model deﬁned by Eqs. (1)–(14) we assume that at each moment the ﬂow, the heat transport, and Ga distribution
© 2004 by CRC Press LLC
Figure 2 (a) Computed axial Ga distribution in the crystal due to the decrease of the melt in the ampoule and the Soret effect for initial Ga concentration c0 (x) ≡ 1. (b) Computed axial Ga distribution in the crystal due to the decrease of the melt in the ampoule and the Soret effect for initial Ga concentration c0 (x) = 1 + 11.78 · exp(−12.78x).
are axisymmetric. This assumption is realistic because Ra = 103 and convenient because it permits reduction of the 3D problem to a 2D problem. We also assume that the ampoule has negligible thermal mass and therefore according to [1] θw = θf . Therefore in Eq. (3) L the term −2 KKa Lδ R (θw − θf ) disappears. In our computation we also neglect the thermotransport and in Eq. (4) we take σ = 0. The dimensionless parameters and representative values used in the computations are given in Table 1. Computations were performed using the software Cosmos/M, which is a ﬁnite element system. For the initial Ga distribution L c0 (r, ϕ, x3 ) ≡ 1, the computed results obtained in [14] for the moments t1 = u1p 2Lg and t2 =
5
1 Lg up L
are plotted in Figs. 3 to 6.
MODELBASED PREDICTION OF THE AXIAL AND RADIAL SEGREGATION IN LOWGRAVITY EXPERIMENTS
Using the computed Ga distribution in the melt c(t, r, x3 ) it is possible to predict the distribution of Ga in the crystal. For this it is convenient to introduce the solidiﬁed fraction ξ = up t − Lg /2L for 1/up · Lg /2L t 1/up · (1 + Lg /2L) and to express the amount of Ga entering into the crystal at the moment t : ccr (t) = k · c(t, r, 1 + Lg /2L − up t) as a function of the solidiﬁed fraction ξ: ccr (ξ, r) = k · c((ξ + Lg /2L)/up , r, 1 − ξ); 0 ξ 1. The computed values are plotted in Fig. 7.
© 2004 by CRC Press LLC
Figure 3 Flow ﬁeld (a), axial velocity component (b), radial velocity component (c) and velocity magnitude (d) in the absence of the precrystallization zone for t1 = 1/up · Lg /2L and Ra = 103 .
Figure 4 Temperature ﬁeld ((a), (b)), and dopant concentration ﬁeld ((c), (d)) in the absence of the precrystallization zone for t1 = 1/up · Lg /2L and Ra = 103 .
Figure 5 Flow ﬁeld (a), axial velocity component (b), radial velocity component (c), and velocity magnitude (d) in the absence of the precrystallization zone for t2 = 1/up · Lg /L and Ra = 103 .
© 2004 by CRC Press LLC
Figure 6 Temperature ﬁeld ((a), (b)), and dopant concentration ﬁeld ((c), (d)) in the absence of the precrystallization zone for t2 = 1/up · Lg /L and Ra = 103 .
Figure 7 Ga concentration in the crystal function of the solidiﬁed fraction in the absence of the precrystallization zone (a) and in presence of the precrystallization zone (b). Plot (c) represents a comparison of the axial Ga concentration in the crystal in the absence/presence of the precrystallization zone.
© 2004 by CRC Press LLC
REFERENCES 1. Ch.J. Chang and R.A. Brown, “Radial segregation induced by naturalconvection and melt solid surface interface shape in vertical Bridgman growth,” J. Crystal Growth 63 (1983), 343. 2. D.J. Larson Jr. and B.S. Dressler, “Shuttle mission 51G, experiment MRS77F075 Flight sample characterization report,” NASA Report, Re753 (1988). 3. P.M. Adornato and R.A. Brown, “Convection and segregation in directional solidiﬁcation of dilute and nondilute binary alloys — effects of ampoule and furnace design,” J. Crystal Growth 80 (1987), 155. 4. C. Wang, Ph.D. thesis, Dept. of Materials Science and Engineering, Massachussetts Institute of Technology, Cambridge (1984). 5. L.L. Zheng, D.J. Larson Jr., and H. Zhang, “Role of thermotransport (Soret effect) in macrosegregation during eutectic/offeutectic directional solidiﬁcation,” J. Crystal Growth 191 (1998), 243. 6. A.M. Balint, E. TulcanPaulescu, and St. Balint, “The effect of the initial dopant distribution in the melt on the axial compositional uniformity of a thin doped crystal grown in strictly zerogravity environment by Bridgman–Stockbarger method,” J. Crystal Growth 247 (2003), 313. 7. S.R. Coriell, R.G. Boisvert, R.G. Rehm, and R.F. Sekerka, “Lateral solute segregation during unidirectional solidiﬁcation of a binary alloy with curved solid–liquid interface. 2. Large departures from planarity,” J. Crystal Growth 54 (1981), 167. 8. A.M. Balint, M.M. Mihailovici, D.G. Baltean, and St. Balint, “A modiﬁed Chang– Brown model for the determination of the dopant distribution in a Bridgman– Stockbarger semiconductor crystal growth system,” J. Crystal Growth 230 (2001), 195. 9. J.A. Burton, R.C. Prim, and W.P. Slichter, “The distribution of solute in crystals grown from melt. 1. Theoretical,” J. Chem. Phys. 21 (1953), 1987. 10. W.A. Tiller, K.A. Jackson, J.W. Rutter, et al., “The redistribution of solute atoms during the solidiﬁcation of metals,” Acta Metall. Mater. 1(4) (1953), 428. 11. A.F. Witt, H.C. Gatos, M. Lichtensteiger, et al., “Crystal growth and steadystate segregation under zerogravityINSB,” J. Electrochem. Soc. 122(2) (1975), 276. 12. P.R. Grifﬁn and S. Motakef, “Inﬂuence of nonsteady gravity on natural convection during microgravity solidiﬁcation of semiconductors. I. Time scale analysis,” Appl. Microgravity Tech. 2(3) (1989), 121. 13. P.R. Grifﬁn and S. Motakef, “Inﬂuence of nonsteady gravity on natural convection during microgravity solidiﬁcation of semiconductors. II. Implications for crystal growth experiments,” Appl. Microgravity Tech. 2(3) (1989), 128. 14. H.M. Ettouney and R.A. Brown, “Effect of heat transfer on melt solid interface shape and solute segregation in edgedeﬁned ﬁlmfed growth — ﬁnite element analysis,” J. Crystal Growth 58 (1982), 313. 15. J.P. Kalejs, L.Y. Chin, and F.M. Carlson, “Interface shape studies for silicon ribbon growth by the EPG technique. 1. Transport phenomena modeling,” J. Crystal Growth 61 (1983), 473. 16. M.M. Mihailovici, A.M. Balint, and St. Balint, “The axial and radial macrosegregation due to the thermoconvection, the decrease of the melt in the ampoule and the effect of the precrystallization zone in the semiconductor crystals grown in a Bridgman– Stockbarger system in a low gravity environment,” J. Crystal Growth 237–239 (2002), 1752.
© 2004 by CRC Press LLC
13 Identiﬁcation of Stiffness Matrices of Structural and Mechanical Systems from Modal Data Firdaus E. Udwadia Aerospace and Mechanical Engineering, Civil Engineering, Mathematics, and Information and Operations Management, University of Southern California, Los Angeles, CA
In this chapter we present a simple method for the identiﬁcation of stiffness matrices of structural and mechanical systems from information about some of their natural frequencies and corresponding mode shapes of vibration. The method is computationally efﬁcient and is shown to perform well in the presence of measurement errors in the mode shapes of vibration. The method is applied to the identiﬁcation of the stiffness distribution along the height of a simple vibrating structure. An example illustrating the method’s efﬁcacy in structural damage detection is also given. The efﬁciency and accuracy with which the method yields estimates of the system’s stiffness make it worthy of further exploration for damage detection.
INTRODUCTION Modal testing of structures is an extensive ﬁeld in civil and mechanical engineering. It is generally used to understand/predict the dynamic behavior of a structure when subjected to lowamplitude vibrations. Often modal information is also used to identify/estimate the structural parameters of a system, under the assumption that it has classical normal modes of vibration [1]. Such identiﬁcation leads to improved mathematical models that can be used in either predicting and/or controlling structural response to dynamic excitations. Several different approaches to the parameter identiﬁcation problem have appeared in the literature [2–10]. One approach is the socalled model updating method. Here a suitable analytical model of a structural system is developed using the equations of motion, and its numerical representation is obtained. Validation of the numerical model through modal testing is then sought. Such tests usually provide some of the frequencies of vibration (usually the lower frequencies) and the corresponding mode shapes. When these frequencies and mode shapes obtained from modal testing are compared with those obtained from the numerical model, they generally do not agree with one another. Discrepancies between the results from experimental testing and theoretical modeling arise due to a variety of reasons: simpliﬁcations used in developing the analytical model, uncertainties in things like material properties and boundary conditions in the analytical model, and experimental errors during modal testing. The problem of updating a numerical model so that it is as much as possible
© 2004 by CRC Press LLC
in conformity with experimental modal test data is referred to as the updating problem, and over the years it has received considerable attention. In this paper we investigate a direct approach to structural identiﬁcation through the use of modal test data. No a priori estimates are used. It should be pointed out that such experimental test data is seldom “complete,” i.e., all the mode shapes of vibration and the corresponding natural frequencies are seldom available, for there is a practical limit to the range of frequencies that a structural or mechanical system can be tested for. Hence the idea is to obtain suitable models through the use of incomplete information, i.e., information on only a limited number of mode shapes and frequencies of vibration. We shall illustrate the method assuming that normal classical modes exist.
SYSTEM MODEL Consider a structural system modeled by the linear differential equation ˆ = 0, Mx ¨ + Cˆ x˙ + Kx
(1)
where, x is an n by 1 vector, and M is the n by n symmetric positivedeﬁnite mass matrix, ˆ is the symmetric stiffness matrix, and Cˆ is the damping matrix. We shall assume that K the elements of the mass matrix, M , are sufﬁciently well known, and that the system is classically damped. We could then rewrite Eq. (1) as y¨ + C y˙ + Ky = 0,
(2)
ˆ −1/2 , and C = M −1/2 CM ˆ −1/2 . where K = M −1/2 KM Our intention is to investigate the identiﬁcation of the stiffness matrix from a knowledge of the modal data corresponding to Eq. (2), i.e., from the eigenvectors and eigenvalues of the matrix K. We note that the eigenvectors ϕˆi of ˆ ϕˆi = λi M ϕˆi K
(3)
Kϕi = λi ϕi
(4)
are related to the eigenvectors ϕi of
by the relation ϕˆi = M −1/2 ϕi , for the eigenvalue λi [1]. We shall assume that λi λi+1 , i = 1, 2, ..., (n − 1). Often from the analytical model, the “structure” of the matrix K is known. Let us say that the matrix K is a function of the parameter svector k := [k1 k2 ...ks ]T . Usually, because of the limited connectivity between the subassemblages of a structure, the number of parameters, s, that are required to be identiﬁed in the n by n symmetric matrix K is much less than n(n + 1)/2. We assume that each element of the matrix K is a linear combination of the parameters ki , so that each of the n equations in the equation set Kϕi = λi ϕi is linear in these parameters. Hence, each row of Eq. (4) is linear in both the parameters ki and in the n components of the eigenvector ϕi . One can then rewrite equation (4) as Φi k = λ i ϕi ,
© 2004 by CRC Press LLC
(5)
where the elements of the n by s matrix Φi are linear functions of the components of the nvector ϕi . In what follows, we shall denote the jth component of the nvector ϕi by ϕji . Modal test data provides the measured frequencies and the measured mode shapes m of vibration, yielding λm i , i = 1, 2, ..., r, and the corresponding eigenvectors ϕi , i = 1, 2, ..., r, where some r n. We denote experimental data by the superscript m. Were Eq. (5) to be true for the measured modal data, we would then obtain m m ˜ Φm i k = λ i ϕi ,
i = 1, 2, ..., r,
(6)
giving us an estimate k˜ of the parameter svector k that contains the s parameters ki , i = 1, 2, ..., s, of which the stiffness matrix K is a function. Equation (6) can be expressed in a more compact form, by stacking the measured data from each of the r modes, as B k˜r :=
Φm 1 Φm 2 . . Φm r
k˜r =
m λm 1 ϕ1 m λm ϕ 2 2 . . m λm r ϕr
= br ,
(7)
where the matrix B is rn by s, and the vector b is an rnvector. The minimumnorm leastsquares solution to this system of Eq. (7) is simply given by k˜r = B + br , (8) where B + stands for the Penrose generalized inverse of the matrix B; see Ref. [11] for details on the properties of generalized inverses of matrices, and their computation. The subscripts “r” indicate that measured modal data from r modes is “stacked” together in Eq. (7). To illustrate the above equations, we consider a building structure modeled by a simple ndegreeoffreedom system (see Fig. 1) subjected to horizontal base motion. The mass matrix M is taken to be the identity matrix, and the structure is assumed to be lightly damped. Though effects like soil–structure interaction may be important in understanding the structural dynamics of such building structures, to illustrate our ideas we shall ignore soilstructure interaction and further assume that the structure is resting on a rigid base. We thus focus primarily on our ability to estimate the constant stiffness matrix of the structure from modal data. Furthermore, since numerous mechanical systems are often modeled by such a “chain” of springs and masses, and hence we use this as a prototypical system. ˆ = K is tridiagonal The equation of motion is described by Eq. (1), and the matrix K and has the form k1 + k2 −k2 −k2 k2 + k3 −k3 . . . . (9) K= . . . −kn kn One would like to estimate the parameter nvector k = [k1 k2 ... kn ]T from modal test data.
© 2004 by CRC Press LLC
Figure 1 Model of a simple building structure.
The matrix Φi in Eq. (5) now becomes the upper triangular banded matrix
ϕ1i
Φi = O
−ϕi2,1 ϕi2,1 −ϕi3,2 . . .
O .
ϕin−1,n−2
−ϕn.n−1 i ϕn.n−1 i
,
(10)
= (ϕpi − ϕqi ). In this special case one can explicitly obtain the inverse of the where ϕp,q i matrix Φi and solve Eq. (5) to yield k = λi Φ−1 i ϕi ,
© 2004 by CRC Press LLC
where the upper triangular matrix Φ−1 i is given by 1/ϕ1i 1/ϕ1i 1/ϕ1i . . 2,1 2,1 1/ϕ 1/ϕ . i i . . −1 Φi = .
. . . 1 / ϕn−1,n−2 i
1/ϕ1i 1/ϕi2,1 . . 1/ϕin−1.n−2 1 / ϕn.n−1 i
.
(11)
Were the eigenvalue λi obtained from measurements of the frequency of vibration of the ith normal mode, and the eigenvector ϕi obtained from experimental measurements of the ith mode shape of vibration of the building structure, we would obtain an estimate of the parameter nvector k as m −1 m k˜ = λm ϕi , (12) i [Φi ] where we have replaced all the ϕji ’s in (11) by their measured values, (ϕji )m . While this estimate may be adequate when the mode shapes of vibration are assumed to be noisefree (a situation that never arises in practice), as shown below, it quickly deteriorates in the presence of measurement noise. Were frequency and mode shape data obtained for modes i = 1, 2, ..., r, we would form the matrix B as shown in Eq. (7) and obtain the estimate of the parameters as indicated in Eq. (8). We illustrate this using the following numerical example.
A. NUMERICAL EXAMPLE Consider a building structure modeled by a 9degreeoffreedom system (n = 9) whose true (exact) story stiffnesses are taken to be (we assume the numerical data to be in consistent units): k1 = 2000; k2 = 2050; k3 = 2090; k4 = 2050; k5 = 1950; k6 = 1990; k7 = 1920; k8 = 1900; and, k9 = 1900. The stiffness parameters are purposely chosen to have a small range of numerical values so that the power of the identiﬁcation scheme in discriminating between these close stiffness values is assessed. The fundamental frequencies of vibration corresponding to the 9 by 9 stiffness matrix shown in (9) are then: 1.1810, 3.4637, 5.6707, 7.7577, 9.5685, 11.1865, 12.4205, 13.3962, 14.0528 cycles/sec. We begin by assuming that the experimentally determined frequencies and the modem shapes of vibration from modal testing are accurate (noiseless). Thus, λi = λm i , ϕ i = ϕi , i = 1, 2, ..., r, where r is the total number of resonant frequencies and modeshapes obtained from the test data. Figure 2(a) shows the percentage error in estimates of the parameter vector k whose components are the story stiffnesses. The error in the estimate, k˜i , of the story stiffness ki is deﬁned as (k˜i − ki )/ki . It is assumed that only the lowest four modes of vibration are measured, along with the lowest four modal frequencies. The ﬁgure shows that the stiffness distribution can be very accurately determined using Eq. (12) above, for i equal to 1 to 4. However, the addition of measurement noise alters the situation considerably.
© 2004 by CRC Press LLC
Figure 2 (a) Identiﬁcation (ID) with noisefree data; + shows ID from ﬁrst mode; x shows ID from second mode; * shows ID from third mode; o shows ID from fourth mode.
Figure 2 (b) Identiﬁcation (ID) with 5% noise in modal amplitude data; + shows ID from ﬁrst mode; x shows ID from second mode; * shows ID from third mode; o shows ID from fourth mode.
During modal testing it is customary to assume that the frequencies of vibration are accurately determined, and that it is in the determination of the amplitudes of the mode shapes that the experimental errors arise. This assumption is by and large valid because the frequency of shakers, even at resonance, can be quite accurately controlled. Accordingly, the data is simulated by adding noise to each modal amplitude, and the “measured” component “j” of the ithmode shape is taken to be (ϕji )m = ϕji (1 + αnoise ξ),
© 2004 by CRC Press LLC
(13)
Figure 2 (c) Identiﬁcation (ID) with 5% noise in modal amplitude data using generalized inverses; + shows ID using r = 1; x shows ID using r = 2; * shows ID using r = 3; o shows ID using r = 4.
where ξ is a uniformly distributed random number between −1 and +1. Figure 2(b) shows results with αnoise = 5% indicating the dramatic increase in the error when Eq. (12) is used in estimating k˜ when using noisy mode shape data. We next illustrate the improvement that is created by using Eq. (8) where the estimate is obtained by using the generalized inverse of the matrix B + for values of r ranging from 1 to 4. Figure 2(c) shows the progressive improvement in the estimates of the stiffness with the addition of information about each successive mode of vibration. The simultaneous use of data from all 4 modes (r = 4) shows that there is a substantial reduction in the percentage error in the estimation of the stiffness. The maximum percentage error is now less than 10%.
Iterative Improvements of the Stiffness Estimates in the Presence of Measurement Noise Having obtained the minimum length leastsquares estimate of the parameters by using the generalized inverse of the matrix B, we now attempt to improve this estimate when the mode shape information is corrupted by measurement noise, i.e., when we use the measured modal data, ϕm i , i = 1, 2, ..., r. We assume, as is customary, that the measurement errors in determining the r resonant frequencies are negligible when compared to those incurred in the measurement of the mode shapes; hence λi = λm i , i = 1, 2, ..., r. For convenience we will denote the minimum length leastsquares estimate we have obtained so far by k˜(0) . ˜ (1) , of This estimate, which is obtained from Eq. (8), gives us, in turn, an estimate, K (1) ˜ the matrix K. Hence, a solution of the eigenvalue problem, K ψj = µj ψj , provides: (1) an estimate of all the eigenvalues (frequencies), µj , j = 1, 2, ..., n of vibration (estimates of the eigenvalues of K) and, (2) an estimate of all the modeshapes, ψj , j = 1, 2, ..., n ˜ (1) is symmetric, ψj are (or can be chosen to (estimates of the eigenvectors of K). Since K
© 2004 by CRC Press LLC
be) orthogonal to one another, and can be normalized to have unit length. In summary, through our estimate of the stiffness parameter vector k˜(0) at this point we have obtained estimates, µj , of the actual eigenvalues, λj , of K as well as estimates, ψj , of the actual mode shapes, ϕi . We next use this information to iteratively update the estimate k˜(0) . ˜ (1) form a basis set in Rn , the “measured” noisecorrupted Since the eigenvectors of K m mode shape ϕi can be expanded in terms of the estimated mode shapes ψj , j = 1, 2, ..., n. We then have n m ϕi = δji ψj , i = 1, 2, ..., r, (14) j=1
where
δji
=
ψjT ϕm i .
We then obtain
˜ (1) ϕm ˜ (1) K i =K
n
δi
δji ψj =
j=1
j
˜ (1) ψj = K
j=1
n
δji µj ψj ,
i = 1, 2, ..., r.
(15)
j=1
˜ (1) ϕm as Φm k˜(0) , and Eq. (15) yields As before, we can now express K i i ˜(0) = Φm i k
n
δji µj ψj ,
i = 1, 2, ..., r.
(16)
j=1
However, the measured r lowest frequencies are assumed to be accurately known and we can use this information in Eq. (16) by replacing the µj ’s by the known (i.e., accurately measured) eigenvalues λj for j = 1, 2, ..., r. This additional information when injected into Eq. (16) provides us with the opportunity to update our estimate of the stiffness parameter vector from k˜(0) to k˜(1) , and we obtain ˜(1) = Φm i k
r j=1
δji λj ψj +
n
δji µj ψj ,
i = 1, 2, ..., r.
(17)
j=r+1
This forms the basis of our iterative improvement of the stiffness parameter vector k˜(0) . We note that we have used: (1) our information from the r measured eigenvalues (frequencies of vibration) of the system, which we assume are accurate, (2) our best hereto available estimates, µj , of the remaining (n − r) unmeasured eigenvalues (frequencies of vibration) of the system, and (3) our best hereto available estimates, ψj , j = 1, 2, ..., n derived from the measured eigenvectors ϕm j , j = 1, 2, ..., r, which are corrupted by measurement noise. Equation (17) can be rewritten for convenience as i δ 1 λ1 δ2i λ2 . . m ˜ (1) i := Ψhi , i = 1, 2, ..., r. (18) Φi k = [ψ1 ψ2 ...ψn ] δr λr i δr+1 µ r+1 . . i δn µn
© 2004 by CRC Press LLC
˜ (1) . Lastly, since we have r Here Ψ is the n by n orthogonal matrix of eigenvectors of K measured mode shapes, we can, as before, stack the information again as m Φ1 Ψh1 Φm 2 (1) Ψh2 (1) k˜r = . = b(1) . (19) B k˜r := r , . . Φm Ψhr r so that we again obtain the minimumlength leastsquares solution for k˜r in Eq. (19) as (1)
k˜r(1) = B + b(1) r .
(20)
It should be pointed out that the matrix B in Eq. (19) is the same as it was in Eq. (7); it is obtained from the r measured mode shapes. (1) ˜ (2) of the stiffness matrix, which The estimate k˜r is next used to create an estimate K (2) ˜ in turn yields the new estimate kr , and the iteration continues until the improvement in (i) the vector k˜r is deemed to be negligible, or for a preﬁxed number of iterations. The algorithm that we have developed can be summarized as follows: m 1. Use the measured mode shape ϕm i , to obtain the n by s matrices Φi , i = 1, 2, ..., r.
2. Stack the r matrices Φm i ’s to obtain B. (0)
m 3. Stack the vectors λm i ϕi , i = 1, 2, ..., r, to obtain the vector br .
4. Calculate k˜r
(0)
(0)
= B + br [Eq. (8)].
5. for i = 1 to N iterations, do: ˜ (i) , its eigenvalues µj , and its eigenvectors ψj , j = 1, 2, ..., n; Obtain K Calculate δjp = ψjT ϕm p , j = 1, 2, ..., n; p = 1, 2, ..., r; Calculate the vectors Ψhp , [Eq. 18] using measured λm p , p = 1, 2, ..., r, and µp , p = r + 1, ..., n; (i) Stack the vectors Ψhp , p = 1, 2, . . . , r, to get br [Eq. 19]; (i) (i) Calculate k˜r = B + br ; (i) (i−1) If k˜r − k˜r < ∆, exit do loop; end do loop.
B. NUMERICAL EXAMPLE (CONTINUED) To illustrate the iterative improvement of the estimate of the stiffness parameter vector, we show in Fig. 2(d) the results from the same modal data shown in Figs. 2(b) and 2(c) by using the same realization as before of the random noise that corrupts the mode shape data. The solid line shows the identiﬁcation results obtained after 1000 iterations, the dashed line is the result generated by direct use of the generalized inverse (i.e., Eq. (8)). We note in passing that negligible changes in the stiffness occur after 150 iterations. Since the only experimental information that we assume here is that each component of the measured mode shape (corresponding to each natural frequency of vibration) is
© 2004 by CRC Press LLC
Figure 2 (d) Improvement in estimation caused by iterative scheme.
corrupted with zero mean, uniformly distributed white noise, what is of relevance are the expected values of the stiffness estimate obtained and its standard deviation. To ﬁnd the expected value of the estimate, the simulation was carried out 1000 times with αnoise =5%, each time using a different set of randomly generated noise to corrupt the mode shape measurements. Figure 3(a) shows the expected value of the estimate of the stiffness at each story along with a 2sigma band shown by the dashed lines to indicate the variance in the expected estimate. The true (exact) stiffness is shown by the shaded line and the expected value of the stiffness is shown by the solid dark line. As seen from the plot, the identiﬁcation scheme appears to work remarkably well, and the expected estimate closely tracks the true stiffness. To get a feeling for the expected error in the estimation of the stiffness at each story, i.e., E{k˜ − k}, we show it in Fig. 3(b) along with its ±1sigma band. As seen from the ﬁgure, the estimation is quite accurate, and the expected error in the estimation of the stiffness is very small, the maximum of {mean error +1 sigma} being 5% at story number 2. Lastly, we show in Fig. 3(c) the expected value (with a sample size of 1000) of the estimate along with the ±1sigma bands when only the generalized inverse is used without the iterative improvement in the estimates discussed in this section. Again the shaded line is the true stiffness, the dark solid line is the expected value of the stiffness, and dotted lines indicate the mean ±1sigma band. Comparing Fig. 3(a) with 3(c) we observe that the iteration discussed in this section (Eqs. (17)–(20)) has a deﬁnite advantage, both in terms of the expected value of the stiffness and in reducing the size of the ±1sigma band. We observe that the estimates that we have obtained use information from only 4 modes of vibration of the structure. This is the usual situation, for complete information on all the modes and all the resonant frequencies of structures is seldom experimentally decipherable from measurement data. This is due to limitations in experimental testing. Also, higher modes “sample” smaller spatial domains and our models may be inadequate in representing these spatial heterogeneities.
© 2004 by CRC Press LLC
Figure 3 (a) Expected value of the identiﬁed estimate (dark line), true stiffness (shaded line), and the expected estimate ±standard deviation (dashed lines).
Figure 3 (b) Expected value of the error in the stiffness at each story and the ±1sigma band around it (dashed lines).
We next show the inﬂuence of using more modal information on the identiﬁcation results. We next use measured modal data from 6 modes instead of 4. Clearly, the additional information should improve the estimates of the stiffness distribution and reduce the uncertainty in the expected value of the estimates that are obtained. However, noise in the measurements of the modal amplitudes usually increases with increasing natural frequency, and so to simulate the noisecorrupted mode shape amplitude data we shall use αnoise =10% in Eq. (13), instead of 5% as we did when we used only 4 modes for the identiﬁcation. We exhibit our results in Fig. 4.
© 2004 by CRC Press LLC
Figure 3 (c) Results with no iterative improvement. Expected value of stiffness at each story (dark line), true stiffness (shaded line), expected value ±1sigma band (dashed line).
We observe that even with the sizably larger measurement noise of 10% in the mode shape measurements, the estimates of the 1sigma band in the percentage error in the stiffness estimates cover only ±4%. C. NUMERICAL EXAMPLE OF DAMAGE DETECTION We now assume that the structural (or mechanical) system has, perhaps after being subjected to horizontal ground shaking, suffered deterioration along its height; and the stiffness at the third story has been reduced from its previous value by about 30%. Our aim is to estimate the new stiffness distribution, and investigate how well this drop in stiffness can be captured by modal testing, using again only 4 modes of vibration. We assume that the true stiffness of the damaged structure is given by k1 = 1950; k2 = 1900; k3 = 1460; k4 = 1910; k5 = 1910; k6 = 1930; k7 = 1920; k8 = 1900; and, k9 = 1900. Figure 5(a) shows that the weakened structure can be well identiﬁed, and the location of the drop in the stiffness is quite evident, and accurately determined using noisecorrupted modal information from only 4 modes. The solid line is the true stiffness, the dashed line is the stiffness after 1000 iterations, and the dashdot line is the result of Eq. (8), with no iterative improvement. Figure 5(b) shows the percentage error in estimating the stiffness with (solid line) and without (dashed line) iterative improvements. With the iterative improvements, the percentage error in identiﬁcation is seen to be less than 4%. The expected value of the stiffness estimates are again obtained using 1000 realizations of noisecorrupted “measured” mode shapes. The solid line in Fig. 6(a) shows the expected value of the stiffness estimates at each story, the shaded solid line shows the true
© 2004 by CRC Press LLC
Figure 4 Identiﬁcation using 6 modes. Figure 4(a): Expected value of the stiffness (dark line), true stiffness (shaded line), and the ±1sigma band using the iterative improvement; Figure 4(b): Expected error in estimation of stiffness and ±1sigma error band using the iterative improvement; Figure 4(c): Expected percentage error and ±1sigma error band using iterative improvements; Figure 4(d): Expected value of the stiffness (dark line), true stiffness (shaded line), and the ±1sigma band without iterative improvement.
stiffnesses, and the dashed lines show the ±1sigma bands. Figure 6(b) shows the expected error in the stiffness estimates, and the ±1sigma error bands. We see that the identiﬁcation results are rather accurate and the ±1sigma bands are far smaller than the drop in stiffness at the 3rdstory level, indicating that accurate damage detection can be accomplished in an efﬁcient manner using the simple system identiﬁcation approach describe here. Figure 6(c) shows the expected percentage error in estimation of the stiffness. The ±1sigma bands show that changes of more than ±5% in the stiffness parameters can be rapidly identiﬁed with reasonable reliability. Lastly, Fig. 6(d) shows the expected values of the identiﬁcation results without any iterative improvements. We note that though not as good as those obtained after iteration, these results too are considerably superior to those that are obtainable by other system identiﬁcation techniques when using noisecorrupted data.
CONCLUSIONS AND DISCUSSION We have proposed here a simple and efﬁcient way of estimating the stiffness distribution in a structure from incomplete modal information. It relies on the Penrose generalized
© 2004 by CRC Press LLC
Figure 5 (a): Estimate of stiffness using one realization of noisy modal measurements; (b): Percentage error in stiffness estimates.
Figure 6 A case study of damage detection.
inverse of a matrix, which can be computed rapidly and efﬁciently. A subsequent iterative improvement of the estimates is then carried out. Our numerical computations show that this iterative improvement reduces the variance of the estimates and brings their expected values closer to the true (exact) values. The method appears to work well for estimating
© 2004 by CRC Press LLC
the stiffness distribution of simple building structures. We ﬁnd that the method gives good results in the presence of noisecorrupted mode shape data, and is capable of tracking small changes in the stiffness parameters. Its accuracy in the presence of noise appears to surpass several other more sophisticated identiﬁcation methods in common use (e.g., compared to results in [4, 8, 10]). This makes the method perhaps valuable for rapid damage detection in structures. Thus, we show that it can be used for realtime, online, accurate system identiﬁcation and damage detection. Though in this chapter we have used information from the lowest r mode shapes and frequencies of vibration, we could have chosen any r of the measured mode shapes and r corresponding frequencies of vibration for the identiﬁcation. The identiﬁcation method presented here can be extended to this situation in a simple and obvious manner. It would be useful to compare the method proposed here with those commonly in use, especially the socalled updating methods. We note that in the method proposed herein: 1. The “structure” of the stiffness matrix is assumed to be known, and in addition, the way in which the unknown stiffness parameters enter each element of the stiffness matrix are assumed to be known. 2. The elements of the stiffness matrix are assumed to be linear functions of the parameters to be estimated; this is not as much of a restriction as may appear at ﬁrst sight, for the “assemblage” of the stiffness matrix using, say, FEM models provides this sort of information in a more or less natural way. 3. We use generalized inverses to obtain the actual stiffness estimates, and no a priori stiffness estimates are required as in all the stiffness updating methods. 4. Unlike the usual updating methods that try to estimate each element of the stiffness matrix, thereby increasing the unknowns (usually to n(n + 1)/2), the method proposed here uses the knowledge of the way in which the unknown parameters enter the stiffness matrix; this can reduce the number of unknowns signiﬁcantly. 5. Our approach averts “connectivity” problems that require being addressed by the use of additional constraints when using some of the usual updating methods that employ minimization techniques. 6. Even when there is no measurement noise, the accuracy of the available updating methods is known to rely on the proper choice of “weighting matrices,” and there appears to be no systematic method of choosing these matrices. 7. Updating methods are highly susceptible to measurement errors; we have shown that the method developed here can provide signiﬁcant, highquality information about the stiffness parameters even when the measurements of the mode shapes are incomplete, and corrupted by noise. 8. Unlike most updating methods, no attempt is made to orthonormalize the measured mode shapes. 9. The iterative procedure provides a signiﬁcant advantage both in terms of reducing the variance of the stiffness estimates and in getting the expected estimates closer to the true values. This is achieved by expanding, at each iteration, the noisecorrupted
© 2004 by CRC Press LLC
mode shapes in terms of the best available estimates of the mode shapes and eigenvalues, as well as the measured frequencies of vibration. Despite the promise that this method shows, especially for rapid damage detection, and for providing a priori estimates that can then be used with other identiﬁcation methods, the results presented herein should be considered as being preliminary. The way the method behaves when the number of unknown stiffness parameters becomes large, and when the structure is nonclassically damped still remains to be examined. Some of these issues will be presented in a future communication.
REFERENCES 1. Caughey T.K. and O’Kelley M.E.J., “General theory of vibration of damped linear dynamics systems,” Dynamics Laboratory Report, California Institute of Technology, June 1963. 2. Baruch M. and Bar Itzhak Y., “Optimal weighted orthogonalization of measured modes,” AIAA Journ., 16 (4), pp. 346–351, 1978. 3. Kabe A.M., “Stiffness matrix adjustment using mode data,” AIAA Journ., 23 (9), pp. 1431–1436, 1985. 4. Kalaba R. and Udwadia F., “An associative memory approach to the rapid identiﬁcation of nonlinear structural and mechanical systems,” Journ. Optim. Theory Appl., 76(2), pp. 207–223, 1993. 5. Kenigsbuch R. and Halevi Y., “Generalized reference basis model updating in structural dynamics,” ASME Design Engineering Conference, Sacramento, CA, 1997. 6. Mottershead J.E. and Friswell M.I., “Model updating in structural dynamics: a survey,” Journ. Sound Vibration, 16(2), pp. 347–375, 1993. 7. Udwadia F. and Ghodsi A., “Parameter identiﬁcation problems in structural and geotechnical engineering,” Journ. Eng. Mech., 110, pp. 1409–1432, 1984. 8. Udwadia F. and Proskurowski, W., “A memorymatrixbased identiﬁcation methodology for structural and mechanical systems,” Earthquake Eng. Struct. Dyn., 27, pp. 1465–1481, 1998. 9. Wei F.S., “Structural dynamic model modiﬁcation using vibration data,” IMAC, 7, pp. 562–567, 1989. 10. Koh C.G., Hong B., and Liaw C.Y., “Parameter identiﬁcation of large structural systems in the time domain,” Journ. Struct. Eng., 126(8), pp. 957–963, 2000. 11. Udwadia F., and Kalaba R., Analytical Dynamics, Chapter 2, Cambridge University Press, 1996.
© 2004 by CRC Press LLC
14 A Survey of Applications in Structural–Acoustic Optimization for Passive Noise Control with a Focus on Optimal Shell Contour Steffen Marburg Institut f¨ur Festk¨orpermechanik Technische Universit¨at, Dresden, Germany
This chapter reviews applications in structural–acoustic optimization. The author concentrates on problems in passive noise control. Several dozen papers on structural–acoustic optimization were found. An increasing community of authors indicates optimization of structures whereas a closeup conﬁrms that just a few different variants were investigated. Herein, structural–acoustic optimization is understood as an autonomous (mostly iterative) search to minimize a certain objective function. In a brief survey of the literature, the articles are categorized into academic and more realistic applications. A few of them are discussed in more detail. These three applications have in common that a shell contour is optimized to obtain certain acoustic characteristics. The ﬁrst application is the design of carillon bells, the second example reviews the design of a loudspeaker diaphragm, and the third application consists of the design of sedan body panels. It is emphasized that shell contour optimization accounts for an efﬁcient tool to inﬂuence local structural stiffness without increasing the mass.
1
INTRODUCTION
The decrease of noise emission of machines, vehicles, domestic appliances, etc. will probably become one of the major challenges of engineering in the 21st century. Development of computers enabled a wide range of applications of numerical methods like the ﬁnite element method and the boundary element method during the past three decades. Further and further improvement of hardware systems and simulation methods encouraged researchers to employ methods of numerical optimization for structural–acoustic optimization over the past decade. Herein, we focus on the ﬁeld of structural–acoustic optimization for passive noise control. Structural–acoustic optimization is understood as the improvement of certain acoustic characteristics of a structure that emits sound or noise by mathematically controlled modiﬁcation of the structure. That a multiﬁeld problem has to be solved is the reason that the subject is also considered as a problem of multidisciplinary optimization.
© 2004 by CRC Press LLC
Solution of multiﬁeld problems usually appears as a procedure that is computationally expensive. If structural–acoustic simulation covers a wide frequency range, the multiﬁeld problem has to be solved for a certain number of discrete frequencies. Optimization will then require an outer loop that contains the entire solution of the multifrequency multiﬁeld problem in a single step. This, however, considers analysis only. A number of additional technical problems is encountered if this type of multidisciplinary optimization is actually implemented. These difﬁculties may explain why comparatively few publications exist in the ﬁeld of structural–acoustic optimization. We have not been able to ﬁnd any paper on this subject for time–domain problems. A survey on analysis techniques in structural–acoustic optimization is given in the paper by Christensen et al. [1]. Several dozen papers have been published on the subject since 1990. However we found that often paper titles promise structural and/or acoustic optimization, but close examination reveals that only a number of different variants are compared while optimization in the sense of an (autonomous) iterative minimum search of an objective function (or cost function) is not reported, cf. Refs. [2–5]. Moreover, the ﬁeld of structural–acoustic optimization is also important for problems in active noise control. This subject is excluded from our reﬂections for this chapter. As the acoustic characteristics to be inﬂuenced, we can name the sound power, sound pressure level, directivity patterns, or still other measures. In practical applications, a number of very different objectives must be considered in the design process. Cases of multicriteria optimization are familiar because it often happens that two or more conﬂicting design criteria should be taken into account simultaneously. Then, we have to make a tradeoff among them. Structural modiﬁcations involve almost all possible changes of existing construction. To identify a few of them, modiﬁcations include plate and shell thicknesses, material data, size and location of added masses, damping characteristics, and shell curvature. In general, modiﬁcations must satisfy design and functionality requirements. Any further variation is subject to the engineer’s experience and imagination. Many types of modiﬁcations necessitate certain constraints. These constraints typically comprise mass conditions but others like particular cost functions or even different deﬁnite conditions must be regarded. Optimization involves a number of calculations of the objective function. Depending on the number of variables, the condition of the optimization problem, optimization method, and other features of the particular problem, this calculation is repeated between 10 and 106 times. The upper limit is an estimation and may be exceeded in some cases. This, however, demands efﬁcient analysis techniques for calculation of the objective function. In most applications, the structural models are analyzed using ﬁnite element models. The variety of noise emission evaluation is greater. So, both ﬁnite element and boundary elementbased methods are used. Additionally, several methods based on the Rayleigh integral have been reported. The optimization method will essentially inﬂuence success of the entire procedure. With respect to a high degree of nonlinearity of the objective function in terms of design parameters, only a numerical treatment like an iterative minimum search accomplishes the necessities of this problem. Deterministic and stochastic algorithms are appropriate for these purposes. A number of the problems discussed above has been addressed in the books by Koopmann and Fahnline [6] and Kollmann [7]. One problem that was very little focused in these volumes is the technical completion for complex structural models. This includes mesh
© 2004 by CRC Press LLC
coupling between structure and ﬂuid, parameterization, and data handling. For appreciation of optimization results, this point should not be neglected since it often requires major efforts of manual work. There are software tools for applications in structural–acoustic optimization. In most cases, however, one needs to combine several different tools or a particular code must be developed because generalpurpose applications do not ﬁt together with speciﬁc desiderata of the particular application. In this chapter we will start with a brief review of the literature in the ﬁeld of structural–acoustic optimization for passive noise control. Then, we will focus on optimization of shell curvature. The author has found three applications. First, there are some papers about bell design [8–13], second, one paper on the geometry optimization of a loudspeaker diaphragm was found [14], and, third, a number of papers were written on the optimization of metal sheets for sedan body noise, vibration, and harshness problems in the lowfrequency range [15–22]. 2
BRIEF REVIEW OF CONTRIBUTIONS IN STRUCTURAL–ACOUSTIC OPTIMIZATION
The relatively small number of contributions to this ﬁeld can be categorized into academic examples and more realistic applications. It is the advantage of simple examples that one can experience and investigate many effects of a method. In some cases, however, particular adjustments become necessary if problems and models become more complex. We start with the academic examples. These can be grouped into problems of beams, plates, shells, ducts, and boxes. Few papers considered bafﬂed beams, cf. Naghshineh et al. [23], Jog [24], and Fahnline and Koopmann [6]. The beam structure was likely chosen for simplicity. Some interesting effects could be demonstrated by investigating this example. Moreover, beam structures are suitable to test a method. Plates have been investigated quite often [6, 14, 25–37]. However, the category of academic examples seems to be conﬁned to either rectangular or circular ones. It was possible to ﬁnd one article presenting the example of an engine cover plate [25] that will be discussed with the realistic applications. Choice of circular or rectangular plates could be critical since these structures contain symmetry properties that may have unfavorable effects on sound power radiation. For all of the plate examples, sound radiation problems were investigated. Usually, the sound power was to be minimized. Few examples on circular plates considered different targets. Fahnline and Koopmann [28] sought a circular plate of maximum power–conversion efﬁciency. Christensen et al. [14, 26] optimized the directivity pattern of sound radiation from a circular plate. Tinnsten [36] chose a circular plate to construct an example for experimental veriﬁcation of structural–acoustic optimization. He measured the sound intensity above the plate. Wodtke and Lamancusa [37] optimized a circular bafﬂed plate with an unconstrained damping layer for minimum sound power emission. In most of these examples on circular plates, the structure was excited by a point force [14, 26, 36, 37] or a ring load [37]. One example is based on modal properties and, therefore, independent of the excitation [28]. Rectangular plates were excited by point loads [6, 24, 25, 27, 30, 32–35] or by distributed loads [31, 34]. Lamancusa [30] optimized modal properties to achieve an optimal solution independent of a particular excitation. Plates were mostly clamped. Few examples of different boundary conditions are reported. Cases of clamped and free boundary conditions are compared in [36, 37].
© 2004 by CRC Press LLC
Additionally, elastic support of edges is included in [37]. In [32], the example comprises a plate that is clamped at two edges while the other two edges are free. Ratle and Berry [34] investigated a simply supported rectangular plate. In some of these examples, the authors reported and compared both broadband and singlefrequency excitation [25, 33, 35]. Only one driving frequency or mode is considered (for the plate example) in [14, 26, 27]. Two or more frequencies or modes were included by [32, 34, 36, 37]. Broadband sound radiation including many modes was minimized by Lamancusa and Eschenauer [30, 31] and by Hibinger [29]. Plates seem to be well suited for experimental veriﬁcation as supplied in [6, 29, 32, 33, 36]. In addition to the axisymmetric examples discussed above, symmetry of vibration modes was assumed in some of the examples, cf. [27, 30, 31, 33, 35]. Structural–acoustic optimization of shell structures without an apparent application was addressed in [26, 27, 38, 39]. Hambric’s [38, 40] example was a cylindrical shell with hemispherical end caps. In one test case the shell was simply supported, in another one these supports were omitted. The structure was excited by a ring load. Singlefrequency and broadband excitation cases were investigated for optimization. For this shell structure submerged in water, a full structure–ﬂuid interaction became necessary as well as for the thin conical membrane that was considered by Christensen et al. [26]. This axisymmetric structure clamped at its edge was excited by a point load in the center. The directivity pattern was optimized for a single frequency. Constans et al. [27,39] chose the example of a halfcylindrical shell clamped along the two lower straight edges. The emitted sound power due to a driving force was minimized for the lowest ﬁve modes. Although structural and loading conﬁguration were symmetric the optimal solution is asymmetric. An experimental veriﬁcation was provided [39]. The simple example of a duct was investigated by Shii et al. The paper [41] is a reduced version of the Japanese paper [42]. The authors stated that the onedimensional duct system with a coupled spring–masssystem should represent the rear window of a vehicle but complexity of vehicle structures is much higher than in their example. Sound pressure at a single point is calculated for a narrowfrequency band. Boxes have been addressed for both, internal [43–46] and external [7, 29, 47] noise problems. Starting with the internal problem, a simple box model was investigated by Pal et al. [44–46]. A fully coupled structure–ﬂuid interaction was included. The box was excited by a point load and analyzed for two discrete frequencies. The paper by Marburg et al. [43] addressed modeling techniques of a steel box structure that is reinforced by two frames. Although a frequency range containing about sixty modes was chosen, a reliable structural model was created for a range of about twenty modes. The rootmeansquare sound pressure level over the frequency range was minimized for one test point. The paper by Tinnsten et al. [47] addressed sound radiation. A point load was applied for one or two discrete frequencies. The authors reported an uncoupled analysis. The box example of Hibinger [7, 29] was primarily focused on radiation problems. Sound radiation of this box was not computed, though. Minimization of the sum of meansquare velocities of the box surface accounted for the optimization target. A point load was applied. The frequency range contained the ﬁrst 15 modes. The more realistic applications of structural–acoustic optimization can be grouped into problems of sandwich panels, loudspeaker diaphragm, bells, a cylinder representing an aircraft fuselage, and vehicle noise problems. For the latter, we can further distinguish between body noise problems and others.
© 2004 by CRC Press LLC
Dym et al. [48, 49] investigated optimization of sandwich panels. It was their target to maximize transmission loss of these plates. A large frequency range was analyzed and several, at that time new, ﬁndings were reported. Although they belong to this category of realistic applications we will discuss examples on the design of carillon bells and a loudspeaker diaphragm in the next subsection. A cylinder clamped at both ends accounted for a design problem similar to an aircraft fuselage in several papers by Cunefare, Engelstad et al. [50–52]. The structure was excited by a monopole source outside at a single frequency that coincided with an eigenfrequency of the original model. Optimization included thickness of shell rings [50], crosssection [51, 52] and location [52] of stiffening elements like stringers and frames. The sound pressure was evaluated at certain data recovery points inside the cylinder. There are a few papers on the application to vehicle acoustics not primarily considering body vibrations. All of them address design and support of the engine or components. An engine cover plate accounts for the practical application in the paper by Belegundu et al. [25]. The plate was clamped with respect to rotations and supported on springs of high stiffness. Excitation forces of different phase angle were applied. In the optimization process, sound radiation was minimized with respect to the ﬁrst three structural modes. The paper by Milstedt et al. [53] discussed the design and composition of an engine that consisted of several single substructures. Optimization was accomplished for the frequency range of 0 to 5000 Hz. Their model, however, is known as a concept model. Therefore, the objective of that study was to show the potential of structural optimization and to illustrate the concept. Real engine analysis requires ﬁner models as the authors stated. La Civita and Sestieri [54] described the design of an optimal engine mounting system. They computed the sound pressure level at the driver’s ear in the frequency range up to 120 Hz. The paper regards very different possibilities to construct engine supports, i.e., position and orientation, nonlinear spring characteristics, and, at the same time, they imposed different technical and geometrical constrains. Most of the applied papers in structural–acoustic optimization addressed sedan body construction. These examples can be categorized in terms of the design variables. One category consists of seeking an optimal distribution of piecewise constant shell thickness [55–58]. These applications usually require many variables. The paper of Choi et al. [57] indicates use of 36,000 parameters. Hermans and Brughmans [58] composed the vehicle body as an assembly of several components being investigated independently in a ﬁrst design stage. The other category utilizes modiﬁcation of shell geometry that will be discussed in more detail later in this article [15, 18, 19, 21, 22]. All of these examples on sedan interior noise were aimed at pure reduction. Future work will likely design sound in cars as suggested by Freymann [59]. The importance and necessity of optimization in the body design process to achieve a “design right the ﬁrst time” was emphasized in a paper on optimization of the noise, vibration, and harshness engineering process by Roesems [60]. There are a number of papers that do not actually address structural–acoustic optimization but are focused on structural–acoustic sensitivity analysis. Again, we start with the academic examples. Hahn and Ferri [61] demonstrated sensitivity analysis for a cylindrical shell submerged in water. Shepard and Cunefare [62] used a semiinﬁnite plate. Both papers mainly addressed modeling techniques including the question how coarse or ﬁne certain details should be modeled. A number of papers were found on the interior noise of simple boxes [56, 63–69] and just one on the radiation of a box [70]. The latter
© 2004 by CRC Press LLC
Figure 1 Geometrybased modeling: Shell contour modiﬁcation by shifting a single keypoint.
also presented an example of a cylindrical shell. While most of the papers that consider the interior noise of boxes concentrated on coupled mode sensitivities, the contribution of Bitsie and Bernhard [67] is devoted to an unusual topic in the ﬁeld, namely highfrequency noise problems. Applying energy ﬁnite element methods, they tried to reduce the energy level of structure and ﬂuid due to excitation of either structure or ﬂuid. Sensitivity analysis of vehicle body structures was accomplished in [56, 66, 71]. Some of the papers that consist of two parts discussed theoretical aspects in part one. The review paper of Christensen et al. [1] supplied detailed description of efﬁcient analysis in structural acoustics. The author of this article described theoretical aspects of the optimization of vehicle noise transfer functions in [17]. In another recent paper [20], concepts of parametric description of shell structures were investigated. Geometrybased models were compared with a newly developed modiﬁcation concept that enables a user to directly modify the shell geometry of a ﬁnite element mesh.
3 3.1
OPTIMAL SHELL GEOMETRY Building and Parameterizing a Shell Model
In structural–acoustic optimization, two major problems are usually encountered when designing shell structure geometry parametrically: First, variable geometry may require modiﬁcation of the mesh of the shell and, second, the volume where the sound is emitted changes its shape. For these two reasons, a ﬁniteelement model of the shell structure and a boundaryelement model of the ﬂuid were used for the three cases of shell contour optimization that will be reviewed in this section. Apparently, axisymmetric shell models can be used for analysis of a bell or a loudspeaker diaphragm. In these cases, the structural model to be parameterized and meshed reduces to a line. This line may be deﬁned by the position of a few keypoints. These keypoints are then connected by lines, i.e., by spline interpolation. Finite elements are created by dividing these lines into a certain number of line pieces accounting for axisymmetric shell elements. This technique of modeling is known as geometrybased modeling. In a paper describing optimization of shell geometry for vehicle interior noise problems in general [17], the author proposed geometrybased modeling as a reasonable strategy of modeling and
© 2004 by CRC Press LLC
Figure 2 Direct modiﬁcation of ﬁnite element mesh: Superposition of global and local modiﬁcation yields complete modiﬁcation of shell contour (mesh not visualized).
parameterizing the shell contour. Modiﬁcation of one parameter affects a group of nodes and elements being shifted into normal direction, cf. Fig. 1. An alternative technique to parameterize shell structures is based on direct modiﬁcation of the geometry of a ﬁnite element model. This concept is explained in the author’s paper [20]. It starts with deﬁnition of a modiﬁcation domain Ωm . Two types of modiﬁcation are allowed on this domain: Global and local modiﬁcation. We deﬁne both by socalled modiﬁcation functions. A global modiﬁcation function is a relatively complicated function being deﬁned on the entire modiﬁcation domain. A local modiﬁcation function is deﬁned on a local basis Ωl as part of the modiﬁcation domain. It is a very simple function but the local domain Ωl is variable in size, position, and orientation. One example of both, global and local modiﬁcation, is explained in Ref. [20]. Finally, global and local modiﬁcation functions are superimposed and, if necessary, normalized to complete modiﬁcation. A major advantage of the alternative concept is that a priori existing detailed structures can be parameterized by a few variables. A geometrybased model would require many keypoints to describe detailed shell geometries. Another advantage is the simpliﬁed use of discrete acoustic inﬂuence coefﬁcients [15, 17]. As they depend on this mesh topology, remeshing techniques require extraction of these coefﬁcients for every computation of the objective function. For further discussion, we refer to Ref. [20].
3.2
Design of Carillon Bells
A comprehensive research project on the optimization of carillon bells, in particular the design of a major third one, was initiated and kept up for about two decades by Lehr and Schoofs. An early threepart paper gives an excellent and thorough survey of that problem [8–10]. Structural–acoustic optimization was incorporated by Roozen–Kroon [11]. Although still using a modal approach, she had added modal damping coefﬁcients to the formulation. Another important step in bell design was completed by the dissertation of van Houten [12]. Schoofs and van Campen summarized these and other activities on bell design and optimization in a longer paper [13]. Finally, new generations of bells were designed by numeric optimization.
© 2004 by CRC Press LLC
The objective function that was used for bell design in [11, 12] and in other papers is formulated as F = wD
fD fDopt
2 Nf Nf 2 2 ¯ ¯ −1 + wf,i fi − fopt,i + wa,i ηai − ηaopt,i . i=1
i=1
(1) In this function, wD , wf,i and wa,i denote weighting factors for each single term. Nf is the number of structural eigenfrequencies being considered, f¯i and ηai represent the eigenfrequencies and corresponding acoustic damping values of the current design while f¯opt,i and ηaopt,i account for the target values. Acoustic damping has been deﬁned as a frequencyscaled relation between the radiated sound power and the kinetic energy of the structure. The values of fD and fDopt correspond to a relation between the diameter of the bell and the shell thickness. For that, it was deﬁned fD = Dl f1 , the product of the lip diameter and the ﬁrst eigenfrequency. Excluding this ﬁrst term in Eq. (1), this objective function could also be utilized for autonomous model updating purposes where experimental and simulated data of a modal analysis are compared. In addition to the geometry, variable thickness was optimized. In early papers on optimization of bells, optimization was performed by design of experiments [8,11]. This global approximation concept was used with respect to the computationally expensive analysis. In more recent papers, alternative approximation concepts have been favored, i.e., midrange or local approximation. A thorough discussion on approximation concepts and applications in bell design has been presented by van Houten [12]. This work, however, took advantage of the axisymmetric boundaryelement formulation for sound radiation developed by Kuijpers et al. [72].
3.3
Design of Loudspeaker Diaphragm
Christensen and Olhoff [14] optimized the geometry of the axisymmetric shell contour of a loudspeaker diaphragm. They presented three case studies, one for a clamped ﬂat membrane with added ring masses, one for geometry optimization of the clamped membrane, and a third for geometry optimization combined with variable shell thickness in soft surrounds. A point force excitation was applied in the center of the membrane. Fully coupled ﬂuid structure interaction was considered. The objective function was formulated as the squared deviation from a prescribed directivity pattern curve. In the ﬁrst case, a large mass in the driven center created a point source. Apparently, a point source guarantees a regular directivity pattern. Performance of the ﬁrst case was improved in the second case. The axisymmetric geometry was deﬁned by the variable positions of eight keypoints whereas the contour was created using Bspline interpolation of nine keypoints. (The position of one keypoint was arbitrarily chosen.) The best performance was found for the third case. The prescribed uniform directivity pattern was well adjusted for three different frequencies, 0.5, 10, and 15 kHz. Local differences of about 20 dB were observed for the initial design and less than about 1 dB for the optimal solution. In this case, the diaphragm was put into soft surround. This was necessary to achieve a uniform pattern at low frequencies. Constraint equations became necessary for the ﬁrst case only. For that a mass condition was introduced to avoid an unlimited increase of the mass. Geometry optimization succeeded without any constraints.
© 2004 by CRC Press LLC
Figure 3 New design of sedan body panels: Optimized geometry as modiﬁcation of the original geometry for symmetry half of a dashboard and hatshelf (scales in mm).
Although optimization was limited to three discrete frequencies only, the authors stated that uniform directivity patterns were also observed for other frequencies that they have checked. Finally, the authors emphasized suitability of optimization of the shell contour.
3.4
Design of Sedan Body Panels
In a number of papers on vehicle interior noise, the author of this chapter investigated several different issues of shell contour optimization for realistic applications, cf. Ref. [15– 22]. These applications can be categorized into those that apply geometrybased models and those that use direct modiﬁcation of ﬁniteelement meshes. Geometrybased modeling was tested for a sedan roof [15, 16] and for a dashboard [17, 18]. Geometry of a symmetry half of the roof was deﬁned by 15 keypoints, six of them with parametrically deﬁned position. Simple support at six points and an excitation pressure from above were unrealistic conditions that essentially contributed to the success of this work. In that case, where the elastic modes occur at frequencies much lower than that for rigid body modes, the optimization strategy favors stiffening effects to decrease the average sound pressure level at the driver’s ear. We observed a lowest eigenfrequency of 88 Hz and altogether 39 modes up to 200 Hz for the original structure. A 30cm higher optimal roof had its ﬁrst eigenfrequency at 209 Hz; in other cases it was even increased to about 250 Hz. Since optimization was limited to a frequency range of 0 to 200 or 0 to 300 Hz, amazing decreases of the objective function could be reported. Even for the case of only 10mm modiﬁcation signiﬁcant decreases were observed. In that case, however, stiffness was not increased by maximizing the lowest eigenfrequency but by decreasing the number of modes in the frequency range under consideration. More generally, it became apparent that modes changed from global to local shapes with the tendency of extinguishing each other. This tendency was conﬁrmed and utilized for optimization of a sedan dashboard [18]. Engine excitation, transmission through the body, and maximum feasible design modiﬁcations of 10 mm accounted for more realistic conditions. The geometrybased model consisted of several hundreds of keypoints, lines, and areas. Apparently, only
© 2004 by CRC Press LLC
Figure 4 New design of sedan body panels: Optimized geometry as modiﬁcation of the original geometry for part of a ﬂoor panel and spare wheel well (scales in mm).
few of them could be included in the optimization process. The objective function was more generally deﬁned as [17] 1 F = F˜ n , (2) where F˜ is given by F˜ =
1 ωmax − ωmin
ωmax
Φ {pL (ω)} dω.
(3)
ωmin
The operator Φ{} applied to sound pressure level pL represents a kind of a weighting function. A reasonable weighting function is n for pL > pRef (pL − pRef ) Φ {pL } = (4) 0 for pL ≤ pRef . So, the objective function simply appears as a frequencyaveraged sound pressure level and the exponent n controls the type of average. For n = 1, Eqs. (2)–(4) lead to the mean value where only values higher than a certain reference level pRef are taken into account. Similarly, this form provides the rootmeansquare value for n = 2. The major advantage of this value is that highlevel peaks are higher valued than lowlevel parts of the function. This helps to reduce these highlevel peaks during optimization procedure and avoids deep valleys as compensation of high peaks. If n tends to inﬁnity the objective function will be equal to the maximum value of the noise transfer function. In the case of the dashboard, 44 variables were optimized with the objective to decrease the rootmeansquare sound pressure level at the driver’s ear. New shapes showed wellknown structures. So, a horizontal bead and two bulges could be identiﬁed, cf. Fig. 3. The ﬁrst application of direct modiﬁcation of a ﬁniteelement mesh to ﬁnd an optimal shell geometry for acoustics was reported for a sedan hat shelf [19]. In that case, an alternative to geometrybased modeling was required since the geometrybased model contained too many details. These details made it impossible to optimize keypoint positions. As a consequence, the mesh was directly modiﬁed by using just four design variables. Moreover, the case of stiff supports and excitation in the hat shelf, i.e., subwoofer, belt retractor,
© 2004 by CRC Press LLC
etc., made it possible to circumvent full structural–acoustic optimization and concentrate just on optimization of the lowest eigenfrequency. This was increased from 32 to 101 Hz, whereas the average sound pressure level at the driver’s ear could decrease by between 4.4 and 13.9 dB, depending on the particular excitation and different parameters of the objective function, Eqs. (2)–(4). The new shape of the hat shelf is shown in Fig. 3. A full concept of direct modiﬁcation was presented in [20] and applied to a sedan ﬂoor panel in [21] and a spare wheel well [22]. Similar to the dashboard case, excitation at engine supports and transmission through the body were considered. In both cases, we could not speak of stifﬂy supported structures. Consequently the mode spectrum remained the same, but vibration mode shapes varied. Both panels were stiffened, though. This became obvious for the ﬂoor panel at low frequencies, cf. new geometry in Fig. 4, and for the spare wheel well in the entire frequency range. A strong vibration loop at the bottom of the spare wheel well could be replaced by stiffening beadlike structures, cf. Fig. 4. Due to the weak support of both structures, comparatively small improvements of the objective function can be reported.
4
CONCLUSION
Herein we have reviewed contributions in structural–acoustic optimization. After a brief review of applications in the ﬁeld, we have considered shell contour optimization in more detail. Based on the knowledge of the author, only three different applications were discussed. These were design of carillon bells, loudspeaker diaphragm, and sedan body panels. In all of these cases, it turned out that geometry optimization is an efﬁcient tool for structural–acoustic design. It is one of the major advantages of geometry optimization that a mass condition is not required. Furthermore, it is advantageous compared to optimization of piecewise constant shell thickness that a large variety of geometries of metal sheets can be easily produced by conventional metalforming techniques. Finally, we emphasize again that optimization of the shell contour is an efﬁcient tool to inﬂuence local stiffness of shells while its mass remains constant.
REFERENCES 1. S.T. Christensen, S.V. Sorokin, and N. Olhoff, “On analysis and optimization in structural acoustics – Part I: Problem formulation and solution techniques,” Structural Optimization, vol. 16, pp. 83–95, 1998. 2. F. Dirschmid, H. Troidl, A. Kunert, S. Dillinger, O. von Estorff, E. Negele, and M. Stowasser, “Akustische Optimierung von Getriebegeh¨ausen,” in Berechnung und Simulation im Fahrzeugbau, pp. 633–651, VDI–Report 1283, 1996. 3. T. Bauer and G. Henneberger, “Threedimensional calculation and optimization of the acoustic ﬁeld of the induction furnace caused by electromagnetic forces,” IEEE Transactions on Magnetics, vol. 35, pp. 1598–1601, May 1999. 4. N. Vincent, P. Bouvet, D.J. Thompson, and P.E. Gautier, “Theoretical optimization of track components to reduce rolling noise,” Journal of Sound and Vibration, vol. 193, no. 1, pp. 161–171, 1996. 5. J.C.O. Nielsen, “Acoustic optimization of railway sleepers,” Journal of Sound and Vibration, vol. 231, no. 3, pp. 753–764, 2000.
© 2004 by CRC Press LLC
6. G.H. Koopmann and J.B. Fahnline, Designing Quiet Structures: A Sound Power Minimization Approach. San Diego, London: Academic Press, 1997. 7. F.G. Kollmann, Maschinenakustik. Grundlagen, Messtechnik, Berechnung, Beeinﬂussung. Berlin, Heidelberg: Springer Verlag, 1999. 8. A.J.G. Schoofs, F. van Asperen, P. Maas, and A. Lehr, “A carillon of major–third bells. Part I: Computation of bell proﬁles using structural optimization,” Music Perception, vol. 4, no. 3, pp. 245–254, 1987. 9. A.J.M. Houtsma and H.J.G.M. Tholen, “A carillon of major–third bells. Part II: A perceptual evaluation,” Music Perception, vol. 4, no. 3, pp. 255–266, 1987. 10. A. Lehr, “A carillon of major–third bells. Part III: From theory to practice,” Music Perception, vol. 4, no. 3, pp. 267–280, 1987. 11. P.J.M. RoozenKroon, Structural Optimization of Bells. Dissertation, Technische Universiteit Eindhoven, 1992. 12. M.H. van Houten, Function Approximation Concepts for Multidisciplinary Design Optimization. Dissertation, Technische Universiteit Eindhoven, 1998. 13. A.J.G. Schoofs and D.H. van Campen, “Analysis and optimization of bells systems,” in 11th Carillon World Congress (Mechelen and Leuven, Belgium), August 1998. 14. S.T. Christensen and N. Olhoff, “Shape optimization of a loudspeaker diaphragm with respect to sound directivity properties,” Control and Cybernetics, vol. 27, no. 2, pp. 177–198, 1998. 15. S. Marburg, H.J. Hardtke, R. Schmidt, and D. Pawandenat, “An application of the concept of acoustic inﬂuence coefﬁcients for the optimization of a vehicle roof,” Engineering Analysis with Boundary Elements, vol. 20, no. 4, pp. 305–310, 1997. 16. S. Marburg, H.J. Hardtke, R. Schmidt, and D. Pawandenat, “Design optimization of a vehicle panel with respect to cabin noise problems,” in Proceedings of the NAFEMS World Congress (Stuttgart), pp. 885–896, 1997. 17. S. Marburg, “Efﬁcient optimization of a noise transfer function by modiﬁcation of a shell structure geometry. Part i: Theory.” Structural and Multidisciplinary Optimization, vol. 24, no. 1, pp. 51–59, 2002. 18. S. Marburg and H.J. Hardtke, “Efﬁcient optimization of a noise transfer function by modiﬁcation of a shell structure geometry. Part II: Application to a vehicle dashboard,” Structural and Multidisciplinary Optimization, vol. 24, no. 1, pp. 60–71, 2002. 19. S. Marburg and H.J. Hardtke, “Shape optimization of a vehicle hat–shelf: Improving acoustic properties for different load–cases by maximizing ﬁrst eigenfrequency,” Computers and Structures, vol. 79, no. 20–21, pp. 1943–1957, 2001. 20. S. Marburg, “A general concept for design modiﬁcation of shell meshes in structural– acoustic optimization. Part i: Formulation of the concept,” Finite Elements in Analysis and Design, vol. 38, no. 8, pp. 725–735, 2002. 21. S. Marburg and H.J. Hardtke, “A general concept for design modiﬁcation of shell meshes in structural–acoustic optimization. Part II: Application to a vehicle ﬂoor panel,” Finite Elements in Analysis and Design, vol. 38, no. 8, pp. 737–754, 2002. 22. S. Marburg and H.J. Hardtke, “Investigation and optimization of a spare wheel well to reduce vehicle interior noise,” International Journal for Acoustics and Vibration, vol. 11, no. 3, 2003 (in press). 23. K. Naghshineh, G. H. Koopmann, and A. D. Belegundu, “Material tailoring of structures to achieve a minimum radiation condition,” Journal of the Acoustical Society of America, vol. 92, no. 2, pp. 841–855, 1992.
© 2004 by CRC Press LLC
24. C.S. Jog, “Reducing radiated sound power by minimizing the dynamic compliance,” in IUTAM–Symposium on Designing for Quietness, December 2000 (M.L. Munjal, ed.). Dordrecht / Boston / London: Kluwer Academic Publishers, pp. 215–236, 2002. 25. A.D. Belegundu, R.R. Salagame, and G.H. Koopmann, “A general optimization strategy for sound power minimization,” Structural Optimization, vol. 8, no. 2–3, pp. 113– 119, 1994. 26. S.T. Christensen, S.V. Sorokin, and N. Olhoff, “On analysis and optimization in structural acoustics – Part II: Exempliﬁcations for axisymmetric structures,” Structural Optimization, vol. 16, pp. 96–107, 1998. 27. E.W. Constans, A.D. Belegundu, and G. H. Koopmann, “Design approach for minimizing sound power from vibrating shell structures,” AIAA Journal, vol. 36, no. 2, pp. 134–139, 1998. 28. J.B. Fahnline and G.H. Koopmann, “Design for a highefﬁciency sound source based on constrained optimization procedures,” Acoustical Physics, vol. 41, no. 5, pp. 700– 706, 1995. 29. F. Hibinger, Numerische Strukturoptimierung in der Maschinenakustik. Dissertation, Technische Universit¨at Darmstadt, 1998. 30. J.S. Lamancusa, “Numerical optimization techniques for structural–acoustic design of rectangular panels,” Computers and Structures, vol. 48, no. 4, pp. 661–675, 1993. 31. J.S. Lamancusa and H.A. Eschenauer, “Design optimization methods for rectangular panels with minimal sound radiation,” AIAA Journal, vol. 32, no. 3, pp. 472–479, 1994. 32. K. Nagaya and L. Li, “Control of sound noise radiated from a plate using dynamic absorbers under the optimization by neural network,” Journal of Sound and Vibration, vol. 208, no. 2, pp. 289–298, 1997. 33. R.L. St. Pierre Jr. and G.H. Koopmann, “A design method for minimizing the sound power radiated from plates by adding optimally sized, discrete masses,” Journal of Mechanical Design, vol. 117, pp. 243–251, June 1995. 34. A. Ratle and A. Berry, “Use of genetic algorithms for the vibroacoustic optimization of a plate carrying point–masses,” Journal of the Acoustical Society of America, vol. 104, no. 6, pp. 3385–3397, 1998. 35. R.R. Salagame, A.D. Belegundu, and G.H. Koopmann, “Analytical sensitivity of acoustic power radiated from plates,” Journal of Vibration and Acoustics, vol. 117, pp. 43–48, January 1995. 36. M. Tinnsten, “Optimization of acoustic response — a numerical and experimental comparison,” Structural Optimization, vol. 19, pp. 122–129, 2000. 37. H.W. Wodtke and J.S. Lamancusa, “Sound power minimization of circular plates through damping layer placement,” Journal of Sound and Vibration, vol. 215, no. 5, pp. 1145–1163, 1998. 38. S.A. Hambric, “Approximation techniques for broadband acoustic radiated noise design optimization problems,” Journal of Vibration and Acoustics, vol. 117, pp. 136– 144, January 1995. 39. E.W. Constans, G.H. Koopmann, and A.D. Belegundu, “The use of modal tailoring to minimize the radiated sound power of vibrating shells: Theory and experiment,” Journal of Sound and Vibration, vol. 217, no. 2, pp. 335–350, 1998. 40. S.A. Hambric, “Sensitivity calculations for broadband acoustic radiated noise design optimization problems,” Journal of Vibration and Acoustics, vol. 118, pp. 529–532, July 1996.
© 2004 by CRC Press LLC
41. Q. Shi, I. Hagiwara, A. Azetsu, and T. Ichkawa, “Holographic neural network approximations for acoustic optimization,” JSAE Review, vol. 19, pp. 361–363, 1998. 42. Q. Shi, I. Hagiwara, S. Azetsu, and T. Ichikawa, “Optimization of acoustic problem using holographic neural network,” Transactions of the Society of Automotive Engineers of Japan, vol. 29, pp. 93–97, July 1998 (in Japanese). 43. S. Marburg, R. Rennert, H.J. Beer, J. Gier, H.J. Hardtke, and F. Perret, “Experimental veriﬁcation of structural–acoustic modeling and design optimization,” Journal of Sound and Vibration, vol. 252, no. 4, pp. 591–615, 2002. 44. W. Kozukue, C. Pal, and I. Hagiwara, “Optimization of noise level reduction by truncated model coupled structural–acoustic sensitivity analysis,” Computers in Engineering (ASME), vol. 2, pp. 15–22, 1992. 45. C. Pal and I. Hagiwara, “Dynamic analysis of a coupled structural–acoustic problem. Simultaneous multi–modal reduction of vehicle interior noise level by combined optimization,” Finite Elements in Analysis and Design, vol. 14, pp. 225–234, 1993. 46. C. Pal and I. Hagiwara, “Optimization of noise level reduction by truncated modal coupled structural–acoustic sensitivity analysis,” JSME International Journal, Series C, vol. 37, no. 2, pp. 246–251, 1994. 47. M. Tinnsten, B. Esping, and M. Jonsson, “Optimization of acoustic response,” Structural Optimization, vol. 18, no. 1, pp. 36–47, 1999. 48. M.A. Lang and C.L. Dym, “Optimal acoustic design of sandwich panels,” Journal of the Acoustical Society of America, vol. 57, no. 6, pp. 1481–1487, 1975. 49. S.E. Makris, C.L. Dym, and J. MacGregor Smith, “Transmission loss optimization in acoustic sandwich panels,” Journal of the Acoustical Society of America, vol. 92, no. 6, pp. 1833–1843, 1986. 50. S.P. Crane, K.A. Cunefare, S.P. Engelstad, and E.A. Powell, “Comparison of design optimization formulations for minimization of noise transmission in a cylinder,” Journal of Aircraft, vol. 34, no. 2, pp. 236–243, 1997. 51. K. A. Cunefare, S.P. Crane, S.P. Engelstad, and E.A. Powell, “Design minimization of noise in stiffened cylinders due to tonal external excitation,” Journal of Aircraft, vol. 36, no. 3, pp. 563–570, 1999. 52. S.P. Engelstad, K.A. Cunefare, E.A. Powell, and V. Biesel, “Stiffener shape design to minimize interior noise,” Journal of Aircraft, vol. 37, no. 1, pp. 165–171, 2000. 53. M.G. Milsted, T. Zhang, and R.A. Hall, “A numerical method for noise optimization of engine structures,” Proceedings of the Institution of Mechanical Engineers / Part D: Journal of Automobile Engineering, vol. 207, pp. 135–143, 1993. 54. M. la Civita and A. Sestieri, “Optimization of an engine mounting system for vibro– acoustic comfort improvement,” Proceedings — SPIE The International Society for Optical Engineering, Issue 3727/PT2, pp. 1998–2004, 1999. 55. I. Hagiwara, Z.D. Ma, A. Arai, and K. Nagabuchi, “Reduction of vehicle interior noise using structural–acoustic sensitivity analysis methods,” SAE Technical Paper Series No. 910208, 1991. 56. Z.D. Ma and I. Hagiwara, “Sensitivity analysis methods for coupled acoustic– structural systems, Part II: Direct frequency response and its sensitivities,” AIAA Journal, vol. 29, no. 11, pp. 1796–1801, 1991. 57. K.K. Choi, I. Shim, and S. Wang, “Design sensitivity analysis of structure–induced noise and vibration,” Journal of Vibration and Acoustics, vol. 119, pp. 173–179, April 1997.
© 2004 by CRC Press LLC
58. L. Hermans and M. Brughmans, “Enabling vibro–acoustic optimization in a superelement environment: A case study,” Proceedings — SPIE The International Society for Optical Engineering, Issue 4062/PT2, pp. 1146–1152, 2000. 59. R. Freymann, “Sounddesign und Akustikentwicklung im Automobilbau,” in Maschinenakustik ’99 – Entwicklung l¨arm– und schwingungsarmer Produkte, pp. 47–64, VDI–Report 1491, 1999. 60. D. Roesems, “A new methodology to support an optimized NVH engineering process,” Sound and Vibration, vol. 31, no. 5, pp. 36–45, 1997. 61. S.R. Hahn and A.A. Ferri, “Sensitivity analysis of coupled structural–acoustic problems using perturbation techniques,” Journal of the Acoustical Society of America, vol. 101, no. 2, pp. 918–924, 1997. 62. W.S. Shephard Jr. and K.A. Cunefare, “Sensitivity of structural acoustic response to attachment feature scales,” Journal of the Acoustical Society of America, vol. 102, no. 3, pp. 1612–1619, 1997. 63. F. Scarpa and G. Curti, “A method for the parametric sensitivity of interior acousto– structural coupled systems,” Applied Acoustics, vol. 58, no. 4, pp. 451–467, 1999. 64. F. Scarpa, “Parametric sensitivity analysis of coupled acoustic–structural systems,” Journal of Vibration and Acoustics, vol. 122, pp. 109–115, April 2000. 65. Z.D. Ma and I. Hagiwara, “Development of new mode–superposition technique for truncating lower and/or higher–frequency modes (application of eigenmode sensitivity analysis for systems with repeated eigenvalues),” JSME International Journal, Series C, vol. 37, no. 1, pp. 7–13, 1994. 66. Z.D. Ma and I. Hagiwara, “Sensitivity analysis methods for coupled acoustic– structural systems, Part I: Modal sensitivities,” AIAA Journal, vol. 29, no. 11, pp. 1787– 1795, 1991. 67. F. Bitsie and R.J. Bernhard, “Sensitivity calculations for structural–acoustic efem predictions,” Noise Control Engineering Journal, vol. 46, no. 3, pp. 91–96, 1998. 68. J. Luo and H.C. Gea, “Modal sensitivity analysis of coupled acoustic–structural systems,” Journal of Vibration and Acoustics, vol. 119, pp. 545–550, October 1997. 69. K.K. Choi, I. Shim, J. Lee, and H.T. Kulkarni, “Design sensitivity analysis of dynamic frequency responses of acousto–elastic builtup structures,” in Optimization of Large Structural Systems (G.I.N. Rozvany, ed.), Kluwer Academic Publishers, 1993, vol. 1, pp. 329–343. 70. K.A. Cunefare and G.H. Koopmann, “Acoustic design sensitivities for structural radiators,” Journal of Vibration and Acoustics, vol. 114, pp. 178–186, April 1992. 71. S. Wang, “Design sensitivity analysis of noise, vibration, and harshness of vehicle body structure,” Mechanics of Structures and Machines, vol. 27, no. 3, pp. 317–336, 1999. 72. A.H.W.M. Kuijpers, G. Verbeek, and J.W. Verheij, “An improved acoustic fourier boundary element formulation using fast fourier transform integration,” Journal of the Acoustical Society of America, vol. 102, no. 3, pp. 1394–1401, 1997.
© 2004 by CRC Press LLC
15 Intelligent Control of Aircraft Dynamic Systems with a New Hybrid NeuroFuzzyFractal Approach Patricia Melin and Oscar Castillo Department of Computer Science, Tijuana Institute of Technology, Chula Vista, CA
We describe in this chapter a hybrid method for adaptive modelbased control of nonlinear dynamic systems using Neural Networks, Fuzzy Logic, and Fractal Theory. The new neurofuzzyfractal method combines Soft Computing (SC) techniques with the concept of the fractal dimension for the domain of Nonlinear Dynamic System Control. The new method for adaptive modelbased control has been implemented as a computer program to show that our neurofuzzyfractal approach is a good alternative for controlling nonlinear dynamic systems. It is well known that chaotic and unstable behavior may occur for nonlinear systems. Normally, we will need to control this type of behavior to avoid structural problems with the system. We illustrate in this chapter our new methodology in the case of controlling aircraft dynamic systems. For this case, we use mathematical models for the simulation of aircraft dynamics during ﬂight. The goal of constructing these models is to capture the dynamics of the aircraft, so as to have a way of controlling these dynamics to avoid dangerous behavior of the aircraft dynamic system.
1
INTRODUCTION
We describe in this chapter a new method for adaptive control of nonlinear dynamic systems based on the use of Neural Networks, Fuzzy Logic, and Fractal Theory. The dynamics of realworld systems are often highly nonlinear and difﬁcult to control [4]. The problem of controlling them using conventional controllers has been widely studied [1]. Much of the complexity in controlling any process comes from the complexity of the process being controlled. This complexity can be described in several ways. Highly nonlinear systems are difﬁcult to control, particularly when they have complex dynamics (such as instabilities to limit cycles and chaos [12]). Difﬁculties can often be presented by constraints, either on the control parameters or in the operating regime. Lack of exact knowledge of the process, of course, makes control more difﬁcult. Optimal control of many processes also requires systems that make use of predictions of future behavior. The mathematical models for the dynamic systems are assumed to be differential equations. The goal of having these models is to capture the dynamics of nonlinear processes, so as to have a way of controlling these dynamics for industrial purpose [7].
© 2004 by CRC Press LLC
We need a mathematical model of the nonlinear dynamic system to understand the dynamics of the processes involved in the evolution of the system. For a speciﬁc case, this may require testing several models before obtaining the appropriate mathematical model for the process [9]. For realworld systems with complex dynamics, we may even need several models for different sets of parameter values to represent all of the possible behaviors of the system. A general mathematical model of a dynamic system can be expressed as follows: dx/dt = f1 (x, D, α) − βf2 (x, D, α)
(1)
dp/dt = βf2 (x, D, α) where x ∈ Rn is a vector of state variables, p ∈ Rm is a vector of outputs, β ∈ R is a constant measuring the efﬁciency of the conversion process, D ∈ (0, 3) is the fractal dimension of the process, and α ∈ R is a selection parameter. The fractal dimension is used to characterize the process, for example, in the case of biochemical reactors D represents the fractal dimension of the bacteria used for production [2, 9]. For a complex dynamical system it may be necessary to consider a set of mathematical models to represent adequately all possible dynamic behaviors of the system. In this case, we need a decision scheme to select the appropriate model to use according to the linguistic value of a selection parameter α. We use a new fuzzy inference system for differential equations to achieve fuzzy modeling [3]. We have fuzzy rules of the form: IF α is A1 ... IF α is An
AND D is B1 ... AND D is Bn
THEN M1 ... THEN Mn
(2)
where A1 , ..., An are linguistic values for α, B1 , ..., Bn are linguistic values for the fractal dimension D, and M1 , ..., Mn are mathematical models of the form given by Eq. (1). The selection parameter α can be the temperature for biochemical processes, because temperature changes cause the presence of new bacteria in this case [9]. For the case of aircraft dynamic systems, α can be related to environment parameters. We combine adaptive modelbased control using neural networks with the method for modeling using fuzzy logic and fractal theory, to obtain a new hybrid neurofuzzyfractal method for control of nonlinear dynamic systems. This general method combines the advantages of neural networks (ability for identiﬁcation and control) with the advantages of fuzzy logic (ability for decision and use of expert knowledge) to achieve the goal of robust adaptive control of nonlinear dynamic systems. We also use the fractal dimension to characterize the processes in modeling these dynamical systems. We have developed intelligent control systems using this new method for adaptive control for several applications, to validate our new approach for control. We have obtained very good results in controlling biochemical reactors and chemical reactors with the hybrid approach for control [8]. In this chapter, we describe the application of our new method to the case of controlling aircraft dynamic systems.
2
FUZZY MODELING OF DYNAMICAL SYSTEMS
For a realworld dynamical system it may be necessary to consider a set of mathematical models to represent adequately all of the possible dynamic behaviors of the system [8, 9].
© 2004 by CRC Press LLC
In this case, we need a fuzzy [15] decision procedure to select the appropriate model to use according to the value of a selection parameter vector α. To implement this decision procedure, we need a fuzzy inference system that can use differential equations as consequents. For this purpose, we have developed a new fuzzy inference system that can be considered as a generalization of Sugeno’s inference system [13, 14], in which we are now using differential equations as consequents of the fuzzy rules, instead of simple polynomials like in the original Sugeno’s method. Using this method, a fuzzy model for a general dynamical system can be expressed as follows: IF α1 is A11 IF α1 is A21 . .
AND α2 is A12 AND α2 is A22 . .
. IF α1 is An1
. AND α2 is An2
... AND αm is A1m ... AND αm is A2m
THEN THEN
dy dt dy dt
= f1 (y, α) = f2 (y, α)
= fn (y, α) (3) where Aij is the linguistic value of αj for rule ith, α ∈ Rm and is deﬁned by α = [α1 , ..., αm ], and y ∈ Rp is the output obtained by the numerical solution of the corresponding differential equation. Of course, it is assumed that each differential equation in (3) locally approximates the real dynamical system over a neighborhood (or region) of Rm . The numerical solution of the differential equations can be achieved by the standard Runge–Kutta type method: ... AND αm is Anm
THEN
yn+1 = RK(yn ) + 1/2(k1 + k2 ) k1 = hf (yn , tn ) k2 = hf (yn + k1 , tn+1 ),
dy dt
(4)
where h is the step size of the method and RK can be considered as the Runge–Kutta operator that transforms numerical solutions from time n to time n+1. Numerical solutions are then aggregated by weighted average with weights obtained by the minimum of the ﬁring strengths of the inputs: y=
w1 y1 + w2 y2 + ... + wn yn w1 + w2 + ... + wn
(5)
where: y1 = RK(f1 (y, α)), y2 = RK(f2 (y, α)), . . . , yn = RK(fn (y, α)). The new fuzzy inference system for differential equations can be illustrated as in Fig. 1, where a complex dynamical system is modeled by using four different mathematical models (M1 , M2 , M3 , and M4 ). The decision scheme can be expressed as a singleinput fuzzy model as follows: IF IF IF IF
α is small α is regular α is medium α is large
THEN THEN THEN THEN
dy/dt = f1 (y, α) dy/dt = f1 (y, α) dy/dt = f1 (y, α) dy/dt = f1 (y, α),
(6)
where the output y is obtained by the numerical solution of the corresponding differential equation.
© 2004 by CRC Press LLC
Figure 1 Modeling a complex dynamical system with the new fuzzy system.
3
NEURAL NETWORKS FOR CONTROL
Parametric Adaptive Control is the problem of controlling the output of a system with a known structure but unknown parameters. These parameters can be considered as the elements of a vector p. If p is known, the parameter vector θ of a controller can be chosen as θ∗ so that the plant together with the ﬁxed controller behaves like a reference model described by a difference (or differential) equation with constant coefﬁcients [11]. If p is unknown, the vector θ(t) has to be adjusted online using all the available information concerning the system. Two distinct approaches to the adaptive control of an unknown system are (i) direct control and (ii) indirect control. In direct control, the parameters of the controller are directly adjusted to reduce some norm of the output error. In indirect control, the parameters of the system are estimated as p(t) at any time instant and the parameter vector θ(t) of the controller is chosen assuming that p(t) represents the true value of the system parameter vector. When indirect control is used to control a nonlinear system, the plant is parameterized using a mathematical model of the general form described in Section 1 and the parameters of the model are updated using the identiﬁcation error. The controller parameters in turn are adjusted by backpropagating the error (between the identiﬁed model and the reference model outputs) through the identiﬁed model. A block diagram of such an adaptive system is shown in Fig. 2. The overall structure of the adaptive system proposed in this paper to control a nonlinear dynamical system is the same as shown in Fig. 2 and is independent of the speciﬁc model used to identify the system. The delayed values of the system input and system output form the inputs to the neural network Nc, which generates the feedback control signal to the system. The parameters of the Neural Network Ni are adjusted by backpropagating the identiﬁcation error ei while those of the Neural Network Nc are adjusted by backpropagating the control error (between the output of the reference model and the identiﬁcation model) through the identiﬁcation model. The mathematical model for the nonlinear dynamic system in the time domain is generated by the method of modeling (described in Section 2) using the real data that is measured online in the system. On the other hand, the fractal module is used to charac
© 2004 by CRC Press LLC
Figure 2 General architecture for adaptive neurofuzzyfractal control.
terize the process and this information is used to specify the mathematical model in the time and space domain. This scheme enables dynamic changes of models according to the changes of online process identiﬁcation. Our new method for adaptive modelbased control combining Neural Networks, Fuzzy Logic and Fractal Theory differs from our previous approach of considering only the use of neural networks and models [5, 6]. 4
ADAPTIVE CONTROL OF AIRCRAFT SYSTEMS
The mathematical models of aircraft systems can be represented as coupled nonlinear differential equations [10]. In this case, we can develop a fuzzy rule base for modeling that enables the use of the appropriate mathematical model according to the changing conditions of the aircraft and its environment. For example, we can use the following model of an airplane when wind velocity is relatively small: p = I1 (−q + l),
q = I2 (p + m),
(7)
where I1 and I2 are the inertia moments of the airplane with respect to axis x and y, respectively, l and m are physical constants speciﬁc to the airplane, and p, q are the positions with respect to axis x and y, respectively. However, a more realistic model of an airplane in threedimensional space is as follows: p = I1 (−qr + l),
q = I2 (pr + m),
r = I3 (−pq + n),
(8)
where now I3 is the inertia moment of the airplane with respect to the z axis, n is a physical constant speciﬁc to the airplane, and r is the position along the z axis. Now considering wind disturbances in the model, we have the following equation: p = I1 (−qr + l) − ug ,
q = I2 (pr + m),
r = I3 (−pq + n),
(9)
where ug is the wind velocity. The magnitude of wind velocity is dependent on the altitude of the airplane in the following form: ln(h/510) ug = uwind510 1 + , (10) ln(51)
© 2004 by CRC Press LLC
Table 1 Fuzzy rule base for modeling aircraft dynamic systems
IF
THEN
Wind velocity
Inertia moment
Fractal dimension
Model
Small Small Small Small Large Large Large
Small Small Large Large Small Large Large
Low Medium Low Medium Medium Medium High
M1 M2 M2 M2 M3 M3 M3
where uwind510 is the wind speed at 510 ft altitude (typical value = 20 ft/sec). If we use the models of Eqs. (7)–(9) for describing aircraft dynamics, we can formulate a set of rules that relate the models to the conditions of the aircraft and its environment. Let us assume that M1 is given by Eq. (7), M2 is given by Eq. (8), and M3 is given by Eq. (9). Now using the wind velocity ug and inertia moment I1 as parameters, we can establish the fuzzy rule base for modeling as in Table 1. In Table 1, we are assuming that the wind velocity ug can have only two possible fuzzy values (small and large). This is sufﬁcient to know if we have to use the mathematical model that takes into account the effect of wind (M3 ) for ug large or if we don’t need to use it and simply the model M2 is sufﬁcient (for ug small). Also, the inertia moment (I1 ) helps in deciding between models M1 and M2 (or M3 ). 5
EXPERIMENTAL RESULTS
To give an idea of the performance of our neurofuzzyfractal approach for adaptive control, we show below simulation results for aircraft dynamic systems. First, we show in Fig. 3(a) the fuzzy rule base for a prototype intelligent system developed in the fuzzy logic toolbox of the MATLAB programming language. We show in Fig. 3(b) the nonlinear surface for the problem of aircraft dynamics using fractal dimension and wind velocity as input variables. We show simulation results for an aircraft system obtained using our new method for modeling dynamical systems. In Figs. 4(a) and 4(b) we show results for an airplane with inertia moments: I1 = 1, I2 = 0.4, I3 = 0.05 and the constants are: l = m = n = 1. The initial conditions are: p(0) = 0, q(0) = 0, r(0) = 0. To give an idea of the performance of our neurofuzzy approach for adaptive modelbased control of aircraft dynamics, we show below (Fig. 5) simulation results obtained for the case of controlling the altitude of an airplane for a ﬂight of 6 hours. We assume that the airplane takes about one hour to achieve the cruising altitude 30,000 ft, then cruises along for about three hours at this altitude (with minor ﬂuctuations), and ﬁnally descends for about two hours to its ﬁnal landing point. We will consider the desired trajectory as
© 2004 by CRC Press LLC
Figure 3 (a) Fuzzy rule base. (b) Nonlinear surface for aircraft dynamics.
Figure 4 (a) Simulation of position q. (b) Simulation of position p.
follows:
for 0 t 1 30t + sin 2t 30 + 2 sin 10t for 1 < t 4 hd = 90 − 15t for 4 < t 6.
Of course, a complete desired trajectory for the airplane would have to include the positions for the airplane in the x and y directions. However we think that here for illustration purposes it is sufﬁcient to show the control of the altitude h for the airplane. We used threelayer neural networks (with 10 hidden neurons) with the Levenberg– Marquardt algorithm and hyperbolic tangent sigmoidal functions as the activation functions for the neurons. We show in Fig. 5 the function approximation achieved by the neural network for control after 800 epochs of training with a variable learning rate. The identiﬁcation achieved by the neural network (after 800 epochs) can be considered very good because the error has been decreased to the order of 10−1 . Still, we can obtain a better approximation by using more hidden neurons or more layers. In any case, we can see clearly (from Fig. 5) how the neural network learns to control the aircraft, because it is able to
© 2004 by CRC Press LLC
Figure 5 Function approximation of the neural network for control of an airplane.
follow the arbitrary desired trajectory. We have to mention here that these simulation experiments for the case of a speciﬁc ﬂight for a given airplane show very good results. We have also tried our approach for control with other types of ﬂights and airplanes with good results.
6
CONCLUSIONS
We have developed a general method for adaptive modelbased control of nonlinear dynamic systems using Neural Networks, Fuzzy Logic, and Fractal Theory. We illustrated our method for control by the case of controlling aircraft dynamics. In this case the models represent the aircraft dynamics during ﬂight. We also described in this chapter an adaptive controller based on the use of neural networks and mathematical models for the system. The proposed adaptive controller performs rather well considering the complexity of the domain being considered in this research work. We have shown that our method can be used to control chaotic and unstable behavior in aircraft systems. Chaotic behavior has been associated with the “ﬂutter” effect in real airplanes, and for this reason it is very important to avoid this kind of behavior. We can say that combining Neural Networks, Fuzzy Logic, and Fractal Theory, using the advantages that each of these methodologies has, can give good results for this kind of application. Also, we believe that our neurofuzzyfractal approach is a good alternative for solving similar problems.
REFERENCES 1. P. Albertos, R. Strietzel, and N. Mart, Control Engineering Solutions: A Practical Approach, IEEE Computer Society Press, 1997. 2. O. Castillo and P. Melin, “Developing a new method for the identiﬁcation of microorganisms for the food industry using the fractal dimension,” Journal of Fractals, 2, No. 3 (1994), 457–460.
© 2004 by CRC Press LLC
3. O. Castillo and P. Melin, “A new fuzzy inference system for reasoning with multiple differential equations for modeling complex dynamical systems,” in Proceedings of CIMCA’99, IOS Press, Vienna, Austria, 1999, pp. 224–229. 4. P. Melin and O. Castillo, “Modeling and simulation for bacteria growth control in the food industry using artiﬁcial intelligence,” in Proceedings of CESA’96, Gerf EC Lille, Lille, France, 1996, pp. 676–681. 5. P. Melin and O. Castillo, “An Adaptive modelbased neural network controller for biochemical reactors in the food industry,” in Proceedings of Control’97, Acta Press, Canada, 1997, pp. 147–150. 6. P. Melin and O. Castillo, “An adaptive neural network system for bacteria growth control in the food industry using mathematical modeling and simulation,” in Proceedings of IMACS World Congress’97, Berlin, Germany, 1997, Vol. 4, pp. 203–208. 7. P. Melin and O. Castillo, “Automated mathematical modeling and simulation for bacteria growth control in the food industry using artiﬁcial intelligence and fractal theory,” Journal of Systems, Analysis, Modeling and Simulation, Gordon and Breach, 1997, pp. 189–206. 8. P. Melin and O. Castillo, “An adaptive modelbased neurofuzzyfractal controller for biochemical reactors in the food industry,” in Proceedings of IJCNN’98, IEEE Computer Society Press, Anchorage, AK, 1998, Vol. 1 pp. 106–111. 9. P. Melin and O. Castillo, “A new method for adaptive modelbased neurofuzzyfractal control of nonlinear dynamic plants: the case of biochemical reactors,” in Proceedings of IPMU’98, EDK Publishers, Paris, France, 1998, Vol.1, pp. 475–482. 10. P. Melin and O. Castillo, “A new method for adaptive modelbased neurofuzzyfractal control of nonlinear dynamical systems,” in Proceedings of the International Conference of Nonlinear Problems in Aviation and Aerospace’98, European Publications, Daytona Beach, FL, 1998, Vol. 2, pp. 499506. 11. K.S. Narendra and A.M. Annaswamy, Stable Adaptive Systems, PrenticeHall, Englewood Cliffs, NJ, 1989. 12. S. Rasband, Chaotic Dynamics of Nonlinear Systems, John Wiley & Sons, New York, 1990. 13. M. Sugeno and G.T. Kang, “Structure identiﬁcation of fuzzy model,” Fuzzy Sets and Systems, 28 (1988), 15–33. 14. T. Takagi and M. Sugeno, “Fuzzy identiﬁcation of systems and its applications to modeling and control,” IEEE Transactions on Systems, Man and Cybernetics, 15 (1985), 116–132. 15. L.A. Zadeh, “The concept of a linguistic variable and its application to approximate reasoning,” Information Sciences, 8 (1975), 43–80.
© 2004 by CRC Press LLC
16 ClosedForm Solution of ThreeDimensional Ideal Proportional Navigation †
ChiChing Yang† and HsinYuan Chen∗ Department of Electrical Engineering, Hsiuping Institute of Technology, Taichung, Taiwan Department of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan
∗
The main goal of this research is to derive the complete closedform solutions of threedimensional ideal proportional navigation (3D IPN) for nonmaneuvering and maneuvering targets. IPN is a guidance law by which the commanded missile acceleration is applied in the direction perpendicular to the relative velocity between the missile and the target, and the acceleration magnitude is proportional to the product of lineofsight (LOS) rate and relative speed. Three coupled nonlinear secondorder dynamic equations modeling the relative motion are solved analytically without any linearization for performance and trajectory analysis. Properties of 3D IPN such that capture region, rangetogo, timetogo, and two aspect angles within the spherical coordinates are all obtained in closed form.
I
INTRODUCTION
Proportional navigation (PN) has been proved to be a useful guidance scheme in many airtoair and surfacetoair homing systems for interception of airborne targets. Proportional navigation is so named because the magnitude of commanded acceleration is generated proportional to the measured rate of rotation of the missiletarget line of sight (LOS). In traditional proportional navigation, pure proportional navigation (PPN) and true proportional navigation (TPN) have been widely discussed and thoroughly developed for homing systems of missiles. The earlier study started from Murtaugh and Criel [15]. During the three decades that followed, researchers such as Guelman [9, 10, 11, 12], Yang and Yeh [21–22], Yang and Yang [27], Shukla and Mahapatra [18, 19], Mahapatra and Shukla [14], Becker [2], Yuan and Chern [28–32], Dhar and Ghose [5], Ghose [6,7,8], Rao [16], Vathsal and Rao [19], have explored proportional navigation in many different ways, such as true proportional navigation (TPN), pure proportional navigation (PPN), generalized proportional navigation (GPN), realistic true proportional navigation (RTPN), and ideal proportional navigation (IPN), etc. Because of the difﬁculties in analysis, most analytical investigations on PN guidance laws in the past used twodimensional (2D) models, although practical pursuitevasion motion occurred in a threedimensional (3D) environment. As to 3D studies, earlier work was done by Adler [1]. The main viewpoint of his study was that the 3D relative motion can be
© 2004 by CRC Press LLC
Figure 1 Threedimensional pursuit geometry.
described by the principal plane deﬁned by the instantaneous line of sight (LOS) and by the missile’s velocity. Shinar et al. [17] used the linearized 3D model to analyze guidance laws. Guelman [13] extended the optimal guidance law in the plane to 3D models and obtained a preliminary version of 3D optimal guidance law. Cochran et al. [4], considered the 3D RTPN guidance problem, and revealed a partial solution. More recently, Yang and Yang [24–26] proposed a systematic solution technique for 3D guidance laws and the technique has been applied to generalized 3D PN, 3D RTPN, and 3D TPN problems, where complete solutions of the three coupled nonlinear differential equations within spherical coordinates were obtained without any linearization. The discussion of 2D PN guidance law had used moving polar coordinates wherein the origin is ﬁxed on the missile to characterize the relative motion. In 3D models, seeker measurements are spherical in nature [3] (range and angles), and are, therefore, nonlinear functions of the states in Cartesian coordinates. The nonlinear transformation of states could be avoided if the guidance laws were formulated in spherical coordinates. Thus, moving spherical coordintes seem to be most natural in characterizing 3D relative motions. We will use spherical coordinates to describe the relative motion of a target with respect to the missile, which is guided by 3D IPN. The strategy of IPN guidance law is to turn the relative velocity in most efforts to the direction of LOS. In the IPN scheme the commanded acceleration is applied in the direction normal to the relative velocity between the missile and the target, and its magnitude is proportional to the product of the LOS rate and the relative velocity. Planar motion using IPN has been studied by Yuan and Chern (30). It was reported that with some more energy consumption, the IPN guidance law has a larger capture area and is much more effective than the other schemes. The 3D relative motion resulting from the 3D IPN guidance law is governed by three secondorder nonlinear differential equations whose solution has never been discussed before. Based on the systematic framework [24–26] for analyzing 3D guidance laws, we have made a breakthrough in obtaining the analytical solutions of these coupled nonlinear differential equations. The solutions are derived in terms of the unit angular momentum deﬁned by the relative motion between the missile and the target. Unit angular momentum in the space is a performance index measuring the departure tendency of the 3D relative motion from a ﬁxed plane. If the relative motion occurs wholly within a ﬁxed plane, then the direction of the relative angular momentum will remain constant during the course of interception. Besides as an index for measuring 3D relative motions, unit angular momentum has a decoupling effect such that the relative distance and azimuths can be solved independently. A systematic framework for analyzing 3D guidance laws is introduced in Section II.
© 2004 by CRC Press LLC
3D IPN is deﬁned in Section III where three nonlinear differential equations describing the relative motion with respect to a nonmaneuvering target are solved analytically, and the trajectory properties such as capture region, timetogo, rangetogo, and aspect angles are all derived in closed forms. In Section IV, we solve the twoplayergame problem wherein a missile guided by 3D IPN is designed to pursue a target guided by 3D RTPN, and vice versa. Numerical results and more physical insight about 3D IPN guidance law are discussed in Section V.
II
GENERAL ANALYSIS OF 3D RELATIVE MOTION
Consider the spherical coordinates (r, θ, φ) with origin ﬁxed at the missile, where r is the relative distance between the missile and the target, and θ and φ are azimuths. Let (er , eθ , eφ ) be unit vectors along the coordinate axes (see Fig. 1). According to the principles of kinematics, the three relative acceleration components (ar , aθ , aφ ) can be expressed by the following set of secondorder nonlinear differential equations: r¨ − rφ˙ 2 − rθ˙2 cos2 φ = aTr − aMr ≡ ar rθ¨ cos φ + 2r˙ θ˙ cos φ − 2rφ˙ θ˙ sin φ = aTθ − aMθ ≡ aθ rφ¨ + 2r˙ φ˙ + rθ˙2 cos φ sin φ = aTφ − aMφ ≡ aφ ,
(1a) (1b) (1c)
where aTr ,aTθ , and aTφ are the acceleration components of the target; aMr ,aMθ , and aMφ are the acceleration components of the missile. To analyze these coupled nonlinear equations, we ﬁnd that the exploiting of the angular momentum of unit mass is very helpful. The unit angular momentum h for the missiletarget relative motion is deﬁned as h = r × r˙,
(2)
where r is the relative displacement along the line of sight (LOS), and r˙ is the relative velocity. Accordingly, if the relative motion occurs within a ﬁxed plane, i.e., r and r˙ are in the same plane during the interception, then the direction of h will be constant, always being perpendicular to the plane spanned by r and r˙ . Hence, the variation of the direction of h is a natural measure for departure tendency of the relative motion from a ﬁxed plane. r and r˙ can be expressed by the spherical unit vectors (er , eθ , eφ ) as r = rer ˙ eφ . r˙ = r ˙ er + rθ˙ cos φeθ + rφ
(3a) (3b)
Substituting Eq. (3) into Eq. (2) yields the expression for h: ˙ eθ + θ˙ cos φeφ ), h = heh = r2 (−φ where h = r2
φ˙ 2 + θ˙2 cos2 φ
(4a)
(4b)
is the magnitude of h, and eh =
© 2004 by CRC Press LLC
r2 ˙ eθ + θ˙ cos φeφ ) (−φ h
(4c)
is the unit vector along the direction of h. According to the deﬁnition of eh , we know that eh is perpendicular to er . We can ﬁnd another unit vector e⊥ h that is perpendicular to both er and eh . It is straightforward to verify that e⊥ h =
r2 ˙ ˙ eφ ). (θ cos φeθ + φ h
(5)
The set of unit vectors (er , eh , e⊥ h ) constitutes a new moving coordinate system that is more convenient for describing 3D guidance laws than the conventional spherical coordinates (er , eθ , eφ ), as will be seen later. We now proceed to derive the various physical components under the new coordinate system (er , eh , e⊥ h ). The differentiation of Eq. (2) gives dh = r × r¨, dt
(6)
where r¨ is the relative acceleration deﬁned by r¨ = ar er + aθ eθ + aφeφ .
(7)
Using Eq. (3a) and Eq. (7), the righthand side of Eq. (6) can be rewritten as r × r¨ = (−raφ )eθ + (raθ )eφ ,
(8)
and the lefthand side of Eq. (6) becomes d ˙ eh + he˙ h , (h) = h dt
(9)
where h˙ and e˙ h can be obtained from the differentiation of Eq. (4b) and Eq. (4c) with respect to time t. Upon differentiating h and eh , we need the relations : × er , e˙ r = ω
e˙ θ = ω × eθ ,
e˙ φ = ω × eφ ,
where ω is the angular velocity of the moving coordinates (er , eθ , eφ ), which can be derived as ˙ eθ + θ˙ cos φeφ . ω = θ˙ sin φer − φ (10) h˙ and e˙ h are then derived as follows: ¨ 2rr( ˙ φ˙ 2 + θ˙2 cos2 φ) + r2 (θ¨θ˙ cos2 φ − φ˙ θ˙2 sin φ cos φ + φφ) h˙ = φ˙ 2 + θ˙2 cos2 φ
(11a)
1 e˙ h = φ˙ 2 + θ˙2 cos2 φ ˙ φ˙ φ¨ + θ˙θ¨ cos2 φ − θ˙2 φ˙ sin φ cos φ) φ( 2 − θ˙ sin φ cos φ eθ × −φ¨ + φ˙ 2 + θ˙2 cos2 φ θ˙ cos φ(φ˙ φ¨ + θ˙θ¨ cos2 φ − θ˙2 φ˙ sin φ cos φ) ¨ ˙ ˙ ˙ ˙ + θ cos φ − θφ sin φ − − θφ sin φ eφ . φ˙ 2 + θ˙2 cos2 φ (11b)
© 2004 by CRC Press LLC
Equation (11) is substituted into Eq. (9) to obtain d (h) = −r(rθ¨ cos φ + 2r˙ θ˙ cos φ − 2rφ˙ θ˙ sin φ)eθ dt + r(rφ¨ + 2r˙ φ˙ + rθ˙2 cos φ sin φ)eφ .
(12)
If we equate Eq. (8) with Eq. (12), we go back to the results of Eq. (1b) and Eq. (1c); however, this fact manifests that the roles of Eq. (1b) and Eq. (1c) can be replaced by the roles of Eq. (11) in describing the relative motion. Equation (11) can be further simpliﬁed to the following form: h˙ = r · (r¨ × eh ) r3 ˙ = aφ φ + aθ θ˙ cos φ h ˙eh = r (r¨ · eh )e⊥ h h r3 = 2 (aθ φ˙ − aφ θ˙ cos φ)e⊥ h. h
(13a)
(13b)
The remaining equation describing the relative motion comes from Eq. (1a), which can be rewritten in terms of h as h2 r¨ − 3 = ar . (13c) r Equation (13) is equivalent to Eq. (1), but Eq. (13) has the advantage of decoupling the radial motion from the tangential motion as can be seen in the following sections. III
3D IPN WITH NONMANEUVERING TARGETS
Ideal proportional navigation is a guidance law whose commanded acceleration is applied in the direction normal to the relative velocity (r˙ ) between the missile and its target, and its magnitude is proportional to the product of the LOS rate and relative speed. A 2D version of IPN has been proposed by Yuan and Chern [26]. Commanded acceleration for a 3D IPN can be expressed in the following vector form: aM = λr˙ × Ω.
(14a)
of LOS can be found from the relation The angular velocity Ω ˙ = r × r = h Ω 2 r r2 ˙ eθ + θ˙ cos φeφ , = −φ
(14b)
where Eq. (4a) has been used to obtain Eq. (14b). Therefore, missile acceleration becomes aM = λr˙ × Ω h = λr˙ × 2 r h2 h = λ( 3 er − r˙ 2 e⊥ ). r r h
© 2004 by CRC Press LLC
(15)
In this section a nonmaneuvering target is assumed, i.e., aT = 0. After substituting Eq. (15) into Eq. (13), we have h2 h2 = −λ 3 3 r r r ˙ h˙ = λh r e˙ h = 0.
r¨ −
(16a) (16b) (16c)
Equation (16a) and Eq. (16b) corresponding to radial motion can be solved together, leading to the following results: h = h0
r r0
λ
r˙ = − r˙02 +
(17) h20 h20 2(λ−1) − r . r02 r02λ
(18)
Combining Eq. (14b) and Eq. (17) yields h h0 = 2 r2 r0
Ω=
r r0
λ−2 .
(19)
In order to maintain ﬁnite commanded acceleration, the navigation constant λ must be greater than 2. It is seen from Eq. (18) and Eq. (19) that the capture area for the IPN is simply λ ≥ 2. That is, no matter what the initial condition of r˙ is, the interception can always be achieved successfully when λ ≥ 2. Since the commanded acceleration (15) is normal to r˙ , we can show that the relative speed V (t) = r˙  remains constant during the interception. This can be checked by using the deﬁnition of r˙ from Eq. (3b): V (t) = r˙  =
r˙ 2 +
h2 . r2
Substituting Eq. (17) and Eq. (18) into the above equation, yields
V (t) =
r˙02 +
h20 = V0 = constant. r02
(20)
Equation (1) consists of three secondorder differential equations, and six initial conditions shall be assigned to determine the solution uniquely. These six initial conditions are chosen as r0 , r˙0 , θ0 , θ˙0 , φ0 , and φ˙ 0 . In general, we can choose the inertia reference line as the initial LOS such that θ0 = φ0 = 0. In order to have the analytical solution independent of the initial conditions, a nondimensionalized process is required before we proceed further. Dimensionless variables are deﬁned as follows: ρ=
© 2004 by CRC Press LLC
r , r0
τ=
t , r0 /V0
¯= h
h . r0 V 0
(21)
Using these dimensionless variables, Eq. (17) and Eq. (18) can be rewritten as ¯=h ¯ 0 ρλ h
(22)
−dρ . dτ = ¯ 2 ρ2λ−2 1−h
(23)
0
Equation (23) is integrated to obtain the time history of the relative distance ρ:
ρ τ =− 1
dρ , ¯ 2 ρ2λ−2 1−h 0
(24)
and the duration of interception can be found as
1 τf = 0
dρ . ¯ 2 ρ2λ−2 1−h 0
(25)
The commanded missile acceleration can be calculated from Eq. (15) as
2 2 h2 rh ˙ + 3 r r2 h h2 =λ 2 + r˙ 2 r r2 h = λV0 2 . r
aM = λ
The dimensionless form becomes aM ¯ 0 ρλ−2 . = λh V02 /r0
(26)
The total cumulative velocity increment ∆V of the missile during interception is deﬁned as
τf aM dτ. ∆V = 2 (27) V0 /r0 0
Substituting Eq. (23) and Eq. (27) into Eq. (28), we have
1
∆V = 0
¯ 0 ρλ−2 dρ λh . ¯ 2 ρ2(λ−1) 1−h
(28)
0
There are three time functions ρ(t), θ(t), and φ(t) needed to describe the 3D relative motion within the spherical coordinates. Up to this stage, only relative distance ρ(t) for radial motion is obtained. The following work is to ﬁnd the azimuths θ(t) and φ(t) for tangential motion. This step is harder than the solution of r. The attack on the problem
© 2004 by CRC Press LLC
starts from Eq. (16c). Substituting Eq. (11b) into Eq. (16c) and using the transformation between the two coordinate systems (er , eθ , eφ ) and (er , eh , e⊥ h ), we obtain the key equation characterizing the behavior of θ(t) and φ(t) as ¨ θ¨ cos φ − θ˙φ˙ sin φ θ˙ cos φ(θ¨θ˙ cos2 φ − φ˙ θ˙2 sin φ cos φ + φφ) + + θ˙ sin φ = 0. ˙ φ˙ 2 + θ˙2 cos2 φ) −φ˙ φ(
(29)
This equation can be simpliﬁed to a totally differentiable form: d(θ˙ cos φ) d(φ˙ 2 + θ˙2 cos2 φ) + + tan φdφ = 0, −θ˙ cos φ 2(φ˙ 2 + θ˙2 cos2 φ)
(30)
and the integration gives 2 2 φ˙ 2 + θ˙2 cos2 φ φ˙0 + θ˙0 cos2 φ0 = l2 , = 2 ˙θ2 cos4 φ 4 ˙ θ0 cos φ0
(31)
where l is an integration constant. The above equation can be solved for φ˙ 2 in terms of θ˙2 :
dφ dτ
2
= (l2 cos4 φ − cos2 φ)
dθ dτ
2 .
The other relation between dφ/dτ and dθ/dτ comes from Eq. (4b): 2 2 dφ dθ 2 4 2 ¯ =ρ h + cos φ . dτ dτ
(32)
(33)
Substituting Eq. (22) and Eq. (32) into Eq. (33) yields
dθ dτ
2
¯ 2 h0 = ρ2λ (ρ cos φ)−4 , l
(34)
which, in turn, is substituted into Eq. (32) to obtain the expression for dφ/dτ as dφ ¯ 0 ρλ−2 1 − l−2 cos−2 φ. = sign(φ˙ 0 )h dτ
(35)
The relation between φ and ρ can be obtained by integrating Eq. (35) with the help of Eq. (23):
φ
1 ¯ λ−2 dφ dρ h0 ρ ˙ . (36) = sign(φ0 ) 2 −2 −2 ¯ 1 − l cos φ 1 − h0 ρ2λ−2 ρ
0
The remaining step is to relate θ to φ by integrating Eq. (32): θ = sign
© 2004 by CRC Press LLC
θ˙0 φ˙ 0
φ 0
cos2
dφ . φ l2 − cos−2 φ
(37)
It can be observed from Eq. (35) and Eq. (37) that both θ(τ ) and φ(τ ) are monotonically increasing when θ˙0 > 0 and φ˙ 0 > 0, while both θ(τ ) and φ(τ ) are monotonically decreasing when θ˙0 < 0 and φ˙ 0 < 0. Without loss of generality, we will assume θ˙0 > 0 and φ˙ 0 > 0 in the following discussion. These initial conditions can be naturally achieved by choosing the inertia reference line as the initial LOS and by deﬁning the positive directions of θ and φ as the directions of θ˙0 and φ˙ 0 . To express the solutions more explicitly, we ﬁnd that the introduction of the following auxiliary azimuths ψ and η is very helpful. The ﬁrst auxiliary azimuth ψ is deﬁned as cos ψ = ρλ−1 , cos ψ0
(38)
¯ 0 . It is noted that according to the deﬁnition where ψ0 = cos−1 h ¯ 0 = h0 = h0 , h r0 V 0 h20 + r02 r˙02
(39)
¯ 0 ≤ 1. The second auxiliary azimuth η is deﬁned as we have 0 ≤ h 1 cos η = , cos η0 cos φ
(40)
where cos η0 = 1/l = θ˙0 / θ˙02 + φ˙ 20 ≤ 1. Instead of the timetogo τ and the rangetogo ρ, we ﬁnd that ψ and η are more appropriate to use as independent variables to describe the ¯ ρ, τ , φ, and θ can all be expressed by ψ and η. From Eq. (22) and 3D relative motion. h, Eq. (38), we have λ λ−1 ¯=h ¯ 0 cos ψ h . (41) cos ψ0 The relation between time τ and ψ is established by substituting Eq. (38) into Eq. (24). After some manipulations, we have 1 τ= (λ − 1) cos ψ0
ψ ψ0
cos ψ cos ψ0
2−λ λ−1 dψ.
(42)
The required time of interception is obtained by setting ψ = π/2 (i.e., ρ = 0) in the upper limit of the above integration: 1 τf = (λ − 1) cos ψ0
π/2 ψ0
cos ψ cos ψ0
2−λ λ−1 dψ.
(43)
¯ 0 = 1, i.e., ψ0 = 0. In this case, τf has a When the initial closing rate r˙0 is zero, we have h cleaner expression: 1 √ Γ π 2λ − 2 , τf = (44) λ 2λ − 2 Γ 2λ − 2
© 2004 by CRC Press LLC
where Γ(·) is gamma function. φ is related to ψ by substituting Eq. (38) into Eq. (36), leading to the following identity:
φ
ψ − ψ0 = (λ − 1) 0
dφ 1−
l−2
cos−2 φ
.
(45a)
The integration involved in Eq. (45a) can be evaluated in terms of elementary functions as ψ − ψ0 . (45b) 1 − l−2 sin φ = sin−1 λ−1 On the other hand, θ is related to η by substituting Eq. (40) into Eq. (37).
η θ=−
cos ηdη
cos2 η − l−2 η0 sin η . = cos−1 sin η0
(46)
Finally, the total cumulative velocity increment ∆V shown in Eq. (29) can be integrated to the following form: λ π ∆V = (47) − ψ0 . λ−1 2 The auxiliary azimuth η has a concrete physical meaning. Combining Eq. (31) and Eq. (40), we can show that sin η =
φ˙ φ˙ 2 + θ˙2 cos2 φ
,
cos η =
θ˙ cos φ φ˙ 2 + θ˙2 cos2 φ
.
(48)
Substituting the above equation into Eq. (4c) yields eh = − sin η eθ + cos η eφ .
(49)
This reveals that η is the angle between eh and eφ as shown in Fig. 2. As mentioned earlier, eh is normal to the instantaneous plane spanned by r and r˙ . Consequently, when the relative motion remains within a ﬁxed plane in the 3D space, eh will not change. This is what we have derived in Eq. (16c). In the following we will show that when a nonmaneuvering target is pursued by a missile guided by the 3D IPN guidance law, the whole relative motion indeed occurs within a ﬁxed plane, but this plane may not be limited to the horizontal x − y plane. To show eh is ﬁxed in the 3D space, the most natural way is to express eh in terms of the inertia coordinates (i, j, k), and to show that the components of eh in the i, j, and k directions are all constant. Exploiting the transformation between the spherical coordinates and the inertia Cartesian coordinates, eθ = − sin θ i + cos θ j eφ = − sin φ cos θ i − sin φ sin θ j + cos φ k,
© 2004 by CRC Press LLC
Figure 2 The direction of eh .
we can express eh of Eq. (49) in terms of the inertia Cartesian coordinates as eh = (sin η sin θ − cos η sin φ cos θ) i + (− sin η cos θ − cos η sin φ sin θ) j + (cos η cos φ) k.
(50)
The simpliﬁcation of Eq. (50) needs two relations coming from Eq. (40) and Eq. (46): cos θ =
sin η , sin η0
sin θ = cot η0 tan φ.
(51)
Using Eq. (40) and Eq. (51), we can establish the following identities: sin η sin θ − cos η sin φ cos θ = 0 − sin η cos θ − cos η sin φ sin θ = − sin η0 cos η cos φ = cos η0 .
(52a) (52b) (52c)
Therefore, Eq. (50) becomes eh = − sin η0 j + cos η0 k.
(53)
As expected, eh is ﬁxed in the inertia coordinates and the orientation of the ﬁxed plane where the relative motion occurs is determined by η0 , which is deﬁned by the initial angular rates as
θ˙02 η0 = cos−1 . θ˙2 + φ˙ 2 0
0
Since we have chosen θ0 = 0 and φ0 = 0, the moving spherical coordinates coincide with the inertia Cartesian coordinates at the start of the interception, i.e., er = i, eθ = j, and eφ = k . Hence, at the beginning of interception Eq. (49) is reduced to Eq. (53). As the
© 2004 by CRC Press LLC
Figure 3 Angular rate for line of sight.
interception proceeds, η changes from η0 and the spherical coordinates (er , eθ , eφ ) rotate continuously in such a way that the orientation of eh = − sin η eθ + cos η eφ keeps ﬁxed in the space. The physical meaning of the auxiliary azimuth ψ can be revealed by combining Eq. (19) and Eq. (38). It is easy to show that ¯ h = V0 cos ψ ρ dρ radial speed = V0 = V0 sin ψ. dτ
tangential speed = V0
(54a) (54b)
Therefore, ψ is the angle between the tangential relative velocity and the total relative velocity, as shown in Fig. 4.
IV
NUMERICAL RESULTS
The analysis of the 3D IPN guidance law can be greatly simpliﬁed by using the above explicit formulae. Some inherent properties of the 3D IPN, gained from numerical results, are discussed in the following.
Inﬂuence of Navigation Constant Figures 3 to 6 show the variations of Ω, dρ/dτ , φ, and θ with respect to the dimensionless relative distance ρ for various values of navigation constant λM and for various initial conditions. The relative motions can be divided into three categories according to the value of λM :
© 2004 by CRC Press LLC
Figure 4 Closing speed for 3D IPN
Figure 5 Azimuth φ for 3D IPN with varying l.
• (1) Divergent relative motion: When λM < λT + 2, the angular rate of LOS, the azimuths θ and φ, and the maximum required missile acceleration all grow without bound, indicating a failure of interception. • (2) Steadystate relative motion: When λM = λT + 2, , we have Ω = Ω0 =constant. In this case, the missile approaches the target steadily, and the required interception time is evaluated as π/2 − ψ0 . τf = ¯ 2 cos ψ0 1 − λT h 0
© 2004 by CRC Press LLC
Figure 6 Azimuth θ for 3D IPN with varying l.
• (3) Convergent relative motion : When λM > λT + 2, all the physical variables converge as the collision point approaches, and the converging speed of the relative motion increases as λM is increased.
V
CONCLUSIONS
This paper is devoted to the study of threedimensional pursuitevasion motion. Three secondorder nonlinear differential equations which describe the motion of missiles guided by threedimensional ideal proportional navigation (3D IPN) are solved analytically without any linearization. Trajectory and performance analysis are described in closed form for nonmaneuvering targets. The general analysis proposed here can be applied to other 3D relative motions using guidance laws other than IPN in uniﬁed framework.
Acknowledgment This research was supported by the National Science Council under Grant No. NSC 902213E164002.
REFERENCES 1. Adler, F.P., “Missile guidance by threedimensional proportional navigation,” Journal of Applied Physics, 27 (1956) 500–507. 2. Becker, K., “Closedform solution of pure proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 26 (1990) 526–533.
© 2004 by CRC Press LLC
3. Cloutier, J.R., Evers, J.H., and Feeley, J.J., “Assessment of airtoair missile guidance and control technology,” IEEE Control Systems Magazine, October (1989) 27–34. 4. Cochran, J.E., No, T.S., and Thaxton, D.G., “Analytical solutions to a guidance problem,” Journal of Guidance, Control, and Dynamics, 14 (1991) 117–122. 5. Dhar, A. and Ghose, D., “Capture region for a realistic TPN guidance law,” IEEE Trans. Aerospace and Electronic Systems, 29 (1993) 995–1003. 6. Ghose, D., “True proportional navigation with maneuvering target,” IEEE Trans. Aerospace and Electronic Systems, 30 (1994) 229–237. 7. Ghose, D., “Capture region for true proportional navigation guidance with nonzero missdistance,” Journal of Guidance, Control, and Dynamics, 17 (1994) 627–628. 8. Ghose, D., “On the generalization of true proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 30 (1994) 545–555. 9. Guelman, M., “A qualitative study of proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 7 (1971) 638–643. 10. Guelman, M., “Proportional navigation with a maneuvering target,” IEEE Trans. Aerospace and Electronic Systems, 8 (1972) 364–371. 11. Guelman, M., “Missile acceleration in proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 9 (1973) 462–463. 12. Guelman, M., “The closed form solution of true proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 12 (1976) 472–482. 13. Guelman, M. and Shinar, J., “Optimal guidance law in the plane,” Journal of Guidance, Control, and Dynamics, 7 (1984) 471–476. 14. Mahapatra, P.R. and Shukla, U.S., “Accurate solution of proportional navigation for maneuvering targets,” IEEE Trans. Aerospace and Electronic Systems, 25 (1989) 81– 89. 15. Murtaugh, S.A. and Criel, H.E., “Fundamental of proportional navigation,” IEEE Spectrum, 3 (1966) 75–85. 16. Rao, M.N., “New analytical solutions for proportional navigation,” Journal of Guidance, Control, and Dynamics, 16 (1993) 591–594. 17. Shinar, J., Rotsztein, Y., and Bezner, E., “Analysis of threedimensional optimal evasion with linearized kinematics,” Journal of Guidance, Control, and Dynamics, 2 (1979) 353–360. 18. Shukla, U.S. and Mahapatra, P.R., “Optimalization of biased proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 25 (1989) 73–80. 19. Shukla, U.S. and Mahapatra, P.R., “The proportional navigation dilemma — pure or true?” IEEE Trans. Aerospace and Electronic Systems, 26 (1990) 382–392. 20. Vathsal, S. and Rao, M.N., “Analysis of generalized laws for homing missiles,” IEEE Trans. Aerospace and Electronic Systems, 31 (1995) 514–521. 21. Yang, C.D., Yeh, F.B., and Chen, C.H., “The closedform solution of generalized proportional navigation,” Journal of Guidance, Control, and Dynamics, 10 (1987) 216– 218. 22. Yang, C.D. and Yeh, F.B., “Closedform solution of a class of guidance laws,” Journal of Guidance, Control, and Dynamics, 10 (1987) 412–415. 23. Yang, C.D., Hsiao, F.B., and Yeh, F.B., “Generalized guidance law for homing missiles,” IEEE Trans. Aerospace and Electronic Systems, 25 (1989) 197–212. 24. Yang, C.D. and Yang, C.C., “Analytical solution of generalized threedimensional
© 2004 by CRC Press LLC
25.
26. 27. 28. 29.
30. 31. 32.
proportional navigation,” Journal of Guidance, Control, and Dynamics, 19 (1996) 721–724. Yang, C.D. and Yang, C.C., “Analytical solution of threedimensional realistic true proportional navigation,” Journal of Guidance, Control, and Dynamics, 19 (1996) 569– 577. Yang, C.D. and Yang, C.C., “Analytical solution of threedimensional true proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 32 (1997) 1509–1522. Yang, C.D. and Yang, C.C., “A unifed approach to proportional navigation,” IEEE Trans. Aerospace and Electronic Systems, 33 (1997) 557–567. Yuan, P.J. and Chern, J.S., “Analytic study of biased proprotional navigation,” Journal of Guidance, Control, and Dynamics, 15 (1992) 185–190. Yuan, P.J. and Chern, J.S., “Solutions of true proportional navigation for maneuvering and nonmaneuvering targets,” Journal of Guidance, Control, and Dynamics, 15 (1992) 268–271. Yuan, P.J. and Chern, J.S., “Ideal proportional navigation,” Journal of Guidance, Control, and Dynamics, 15 (1992) 1161–1165. Yuan, P.J. and Hsu, S.C., “Exact closedform solution of generalized proportional navigation,” Journal of Guidance, Control, and Dynamics, 16 (1993) 963–966. Yuan, P.J. and Hsu, S.C., “Solutions of generalized proportional navigation with maneuvering and nonmaneuvering targets,” IEEE Trans. Aerospace and Electronic Systems, 31 (1995) 469–474.
© 2004 by CRC Press LLC
17 Guidance Design with Nonlinear H2/H∞ Control HsinYuan Chen Department of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan
A mixed H2 /H∞ optimal control design and its application to a missile guidance problem of homing phase are studied. The problem consists of combining the performance requirements of quadratic optimal controllers with the robustness properties of H∞ controllers. Our approach has ﬁve features. (1) The complete nonlinear kinematics of the pursuitevasion motion are considered; neither linearization nor small angle assumption is made here. (2) The nonlinear H2 /H∞ guidance law is derived analytically and expressed in a very simple form; neither iterative approximation nor complicated numerical computation is required. (3) Unlike adaptive and neural guidance laws, the implementation of the proposed robust H2 /H∞ law does not need information on target acceleration. (4) Under appropriate constraints, the pair of crosscoupled HamiltonJacobi Partial Inequality (HJPDI) can be easily solved. (5) The derived nonlinear H2 /H∞ guidance law exhibits strong robustness and excellent performance against maneuvering targets.
I
INTRODUCTION
The objective of optimal control design is minimizing or maximizing some index performance by which various designs can be compared. The traditional H2 optimization attempts to minimize the energy of the system output when the system is faced with white noise inputs. The result is a controller adept at handling the problem with white noises but potentially weak in robustness characteristics and tracking performance. The H∞ optimization, on the other hand, attempts to minimize the system output energy to unknown but bounded energy inputs. The optimal H∞ optimal controller results in a highly robust system but one that can be notably deﬁcient in handling the performance of the system. The design method explored in this chapter is the mixed H2 /H∞ optimization [1] that handles performance and bounded energy input simultaneously and establishes a link between H2 and H∞ optimization. Through this method, the designer can determine the tradeoff between the performance (H2 ) and robust stability characteristics (H∞ ). Recently, there has been a great deal of interest in formulating a mixed H2 /H∞ optimal control problem. The mixed H2 /H∞ optimal control has been studied for linear systems [2–7]. Although this problem can be stated and motivated quite easily, solving it has turned out to be difﬁcult. So, very powerful algorithms and associated theory have recently been developed for convex optimization in Refs. [8] and [9]. As a result of this
© 2004 by CRC Press LLC
development, many convex optimization problems for which no traditional analytic solutions are known can be solved rapidly. In particular, the ﬁrst application of the theory of Nash games to H2 /H∞ control problems can be found in Ref. [10]. Based on this development, the mixed H2 /H∞ controllers have been extensively applied to aircraft control in Ref. [11]. Reference [12] contains examples of gain scheduling and H2 /H∞ optimal control of aircraft and missiles. In Ref. [13], it is shown that robotic systems can be robustly controlled by the H2 /H∞ performance even if the different noises are causally dependent. Additionally, in many ﬁltering problems [14,15], the applicability of this H2 /H∞ control theory is illustrated by designing a ﬁlter to estimate the states of an aircraft ﬂying through a downburst and ﬂexible structures. To our knowledge, however, this design method has not yet been discussed for nonlinear systems and even not applied to tear practical (physical) missile systems. The design objective is achieve the optimal H2 optimal performance under a desired H∞ disturbance rejection constraint. In this study, a mixed nonlinear H2 /H∞ control design is proposed for the homing missile system under unpredictable accelerations of the target (unpredictable disturbances). This design scheme is useful for robust performance design of the system under external disturbances. The solution to this nonlinear mixed H2 /H∞ control problem is explicitly presented by combining dissipative theory [16] with the LQ optimal control method. A pair of coupled nonlinear HamiltonJacobi Partial Inequality (HJPDI) must ﬁrst be solved. This subsequently leads to a sufﬁcient condition for the solvability of the missile system of the maneuvering targets via a mixed H2 /H∞ control method. Next, two nonlinear coupled HJPDIs can be transformed to corresponding algebraic inequalities via an adequate choice of the solution of HJPDI. Finally, these two coupled HJPDI can be easily solved and a simpliﬁed mixed H2 /H∞ guidance law is ultimately obtained for missile guidance systems. The robust H2 /H∞ guidance law obtained from the solution of HJPDI exhibits robustness properties against target maneuvers [17] and in the meanwhile, maintains some special performances. Comparisons with conventional proportional navigation [18] schemes show that the conventional PN schemes do perform very well in the absence of the target’s maneuvers. However, when target maneuvers are present, their performance may degrade dramatically. On the other hand, the H2 /H∞ robust guidance law does not only demonstrate strong performance robustness under the variation of target maneuvers, but also gives a better performance than the H∞ robust guidance law [19]. This chapter is organized as follows. In Sec. II, we brieﬂy survey some preliminaries of the nonlinear H2 /H∞ control theory. In Sec. III, we formulate the missile guidance problem as a nonlinear disturbance attenuation H2 /H∞ control problem. Then we solve analytically the associated HJPDIs in Sec. IV. Finally, comparison with PN schemes, robust H∞ guidance law and the robustness and performance of the H2 /H∞ guidance laws against target maneuvers are illustrated numerically in Sec. V.
II
DISSIPATION AND THE NONLINEAR H2 /H∞ CONTROL PROBLEM
In this section, we apply dissipative theory [6] to derive the nonlinear mixed H2 /H∞ control theory for later use. Consider a nonlinear control system governed by a set of differen
© 2004 by CRC Press LLC
tial equations of the form x˙ = f (x) + g1 (x)w + g2 (x)u h1 (x) z= , ρu
(1a) (1b)
where x = [x1 · · · xn ]T consists of the n independent generalized coordinates describing the motion of the system; w(t) is the input vector; z(t) is the penalized output vector, and u is the control to be designed. System (1) is said to be dissipative with respect to the supply rate r(·, ·), if there exists an energy storage function V (x) > 0, ∀x = 0, satisfying the following inequality: t 0 ≤ V (x(t)) ≤
r(w(ξ), z(ξ))dξ,
(2)
0
for all t and for all x(·), z(·), and w(·) satisfying Eq. (1). For a differentiable function V (x), an equivalent statement of dissipative condition (2) is D˙ = r(w(t), z(t)) − V˙ (x(t)) ≥ 0.
(3)
The energy dissipative rate D˙ is recognized as the energy supply rater minus the energy ∂V ∂V ∂V ∂V storage rate V˙ . The term dV (x(t))/dt = ∂x x˙ = ∂x1 ∂x2 · · · ∂x x˙ denotes the total n time derivative of V (x(t)) along the state trajectory x(t). The supply rate r∞ relating to H∞ performance is deﬁned as 2 wT w − z T z. r∞ (w(t), z(t)) = γ∞
(4)
Using r∞ in Eq. (2), we have T
T
z(t) z(t)dt ≤ 0
T
2 γ∞
T
wT (t)w(t)dt,
∀w ∈ L2 ,
(5)
0
T
where 0 z T zdt/ 0 wT wdt is known as the L2 gain of the system. Equation (5) implies that system (1) has a L2 gain lower than or equal to γ∞ . A special case of Eq. (1) is the familiar linear system: x˙ = Ax + B1 w + B2 u C1 x . z= ρu
(6a) (6b)
For linear systems, it can be shown that the L2 gain constraint (5) is equivalent to the ∞norm constraint: Hzw (s)∞ ≤ γ∞ , (7) where Hzw (s)∞ is the ∞norm of the transfer function from the input w to output z. The equivalence between Eq. (3) and Eq. (5) indicates that a system with L2 gain ≤ γ∞ is a dissipative system, and vice versa. Hence, with respect to dissipation, we can say that
© 2004 by CRC Press LLC
the nonlinear H∞ control is an essential means to make a nonlinear system dissipative. In terms of an energy storage function V = W and the supply rate r = r∞ , the H∞ constraint (5) can be rewritten by using Eq. (3) as ∂W ∂W 2 D˙ ∞ (w, u) = γ∞ wT w − g1 (x)w − ρuT u + g2 (x)u ∂x ∂x ∂W (8) − f (x) − h1 (x)T h1 (x) ≥ 0, ∀w ∈ L2 . ∂x On the other hand, the supply rate r2 relating to H2 performance is deﬁned as r2 (w(t), z(t)) =
γ22 − z T z. T
(9)
Substituting this r in Eq. (2), we have T
z(t)T z(t)dt ≤ γ22 .
(10)
0
For the linear system (6), Eq. (10) reduces to the familiar quadratic performance index: T
(xT Qx + uT Ru)dt ≤ γ22 ,
(11)
0
where Q = C1T C1 and R = ρ2 I. In terms of energy storage function V = U and supply rate r = r2 , the quadratic constraint (10) can be rewritten by using Eq. (3) as γ22 ∂U ∂U ∂U 2 T ˙ D2 (w, u) = − g1 w − ρ u u + g2 (x)u − f (x) − h1 (x)T h1 (x) ≥ 0. T ∂x ∂x ∂x (12) The purpose of the nonlinear mixed H2 /H∞ control is to ﬁnd the control u such that the H∞ criterion (5) and the H2 criterion (10) are satisﬁed simultaneously. In terms of the dissipative condition (3), the solvability of the mixed H2 /H∞ controller relies on the existence of a controller u0 and two positive energy storage functions W and U satisfying Eq. (8) and Eq. (12), i.e., D˙ ∞ (w, u0 ) ≥ 0, D˙ 2 (w, u0 ) ≥ 0,
∀w ∈ L2 ∀w ∈ L2 .
(13a) (13b)
Conditions (13a) and (13b) are known as robust H∞ constraint and robust H2 constraint, respectively. Condition (13a) can be met by requiring u0 to satisfy D˙ ∞ (w∗ , u0 ) = min D˙ ∞ (w, u0 ) ≥ 0, w
(14)
where w∗ is called the worstcase disturbance. Using this worstcase disturbance w∗ in D˙ 2 results in the worstcase H2 performance D˙ 2 (w∗ , u0 ). In conjunction with condition (13b),
© 2004 by CRC Press LLC
controller u0 need be chosen to make D˙ 2 (w∗ , u0 ) as large as possible. The maximum value that can be achieved by u0 is maxu0 D˙ 2 (w∗ , u0 ). Hence, condition (13b) requires D˙ 2 (w∗ , u∗ ) = max D˙ 2 (w∗ , u0 ) ≥ 0. u0
(15)
Combining Eq. (14) with Eq. (15), we can replace conditions (13) by D˙ ∞ (w∗ , u∗ ) ≥ 0 D˙ 2 (w∗ , u∗ ) ≥ 0,
(16a) (16b)
where the extreme functions w∗ and u∗ are deﬁned as w∗ = arg min D˙ ∞ (w, u), for ﬁxed u
(17a)
u∗ = arg max D˙ ∞ (w∗ , u).
(17b)
w
u
In the following step, we will express the worstcase disturbance w∗ and the optimal control u∗ in terms of the energy function W and U . From Eq. (8) it can be seen that D˙ ∞ is a quadratic function in w, and D˙ ∞ has a global minimum at T 1 T ∂W ∗ . (18) w (x) = 2 g1 (x) 2γ∞ ∂x By substituting this w∗ into Eq. (12), D˙ 2 (w∗ , u) becomes T γ22 ∂U ∂W 1 ∂U ∗ T 2 T ˙ − ρ u u+ D2 (w , u) = − 2 g1 g1 g2 (x)u T 2γ∞ ∂x ∂x ∂x ∂U − f (x) − h1 (x)T h1 (x) ≥ 0, ∂x
(19)
which has a global maximum at 1 u (x) = − 2 g2T (x) 2ρ ∗
∂U ∂x
T .
(20)
Substituting u∗ and w∗ into Eq. (8) and Eq. (12), we can express the conditions (16) in terms of the two unknown functions W (x) and U (x) as T T 1 ∂W 1 ∂W ∂U ∂U ∂U f− 2 g2 g2T g2 g2T + 2 ∂x 2ρ ∂x ∂x 4ρ ∂x ∂x T ∂W ∂W 1 g1 g1T + 2 + hT1 h1 ≤ 0 (21a) 4γ∞ ∂x ∂x T T 1 ∂U ∂W ∂U ∂U ∂U 1 f+ 2 g1 g1T g2 g2T − 2 ∂x 2γ∞ ∂x ∂x 4ρ ∂x ∂x γ2 +hT1 h1 − 2 ≤ 0. (21b) T The above two inequalities are known as HJPDIs whose solutions characterize the admissible storage functions W (x) and U (x). Once W and U are obtained, the mixed H2 /H∞
© 2004 by CRC Press LLC
Figure 1 Planar pursuit geometry.
control law u∗ , and the worstcase disturbance w∗ can then be determined according to Eq. (20) and Eq. (18). The corresponding HJPDIs relating to the missile guidance problem will be derived in the next section. Before closing this section, it is worth noting that for the linear system (6) the two coupled HJPDIs have a simple solution in quadratic form: W (x) = xT P x,
U (x) = xT Qx,
(22)
where P and Q are two constant positive deﬁnite matrices to be determined. Substituting Eq. (22) into Eq. (21), and replacing f (x), g1 (x), g2 (x), h1 (x) with Ax, B1 , B2 , C1 x, respectively, we have
1 P B2 B2T Q + QB2 B2T P − QB2 B2T Q 2 2ρ 1 + 2 P B1 B1T P + 2C1 C1 ≤ 0 2γ∞
1 1 QA + AT Q + 2 P B1 B1T Q + QB1 B1T P − 2 QB2 B2T Q 2γ∞ 2ρ P A + AT P −
+2C1 C1 − 2γ22 /T ≤ 0,
(23a)
(23b)
which are recognized as the coupled Riccati inequalities that appeared in the linear mixed H2 /H∞ control theory.
III
FORMULATION OF H2 /H∞ GUIDANCE PROBLEMS
The pursuit geometry of relative motion between missile and target is described by the polar coordinate (r, θ) (refer to Fig. 1). Now, we consider the relation between missile and target as point mass models, and only kinematics are considered. The governing equation
© 2004 by CRC Press LLC
in Fig. 1 can be derived as r¨ − rθ˙2 = wr − ur rθ¨ + 2r˙ θ˙ = wθ − uθ ,
(24a) (24b)
where r is the relative distance between missile and target; θ is the aspect angle of the line of sight (LOS) with respect to an inertial reference line; wr and wθ are the target’s acceleration components, which are assumed to be an unpredictable disturbance; ur and uθ are the missile’s acceleration components that are to be found. Now to transfer the governing equation (Eq. (24)) into the standard statespace equation as Eq. (1) by selecting the state variables r, Vr , and Vθ , and the transformed equations as x˙ = f (x) + g1 (x)w + g2 (x)u,
(25)
where Vr 2 f (x) = Vrθ , −Vr Vθ r
r x = Vr , Vθ
The penalized output z is chosen as z=
0 g1 (x) = −g2 (x) = 1 0
h1 (r, Vθ ) , ρu
0 0 . 1
(26)
(27)
where
1 2 (28) V = rθ˙2 r θ is a measure of guidance performance, and ρ is a weighting coefﬁcient concerning the tradeoff between guidance performance and acceleration command. It can be seen from Eq. (29) that the missile is close to the target (small r), and/or the LOS angular rates θ˙ are small if h1 can be kept small. By choosing the weighting coefﬁcient ρ properly, it is possible to obtain an acceptable small h1 without consuming a lot of acceleration u. The problem H2 /H∞ robust guidance law design can be stated as follows: To ﬁnd T the missile acceleration command u = [ur uθ ] such that the following mathematical expressions can be satisﬁed: T T T z zdt (h1 2 + ρ2 uT u)dt 2 0 , ∀w ∈ L2 (29) = 0T ≤ γ∞ T T wdt 2 + w 2 )dt w (w r θ 0 0 h1 (r, Vθ ) =
T
T
T
z zdt = 0
IV
(h1 2 + ρ2 uT u)dt ≤ γ22 .
(30)
0
SOLUTIONS OF THE HAMILTONJACOBI PDI
In this section, the associated HJPDIs for the missile guidance problem will be derived, and the corresponding solutions are derived analytically. Now, to substitute f (x), g1 (x),
© 2004 by CRC Press LLC
g2 (x), and h1 (x) from Eq. (26), Eq. (27), and Eq. (28) into Eq. (21), resulting in ∂W ∂U ∂W ∂U Vr Vθ ∂W 1 ∂W 1 ∂W + Vr − − 2 + Vθ2 ∂r r ∂Vr r ∂Vθ 2ρ ∂Vr ∂Vr ∂Vθ ∂Vθ 2 2 2 2 1 1 ∂U ∂W 1 ∂U ∂W + 2 + 2 Vθ4 ≤ 0. + + + 2 4ρ ∂Vr ∂Vθ 4γ∞ ∂Vr ∂Vθ r (31) The second set of HJPDI can be obtained by substituting f (x), g1 (x), g2 (x), and h1 (x) from Eq. (26), Eq. (27), and Eq. (28) into Eq. (21), resulting in ∂U ∂W ∂U ∂W ∂U −Vr Vθ ∂U 1 1 2 ∂U + + + 2 + Vθ Vr ∂r r ∂Vr r ∂Vθ 2γ∞ ∂Vr ∂Vr ∂Vθ ∂Vθ 2 2 ∂U 1 ∂U 1 + 2 Vθ4 − γ22 /T ≤ 0. + − 2 4ρ ∂Vr ∂Vθ r (32) Both are nonlinear ﬁrstorder partial differential inequalities with the unknown function U (r, Vr , Vθ ) and W (r, Vr , Vθ ). Once qualiﬁed U and W are determined, the nonlinear H2 /H∞ robust guidance law can be given by Eq. (18) as T ∂U 1 1 ∂U/∂Vr . (33) = 2 u(r, Vr , Vθ ) = − 2 g2T ρ ∂x ρ ∂U/∂Vθ And ur = ρ12 ∂U/∂Vr , and uθ = ρ12 ∂U/∂Vθ . Here, we assume the optimal pursuing decision confronts the optimal escaping decision, i.e., assuming U = W . Both can be simpliﬁed as 2 2 ∂U 1 2 ∂U Vr Vθ ∂U 1 ∂U ∂U Vr − − 2 + + Vθ ∂r r ∂Vr r ∂Vθ 4ρ ∂Vr ∂Vθ 2 2 ∂U 1 1 ∂U + 2 Vθ4 ≤ 0 + 2 + (34) 4γ∞ ∂Vr ∂Vθ r and
2 2 ∂U ∂U Vr Vθ ∂U 1 ∂U 1 2 ∂U Vr − + 2 + + Vθ ∂r r ∂Vr r ∂Vθ 2γ∞ ∂Vr ∂Vθ 2 2 ∂U 1 1 ∂U + 2 Vθ4 − γ22 /T ≤ 0. − 2 + 4ρ ∂Vr ∂Vθ r
(35)
It is noted that a function U satisfying Eq. (34) and Eq. (35) is also a solution of the HJPDIS in Eq. (31) and Eq. (32), but it is not unique. Now, we can see these HJPDIs clearly, and ﬁnd that the solution of HJPDIs can be substituted by the solution of the following inequality: 2 2 2 1 ∂U ∂U 1 2 ∂U Vr Vθ ∂U 1 1 ∂U Vr + 2 Vθ4 ≤ 0. − + − 2 + + Vθ 2 ∂r r ∂Vr r ∂Vθ 4 γ∞ ρ ∂Vr ∂Vθ r (36)
© 2004 by CRC Press LLC
This solution is developed by the dissipative rate concept. This solution of Eq. (34) and Eq. (35) can be found readily as U (r, Vr , Vθ ) = W =
λVr Vθ2 ; r
γ2 ≥ 0.
(37)
When Vr < 0, λ < 0, and Vr > 0, λ > 0. To check whether the U given by Eq. (37) is feasible or not, we substitute Eq. (37) into the HJPDI (36) to obtain the expression (1) Vr < 0 2 2 2 1 1 λ 2 λ2 − 3λ V + λ + 1 Vθ2 ≤ 0. − + − r 2 2 γ∞ ρ2 4 γ∞ ρ2
(38)
This condition is satisﬁed for arbitrary Vr (t), and Vθ (t), if each leading coefﬁcient of the two terms in Eq. (38) is negative, i.e.,
λ2 4
1 2 − 3λ ≤ 0 − 2 γ∞ ρ2 2 1 + λ + 1 ≤ 0. − 2 γ∞ ρ2
λ2
(39a) (39b)
We ﬁnd that the intersection of the two inequalities exists if and only if γ∞ > this condition, the intersection gives two admissible ranges for λ:
√
2ρ. Under
−2 √
−1/2 2 − 2 1 + ρ−2 − 2γ∞ (A) λ ≤ , when γ∞ ≥ 2 ρ−2 − 21/4 , ρ < 4/21 −2 −2 ρ − 2γ∞ (40a) √ −2 √
−1/2 3 2ρ < γ∞ < 2 ρ − 21/4 , ρ < 4/21. (B) λ ≤ − −2 −2 , when (ρ − 2γ∞ ) (40b) (2) Vr > 0 Under this condition, the range will become as follows: 0≤λ≤
2+2
−2 1 + ρ−2 − 2γ∞ −2 ρ−2 − 2γ∞
when γ∞ ≥
√
2ρ.
(41)
As mentioned earlier, γ∞ is an index for the disturbance attenuation level. Hence, a smaller γ∞ means that the interceptive performance is less sensitive to target maneuvers. From Eq. (38) we can see that the γ∞ achieved by using the navigation gain (A) is larger than that by using the navigation gain (B). Up to now, we have proved that the functions U and W in Eq. (29) and Eq. (30), respectively, are truly solutions of the HJPDIs. Once the storage functions are determined, the resulting H2 /H∞ guidance law can be obtained from Eq. (35) as Vθ2 r 2 V r Vθ uθ = λ 2 , ρ r ur = λ
© 2004 by CRC Press LLC
(42) (43)
and then from Eq. (16), the worstcase maneuvering target turned out to be at λV 2 ∂U 1 = 2θ wr = 2 γ∞ ∂Vr γ∞ r 2λVr Vθ 1 ∂U = . wθ = 2 2 r γ∞ ∂Vθ γ∞
(44) (45)
γ∞ is not only an index of disturbance attenuation level but also an important parameter describing the worstcase target’s maneuvering ability. It can be seen that as γi −→ ∞, the effect of disturbance attenuation entirely disappears, and the worstcase target that the missile can handle becomes a nonmaneuvering target, i.e., wr = wθ = wφ = 0. Decreasing γ∞ means that the missile’s robustness is increased and the allowable target acceleration that can be captured by the missile is also increased. V
NUMERICAL RESULTS
In this section, maneuvering targets are used, and the mean value of this target is used to test the robustness and robust performance of the nonlinear 2D H2 /H∞ robust guidance law. In this simulation, we need to normalize the governing equations to their dimensionless forms. (1) The relative velocities V r and Vθ are normalized to V r = Vr /V0 , and V θ = Vθ /V0 , respectively, where V0 =
Vr20 + Vθ20 .
(2) The acceleration components ur and uθ are normalized to ur = ur /(V02 /r0 ), and uθ = uθ /(V02 /r0 ), respectively. Now the performance and robustness of four missile guidance laws will be tested and compared. Their acceleration commands are listed as below. 2
(M1) H∞ Guidance Law [18] ur = λV θ /(ρ2 r), and uθ = 2λV r V θ /(ρ2 r), where λ is a constrained variable. (M2) Pure proportional navigation [19] (PPN) guidance law: ur = −λV M (dθ/dτ ) sin(kθ + α0 ), uθ = λV M (dθ/dτ ) cos(kθ + α0 ), where V M is the normalized missile’s velocity; k = λ − 1 with λ being the navigation gain; α0 is the initial aspect M with respect to the inertial reference line (see Fig. 1). angle of V 2
(M3) Nonlinear H2 /H∞ Guidance Law: ur = λV θ /(ρ2 r), and uθ = 2λV r V θ /(ρ2 r). The target commands are the following: (T1) Sinusoidal target: wr = wθ = λT sin(τ dθ/dτ ) (T2) Ramp target: wr = wθ = λT τ (T3) Step target: wr = 0, wθ = λT The robustness and desired performance of the nonlinear H2 /H∞ guidance law can be illustrated in the following aspects: 1. Robustness properties 2. Comparison of the quadratic performance
© 2004 by CRC Press LLC
Figure 2 Robustness of L2 gain for nonlinear H2 /H∞ guidance law.
(1) Robustness properties Depending on the deﬁnition of the system’s L2 gain, we know that γ∞ is small, and indicated that good robustness can be preserved in the presence of unknown target maneuvers. We ﬁnd the H∞ guidance law has better robustness than the H2 /H∞ guidance law, but the H2 /H∞ has better expression. In Fig. 2, when the navigation of the target’s acceleration, its associated initial conditions (Fig. 3) increase, H2 /H∞ guidance law (M3) shows its robustness against the maneuvering target, which can keep the L2 gain smaller than γ∞ .
(2) Comparison of the quadratic performance Here, the quadratic performance is used as an index to reﬂect the interceptive performance. It is shown in Fig. 4 that as the navigation of the target’s acceleration increases, the H2 /H∞ guidance law maintains excellent performance with output energy (i.e., z2 ) smaller than 0.4, but the performance of the PPN becomes worse as the variation of target acceleration changes.
VI
CONCLUSIONS
A nonlinear mixed H2 /H∞ control has been proposed in this chapter for the ﬁrst time for robustness of missile guidance design under maneuvering targets. This nonlinear mixed H2 /H∞ robust control design problem must ﬁrst solve two coupled nonlinear HamiltonJacobi Partial Inequalities. In order to avoid the difﬁculty of solving these two coupled nonlinear HJPDIs, some adequate assumptions are also made for the solution of coupled
© 2004 by CRC Press LLC
¯ 0 for four guidance laws. Figure 3 System L2 gain vs. h
Figure 4 Comparison of performances for four guidance laws.
HJPDIs to obtain the mixed H2 /H∞ guidance law. This solution is optimal and of closed form, and a simple design procedure is also proposed.
© 2004 by CRC Press LLC
Acknowledgment I sincerely thank the National Scientiﬁc Council, Taiwan, for its kind support (NSC 892218E035022).
REFERENCES 1. Lin, W., “Mixed H2 /H∞ control via state feedback for nonlinear systems,” International Journal of Control, 1996, Vol. 5, pp. 899–922. 2. Berstein, D.S., and Haddad, W.M., “LQG control with an H∞ performance bound: a Riccati equation approach,” IEEE Transactions on Automatic Control, 1989, Vol. 34, pp. 293–305. 3. Doyle, J., Glover, K., Khargonekar, P.P., and Francis, B.A., “Statespace solutions to standard H2 and H∞ control problem,” IEEE Transactions on Automatic Control, 1989, Vol. 34, pp. 831–847. 4. Doyle, J., Zhou, K., Khargonekar, P.P., and Francis, B.A., “Mixed H2 and H∞ performance objectives II: optimal control,” IEEE Transactions and Automatic Control, 1994, Vol. 39, pp. 1575–1587. 5. Khargonekar, P.P. and Rotea, M.A., “Mixed H2 /H∞ control: a convex optimization approach,” IEEE Transactions on Automatic Control, Vol. 36, 1991, pp. 824–837. 6. Mosca, E., Casavola, A., and Giarre, L., “Minimax LQ stochastic tracking and servo problems,” IEEE Transactions on Automatic Control, 1990, Vol. 35, pp. 95–97. 7. Zhou, K., Glover, K., Bodenheimer, B., and Doyle, J., “Mixed H2 and H∞ performance objectives I: robust performance analysis,” IEEE Transactions on Automatic Control, 1994, Vol. 39, pp. 15641574. 8. Hardt, M., Helton, J.W., and Delgado, K.K., “Numerical solution of nonlinear H2 and H∞ control problems with application to jet engine control,” Proceedings the 36th Conference on Decision and Control, 1997, Vol. 3, pp. 2317–2322. 9. Fan, Y., Cliff, M.M., Lutze, F.H., and Anderson, M.R., “Mixed H2 /H∞ optimal control for an elastic aircraft,” Journal of Guidance, Control, and Dynamics, 1996, Vol. 19, pp. 650–655. 10. Limebeer, D.N., Anderson, B.D.O., and Handel, B., “Nash game approach to mixed H2 /H∞ control,” IEEE Transactions on Automatic Control, 1994, Vol. 39, pp. 69–82. 11. Ozbay, H. and Bachmann, G.R., “H2 /H∞ controller design for a twodimensional thin airfoil ﬂutter suppression,” Journal of Guidance, Control, and Dynamics, 1994, Vol. 4, pp. 722–728. 12. Shue, S.P., Sawan, M.E., and Rokhsaz, K., “Mixed H2 /H∞ method suitable for gain scheduled aircraft control,” Journal of Guidance, Control, and Dynamics, 1997, Vol. 4, pp. 699–706. 13. Chen, B.S. and Chang, Y.C., “Nonlinear mixed H2 /H∞ control for robust tracking design of robotic systems,” International Journal of Control, 1997, Vol. 6, pp. 837– 857. 14. Rotstein, H., Sznaier, M., and Idan, M., “H2 /H∞ ﬁltering theory and an aerospace application,” International Journal of Robust and Nonlinear Control, 1996, Vol. 6, pp. 347–366.
© 2004 by CRC Press LLC
15. Gawronski, W. and Lim, K.B., “Frequency weighting for the H∞ and H2 control design of ﬂexible structures,” Journal of Guidance, Control, and Dynamics, 1998, Vol. 21, pp. 664–666. 16. Willems, J.C., “Dissipative dynamical systems part I: general theory,” Arch. Rational Mechanics and Analysis, 1972, Vol. 5, pp. 321–351. 17. Austin, F., Carbone, G., Falco, M., and Hinz, H., “Game theory for automated maneuvering during airtoair combat,” Journal of Guidance, Control, and Dynamics, 1990, Vol. 6, pp. 1143–1149. 18. Yang, C.D. and Yang, C.C., “A uniﬁed approach to proportional navigation,” IEEE Transactions on Aerospace and Electronic Systems, 1997, Vol. 2, pp. 557–567. 19. Yang, C.D. and Chen, H.Y., “Nonlinear H∞ robust guidance law for homing missiles,” Journal of Guidance, Control, and Dynamics, 1998, Vol. 6, pp. 882–890.
© 2004 by CRC Press LLC
18 A Control Algorithm for Nonlinear Multiconnected Objects V.Yu. Rutkovsky, S.D. Zemlyakov, V.M. Sukhanov, and V.M. Glumov Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow, Russia
The control of mechanical objects is considered here. A mathematical model of such objects is described by nonlinear Lagrange equations of the second type. The control algorithms, which under the conditions of restricted control actions ensure the separate control for each of the generalized coordinates, monotonic character and guaranteed response speed of the transient processes are synthesized. Such severe requirements for the motion of a nonlinear multiconnected control system are fulﬁlled when the mathematical model complies with some conditions. But these conditions take place for the considered sets of space vehicles.
1
THE PROBLEM MOTIVATION
Airplanes, space vehicles, abovewater and submarine aids, means of transport, robotsmanipulators, and others as the control objects are mechanical systems. As a rule, a mathematical model (MM) of such objects is derived on the basis of Lagrange equations of the second type: d ∂T ∂T − = Q, dt ∂ q˙ ∂q
(1)
where q, Q = Q(q, q, ˙ t) ∈ Rn are the vectors of generalized coordinates and generalized forces; T = T (q, q) ˙ is the kinetic energy of the mechanical system. In particular the equation of freeﬂying space robotic module [1, 2] is as follows: A(q)¨ q+
n
˙ t), q˙T Ds (q)q˙ es = Q(q, q,
(2)
s=1
where A(q) is the positive deﬁnite matrix; Ds (q) is the matrix of Christoffel’s deltas of the T ﬁrst type; es = (e1 , e2 , ..., en ) – vector with zero components with the exception of only one component es = 1; T is the symbol of transposition. Equation (2) represents the nonlinear multiconnected, in a general case, nonstationary MM of a mechanical system. The control of this object is realized with the help of purposeful action on the vector of generalized forces Q(q, q, ˙ t).
© 2004 by CRC Press LLC
For many mechanical objects, for example, which were mentioned above, a high guaranteed accuracy of control is required. Thus the freeﬂying space robotic module that is intended for servicing the manned orbital station must realize all operations not only with very high accuracy but with the predicted dynamics (predicted dynamics is the preassigned form of the transient processes for all variable coefﬁcients of the module’s nonstationary mathematical model). This requirement is dictated both by the safety of the manned orbital station, other space objects, and by the safety of the module itself. In this chapter the task of the design of the control algorithms for the object with a mathematical model (2) is considered. The algorithms must realize monotonic tracking with guaranteed response speed of each generalized coordinate qi = qi (t) (i = 1, n) after the prescribed function qipr (t). Moreover this tracking must be executed at the restricted control action and independently on the presence or the absence of the control actions for the other generalized coordinates. It is clear that this task is very difﬁcult for such a complicated object. For the solution of this task, mathematical model (2) will be accepted with the following simpliﬁcations: – It is assumed that the system has for each coordinate qi (i = 1, n) the independent control action Mi and this control action is restricted, i.e., Mi  Mimax , Mimax = const > 0. – It is assumed that the velocities q˙i (i = 1, n) are small and the second term on the lefthand side of Eq. (2) can be neglected (in many mechanical systems the special automatic limiters are introduced for the restriction of the velocities q˙i ). – It is assumed that the object moves under the conditions of the weightlessness and the damping forces’ absence in the airless space (for example, freeﬂying space robotic module [2]). – Let us introduce a further assumption. In this chapter the external disturbances will not be taken into account. This assumption is not a matter of principle and is introduced only for the brevity of the mathematical calculations (for the freeﬂying space robotic module these disturbances are very small). The aforementioned assumptions allow one to simplify the mathematical model (2) of the mechanical system rather essentially and to represent it as A(q)¨ q = D(q)M,
(3)
where M is the vector of control actions, which are created by the control system, i.e., M T = (M1 , M2 , . . . , Mn ),
(4)
D(q) is the matrix, which deﬁnes the interconnection of the generalized coordinates’ motion by means of control actions. Let us denote by M max the vector of the restrictions on the control actions, i.e., T
M max = (M1max , M2max , . . . , Mnmax ).
(5)
It is important to note that the mechanical system with the model (3) in spite of the simpliﬁcation is nonlinear and multiconnected.
© 2004 by CRC Press LLC
2
FORMULATION OF THE PROBLEM
It will be supposed that the components of the vectors q(t) and q˙ = q(t) ˙ are measured. The control law M = M (q, q, ˙ t) must be synthesized so that it could realize for each coordinate monotonic motion with guaranteed response speed and these requirements must be fulﬁlled by properly choosing only one control action, i.e., Mi = Mi (qi (t), q˙i (t), t).
(6)
Deﬁnition 1. The motion qi = qi (t) (i = 1, n) is called to be monotonic if the coordinate qi moves from the initial point (qi (t0 ) = qi0 , q˙i (t0 ) = 0) at t = t0 into the new preassigned position (qi (t1 ) = qi1 , q˙i (t1 ) = 0) at t = t1 (t1 > t0 ) without changing the sign of the coordinate’s velocity q˙i (t) on the time interval [t0 , t1 ]. Deﬁnition 2. The motion qi (t) is called as one with guaranteed response speed if the coordinate qi (i = 1, n) gets over from the initial point (qi (t0 ) = qi0 , q˙i (t0 ) = 0) at t = t0 into the new preassigned√position (qi (t1 ) = qi1 , q˙i (t1 ) = 0) at t = t1 (t1 > t0 ) during the time Ti (∆qi ) λi ∆qi , where ∆qi = qi (t1 ) − qi (t0 ), λi = const > 0 is the preassigned constant value.
3
AUTONOMOUS TECHNICAL CONTROLLABILITY OF AN OBJECT
The solving of the formulated task is essentially based on the concept of autonomous technical controllability (AT controllability). Deﬁnition 3. The control object with the mathematical model (3) is called autonomous technical controllable (AT controllable) with respect to the coordinate qi = qi (t) (i = 1, n) at the position q = q ∗ of phase space q, if with the applying at any moment t = t∗ the control action Mi , which is equal to its maximum value, i.e., Mi (t∗ ) = Mimax , the second derivative q¨i of the coordinate qi obtains the value  q¨i (t∗ )  ρ0i , ρ0i = const > 0 and the sign of the q¨i (t∗ ) is the same as the sign of the control action Mi (t∗ ) independently of the presence or absence of the control actions applying to the other coordinates. Let us ﬁnd out the conditions at which the object (3) is AT controllable. Beforehand we will introduce the following. Deﬁnition 4. Quadratic matrix B is called the matrix with diagonal prevalence if its elements bij (i, j = 1, m) are real and possess the following properties: 1) bii > 0; bij 0 (j = 1, m; j = i); 2) There exists the positive vector γ ∈ Rm , γ T = (γ1 , γ2 , . . . , γm ), m γ > 0 such that γi bii > γj bij  . j = 1; (j =i)
In this chapter we consider any matrix B = (bij ) (i = 1, m; j = 1 , n) with the real elements to be positive (nonnegative), if bij > 0 (bij 0 ) [3]. On the basis of deﬁnition 4 the following lemma can be formulated. Lemma. The matrix B = (bij ) (i, j = 1, m ) with the diagonal prevalence possesses an inverse matrix and B −1 0 .
© 2004 by CRC Press LLC
The lemma is proved in the Appendix. Let us represent model (3) in the following form: q¨ = R(q)M,
(7)
where R(q) = A−1 (q)D(q), R(q) = (rij (q)) (i, j = 1, n). Let us consider the matrix S(q) = (sij (q)) with elements sii (q) = rii (q), sij (q) = − rij (q) j = 1, n; j = i . According to deﬁnition 3, a necessary condition of AT controllability is the requirement of rii (q) > 0 (i = 1, n) but the elements rij (q) (j = 1, n; j = i ) can be positive, negative, and zero. It is evident that the requirement of the object’s AT controllability is reduced to the following relation: S(q ∗ )M max ρ0 , (8) i.e., to the existence of the positive vector M max > 0 for each positive vector ρ0 at which the inequality (8) occurs. On the basis of the lemma the following theorem can be proved. Theorem. The necessary and sufﬁcient conditions for AT controllability of the object with the mathematical model (3) in the position q = q ∗ is the belonging of the matrix S(q ∗ ), which is obtained from the matrix R(q ∗ ) to the set of the matrices with diagonal prevalence. The theorem is proved in the Appendix. From the theorem the importance for practice consequence can be obtained: Consequence. Among all the vectors M max complying with the condition of AT controllability of the object with mathematical model (3) at the position q = q ∗ , the vector M max deﬁned by the equality M max = S −1 (q ∗ ) ρ0 is minimal with respect to the norm.
4
WORKING DOMAIN OF THE OBJECT’S MOTION
According to the assumption on the restrictions of the coordinates’ velocities, q˙i  q˙i0 , q˙i0 = const > 0
(i = 1, n),
(9)
the domain of the deﬁnition for the object with mathematical model (3) in the space {q, q} ˙ is limited by a cylindrical one: G(q, q) ˙ = q, q˙ q˙i  q˙i0 , i = 1, n . (10) A formulated theorem enables one for every position q = q ∗ , q ∗ ∈ G (q, q) ˙ to answer the question: Is the object at this position AT controllable or not? Let us notice that according to the theorem the object’s AT controllability in the position q ∗ ∈ G (q, q) ˙ does not depend on the control actions Mi = Mi (q, q, ˙ t). It is deﬁned only by the object’s constructive parameters, i.e., the concrete object is AT controllable in the position q* or it is not AT controllable. In the latter case it is impossible to make the object to be AT controllable by the variations of the vector M max of control restrictions. For this purpose it is necessary to change the constructive parameters of the object. If the object is AT controllable we can get any desired vector ρ0 (it is the vector of AT controllability degree) by changing of vector’s M max control restrictions.
© 2004 by CRC Press LLC
Thus in the domain of the deﬁnition of the object’s motion G (q, q), ˙ it is possible to select a domain G0 (q, q) ˙ ⊂ G (q, q) ˙ in which the object is AT controllable in each position q ∗ ∈ G0 (q, q). ˙ But in each position q ∗ ∈ G0 (q, q) ˙ the different vector M max of control restrictions (5) exists guaranteeing the desired degree of AT controllability. From the practice point of view an inverse task can be more important: for the object with the mathematical model (3) the certain vector M max = Momax of control restrictions and the certain vector of AT controllability degree ρ0 are given. Can the preassigned vector Momax guarantee the preassigned degree ρ0 of AT controllability in every position q∗ ∈ G0 (q, q)? ˙ If not then does a domain GW (q, q) ˙ ⊂ G0 (q, q) ˙ with the aforementioned property exist? Deﬁnition 5. We denote as the working domain Gw (q, q) ˙ for the object with the mathematical model (3) a set of points of the space {q, q} ˙ including the point (0,0) for which the same constant vector M max = Momax , Momax > 0 of control restrictions guaranteeing the preassigned degree of AT controllability exists, i.e., in any position q ∗ ∈ Gw (q, q) ˙ the degree of AT controllability ρ(q ∗ , M0max ) ρ0 , where ρ0 > 0 is the preassigned constant vector. In this chapter we will not consider the very interesting and rather nontrivial question about the existence and the methods of ﬁnding out the object’s working domains Gw (q, q), ˙ but we will assume that the working domain exists.
5
THE SOLVING OF THE FORMULATED PROBLEM
The solving of the problem formulated in Section 2 is given by the following. Statement. The object with the mathematical model (3) and with the control law: Mi = Mimax sign (ui );
ui = [(qi0 − qi ) −
(q˙i )2 sign (q˙i )] (i = 1 , n), 2ρ0i
(11)
where qi0 = const is the required value of the coordinate qi , under the condition that the points of the interval [qi (t0 ) , qi0 ] belong to the object’s working domain Gw (q, q), ˙ has monotonic transient processes for each coordinate qi with the guaranteed time Ti (∆qi ) of motion from the position (qi (t0 ) = qi0 ; q˙i (t0 ) = 0 ) into the position (qi (t1 ) = qi0 ; q˙i (t1 ) = 0 ) (t1 > t0 ), which is deﬁned by inequality Ti (∆qi ) Ti0 (∆qi ), ∆qi = 0 0  qi − qi (t0 ) , Ti (∆qi ) = 4∆qi /ρ0i . In the opinion of the authors the proof of this statement includes some interesting facts. So we will consider it in more detail. Let us write the motion equations deﬁning the control system of the object (7) in the following form: q¨i =
n
rij (q)Mj
(i = 1, n),
j=1
Mj = Mjmax sign (uj ), uj = (qj0 − qj ) −
© 2004 by CRC Press LLC
2
(q˙j ) sign (q˙j ). 2ρ0j
(12)
Let us introduce the notation xj = qj0 − qj
(j = 1, n),
(13)
and taking into account qj0 = const we rewrite the system of Eq. (12) in the following form: x ¨i = −rii (q)Mimax sign (ui ) −
n
rij (q)Mjmax sign (uj )
(i = 1, n),
(14)
j=1 (j=i)
or in the other form:
n max x ¨i = − (q)M + rij (q)Mjmax sign (uj )sign (ui ) r ii i · sign (ui )
(i = 1, n).
j=1 (j=i)
(15) Let us introduce the notation ρi (t) = rii (q)Mimax +
n
rij (q)Mjmax sign (uj )sign (ui )
(i = 1, n),
(16)
j=1 (j=i)
and again we rewrite the system (15) as x ¨i = −ρi (t)sign (ui ), ui = xi +
(x˙ i )2 sign (x˙ i ) (i = 1, n). 2ρ0i
(17)
In accordance with the assumption that the motion of the control system described by ˙ the following inequality occurs: Eq. (15) takes place in the working domain Gw (q, q), ρi (t) ρ0i (i = 1, n).
(18)
Let us note a very important result: nonlinear multiconnected system (12) from the mathematical point of view is decomposed on separate subsystems for each generalized coordinate qi (i = 1, n) and moreover each subsystem is described by an analogous nonlinear and nonstationary equation from the system of Eq. (17). Further we will prove that in each subsystem with the equation of motion of the type (17) under condition (18), the transient processes are monotonic with preassigned response speed. Let us consider the case: ρi (t) ≡ ρ0i (i = 1, n),
(19)
and represent for this case the system (17) as follows: x˙ i1 = xi2 , x˙ i2 = −ρ0i sign (ui ), ui = xi1 +
© 2004 by CRC Press LLC
x2i2 sign (xi2 ), 2ρ0i
(20)
where the notations xi = xi1 , x˙ i = xi2 are introduced. On the phase plane with coordinates xi1 , xi2 the representative point of the system (20) is moving from the position xi1 (0) = x0i1 , xi2 (0) = 0 into the origin of coordinates xi1 = 0, xi2 = 0 along the optimal trajectory with minimal control time [4]: 4x0i1  Ti0 (x0i1 ) = , (21) ρ0i and the transient process is monotonic. Now let us consider the case ρi (t) ρ0i .
ρi (t) ≡ / 0,
(22)
For this case we will prove that the transient process is monotonic and the control time is Ti (x0i1 ) Ti0 (x0i1 ). (23) According to the works [5, 6] for the system, x˙ i1 = xi2 , x˙ i2 = −ρi (t)sign (ui ), ui = xi1 +
x2i2 sign (xi2 ); 2ρ0i
(24)
we will consider Lyapunov’s function Vi (xi1 , xi2 ) =
1 2 S (xi1 , xi2 ), 2 i
where Si (xi1 , xi2 ) = xi1 +
x2i2 sign (xi2 ). 2ρ0i
(25)
(26)
Lyapunov’s function is positive at any point of the phase plane {xi1 , xi2 } with the exception of the points of the switching line: Si (xi1 , xi2 ) = 0,
(27)
in which it is zero. The derivative of Lyapunov’s function (25) by virtue of the system (24) for any pointofphase plane {xi1 , xi2 } with the exception of the points of the switching line (27) is as follows: d dSi (xi1 , xi2 ) Vi (xi1 , xi2 ) = Si (xi1 , xi2 ) . (28) dt dt At ﬁrst let us consider the halfplane xi2 > 0. For the derivative
dSi (xi1 ,xi2 ) dt
we will have the correlation ρi (t) dSi (xi1 , xi2 ) = xi1 1 − 0 sign (Si (xi1 , xi2 )) . dt ρi
© 2004 by CRC Press LLC
(29)
(30)
Taking into account the condition (29) and the inequality ρi (t) > 1, ρ0i
(31)
from the correlation (30) we obtain two inequalities: dS < 0 for S > 0, dt dS > 0 for S < 0. dt
(32)
Further we will consider the halfplane xi2 < 0. On this halfplane the following correlation takes place: dSi (xi1 , xi2 ) ρi (t) = xi2 1 + 0 sign (Si (xi1 , xi2 )) . dt ρi
(33)
(34)
For this case under conditions (31) and (33) from equality (34) we also can obtain two inequalities of form (32). Thus for any pointofphase plane {xi1 , xi2 } with the exception of the points on the switching line (27) the inequalities (32) take place. These inequalities are the conditions of arising of the sliding mode on the switching line (27). Since the instant when the representative point gets on this switching line the motion goes along the line (27) into the origin of the coordinate, which is the stable equilibrium. Because the trajectories of the system (24) on the phase plane {xi1 , xi2 } having the beginning in any initial point (xi1 (0) = x0i1 , xi2 (0) = 0) by all means intersect the switching line (27) without changing the coordinate’s xi2 sign, the monotonic character of the transient processes is proved. Now let us prove that the transient processes have the guaranteed response speed. As the transient processes are monotonic the representative points of the systems (24) and (20) overcome the equal distance x0i1 at their motion from the initial position xi1 (0) = x0i1 , xi2 (0) = 0 to the origin of coordinates. But it is obvious that under condition (31) the module of the velocity, xi2 , of the system’s (24) representative point in any moment of the transient process is more than the module of the velocity, xi2 , of the representative point of the system (20), i.e., the system (24) overcomes the required distance more quickly. However for system (20) the control time is determined by expression (21). So we have proved that the control time in system (24) satisﬁes condition (23). Thus the statement is proved completely.
6
CONCLUSION
For nonlinear multiconnected objects with mathematical model (3) and with the restricted control actions the control law of form (11) is suggested. Under certain conditions formulated in this chapter the system with such control law is separated into some subsystems
© 2004 by CRC Press LLC
with mathematical models of form (17). Transient processes in each subsystem are monotonic and have a guaranteed response speed. We would like to note another result that is important for practice: giving vector M max of control restrictions we can point out the guaranteed response speed of all the subsystems and vice versa being given by the desired response speed of the subsystems we can ﬁnd the minimal values of the control vector elements restrictions at which this response speed will take place.
APPENDIX The proof of the lemma. Let us introduce the matrices Γ and L, Γ = diag(γ1 , γ2 , . . . , γm ) and L = diag (b11 , b22 , . . . , bmm ) where bii (i = 1, m) are the diagonal elements of the matrix B, and write the matrix product in the form BΓ = ΓL(E − Q), where E is the identity matrix, Q = (qij ) ( i, j = 1, m), qii = 0 ; qij = γj bij /γi bii (j = m 1, m; j = i ). Matrix Q is nonnegative and j = 1 qij < 1 for any i = 1, m. In this case according to the Perron–Frobenius theorem [3] the matrix (E − Q)−1 exists, and moreover (E − Q)−1 0. From the correlation BΓ = ΓL(E − Q) we obtain B −1 = Γ(E − Q)−1 (ΓL)−1 , B −1 0 . The proof of the theorem. The proof of the necessity. Let the object with a mathematical model of form (3) be AT controllable in the position q = q ∗ . Given vector M max as the vector γ in deﬁnition 4 we obtain that the matrix S(q ∗ ) is relative to the class of the matrices with diagonal prevalence. The proof of the sufﬁciency. Let matrix S(q ∗ ) be relative to the class of the matrices with diagonal prevalence. Then according to the lemma the matrix S −1 (q ∗ ) 0 exists. In such case for any ρ0 > 0 the vector M max > 0 exists, determined by the correlation M max = S −1 (q ∗ ) ρ0 , which satisﬁes the condition (8) of AT controllability. The theorem is proved.
Acknowledgments The work is supported by the Department of Energetics, Machinebuilding, Mechanics and Control Processes of RAS (project 14) and by the Russian Foundation for Basic Research (project 030100062) for which the authors are grateful.
REFERENCES 1. Kulakov F.M. Supervisory Control of Manipulator Robots, Moscow: Nauka, 1980 (in Russian). 2. Rutkovsky V.Yu. and Sukhanov V.M. “A dynamic model of a freeﬂying space robotic engineering module,” Automation and Remote Control, Vol 61, N 5, Part 1, 2000, pp. 749–767. 3. Gantmakher F.R. Matrices’ Theory, Moscow: Nauka, 1967 (in Russian). 4. Boltjansky V.G. Mathematical Methods of Optimal Control, Moscow: Nauka, 1973 (in Russian).
© 2004 by CRC Press LLC
5. Emelyanov S.V. Automatic Control Systems with Variable Structure, Moscow: Nauka, 1967 (in Russian). 6. The Theory of the System with Variable Structure, Ed. S.V. Emelyanov. Moscow: Nauka, 1970 (in Russian).
© 2004 by CRC Press LLC
19 Optimal Control and Differential Riccati Equations under Singular Estimates for eAtB in the Absence of Analyticity Irena Lasiecka and Roberto Triggiani Department of Mathematics, University of Virginia, Charlottesville, VA
In this chapter, we study the quadratic optimal control problem over a ﬁnite time horizon in the case where the free dynamics operator A and the control operator B yield a singular estimate for eAt B. Here, eAt is the corresponding s.c. semigroup which, by assumption, is not analytic. The resulting abstract model covers systems of coupled Partial Differential Equations, which possess an analytic component, but which are not themselves analytic. Several applications are given to hyperbolic/parabolic structural acoustic problems, to thermoelastic problems, and to sandwich beam problems. As to structural acoustic problems, we have that a hyperbolic PDE (a wave equation within an acoustic chamber) is coupled with a parabolic PDE (the ﬂexible wall), which is either modeled by an elastic equation with structural damping [1], [5], or else by a thermoelastic equation with no rotational inertia [18–21], [23], [16, Vol. I], or else by a sandwich beam [11], [28].
1
MATHEMATICAL SETTING AND FORMULATION OF THE CONTROL PROBLEM
Dynamical model Let U (control), Y (state) be separable Hilbert spaces. In this chapter, we consider the following abstract state equation: y(t) ˙ = Ay(t) + Bu(t) + w(t) on, say, [D(A∗ )] ;
(1.1)
subject to the following assumptions, to be maintained throughout the chapter. (H.1) A : Y ⊃ D(A) → Y is the inﬁnitesimal generator of a strongly continuous (s.c.) semigroup eAt on Y . Without loss of generality, at the price of replacing A with a suitable translation of A, we may assume that A−1 ∈ L(Y ). (H.2) B is a linear operator U = D(B) → [D(A∗ )] , the dual space of the domain D(A∗ ), with respect to the pivot space Y . Here A∗ is the adjoint of A in Y . Thus, eAt can be extended as an s.c. semigroup on [D(A∗ )] as well.
© 2004 by CRC Press LLC
(H.3) There exist constants 0 ≤ γ < 1 and T > 0, such that the following singular estimate holds true: At cT ∗ A∗ t e B = e ≤ γ , ∀ 0 < t ≤ T, (1.2) B L(U ;Y ) t L(Y ;U ) where (Bu, v)Y = (u, B ∗ v)U , u ∈ U , v ∈ D(B ∗ ) ⊃ D(A∗ ). (H.4) The function w is a deterministic disturbance, satisfying w ∈ L2 (0, ∞; Y )
(1.3)
R ∈ L(Y ; Z).
(1.4)
to be kept ﬁxed throughout. (H.5)
Optimal Control Problem on [0, T ] With the dynamics (1.1), we associate the following quadratic cost functional over a ﬁnite time horizon: T Jw (u, y) = [Ry(t)2Z + u(t)2U ]dt, (1.5) 0
and the corresponding Optimal Control Problem is: For a ﬁxed w ∈ L2 (0, T ; Y ), minimize Jw (u, y) over all u ∈ L2 (0, T ; U ), where y is the solution of (1.1) due to u (and w).
(1.6)
Remark 1.1. What makes the above optimal control problem different from those studied in the literature [4], [15], [16], is, of course, the new set of assumptions imposed on model (1.1); in particular, the presence of the singular estimate of hypothesis (H.3) = (1.2), while, however, the semigroup eAt is only assumed to be strongly continuous. Explicitly, eAt is not assumed to be analytic. In [16, Chapters 1, 2, 6], the singular estimate (H.3) = (1.2) was also available in the treatment of those chapters; however, it was so a posteriori as a consequence of the original assumption that the s.c. semigroup eAt be, moreover, analytic. Thus, writing eAt B = (−A)γ eAt (−A)−γ B would at once yield estimate (1.2) under the two assumptions in those aforementioned chapters: (i) analyticity of the semigroup of eAt on Y , and (ii) the property (−A)−γ B ∈ L(U ; Y ). In the present paper, by contrast, the s.c. semigroup eAt is not assumed to be analytic. The above set of assumptions on (1.1) are motivated by the structural acoustic problem, with hyperbolic/parabolic coupling, where these assumptions are, in fact, properties of the coupled dynamics: see the examples in Section 4 below, with both γ < 12 (the easier case: Example 4.1 of Section 4.1) and 1 2 < γ < 1 (the more challenging case of Example 4.2 in Section 4.1; and of the examples in Sections 4.2–4.4). Similarly, in line with the structural acoustic problem, we are taking a pure distributed L2 disturbance w (that is G = Identity in the notation of [16, Eq. (6.1.1.1) of Chapter 6]. The corresponding optimal control problem over an inﬁnite time horizon,
© 2004 by CRC Press LLC
T = ∞, was studied in [13], [17] (deﬁnite cost) and [26] (nondeﬁnite cost and minmax problem). Preliminaries. As in previous work [16], we shall have to consider the corresponding Optimal Control Problem over the interval [s, T ], 0 ≤ s < T , rather than just over [0, T ]; thus with s, rather than 0, as the initial time. Let s, 0 ≤ s < T be ﬁxed throughout as an initial time with initial datum y0 ∈ Y . The corresponding solution to (1.1) is then y(t, s; y0 ) = eA(t−s) y0 + (Ls u)(t) + (Ws w)(t);
t (Ls u)(t) =
eA(t−τ ) Bu(τ )dτ,
s ≤ t ≤ T,
(1.7)
(1.8a)
s
(1.8b) :continuous L2 (s, T ; U ) → L2 (s, T ; Y ) :continuous C([s, T ]; U ) → C([s, T ]; Y ), with operator norms that are uniform with respect to s, 0 ≤ s ≤ T ; (1.8c)
t
eA(t−τ ) w(τ )dτ
(Ws w)(t) =
(1.9a)
s
:continuous L2 (s, T ; Y ) → C([s, T ]; Y ), with operator norms that are uniform with respect to s, 0 ≤ s ≤ T . (1.9b) Again, the above regularity properties are due to the Young’s inequality [25] and (H.3) = (1.2). They are formalized in Proposition 2.1.2 below. The L2 adjoints of Ls and Ws are (L∗s f )(t)
T =
B ∗ eA
∗
(τ −t)
f (τ )dτ,
s ≤ t ≤ T;
(1.10a)
t
(1.10b) :continuous L2 (s, T ; Y ) → L2 (s, T ; U ) :continuous C([s, T ]; Y ) → C([s, T ]; U ), with operator norms that are uniform with respect to s, 0 ≤ s ≤ T ; (1.10c)
(Ws∗ v)(t)
T =
eA
∗
(τ −t)
v(τ )dτ,
s≤t≤T
(1.11a)
t
:continuous L2 (s, T ; Y ) → C([s, T ]; Y ), with operator norms that are uniform with respect to s, 0 ≤ s ≤ T . (1.11b) Henceforth, for s = 0, we shall use the notation L, W , L∗ , W ∗ instead of L0 , W0 , L∗0 , W0∗ .
© 2004 by CRC Press LLC
Statement of main results The main result of this chapter is the following counterpart of [21, Theorem 2.1] as it pertains to the ﬁnite horizon case. Theorem 1.1. Assume hypotheses (H.1)–(H.5) above. Then: (a1) for each y0 ∈ Y and w ∈ L2 (0, T ; Y ) ﬁxed, there exists a unique optimal pair 0 {u0w ( · , 0; y0 ), yw ( · , 0; y0 )} of the optimal control problem (1.5), (1.6), for the dynamics (1.1), or (1.7), which satisﬁes the following properties: u0w ( · , 0; y0 ) ∈ C([0, T ]; Y );
0 yw ( · , 0; y0 ) ∈ C([0, T ]; Y ).
(1.12)
(a2) The operator Φ(t, τ ) ∈ L(Y ) deﬁned by 0 Φ(t, τ )x =yw=0 (t, τ ; x) ∈ C([0, T ]; Y ),
x∈Y (1.13a) :solution to optimal control problem (1.1), (1.5), (1.6), with w = 0, over the interval [τ, T ], 0 ≤ τ < T , (1.13b)
is an evolution operator: Φ(t, t) = I,
Φ(t, s) = Φ(t, τ ), Φ(τ, s), 0 ≤ s ≤ τ ≤ t ≤ T.
(1.13c)
(a3) The bounded operator P (t) ∈ L(Y ) deﬁned on Y by T P (t)x =
eA
∗
(τ −t)
0 R∗ Ryw=0 (τ, t; x)dτ
(1.14a)
∗
(τ −t)
R∗ RΦ(τ, t)x,
(1.14b)
t
T =
eA
x∈Y
t
: continuous Y → C([0, T ]; Y ),
(1.14c)
is nonnegative, selfadjoint on Y : P (t) = P ∗ (t), T 0 0 (Ryw=0 (P (t)x1 , x2 )Y = (τ, t; x1 ), Ryw=0 (τ, t; x2 ))Z t
+ (u0w=0 (τ, t; x1 ), u0w=0 (τ, t; x2 ))U dτ.
(1.15)
In particular, 0 0 Jw=0 (t, y0 ) = J(u0w=0 ( · , t; y0 ), yw=0 ( · , t; y0 )) = (P (t)y0 , y0 )Y .
(1.16)
(a4) The gain operator B ∗ P (t) is bounded Y → U at each t and, in fact, B ∗ P (t) : continuous Y → C([0, T ]; U ).
© 2004 by CRC Press LLC
(1.17)
(a5) The operator AP (t) = A − BB ∗ P (t),
D(AP (t)) = {x ∈ Y : [I − A−1 B(B ∗ P (t))] ∈ D(A)}, (1.18) satisﬁes for any x ∈ Y and any t, s < t < T : ∂Φ(t, s)x = AP (t)Φ(t, s)x ∈ C([0, T ]; [D(A∗ )] ); ∂t ∂Φ(t, s)x = −Φ(t, s)AP (s)x ∈ γ C([s, T ]; Y ), x ∈ D(A). ∂s
(1.19) (1.20)
(a6) The following singular estimate holds true for the evolution operator Φ(t, s): Φ(t, s)BL(U ;Y ) = B ∗ Φ∗ (t, s)L(Y ;U ) ≤
cT , (t − s)γ
0 ≤ s < t ≤ T,
(1.21)
where γ < 1 is the constant in assumption (H.3) = (1.2). (a7) The operator P (t) in (1.14) satisﬁes the following Differential Riccati Equation: (P˙ (t)x, z)Y = −(Rx, Rz)Z − (P (t)x, Az)Y − (P (t)Ax, y)Y (1.22) + (B ∗ P (t)x, B ∗ P (t)z)U P (T ) = 0, for all x, z ∈ D(A). 0 (t, s; y0 )} satisﬁes (a8) The optimal pair {u0w (t, s; y0 ), yw 0 (t, s; y0 ) + rw (t, s)] u0w (t, s; y0 ) = −B ∗ pw (t, s; y0 ) = −B ∗ [P yw 0 = −B ∗ P (t)yw (t, s; y0 ) − B ∗ rw (t, s); 0 B ∗ P (t)yw (t, s; y0 ) ∈ C([s, T ]; U );
(1.23)
B ∗ rw (t, s) ∈ C([s, T ]; U ).
(1.24)
Here, rw (t, s) is the function deﬁned by 0 (t, s; y0 = 0) ∈ C([s, T ]; Y ), rw (t, s) ≡ pw (t, s; y0 = 0) − P (t)yw
(1.25)
(a9) such function satisﬁes the equation r˙w (t, s) = −A∗P (t)rw (t, s) − P (t)w(t) ∈ [D(A∗ )] a.e. in t,
(1.26)
so that rw (t, s) is also explicitly given by T rw (t, s) =
Φ∗ (τ, t)P (τ )w(τ )dτ.
(1.27)
t
Dedication. This chapter is dedicated to our friend, Professor A. V. Balakrishnan, who spurred our interest in optimal control problems for PDEs with boundary control, through his pioneering work [2, Sect. 4.12], [3], which we gladly acknowledge in details in [16, Vol. 1, pp. 420–424].
© 2004 by CRC Press LLC
2
PROOF OF THEOREM 1.1
2.1
Existence of a Unique Optimal Pair, Characterization, and Regularity Properties: Proof of (a1)
The following result is the perfect counterpart of [21, Proposition 3.1.1] in the case T = ∞, and it is proved in the same way. Proposition 2.1.1. Assume (H.1)–(H.5). Let y0 ∈ Y . Then: (i) for any w ∈ L2 (s, T ; Y ) ﬁxed, the optimal control problem (1.5), (1.6), over [s, T ] rather than [0, T ], for the dynamics (1.1), or (1.7), admits a unique optimal pair 0 {u0w ( · , s; y0 ), yw ( · , s; y0 )} satisfying the optimality condition 0 ( · , s; y0 ) ∈ L2 (s, T ; U ). u0w ( · , s; y0 ) = −L∗s R∗ Ryw
(2.1.1)
(ii) Moreover, the optimal pair is given explicitly by the following formulae: u0w ( · , s; y0 ) = −[Is + L∗s R∗ RLs ]−1 L∗s R∗ R[eA( · −s) y0 + Ws w] ∈ L2 (s, T ; U ) (2.1.2a) = u0w=0 ( · , s; y0 ) + u0w ( · , s; y0 = 0)
(2.1.2b)
0 yw ( · , s; y0 ) = [Is + Ls L∗s R∗ R]−1 [eA( · −s) y0 + Ws w] ∈ L2 (s, T ; Y )
= =
{Is − Ls [Is + L∗s R∗ RLs ]−1 L∗s R∗ R}[eA( · −s) y0 0 0 yw=0 ( · , s; y0 ) + yw ( · , s; y0 = 0)
+ Ws w]
0 Ryw ( · , s; y0 ) = [Is + RLs L∗s R∗ ]−1 R[eA( · −s) y0 + Ws w] ∈ L2 (s, T ; Z)
=
0 Ryw=0 ( · , s; y0 )
+
0 Ryw ( · , s; y0
= 0),
(2.1.3a) (2.1.3b) (2.1.3c) (2.1.4a) (2.1.4b)
where, as usual, the inverse operators in the above formulas are well deﬁned as bounded operators on all of L2 (s, T ; · ) [16, Chapter 2, Appendix 2A for (2.1.3a)]. Moreover, with y0 ∈ Y , the corresponding optimal dynamics is 0 (t, s; y0 ) = eA(t−s) y0 + {Ls u0w ( · , s; y0 )}(t) + {Ws w}(t) ∈ L2 (s, T ; Y ). yw
(2.1.5)
(iii) For y0 ∈ Y , the optimal cost satisﬁes the following relations: 0 ( · , s; y0 )) Jw0 (s; y0 ) = Jw (u0w ( · , s; y0 ), yw 0 = Jw=0 (s; y0 ) + Jw0 (s; y0 = 0) + Xw (s; y0 ); 0 (s; y0 ) Jw=0 0 Jw (s; y0 = 0)
= =
A( · −s)
(Re y0 , [Is + RLs L∗s R∗ ]−1 ReA( · −s) y0 )L2 (s,T ;Z) ; (w, Ws∗ R∗ [Is + RLs L∗s R∗ ]−1 RWs w)L2 (s,T ;Y ) ; 2(ReA( · −s) y0 , [Is + RLs L∗s R]−1 RWs w)L2 (s,T ;Z)
Xw (s; y0 ) = = linear in w.
(2.1.6) (2.1.7) (2.1.8) (2.1.9)
The next result is the perfect counterpart of [21, Proposition 3.1.2] in the case T = ∞.
© 2004 by CRC Press LLC
Proposition 2.1.2. Assume (H.1), (H.2), (H.3), (H.5). Then, with reference to the operators Ls and L∗s in (1.8) and (1.10), we have the following boundedness (continuity) results, all uniformly in the operator norm, with respect to the parameter s, 0 ≤ s ≤ T , but with bounds that will generally depend on γ and T : (i) Ls : continuous L2 (s, T ; U ) → L2 (s, T ; Y ) : continuous C([s, T ]; U ) → C([s, T ]; Y );
(2.1.10a) (2.1.10b)
L∗s : continuous L2 (s, T ; Y ) → L2 (s, T ; U ) : continuous C([s, T ]; Y ) → C([s, T ]; U );
(2.1.11a) (2.1.11b)
[Is + L∗s R∗ RLs ]−1 ∈ L(L2 (s, T ; U )) ∩ L(C([s, T ]; U ));
(2.1.12)
[Is + Ls L∗s R∗ R]−1 ∈ L(L2 (s, T ; Y )) ∩ L(C([s, T ]; Y ));
(2.1.13)
(ii)
(iii) (iv) (v)
1 . (2.1.14) 2 (vi) Generally, let r0 = 2, and let rn , n = 1, 2, . . . be arbitrary positive numbers such Ls : continuous L2 (s, T ; U ) → C([s, T ]; Y ), if γ <
that 2 < r1 < · · · < rn <
2 , 2nγ − (2n − 1)
n = 1, 2, . . . for
2n − 1 ≤ γ < 1. (2.1.15) 2n
Then, for n = 0, 2, 4, . . . we have Ls : continuous Lrn (s, T ; U ) → Lrn+1 (s, T ; Y ),
(2.1.16)
where for 0 ≤ γ < [2(n + 1) − 1]/[2(n + 1)] we may take rn+1 = ∞; and L∗s : continuous Lrn+1 (s, T ; Y ) → Lrn+2 (s, T ; U ),
(2.1.17)
where rn+1 in (2.1.17) is the same as in (2.1.16), and where for 0 ≤ γ < [2(n + 2) − 1]/[2(n + 2)] we may take rn+2 = ∞. (vii) for p > 1/(1 − γ), Ls : continuous Lp (s, T ; U ) → C([s, T ]; Y ).
(2.1.18)
(viii) Thus, a fortiori, there exists a positive integer n0 = n0 (γ), depending on γ, such that for all positive integers n ≥ n0 (γ), we have (L∗s R∗ RLs )n : continuous L2 (s, T ; U ) → C([s, T ]; Y ).
(2.1.19)
The next result is the present counterpart of [21, Proposition 3.1.3] in the case T = ∞.
© 2004 by CRC Press LLC
Proposition 2.1.3. Assume (H.1)–(H.5). Then, for any y0 ∈ Y and any w ∈ L2 (0, T ; Y ) (ﬁxed), the optimal pair established in Proposition 2.1.1 satisﬁes the additional regularity properties: u0w ( · , s; y0 ) ∈ C([s, T ]; U );
0 yw ( · , s; y0 ) ∈ C([s, T ]; Y ),
continuously in y0 ∈ Y , and uniformly in s.
(2.1.20)
Thus, the proof of property (a1) of Theorem 1.1 is complete. 2.2
The Evolution Operator Φ(t, s); the Functions pw (t, s; y0 ), rw (t, s); the Operator P (t) ∈ L(Y ). Proof of (a2), (a3), (a4), (a8)
The operator Φ(t, s). Lemma 2.2.1. Assume (H.1), (H.2), (H.3), (H.5). The operator Φ(t, s), deﬁned in (1.13a), has the following properties: 0 Φ(t, s)y0 = yw=0 (t, s; y0 ) = {[Is + Ls L∗s R∗ R]−1 eA( · −s) y0 }(t) ∈ C([s, T ]; Y ) uniformly in s, 0 ≤ s ≤ T
(2.2.1a) (2.2.1b)
Φ(t, s)y0 ∈ C([0, t]; Y ); Φ(t, s)L(Y ) ≤ cT , 0 ≤ s ≤ t ≤ T.
(2.2.1c)
(in t)
(in s)
(ii) Moreover, Φ(t, s) is an evolution operator Φ(t, t) = I,
Φ(t, s)x = Φ(t, τ )Φ(τ, s)x,
x ∈ Y, 0 ≤ s ≤ τ ≤ t ≤ T.
(2.2.2)
Proof. As in [16, Lemma 1.4.6.2 of Chapter 1, p. 59]. The regularity noted in (2.2.1b): t → Φ(t, s)y0 continuous in Y , for s ﬁxed, uniformly in s, s ≤ t ≤ T , follows from Proposition 2.1.3, Eq. (2.1.20). Then, such continuity in the ﬁrst variable implies continuity in the second variable: s → Φ(t, s)y0 for t ﬁxed, 0 ≤ s ≤ t, as seen in the proof of [16, Lemma 1.4.6.2]. The function pw (t, s; y0 ), and the operator P (t) ∈ L(Y ). For y0 ∈ Y , we deﬁne [in line with, say, [16, Eq. (6.2.21) of Chapter 6]: T pw (t, s; y0 ) =
eA
∗
(τ −t)
0 R∗ Ryw (τ, s; y0 )dτ ∈ C([s, T ]; Y ),
(2.2.3)
t
which is the unique solution of the problem: 0 (t, s; y0 ) ∈ [D(A)] p˙w (t, s; y0 ) = −A∗ pw (t, s; y0 ) − R∗ Ryw ]pw (T, s; y0 ) = 0,
(2.2.4a) (2.2.4b)
with zero initial condition at t = T . Moreover, we introduce the operator P (t) ∈ L(Y ), as in, say, [16, (6.2.25) of Chapter 6], by setting T P (t)x =
eA
∗
(τ −t)
R∗ RΦ(τ, t)x dτ,
x∈Y
(2.2.5a)
t
: continuous Y → C([0, T ]; Y ), P (T ) = 0,
© 2004 by CRC Press LLC
(2.2.5b)
where the claimed regularity in (2.2.5b) stems, via property (2.2.1c), as in obtaining [16, Eq. (1.2.25) of Chapter 1]. The next result is the present counterpart of [21, Lemma 3.2.3] for T = ∞. Lemma 2.2.2. Assume (H.1)–(H.5). Then, for y0 ∈ Y , with reference to (2.2.3) and (2.2.5), we have (i) pw (t, s; y0 ) = pw=0 (t, s; y0 ) + pw (t, s; y0 = 0) ∈ C([s, T ]; Y ); T pw=0 (t, s; y0 ) =
eA
∗
(τ −t)
0 R∗ Ryw=0 (τ, s; y0 )dτ
∗
(τ −t)
R∗ RΦ(τ, s)y0 dτ
∗
(τ −t)
R∗ RΦ(τ, t)Φ(t, s)y0 dτ
(2.2.6) (2.2.7)
t
T =
eA t
T =
eA t
= P (t)Φ(t, s)y0 ∈ C([s, T ]; Y );
(2.2.8)
(ii) 0 −u0w (t, s; y0 ) = {L∗s R∗ Ryw ( · , s; y0 )}(t)
T =
B ∗ eA
∗
(τ −t)
0 R∗ Ryw (τ, s; y0 )dτ
(2.2.9)
t
= B ∗ pw (t, s; y0 ) ∈ C([s, T ]; U );
(2.2.10)
(iii) P (t) ∈ L(Y ), and moreover, B ∗ P (t) : continuous Y → C([0, T ]; U ), B ∗ P (t)x =
T
B ∗ eA
∗
(τ −t)
R∗ RΦ(τ, t)x dτ ;
(2.2.11) (2.2.12)
t
(iv) −u0w=0 (t, s; y0 ) = B ∗ pw=0 (t, s; y0 ) = B ∗ P (t)Φ(t, s)y0 ∈ C([s, T ]; U ).
(2.2.13)
Proof. (i) The decomposition of pw in (2.2.6) is a consequence of using the decom0 position of yw in (2.1.3c) in the deﬁnition of (2.2.6). The steps from (2.2.7) to (2.2.8) use: (2.2.1a) for Φ, its evolution property (2.2.2), and deﬁnition (2.2.5a) for P (t). (ii) The step from (2.2.9) to (2.2.10) is selfexplanatory, once one recalls the optimality condition (2.1.1), and then (1.10) for L∗s and (2.2.3) for pw . (iii) Equation (2.2.12) follows from (2.2.5a): it is here that one uses assumption (H.3) ∗ = (1.2) on the singular estimate of B ∗ eA t to obtain B ∗ P (t) ∈ L(Y ) at each t, as in
© 2004 by CRC Press LLC
[21, (3.2.15)]. Then, property (2.2.1c) yields, in fact, (2.2.11) as in, say, [16, Proposition 1.4.6.5]. (iv) Equation (2.2.13) follows from (2.2.10) and (2.2.8). The function rw (t, s). Proof of (a8). Next, the present counterpart of [164, Lemma 6.2.2.1, Chapter 6, p. 565], or of [21, Proposition 3.6.1] is: Proposition 2.2.3. Assume (H.1)–(H.5). (i) With reference to pw in (2.2.3), the following identity holds true: 0 pw (t, s; y0 ) = P (t)Φ(t, s)y0 + pw (t, s; y0 = 0) = P (t)yw=0 (t, s; y0 ) + pw (t, s; y0 = 0) (2.2.14) 0 = P (t)yw (t, s; y0 ) + rw (t, s),
(2.2.15)
where the function rw (t, s) is deﬁned by 0 rw (t, s) ≡ pw (t, s; y0 = 0) − P (t)yw (t, s; y0 = 0).
(2.2.16)
(ii) For y0 ∈ Y , the optimal control is written in feedback synthesis as 0 −u0w (t, s; y0 ) = B ∗ pw (t, s; y0 ) = B ∗ [P (t)yw (t; s; y0 ) + rw (t, s)] ∈ C([s, T ]; U ) (2.2.17) 0 = B ∗ P (t)yw (t, s; y0 ) + B ∗ rw (t, s) ∈ C([s, T ]; U );
(2.2.18)
0 (t, s; y0 ) and B ∗ rw (t, s) ∈ C([s, T ]; U ). B ∗ P (t)yw
(2.2.19)
(iii) The optimal dynamics may thus be rewritten as
0 (t, s; y0 ) y˙ w 0 yw (s, s; y0 )
0 = AP (t)yw (t, s; y0 ) − BB ∗ rw (t, s) + w(t) ∈ [D(A∗ )] = y0 ∈ Y ;
AP (t) = A − BB ∗ P (t)
(2.2.20)
(2.2.21a)
D(AP (t)) = {x ∈ Y : [I − A
−1
∗
BB P (t)]x ∈ D(A)},
(2.2.21b)
in the sense that, for all x ∈ D(A∗ ): 0 0 (y˙ w (t, s; y0 ), x)Y = ([I − A−1 BB ∗ P (t)]yw (t, s; y0 ), A∗ x)Y
− (B ∗ rw (t, s), B ∗ A∗−1 A∗ x)U + (w(t), x)Y ;
0 yw (t, s; y0 )
t =e
A(t−s)
y0 −
(2.2.22)
Φ(τ, s)BB ∗ rw (τ, s)dτ
s
t Φ(τ, s)w(τ )dτ ∈ C([s, T ]; Y ).
+ s
© 2004 by CRC Press LLC
(2.2.23)
Proof. (i) Using the decomposition (2.1.3c) in deﬁnition (2.2.3), and recalling (2.2.1a), (2.2.2), (2.2.5a), (2.2.3), (2.1.3c), yields T pw (t, s; y0 ) =
eA
∗
(τ −t)
0 0 R∗ R[yw=0 (τ, s; y0 ) + yw (τ, s; y0 = 0)]dτ (2.2.24)
∗
(τ −t)
R∗ RΦ(τ, t)Φ(t, s)y0 dτ
∗
(τ −t)
0 R∗ Ryw (τ, s; y0 = 0)]dτ
t
T (by (2.2.1), (2.2.2)) =
eA t
T =
eA t
(by (2.2.5a), (2.2.3)) = P (t)Φ(t, s)y0 + pw (t, s; y0 = 0) 0 0 (by (2.1.3c)) = P (t)[yw (t, s; y0 ) − yw (t, s; y0 = 0)] + pw (t, s; y0 = 0) 0 = P (t)yw (t, s; y0 ) 0 + [pw (t, s; y0 = 0) − P (t)yw (t, s; y0 = 0)],
(2.2.25)
and (2.2.25) yields (2.2.15) via the deﬁnition (2.2.16) for rw (t, s). 0 (ii) To get (2.2.17), we use (2.2.15) in (2.2.10). The regularity of B ∗ P (t)yw (t, s; y0 ) ∗ 0 in (2.2.19) follows by combining (2.2.11) for B P (t) with (2.1.20) for yw (t, s; y0 ). Then, 0 recalling likewise (2.2.18) for yw (t, s; y0 ), we obtain 0 (t, s; y0 ) ∈ C([s, T ]; U ), B ∗ rw (t, s) = −u0w (t, s; y0 ) − B ∗ P (t)yw
(2.2.26)
where we could, equally well, set y0 = 0 in (2.2.26) by (2.2.16). 2.3
Singular Estimate for Φ(t, s)B: Proof of (a5), (a6)
Orientation. The present section is the counterpart of [21, Section 3.3] in the case T < ∞. We shall transfer the assumed singularity for eAt B in (H.3) = (1.2) into the same singularity for Φ(t, s)B. We shall make critical use of [21, Proposition 3.3.1], which is repeated below for convenience. Brieﬂy, as in the proof of, say, [16, Theorem 6.26.3.1], in order to derive that the operator P (t) deﬁned in (2.2.5) satisﬁes the Differential Riccati Equation (1.22) on D(A), we need to differentiate strongly Φ(t, s) on D(A). This, in turn, is accomplished if we can establish the same singular estimate for Φ(t, s)B that holds true under (H.3) for eAt B. The resulting Theorem 2.3.2 below is a delicate new point of the present development, which was not explicitly needed in our treatment of the abstract analytic or parabolic case of [16, Chapters 1, 2, and 6]. We begin by collecting [21, Proposition 3.3.1]. Proposition 2.3.1. Assume (H.1) through (H.5). With reference to the operators L and L∗ in (1.8) and (1.10), and recalling that 0 ≤ γ < 1 as postulated in assumption (H.3), we have (i) Let 0 < r < 1. Then for any 0 < T ≤ ∞, L : continuous r C([0, T ]; U ) →(r+γ−1) C([0, T ]; Y ).
© 2004 by CRC Press LLC
(2.3.1)
(ii) Let r > 0, and > 0 arbitrary, L∗ : continuous r C([0, T ]; Y ) →(r+γ−1+) C([0, T ]; U ).
(2.3.2)
(iii) Let 0 < r < 1. Then, there exists a positive integer m = m(r) such that (L∗ R∗ RL)m : continuous r C([0, T ]; U ) → C([0, T ]; U ).
(2.3.3)
[I + L∗ R∗ RL]−1 ∈ L(γ C([0, T ]; Y )).
(2.3.4)
(iv) (v)
[I + LL∗ R∗ R]−1 = I − L[I + L∗ R∗ RL]−1 L∗ R∗ R ∈ L(γ C([0, T ]; Y )),
(2.3.5)
where we recall that, if X is a Banach space, then r C([0, T ]; X) ≡ f (t) ∈ C((0, T ]; X) : f = sup t f (t) < ∞ , r X r C([0,T ];X) 0 12 (Example 4.2). In Subsection 4.2 we provide a thermoelastic plate model with boundary control and γ = 34 + , > 0. Furthermore, in Sections 4.3, 4.4, we provide a sandwich beam model with boundary control and γ = 34 + , > 0. In our present treatment, which is after [27], and unlike [11], the sandwich beam is reduced to the thermoelastic model of Subsection 4.2. This explains why we obtain the same value of γ. This reduction, then, permits consideration of the structural acoustic chamber of Example 4.2, and replace the thermoelastic ﬂexible wall with a sandwich beam. This is noted in Subsection 4.5. Naturally, we obtain the same value γ = 34 + as in Example 4.2. 4.1
Two classes of structural acoustic problems
Example 4.1: A class of structural acoustic problems with constant 0 < γ <
1 2
The model. We consider the following class of structural acoustic problems, where Ω is an acoustic chamber with a ﬂexible (elastic) wall Γ0 , an assumed ﬂat and rigid wall Γ1 . Thus, let Ω ⊂ Rn , n = 2, 3, be an open bounded domain with boundary Γ = Γ0 ∪ Γ1 , where Γ0 and Γ1 are open, connected, and disjoint parts, Γ0 ∩ Γ1 = φ, in Rn−1 , and Γ0 is ﬂat. We allow either Γ to be sufﬁciently smooth (say, C 2 ), or else Ω to be convex: this assumption will then guarantee that solutions to classical elliptic equations with L2 (Ω)nonhomogeneous terms be in H 2 (Ω) [7]. Let z denote the velocity potential of the acoustic medium within the chamber. For simplicity of notation, we take equal to 1 both the density of the ﬂuid and the speed of sound in the ﬂuid. Then zt is the acoustic pressure. Let v denote the displacement of the ﬂat ﬂexible wall Γ0 , modeled by an elastic beam or plate equation (n = 2, or n = 3). The structural acoustic model considered here is as follows: acoustic chamber:
ztt = ∆z − d1 zt + f ∂z + d 2 zt = 0 ∂ν ∂z + d3 zt = ±vt ∂ν
in (0, T ] × Ω, in (0, T ] × Γ1 , in (0, T ] × Γ0 ,
Mk vtt + Avt + Av ± zt elastic 1 = Bu in (0, T ] × [D(A 2 )] wall z(0, · ) = z0 , zt (0, · ) = z1 in Ω; v(0, · ) = v0 , vt (0, · ) = v0 in Γ0 .
(4.1.1a) (4.1.1b) (4.1.1c) (4.1.1d) (4.1.1e)
Assumptions. In (4.1.1a–c), f ∈ L2 (0, T ; L2 (Ω)) denotes the deterministic external noise within the chamber; and the nonnegative constant di , when positive, introduce
© 2004 by CRC Press LLC
interior/boundary damping in the model. Equation (4.1.1d) is an abstract version encompassing several “concrete” elastic models, as documented below. At the abstract level, we make the following assumptions: (a1) Mk : L2 (Γ0 ) ⊃ D(Mk ) → L2 (Γ0 ); A : L2 (Γ0 ) ⊃ D(A) → L2 (Γ0 ) (4.1.2) are two positive, selfadjoint operators (the stiffness operator, and the elastic operator, respectively, the ﬁrst depending on a nonnegative parameter k ≥ 0; in concrete situations, if k > 0, the elastic model on Γ0 accounts for rotational forces); (a2) 1 1 (4.1.3) D(A 2 ) ⊂ D(Mk2 ); (a3) there is a positive constant ρ <
1 2
such that
A−ρ B ∈ L(U; L2 (Γ0 )), equivalently B : continuous U → [D(Aρ )] ,
(4.1.4)
where [ ] denotes duality with respect to L2 (Γ0 ) as a pivot space, and U is the control Hilbert space. (a4) Moreover, the parameter ρ in (4.1.4) satisﬁes either one of the following additional assumptions: (a4i) either 5 , if Ω is a smooth domain, ρ ≤ 12 (4.1.5a) ρ ≤ 7 , if Ω is a parallelopiped, 16 (a4ii) or else 1 1 (4.1.5b) D(Mk2 ) ⊂ H 3 (Γ0 ). Remark 4.1.1 (on the control operator B). In concrete PDE examples of the structural acoustic problems, such as they arise in smart material technology, all of the above assumptions are satisﬁed. First, in this case, the control operator B is given by Bu =
J
ai ui δξ j ,
u = [u1 , . . . , uJ ] ∈ RJ = U,
(4.1.6)
j=1
where: (i) if dim Γ0 = 1 (dim Ω = 2), then ξj are points on Γ0 , aj are constants, and δξ j are derivatives of the Dirac distribution supported at ξj ; (ii) if dim Γ0 = 2 (dim Ω = 3), then ξj denote closed regular curves on Γ0 , aj are smooth functions and δξ j denotes the normal derivative supported at ξj :
(δξ j , φ)L2 (Γ0 ) =
−φ (ξj ), dim Γ0 = 1 −
© 2004 by CRC Press LLC
∇φ · ν dξj , dim Γ0 = 2; ξj
(4.1.7a) ∀φ ∈ H
3 2 +
(Γ0 ), (4.1.7b)
where ν is the unit outer normal vector to the closed curve ξj , and > 0 is arbitrary. Thus, 3 from (4.1.7) we have that, a fortiori, (δξ j , A−( 8 + 4 ) ψ)L2 (Γ0 ) is well deﬁned ∀ ψ ∈ L2 (Γ0 ), 3
3
3
since A−( 8 + 4 ) ψ ∈ D(A 8 + 4 ) ⊂ H 2 + (Γ0 ). Hence the operator B deﬁned by (4.1.6) satisﬁes assumption (a3) = (4.1.4) with ρ = 38 + , ∀ > 0 small [13], [16, vol. 2, p. 907]. Then, such ρ = 38 + satisﬁes both conditions (a4i) = (4.1.5a) as well. Remark 4.1.2 (on the stiffness operator Mk ). In concrete PDE examples where the parameter k > 0 (and so the elastic beam/platemodel on Γ0 accounts for rotational forces), Mk is the translation by the identity of the realization of (−∆) on Γ0 subject to 1
1
appropriate boundary conditions. Thus, D(Mk2 ) ⊂ H 1 (Γ0 ) ⊂ H 3 (Γ0 ) and assumption (a4ii) = (4.1.5b) is satisﬁed as well. The above structural acoustic model (4.1.1), subject to assumptions (a1) through (a4), satisﬁes assumption (H.3) = (1.2) with γ = ρ < 12 . The following claims are shown in [13], [1]. The structural acoustic problem (4.1.1) can be rewritten in the abstract form (1.1), with operators A and B explicitly identiﬁed, and w = [0, f, 0, 0]. Moreover, the operator A is the generator of an s.c. contraction semigroup eAt on an appropriate ﬁnite energy space Yk given by 1 1 Yk ≡ H 1 (Ω) × L2 (Ω) × D(A 2 ) × D(Mk2 ), (4.1.8) for the variables [z(t), zt (t), v(t), vt (t)] = eAt [z0 , z1 , v0 , v1 ]. Finally, such operators A and B do satisfy assumption (H.3) = (1.2) for γ = ρ < 12 , the constant in assumptions (a3) and (a4). In particular, if the control operator B is deﬁned by (4.1.6), then we have γ = ρ = 38 + , ∀ > 0. Remark 4.1.3 (on the uniform stability of problem (4.1.1)). There are several conﬁgurations — that is, choices of the damping constants di in (4.1.1a–c) and corresponding geometrical conditions — which ultimately yield uniform stability on Yk of the associated s.c. semigroup eAt [12]. They include the following cases: (1) d2 = d3 = 0, d1 > 0 (viscous damping); (2) d1 = d3 = 0, d2 > 0 (boundary damping on rigid wall Γ1 ), with no geometrical conditions; (3) d1 = d2 = 0, d3 > 0 (boundary damping on ﬂexible wall Γ0 ) under the geometrical condition that Ω is convex and there exists a point x0 ∈ Rn such that (x − x0 ) · ν(x) ≤ 0, ∀ x ∈ Γ1 [22, Appendices]. Additional cases are also possible [13]. Remark 4.1.4. At the price of introducing heavy notation, it would be possible to include into problem (4.1.1), also the case where the elastic wall Γ0 is curved and, accordingly, modeled by a shell equation to be written abstractly as in (4.1.1d), see [1.3]. Concrete illustrations of the abstract elastic equation (4.1.1d). As canonical illustrations of the abstract elastic equation (4.1.1d) — say, with no coupling term zt and with no control: u ≡ 0 — we may take the classical EulerBernoulli equation on Γ0 (k = 0) or the corresponding Kirchhoff equation on Γ0 (k > 0): vtt − k∆vtt + ∆2 v + ∆2 vt = 0
on (0, T ] × Γ0 ,
(4.1.9)
under a variety of B.C. on (0, T ] × ∂Γ0 : hinged, clamped, free B.C., etc. [1], [13], [21]. Then, A is the realization of ∆2 on L2 (Γ0 ) subject to the appropriate B.C. Finally, Mk = I + kA1 where A1 is the realization of (−∆) on L2 (Γ0 ) under suitable B.C.
© 2004 by CRC Press LLC
Conclusion. Under the above assumptions, including those of Remark 4.1.3, model (4.1.1) is covered by the abstract theory of Sections 1 and 2, with γ < 12 . Remark 4.1.5. Example 4.1 can be extended to replace, in Eq. (4.1.1d), the damping term Avt with a damping term Aα vt , 12 ≤ α ≤ 1, the range of analyticity of the free vproblem. New interesting phenomena arise, which are analyzed in detail in [5]. In particular, the singular estimate for eAt B continues to hold true, with a parameter γ that depends on α. Example 4.2: A class of structural acoustic problems with constant
1 2
0
(4.1.11a) (4.1.11b)
where, for k > 0, H 1 (Γ0 ; k) is topologized by ψ2H 1 (Γ0 ;k) ≡ ψ2L2 (Γ0 ) + k∇ψ2L2 (Γ0 ) .
© 2004 by CRC Press LLC
(4.1.11c)
In (4.1.10b–c), ν is the unit outward vector to the boundary Γ = ∂Ω. By contrast, below, when dim Ω = 3 and so dim Γ0 = 2, we shall let ν˜ be the unit outward normal to ∂Γ0 as the boundary of Γ0 ; and, τ˜ be the unit tangent vector to ∂Γ0 , oriented counterclockwise. System (4.1.10) is supplemented with Boundary Conditions (B.C.). We shall explicitly consider three sets of B.C. for the thermoelastic component: Hinged B.C.: v = ∆v = θ = 0 on (0, T ] × ∂Γ0 ; (4.1.12) Clamped B.C.: ∂v = θ = 0 on (0, T ] × ∂Γ0 . ∂ ν˜ Free B.C. when dim Ω = 3, dim Γ0 = 2: ∆v + B1 v + θ = 0 ∂ ∂ ∂θ ∆v + B2 v − γ vtt + =0 on (0, T ] × ∂Γ0 ∂ ν ˜ ∂ ν ˜ ∂ ν˜ ∂θ + λθ = 0, λ > 0, ∂ ν˜ v=
(4.1.13)
(4.1.14a) (4.1.14b) (4.1.14c)
where, with constant 0 < µ < 1, we have B1 v = (1 − µ)(2ν1 ν2 vxy − ν12 vyy − ν22 vxx ), ν˜ = [ν1 , ν2 ]; ∂ 2 2 B2 v = (1 − µ) [(ν − ν2 )vxy + ν1 ν2 (vyy − vxx )] + v , > 0. ∂ τ˜ 1
(4.1.14d) (4.1.14e)
Remark 4.1.6. The parameter k ≥ 0 in (4.1.10d) is critical in describing the character of the dynamics of the uncoupled free thermoelastic system (4.1.10d–e) [that is, with no coupling term zt and with u ≡ 0]: for k = 0, such a thermoelastic problem generates an s.c. analytic semigroup (“parabolic” case) [16, Chapter 3, Appendices 3E3I], [18–20], while for k > 0 the corresponding s.c. semigroup is “hyperbolicdominated” in a technical sense [17]. Throughout this example, we let A = realization in L2 (Γ0 ) of ∆2 subject to hinged, or clamped, or free homogeneous B.C.
(4.1.15)
Regarding the control operator B in (4.1.10d), we shall assume the same hypothesis (A3) = (4.1.4), here restated as (h1) there exists a positive constant ρ < 12 , such that A−ρ B ∈ L(U; L2 (Γ)), equivalently B : continuous U → [D(Aρ )]
(4.1.16)
[an assumption satisﬁed with ρ = 38 + , if B is the operator deﬁned in (4.1.6)]. In addition, we make the following assumption on the tangential positive selfadjoint operator D occurring in (4.1.10c): 1 (h2) With ρ as in (4.1.16), there exist positive constants δ1 , δ2 , such that D(D 2 ) = D(Aρ0 ) and δ1 z2D(Aρ0 ) ≤ (Dz, z)L2 (Γ0 ) ≤ δ2 z2D(Aρ0 ) ,
© 2004 by CRC Press LLC
1
∀ z ∈ D(Aρ0 ) ⊂ D(D 2 ),
(4.1.17)
1 1 1 1 if ≤ ρ < , then: ρ − ≤ ρ0 ≤ ; (4.1.18a) 4 2 4 4 1 if ρ < , (4.1.18b) then: ρ0 = 0. 4 A typical example of such tangential operator D is a realization of the LaplaceBeltrami operator on L2 (Γ0 ). where
Remark 4.1.7 (on assumption (h2)). (a) If the constant ρ in (4.1.16) satisﬁes ρ < 14 , then the damping operator D may be taken to be the identity operator on L2 (Γ0 ). (b) If, however, 14 ≤ ρ < 12 [as in the case of the control operator B in (4.1.6)], then a stronger, unbounded damping operator D is needed. More precisely: (b1) Let 14 ≤ ρ ≤ 38 so that ρ − 14 ≤ 18 and we can take ρ0 = 18 . Then 4ρ0 − 12 = 0 1 1 and then D(Aρ0 ) = D(A 8 ) is topologically equivalent to H 2 (Γ0 ) subject to appropriate 1
1
2 (Γ0 )]. B.C. [e.g., in the case of hinged or clamped B.C., then D(A 8 ) = H00 3 1 1 3 1 1 1 (b2) Let 8 < ρ < 2 so that 8 = 8 − 4 < ρ − 4 < 4 , and we can take ρ0 satisfying 1 1 1 1 8 < ρ0 ≤ 4 . Then, (*): 0 < 4ρ0 − 2 ≤ 2 . Then, the following two subcases need to be considered. (b2i) Assume either hinged or clamped B.C. for the operator A in (4.1.15), see (4.1.12) or (4.1.13). Then (*) above implies [6] that
D(Aρ0 ) = H04ρ0 (Γ0 ) ⊂ H 1 (Γ0 ).
(4.1.19)
Thus, in this case, Dzt welldeﬁned requires, by assumptions (4.1.17) and (4.1.19), that z∂Γ0 ≡ 0. To ensure this, we then take α = 0 in (4.1.10b), so that the zproblem is endowed with Dirichlet, rather than Robin, B.C. (b2ii) Assume now free B.C. for the operator A in (4.1.15). Then, (*) above implies that D(Aρ0 ) = H 4ρ0 (Γ0 ) ⊂ H 1 (Γ0 ), (4.1.20) and then we can allow α ≥ 0 in (4.1.10b): that is, either Robin or Dirichlet B.C. The above structural acoustic model (4.1.10), subject to assumptions (h1), (h2), and k = 0 (no rotational forces accounted for) satisﬁes assumption (H.3) = (1.3) with γ = 2ρ [13]. Thus, if B is the control operator deﬁned by (4.1.6), then 12 < γ = 2( 38 + ) = 3 4 + 2 < 1. When k = 0 in (4.1.10d) [EulerBernoulli rather than Kirchhoff equation] and assumptions (h1), (h2) above are in force, then the structural acoustic model (4.1.10) can be rewritten in the abstract form (1.1), with operators A and B explicitly identiﬁed. Moreover, the operator A is the generator of an s.c. contraction semigroup eAt on an appropriate ﬁnite energy space Y given by 1
Y ≡ HΓ11 (Ω) × L2 (Ω) × D(A 2 ) × L2 (Γ0 ) × L2 (Γ0 )
(4.1.21)
for the variables [z(t), zt (t), v(t), vt (t), θ(t)]. The abstract deterministic disturbance w in Eq. (1.1) is now w = [0, f, 0, 0, 0], with f the disturbance in (4.1.10a). Finally, the s.c. semigroup eAt is uniformly stable on Y [16, Chapter 3, Sections 3.11–3.13]. (In the case of free B.C., this is due to the term (1−µ)v in (4.1.14e) with coefﬁcient (1−µ) < 0). Conclusion. Under the above assumptions, model (4.1.10) is covered by the abstract theory of Sections 1–3.
© 2004 by CRC Press LLC
4.2
A class of thermoelastic plates with thermal boundary control and γ = ∀>0
3 4
+ ,
In this subsection we consider a class of thermoelastic plate problems with rotational inertia and boundary control in the thermal component. More precisely, let Ω ⊂ R2 be a smooth bounded domain with boundary Γ. [The following model is the counterpart of the thermoelastic component modeling the elastic wall in the structural acoustic problem of Example 4.2.] If w denotes the vertical displacement of the plate and θ denotes the temperature (with respect to the stressfree state), the thermoelastic plate model with boundary thermal control u considered throughout this subsection is wtt − ρ∆wtt + ∆2 w − ∆θ = 0 in (0, T ] × Ω ≡ Q; (4.2.1a) (4.2.1b) θt − ∆θ + ∆wt = 0 in Q;
wΣ ≡ 0, ∆wΣ ≡ 0, θΣ
= u
in (0, T ] × Γ ≡ Σ;
(4.2.1c)
w(0, · ) = w0 , wt (0, · ) = w1 , θ(0, · )
= θ0
in Ω.
(4.2.1d)
The boundary control u acts in the Dirichlet boundary conditions of the thermal component θ. The boundary conditions of the mechanical variable w are “hinged” B.C. The positive constant ρ > 0 is proportional to the square of the thickness of the plate and accounts for the rotational forces on the model. It is known [17] that for ρ > 0, the free dynamics (3.2) with u ≡ 0 is hyperbolicdominated, in the sense that the resulting thermoelastic s.c. semigroup is a compact perturbation, at each t > 0, of a damped s.c. group [17]. Consistently with (4.1.15), we deﬁne the operator A by Af = ∆2 f,
D(A) ≡ H 4 (Ω) ∩ H02 (Ω).
(4.2.2)
Then, with equivalent norms: 1
D(A 2 ) ≡ H02 (Ω),
1
D(A 4 ) ≡ H01 (Ω).
(4.2.3)
The state space of problem (4.2.1) (with ρ > 0) is 1
1
Yρ ≡ D(A 2 ) × D(A 4 ) × L2 (Ω).
(4.2.4)
It is possible to rewrite problem (4.2.1) in the form of the abstract model (1.1), with operators A and B explicitly identiﬁed ([1], [13], [16, Chapter 3], [17]). Moreover, in the same references, it is shown that the resulting operator A is the inﬁnitesimal generator of an s.c. semigroup eAt of contractions on the space Yρ above (4.2.2), so that assumption (H.1) holds true. Further, the resulting operator B satisﬁes assumption (H.2). The key issue is, of course, assumption (H.3): This was established to hold true in [14], with γ = 34 + , with respect to A, B, Y as above, and U = L2 (Γ). Conclusion. Model (4.2.1) satisﬁes the abstract assumptions (H.1), (H.2), (H.3) of Section 1 and the abstract theory of Sections 1–3 applies. 4.3
A class of composite beams models
In this subsection we focus on a model, due to [9], of a sandwich beam consisting of a thin, compliant middle layer and two identical stiff outer layers. In this simpliﬁed model, the
© 2004 by CRC Press LLC
middle layer resists shear, but not bending, and the thickness is assumed to be small enough, so that the mass may be neglected or included in the outer layers. Damping is included in the middle layer in such a way that shear motions are resisted by a force proportional to the rate of shear. The outer layers obey the usual assumptions of the EulerBernoulli beam theory. The original model. For 0 < x < 1 and 0 < t < T , let w(t, x) = transverse displacement; and ξ(t, x) = effective rotation angle of the beam. Then the following mathematical model of a sandwich beam is proposed in [9], where Ω = (0, 1): wtt − ρwttxx − (dξx )xx = 0 in (0, T ] × Ω; (4.3.1a)
1 δ β (4.3.1b) + (ξ + wx ) = 0 in (0, T ] × Ω; (ξt + wtx ) − d ξx + wxx 3 4 3 x (4.3.1c) w(0, · ) = w0 , wt (0, · ) = w1 , ξ(0, · ) = ξ0 in Ω, along with appropriate Boundary Conditions (B.C.) at x = 0 and x = 1. We shall consider two cases: CASE 1: Hinged/Neumann B.C. for {w, ξ}. These are wΣ ≡ 0, wxx Σ ≡ 0; ξx Σ ≡ 0,
Σ ≡ (0, T ] × Γ,
(4.3.2)
where throughout this section, Γ = {0, 1} consists of the two end points x = 0 and x = 1. CASE 2: Clamped/Dirichlet B.C. for {w, ξ}. These are wΣ ≡ 0, wx Σ ≡ 0; ξΣ ≡ 0, Σ ≡ (0, T ] × Γ.
(4.3.3)
In (4.3.1ab), we have that: γ represents a rotational moment of inertia; D > 0 is the stiffness of the beam; δ > 0 is the stiffness of the middle layer; and β is the damping parameter for the middle layer. A ﬁrst transformation of model (4.1). We introduce a new variable [11] s(t, x) ≡ ξ(t, x) + wx (t, x)
(4.3.4)
in terms of the effective rotation angle ξ and the transverse displacement w. Physically, s is proportional to the shear of the middle layer. Then, model (1.1) in the variables {w, ξ} is transformed into the following system in the variables {w, s}, by virtue of (4.3.4): wtt − ρwttxx + dwxxxx − dsxxx ≡ 0 in (0, T ] × Ω; (4.3.5a) βs − 3d s − 3 w (4.3.5b) t xx xxx + δs ≡ 0 in (0, T ] × Ω; 4 w(0, · ) = w0 , wt (0, · ) = w1 ; in Ω, (4.3.5c) s(0, · ) = s0 = ξ0 + wx (0, · ) in Ω, along with the following Boundary Conditions at x = 0 and x = 1. CASE 1: Hinged/Neumann B.C. for {w, ξ} as in (4.3.2) yield Hinged/Neumann B.C. for {w, s}. In fact, direct use of (4.3.2) and (4.3.4) yields wΣ ≡ 0, wxx Σ ≡ 0; sx Σ = [ξx + wxx ]Σ ≡ 0,
© 2004 by CRC Press LLC
Σ = (0, T ] × Γ.
(4.3.6)
CASE 2: Clamped/Dirichlet B.C. for {w, ξ} as in (4.3.3) yield Clamped/Dirichlet B.C. for {w, s}. In fact, direct use of (4.3.2) and (4.3.4) yields wΣ ≡ 0, wx Σ ≡ 0; sΣ = [ξ + wx ]Σ ≡ 0,
Σ ≡ (0, T ] × Γ.
(4.3.7)
A ﬁnal transformation of model (4.1) into a thermoelastic model. With reference to (4.3.4) we introduce a new variable [28] 3 1 θ(t, x) ≡ sx (t, x) − wxx (t, x) ≡ ξx (t, x) + wxx (t, x). 4 4
(4.3.8)
Then, model (4.3.1) in the variables {w, ξ}, or model (4.3.5) in the variables {w, s} are transformed into the following system in the variables {w, θ}, by virtue of (4.3.8):
wtt − ρwttxx + d4 wxxxx − dθxx = 0 in (0, T ] × Ω;
(4.3.9a)
3 3δ βθt − 3dθxx + δθ + βwxxt + wxx in (0, T ] × Ω; 4 4 w(0, · ) = w0 , wt (0, · ) = w1 ; θ(0, · ) = θ0 = ξx (0, · ) + 14 wxx (0, · ) in Ω.
(4.3.9b) (4.3.9c)
Indeed, for instance, starting from (4.3.5): to obtain (4.3.9a) we replace sxxx = θxx + 3 4 wxxxx (obtained from (4.3.8)) into (4.3.5a). To obtain (4.3.9b), we rewrite (4.3.5b) as βst = 3dθx − δs by (4.3.8) and substitute βstx = 3dθxx − δsx = 3dθxx − δ θ + 34 wxx into βθt = βsxt − 34 βwxxt obtained from (4.3.8). With (4.3.9) we associate appropriate B.C. at x = 0 and x = 1. CASE 1: Hinged/Neumann B.C. (for {w, ξ}, hence) for {w, s} yields Hinged/Dirichlet B.C. for {w, θ}. In fact, direct use of (4.3.6), (4.3.2), and (4.3.8), yields 3 1 wΣ ≡ 0, wxx Σ ≡ 0; θΣ = sx − wxx = ξx + wxx ≡ 0, 4 4 Σ Σ
Σ = (0, T ]×Γ.
(4.3.10) CASE 2: Clamped/Dirichlet B.C. (for {w, ξ}, hence) for {w, s} yields Clamped/Neumann B.C. for {w, θ}. In fact, direct use of (4.3.7) with sΣ = 0, hence of st Σ ≡ 0, yields 3 1 wΣ ≡ 0, wx Σ ≡ 0; θx Σ = sxx − wxxx ≡ 0; ξxx + wxxx ≡ 0, 4 4 Σ
(4.3.11)
by virtue also of (4.3.8) and the equation in (4.3.5b). Thus, we recognize the {w, wt , θ}problem in (4.3.9abc) along with either the Hinged/Dirichlet B.C. (4.3.10), or else the clamped/Neumann B.C. (4.3.11), to be a thermoelastic problem of the type of Section 4.2. In Eq. (4.3.9b), the terms θ and wxx which are premultiplied by δ are lowerorder terms. Energy function and state space. With reference to model (4.3.1), we associate its
© 2004 by CRC Press LLC
natural energy function, which is given by 1 2 wt2 (t, x) + ρwtx (t, x) + dξx2 (t, x) Eρ (t) ≡ 0
d 4δ + ξx (t, x) + wxx (t, x)2 + ξ(t, x) + wx (t, x)2 dx. (4.3.12) 3 9 We readily see that the energy function (4.3.12) is welldeﬁned for t ≥ 0, provided that the variables {w(t), wt (t), ξ(t)} belong to the following state space, where Ω = (0, 1), 1 for ρ > 0, Yρ ≡ [H 2 (Ω) ∩ H01 (Ω)] × H0 (Ω) × H 1 (Ω) for ρ = 0, L2 (Ω) in the case of Hinged/Neumann B.C. for {w, ξ} in (4.3.2);
(4.3.13)
1 for ρ > 0, Yρ ≡ H 2 (Ω) × H0 (Ω) × H01 (Ω) for ρ = 0, L2 (Ω) in the case of Clamped/Dirichlet B.C. for {w, ξ} in (4.3.3).
(4.3.14)
In turn, with reference to problem (4.3.5), we see that the variable s deﬁned in (4.3.4) satisﬁes: s = ξ + wx ∈ H 1 (Ω) for {w, ξ} as in (4.3.13), (4.3.14); and hence, ﬁnally, the variable θ in (4.3.8) satisﬁes then θ = sx − 34 wxx ∈ L2 (Ω). We collect these results: 3 s = ξ + wx ∈ H 1 (Ω); θ = sx − wxx ∈ L2 (Ω). 4 4.4
(4.3.15)
Composite beams models under homogeneous hinged mechanical B.C. and boundary control on ξx Σ . Wellposedness. Singular estimates
Nonhomogeneous boundary problem. In this subsection, we continue the study of the sandwich beam model of Section 4.3, originally derived in the original variables {w, ξ} as in (4.3.1a–c); hence in the variables {w, s} as in (4.3.5a–c); ﬁnally, in the “thermoelastic” variables {w, θ} as in (4.3.9a–c), under the following nonhomogeneous B.C.: wΣ ≡ 0,
wxx Σ ≡ 0;
ξx Σ ≡ u,
(4.4.1a)
or sx Σ ≡ [ξx +wxx ]Σ ≡ u;
or
1 3 θΣ ≡ ξx + wxx ≡ sx − wxx ≡ u. (4.4.1b) 4 4 Σ Σ
Following [27], we shall concentrate on the reduction of {w, ξ} to the thermoelastic variables {w, θ}. Wellposedness of {w, θ}problem. In this subsection, we shall exploit the reduction of the original {w, ξ}system to its thermoelastic version of {w, θ} with hinged B.C. in the
© 2004 by CRC Press LLC
mechanical variables w, and thermal boundary control u in the Dirichlet B.C. of θ. To this end, we introduce the following operators, where ρ > 0 throughout: Af = −∆f, 2
2
A f = ∆ f,
D(A) = H 2 (Ω) ∩ H01 (Ω); Aρ = (I + ρA); 2
4
(4.4.2)
D(A ) = {f ∈ H (Ω) : f Γ = ∆f Γ = 0},
(4.4.3)
Y ≡ Yρ ≡ X1 × X2 × X3 ;
(4.4.4)
and the space X1 ≡ D√D/4 (A); x1 2X1
≡
1 2
X2 ≡ D(Aρ );
D Ax1 2L2 (Ω) ; 4
X3 ≡ L2;(√4/3) (Ω); 1
x2 2X2 = Aρ2 x2 2L2 (Ω) ;
4 x3 2L2 (Ω) . 3 Thus, for f ≡ [f1 , f2 , f3 ] ∈ Yρ , g = [g1 , g2 , g3 ] ∈ Yρ , the Yρ inner product is x3 2X3
=
(4.4.5) (4.4.6) (4.4.7)
4 D (4.4.8) (Af1 , Ag1 )L2 (Ω) + ((I + ρA)f2 , g2 )L2 (Ω) + (f3 , g3 )L2 (Ω) . 4 3 Further, we introduce the operator 0 I 0 D −1 2 −1 − A A 0 −A A ρ ρ Aρ = : Yρ ⊃ D(Aρ ) → Yρ , (4.4.9) 4 3δ 3 3D δ A A − A− 4β 4 β β (f, g)Yρ =
with its natural domain. The following result then follows from thermoelastic system [1], [17]. Lemma 4.4.1. Let ρ > 0. (i) The operators 0 0 I d −1 2 d , − 4 Aρ A −1 2 − Aρ A 0 4 0
I
0
0
−A−1 ρ A
,
(4.4.10)
3 A 0 4 with their natural domains, are unitary operators on X1 × X2 , and Yρ = X1 × X2 × X3 , respectively. (ii) For δ = 0, the operator Aρ in (4.4.9) is maximal dissipative, hence the generator of an s.c. contraction semigroup on Yρ . Similarly for its Yρ adjoint A∗ρ . (iii) Let δ > 0. The operator obtained from Aρ in (4.4.9) by removing the left bottom 3δ corner entry 4β A is maximal dissipative, and hence the generator of an s.c. contraction semigroup on Yρ . What is left is the operator 0 0 (4.4.11) 4δ ∈ L(Yρ ), A 0 3β
© 2004 by CRC Press LLC
which is bounded on Yρ . (iv) Thus, the operator Aρ in (4.4.9), being a bounded perturbation of the generator of an s.c. contraction semigroup on Yρ , is itself the generator of an s.c. semigroup on Yρ (noncontraction, however), which is denoted by eAρ t , t ≥ 0. (v) System (4.3.9abc) in {w, θ} with hinged/Dirichlet homogeneous B.C. (4.3.10) (Case 1), can be rewritten abstractly as y(t) ˙ = Aρ y(t),
y(t) = {w(t), wt (t), θ(t)],
(4.4.12)
with Aρ deﬁned by (4.4.9). Thus, for any I.C. y0 = [w0 , w1 , θ0 ] ∈ Yρ , we have y(t) ≡ [w(t), wt (t), θ(t)]
= eAρ t y0 = eAρ t [w0 , w1 , θ0 ] ∈
C([0, T ]; Yρ ).
(4.4.13)
(vi) System (4.3.9abc) on {w, θ} with nonhomogeneous B.C. wΣ = wxx Σ ≡ 0, θΣ = u in (4.4.1), can be rewritten abstractly as y(t) ˙ = Aρ y(t) + Bu ∈ [D(A∗ρ )] ,
(4.4.14)
where Aρ is deﬁned in (4.4.9), and B is deﬁned by
0
ADu Bu = A−1 ρ ADu
where D is the Dirichlet map.
1 ; D : L2 (Γ) → H 2 (Ω);
∆(Dg) = 0 Dg = g
(4.4.15)
in Ω; on Γ,
Lemma 4.4.1 veriﬁes assumptions (H.1), (H.2) of Section 1. Singular estimate for eAρ t B. We now come to the veriﬁcation of the critical assumption (H.3) = (1.2) of Section 1. Proposition 4.4.2. With reference to system (4.3.9abc) in {w, θ} with nonhomogeneous B.C. (4.4.1), or its abstract version (4.4.13), we have: (i) Let δ = 0 in (4.3.9b) or (4.4.9). Then, with corresponding eAρ t B satisﬁes the singular estimate: eAρ t BL(Yρ ) ≤
CT , tγ
0 < t ≤ T, γ =
3 + , ∀ > 0. 4
(4.4.16)
(ii) Now let δ > 0. Then, the corresponding eAρ t B satisﬁes the same singular estimate as in (4.4.16). Explicitly, this means the following: Consider system (4.3.9ab) in {w, θ} with homogeneous B.C. (4.3.10): wΣ ≡ wxx Σ ≡ θΣ ≡ 0, and Initial Conditions (see 4.4.15)): w0 = 0, w1 = A−1 θ0 = ADu, u ∈ R2 . (4.4.17) ρ ADu,
© 2004 by CRC Press LLC
Then, the corresponding solution {w(t), wt (t), θ(t)} satisﬁes the following singular estimate, recalling (4.4.4)–(4.4.8): w(t)H 2 (Ω) + wt (t)H 1 (Ω) + θ(t)L2 (Ω) ≤
CT u, 0 < t ≤ T, tγ 3 γ = + , 4
(4.4.18)
or, in terms of ξ, recalling (4.3.8), 1 CT ξx (t)L2 (Ω) ≤ θ(t)L2 (Ω) + wxx (t)L2 (Ω) ≤ γ , 0 < t ≤ T. 4 t
(4.4.19)
Proof. (i) For δ = 0 (canonical thermoelastic case), the singular estimate (4.4.16) is established in [14], using techniques similar to those used in obtaining singular estimates in Sections 4.1 and 4.2. (ii) For δ > 0, the singular estimate (4.4.16) is retained, since the case δ > 0 yields only a bounded perturbation of the generator (4.4.9) over the case δ = 0. This is true because of the result in Appendix, Proposition A.2. Then, recalling that the space Yρ in (4.4.5) is topologically equivalent to H 2 (Ω) × H 1 (Ω) × L2 (Ω), we readily see that the singular estimate (4.4.16) means precisely the estimate (4.4.18) for problem (4.3.9abc) with initial condition Bu given by (4.4.17), in view of (4.4.15). Then estimate (4.4.19) follows by (4.4.18) and ξx = θ − 14 wxx by (4.3.8). The corresponding cost functional. We consider the optimal control problem of Section 1, for the boundary control problem (4.3.9abc), (4.4.1), with cost functional given by T J(u; {w, wt , ξ}) =
{w(t)2H 2 (Ω) + wt (t)2H 1 (Ω)
0
+ ξ(t)2H 1 (Ω) + u(t)2 }dt.
(4.4.20)
Conclusion. The boundary control system (4.3.9abc), (4.4.1) describing a sandwich beam with boundary control and cost functional (4.4.20) satisﬁes the setting of Section 1, in particular, assumptions (H.1), (H.2), (H.3) with Y ≡ Yρ given by (4.4.4), A = Aρ and B given by (4.4.9) and (4.4.15), and cost given by (4.4.20). Accordingly, the results of Sections 1–3 apply to it.
4.5
A structural acoustic 2D model with a composite (sandwich) beam as ﬂexible wall
In this subsection we return to the 2D model of a structural acoustic chamber of Section 4.1 and replace the elastic ﬂexible wall of Example 4.1, or the thermoelastic ﬂexible wall of Example 4.2 with a composite (sandwich) beam, as in Section 4.3. The resulting coupled PDE system is
© 2004 by CRC Press LLC
ztt = ∆z − d1 zt + f ∂z ∂ν + d2 zt = 0 ∂z ∂ν + d3 zt = ±vt
in (0, T ] × Ω,
(4.5.1a)
in (0, T ] × Γ1 ,
(4.5.1b)
in (0, T ] × Γ0 ,
(4.5.1c)
vtt − ρvttxx − (dξx )xx ± zt = 0 in (0, T ] × Γ , 0 β 1 in (0, T ] × Γ0 , 3 (ξt + vtx ) − d(ξx + 4 vxx )x δ + (ξ + v ) = 0 x 3 on (0, T ] × ∂Γ0 , v = 0, vxx  = 0, ξx  = u
(4.5.1d) (4.5.1e) (4.5.1f)
z(0, · ) = z0 , zt (0, · ) = z1
in Ω,
(4.5.1g)
v(0, · ) = v0 , vt (0, · ) = v1
in Γ0 .
(4.5.1h)
where the ﬁrst system refers to the acoustic chamber, and the second system to the composite beam. Problem (4.5) satisﬁes the abstract assumptions (H.1), (H.2), (H.3) of Section 1, on the state space Y = H 1 (Ω) × L2 (Ω) × X1 × X2 × X3 , for
(4.5.2)
1 z(t), zt (t), v(t), vt (t), θ(t) = ξx (t) + vxx (t) 4
with Xi , i = 1, 2, 3 deﬁned by (4.4.5)–(4.4.7); where γ = 34 + , > 0 as in Section 4.2. All this follows by using Sections 4.3, 4.4 in the analysis of Example 4.2.
APPENDIX: TWO RESULTS ON SINGULAR ESTIMATES In this appendix we collect two results of interest regarding the singular estimate condition (H.3) = 1.2. Proposition A.1. Assume (H.1), with, say, eAt ≤ M eω0 t , t ≥ 0, as well as (H.2), and (H.3), so that [21, Lemma 1.1], the following singular estimate holds true: eAt BL(U ;Y ) ≤
M eωt , tγ
0 < t; 0 < γ < 1,
(A.1)
for a constant ω > ω0 . [If the s.c. semigroup eAt is, moreover, exponentially stable, then the singular estimate (H.3) on 0 < t ≤ T implies that the singular estimate (A.1) on 0 < t
© 2004 by CRC Press LLC
with some ω < 0]. Then, for all λ ∈ C with Re λ > ω, we have n d R(λ, A)Bu n+1 (λ, A)BuL(Y ) = n!R dλn L(Y )
(A.2)
M Γ(1 − γ + n) uU , n = 0, 1, 2, . . . , Re λ > ω, (Re λ − ω)1−γ+n (A.3)
≤
where Γ( · ) is the gamma function. (ii) Conversely, suppose that estimates (A.3) hold true for all λ = Re λ (real) > ω, and 0 < γ < 1. Then, we obtain eAt BL(U ;Y ) ≤
M eωt , tγ
0 < t, 0 < γ < 1.
(A.4)
Proof. First, (A.1) follows from (H.3) as in the elementary proof of [21, Lemma 1.1]. Next,
∞ R(λ, A)Bu =
e−λt eAt Bu dt,
(A.5)
0
where u ∈ U , and thus interpreted at ﬁrst in [D(A∗ )] , we obtain by (A.1) by recalling known results [8, p. 55] on the Laplace transform, with Re λ > ω, γ < 1, dn R(λ, A)Bu = dλn n d R(λ, A)Bu dλn
∞ tn e−λt eAt Bu (−1)n dt; tγ 0
∞
L(Y )
(A.6)
≤ M
e−(Re λ−ω)t tn−γ dtuU
0
=
M Γ(1 − γ + n) uU , (Re λ − ω)1−γ+n
(A.7)
since n − γ > −1, as required, n = 0, 1, . . .. Thus (A.7) establishes (A.2) since dn R(λ, A) = (−1)n n! Rn+1 (λ, A), λ ∈ ρ(A). dλn
(A.8)
(ii) Case ω = 0. Conversely, it sufﬁces to restrict, at ﬁrst, to the case ω = 0. Thus, assume the validity of estimates (A.3) for, say, λ = Re λ > ω = 0. We then seek to establish conclusion (A.4) with ω = 0. Step 1. We shall use the wellknown formula of HillePhillips: eAt Bu = lim
k→∞
© 2004 by CRC Press LLC
k R t
k ,A t
k Bu,
t > 0, u ∈ U,
(A.9)
again interpreted, at ﬁrst, in [D(A∗ )] . We estimate k k k Γ(k − γ) M ,A , Bu ≤ γ kγ R t t t (k − 1)!
k = 1, 2, . . . .
(A.10)
=
kk Rk k , A Bu tk t L(Y )
(A.11)
≤
kk M Γ(k − γ) uU k t (k − 1)! k k−γ
(A.12)
M k γ Γ(k − γ) uU , tγ (k − 1)!
(A.13)
L(Y )
In fact, k k k Bu ,A R t t
L(Y )
(by (A.3))
t
=
t > 0,
where, in (A.12), we have recalled assumption (A.3) with ω = 0 and Re λ = λ = assumed. Step 2. Next, with 0 ≤ γ < 1, we shall show that k γ Γ(k − γ) ≤ const., (k − 1)!
∀ k = 1, 2, . . . ,
as required, after which we then obtain by (A.13), (A.14), k k k C Bu ≤ γ uU , for all 0 < t, ,A R t t t
k t,
as
(A.14)
(A.15)
L(Y )
as desired. To prove (A.14), we set x = k − 1 − γ = K − γ, k − 1 = K, and recall the Stirling formula [Ru, p. 194], to obtain x x √ Γ(k − γ) = Γ(x + 1) ∼ 2πx e (K − γ)K−γ = 2π(K − γ). (A.16) eK−γ Thus, k γ Γ(k − γ) (k − 1)!
∼
(K + 1)γ (K − γ)K−γ 2π(K − γ) K! eK−γ
=
∼
© 2004 by CRC Press LLC
K +1 K −γ K +1 K −γ
γ eγ γ γ
e
(A.17)
(K − γ)K 2π(K − γ) K!eK
(A.18)
2π(K − γ) √ , K K 2πK
(A.19)
(K − γ)K
where in the last step we have again invoked the Stirling formula for √ K! eK ∼ K K 2πK.
(A.20)
Then (A.19) establishes (A.14), as desired, at least for ω = 0. Case = 0. Finally, let (A.3) hold true with ω = 0 and λ = Re λ > ω. Deﬁne: ˜ = λ − ω > 0, so that λI ˜ − A˜ = λI − A and Rn (λ, ˜ A) ˜ ≡ Rn (λ, A). Thus A˜ ≡ A − ωI, λ M Γ(1 − γ + n) ˜ , λ > 0. ˜ (1−γ+n) λ
˜ A)BL(U ;Y ) = Rn (λ, A)BL(U,Y ) ≤ Rn (λ,
(A.21)
Then the previous part with ω = 0 applies to A˜ and we obtain CT , 0 < t; 0 < γ < 1, (A.22) tγ and the desired conclusion (A.4) follows for eAt B. The proof of Proposition A.1 is complete. ˜
e−ωt eAt BL(U ;Y ) = eAt BL(U ;Y ) ≤
The next Proposition shows that the singular estimate (H.3) = (1.2) for eAt B is preserved under a bounded perturbation Π of A. Proposition A.2. Assume (H.1), (H.2), (H.3) = (1.2). Let Π ∈ L(Y ). Then, e(A+Π)t B satisﬁes the same singular estimate (1.2), C , tγ
e(A+Π)t BL(U ;Y ) ≤
0 < t ≤ T, 0 < γ < 1.
(A.23)
Proof. Let y0 ∈ Y . Setting ˙ + (A + Π)y(t) − Πy(t), y(t) ≡ eAt y0 , or y(t)
(A.24)
we may write by the variation of the parameter formula, (A+Π)t
y(t) ≡ e y0 = e At
t y0 −
e(A+Π)(t−τ ) ΠeAτ y0 dτ.
(A.25)
0
Thus, from (A+Π)t
e
t Bu = e Bu + At
e(A+Π)(t−τ ) ΠeAτ Bu dτ,
(A.26)
0
with u ∈ U , and so interpreted, at ﬁrst, in [D(A∗ )] , we estimate for t > 0, u ∈ U , in the L(Y )norm: e(A+Π)t Bu ≤
T eAt Bu +
CT Π eAτ Budτ
(A.27)
0
≤
C
T
tγ
= and (A.29) establishes (A.23).
© 2004 by CRC Press LLC
T + MT 0
1 dτ uU τγ
T 1−γ CT uU , + MT tγ 1−γ
(A.28)
(A.29)
REFERENCES 1. G. Avalos and I. Lasiecka, “Differential Riccati equations for the active control of a problem in structural acoustics,” JOTA 91 (1996), 695–728. 2. A. V. Balakrishnan, Applied Functional Analysis, SpringerVerlag, 1981. 3. A. V. Balakrishnan, “Boundary control of parabolic equations, LQR theory,” Proc. V International Summer School, Control Inst. Math. Mech. Acad. Sci., GDR, Berlin (1977). 4. A. Bensoussan, G. Da Prato, M. Delfour, and S. Mitter, Representation and Control of Inﬁnite Dimensional Systems, vols. 1 and 2, Birkh¨auser, 1993. 5. F. Bucci, I. Lasiecka, and R. Triggiani, “Singular estimates and uniform stability of coupled systems of hyperbolic/parabolic PDEs,” Abstract and Applied Analysis, vol. 7(4) (2002), 169–237. 6. P. Grisvard, “Characterisation del quelques espaces d’interpolation,” Arch. Rational Mech. Anal. 25 (1967), 40–63. 7. P. Grisvard, Elliptic Problems in Nonsmooth Domains, Monographs and Studies in Mathematics, 24, Pitman, 1985. 8. W. Grobner and N. Hofreiter, Integraltafeln, Zweiter Teil, Bestimmte Integral, SpringerVerlag, 1973. 9. S. Hansen, “A model for a twolayered plate with interfacial slip,” in Control and Estimation of Distributed Parameter Systems: Nonlinear Phenomena, Vol. 1SNA118, Editors: W. Desch, F. Kappel, K. Kunish, Birkh¨auser, 143–170. 10. S. Hansen, “Modeling and analysis of multilayer laminated plates,” ESAIM Proc. 4 (1998), 117–135. 11. S. Hansen and I. Lasiecka, “Analyticity, hyperbolicity and uniform stability of semigroups arising in models of composite beams,” Mathematical Models and Methods in Applied Sciences, vol. 10 (2000), 555–580. 12. I. Lasiecka, “Optimization problems for structural acoustic models with thermoelasticity and smart materials,” Discussiones Mathematicae, Differential Inclusions, Control and Optimization, vol. 20 (2000), 113–140. 13. I. Lasiecka, Mathematical Control Theory of Coupled PDEs Systems, SIAM NSFCMBS Lecture Notes, N. 75, 2002. 14. I. Lasiecka, “Wellposedness of optimal control problems for systems with unbounded controls and partially analytic generators,” Control and Cybernetics 31 (2002), 751– 777. 15. I. Lasiecka and R. Triggiani, Differential and Algebraic Riccati Equations with Applications to Boundary/Point Control Problems: Continuous and Approximation Theory, LNICS, Springer Verlag, 1991. 16. I. Lasiecka and R. Triggiani, Control Theory for Partial Differential Equations, vol. I: Abstract Parabolic Equations; Vol II: Abstract hyperboliclike systems over a ﬁnite time horizon. Cambridge University Press, Encyclopedia of Mathematics and its Applications, January 2000. 17. I. Lasiecka and R. Triggiani, “Structural decomposition of thermoelastic semigroups with rotational forces,” Semigroup Forum 60 (2000), 16–66. 18. I. Lasiecka and R. Triggiani, “Two direct proofs on the analyticity of the s.c. semigroup arising in abstract thermoelastic equations,” Advances Diff. Eqns. 3(3) (May 1998), 387–416.
© 2004 by CRC Press LLC
19. I. Lasiecka and R. Triggiani, “Analyticity of thermoelastic semigroups with coupled hinged/Neumann B.C.,” Abstract Appl. Anal. 3(1–2) (1998), 153–169. 20. I. Lasiecka and R. Triggiani, “Analyticity of thermoelastic semigroups with free B.C.,” Annali Scuola Normale Superiore, Pisa, Cl. Sci. (4), vol. 37 (1998), 457–482. 21. I. Lasiecka and R. Triggiani, “Optimal control and algebraic riccati equations under singular estimates for eAt B in the absence of analyticity. Part I: The stable case,” in Differential Equations and Control Theory, Lecture Notes in Pure and Applied Mathematics, Editors: S. Aizicovici and N. Pavel, vol. 225 (2001), Marcel Dekker, 198–219. 22. I. Lasiecka, R. Triggiani, and X. Zhang, “Nonconservative wave equations with unobserved Neumann B.C.: Global uniqueness and observability in one shot,” Contemporary Mathematics, vol. 268, AMS (2000), 227–326. 23. K. Liu and M. Renardy, “A note on equations of a thermoelastic plates,” Appl. Math. Letters 8 (1995), 1–6. 24. W. Rudin, Principles of Mathematical Analysis, 3rd edition, McGrawHill, New York, 1976. 25. C. Sadovsky, Interpolationn of Operators and Singular Integrals, Monographs in Pure and Applied Mathematics, Marcel Dekker, 1979. 26. R. Triggiani, “Minmax game theory and optimal control with indeﬁnite cost under a singular estimate for eAt B in the absence of analyticity.” Invited paper in a special volume: Evolution Equations, Semigroups and Functional Analysis, Editors: A. Lorenzi and B. Ruf, Birkh¨auser Verlag, 2002, 353–380. (In memory of B. Terreni.) 27. R. Triggiani, “Backward uniqueness of the s.c. semigroup arising in the coupled PDEsystem of structural acoustic,” Advances Diff. Eqns., to appear. 28. R. Triggiani, “The coupled PDE system of a composite (sandwich) beam revisited,” Discrete & Continuous Dynamical Systems–Series B; vol. 3(2) (May 2003), 285–298.
© 2004 by CRC Press LLC
20 Nonlinear Problems of Spacecraft Fault Tolerant Control Systems V.M. Matrosov∗† and Ye.I. Somov∗‡
∗
Stability and Nonlinear Dynamics Research Center, Mechanical Engineering Research Institute, Russian Academy of Sciences, Moscow, Russia † Moscow Aviation Institute (State Technical University), Moscow, Russia ‡ Research Institute of Mechanical Systems Reliability, Samara State Technical University, Samara, Russia
A brief overview and new results on nonlinear methods for research of logicdynamic faulttolerant control systems and their applications to precise gyromoment attitude control of information spacecraft are presented in this chapter.
1
INTRODUCTION
During the last 25 years the authors have accumulated substantial experience in research of the spacecraft attitude control systems (ACSs) with high fault tolerance, survivability, and autonomy at the expense of using functional excessibility. The dynamic requirements of ACS of contemporary spacecraft for communication, navigation, and geodetics are: (i) continuous precision 3axis orientation of the spacecraft body, which requires only a minimum number of measurements under possible failures of the ACS onboard equipment, disturbances on optical devices from the Sun, the Moon, etc., and also in executing the spacecraft orbit correction; (ii) the possibility of the spacecraft body reorientation for its orbit correction, as well as autonomous orientation of the solar array panels (SAPs) and each highgain receivingtransmitting antenna (RTA) with respect to the spacecraft body; (iii) robustness to variations of the spacecraft inertial and rigidity characteristics under minimum mass, size, and power expenditures. For remote sensing spacecraft there is the need to orient the line of sight to a predetermined part of the Earth’s surface with the scan in a designated direction and to compensate an image motion at the onboard optical telescope’s focal plane. Moreover, for the loworbit remotesensing spacecraft these requirements are expressed by rapid angular maneuvering and spatial compensative motion with a variable vector of angular rate. Increased requirements of such information spacecraft — exactness of spatial rotation maneuvers (SRMs) with the effective damping of the spacecraft ﬂexible construction oscillations, efﬁciency, fault tolerance, reliability, as well as reasonable mass, size, and energy characteristics, has motivated intensive development the ACS with executive devices in the form of moment gyrocomplex (MGC) based on an excessive number of gyrodynes — singlegimbal control moment gyroscopes (CMGs).
© 2004 by CRC Press LLC
Mathematical aspects of nonlinear spacecraft attitude control [1] were represented in a number of research works [2]. The recent results of B.R. Hoelscher and S.R. Vadali [3] on nonlinear dynamics of gyromoment ACSs are based on methods of optimization and Lyapunov functions, whereas the exact feedback linearization (EFL) technique, suggested by A. Isidori [4], also was applied to this problem by S.N. Singh and T. Bosart [5]. This chapter discusses our newest results on nonlinear logicdynamic faulttolerant gyromoment ACSs for considered classes of information spacecraft.
2
NONLINEAR FAULTTOLERANT FEEDBACK CONTROL
Let there be the nonlinear generalized controlled object O: D+ x(t) = F(x(t), u, p(t, x), γνf (t)), x(t0 ) = x0 ; y(t) = ψ o (x(t), γνf (t));
zo (t) = φo (x(t), y(t), p(t, x)),
(1) (2)
s
where x(t) ∈ H ⊂ Rnν ; x0 ∈ H0 ⊆ H; y(t) ∈ Rrν is an output vector for measurement f and diagnosis of an object’s state, and zo (t) ∈ Rrν is a vector for description of its failure c conditions; u = {uj } ∈ U ⊂ Rrν is a control vector, and p(t, x) ∈ P is the vectorfunction of disturbances in class P; D+ is the symbol of a right derivative with respect to time t ∈ Tt0 = [t0 , ∞); a vector γνf (t) ∈ B m ≡ B×B · · · ×B, where B = {0, 1}, is a vector of logic variables, which are outputs of a “fault” asynchronous logic automaton (ALA) Af : γνf = δ f (κfν , lfν ); κfν+1 = λf (κfν , lfν ), κf0 = κf (0), ν ∈ N0 = [0, 1, 2, . . . ),
(3)
with memory, logic vectors of the object’s state κfν = κf (ν) and input lfν = lf (ν) = gf (zo (tfν )), which are used for representing fault occurrences and damage development depending on the automaton time ν, bound up with the continuous time as t = tfν + (τ f − tfν ); τ f ∈ τνf = [tfν , tfν+1 ), ν ∈ N0 . Moreover lfν (t) = const ∀t ∈ τνf and change of the logic vector γνf in general leads to variation of dimensions for vectors x(t) and y(t) under mappings in time moments t = tfν : x(tfν+ ) = Pνx (x(tfν− )); y(tfν+ ) = Pνy (y(tfν− )). Let there also be a given subsystem for discrete measurement of the object’s state and digital ﬁltering: • for diagnostics of the structural state of object O: d ysd = ψ d (ys ); zdf k = FTu (ys ); k, s ∈ N0 ;
(4)
• for forming the control loop and its reconﬁguration: ysu = ψ u (ys ); ykf = FTu (ysu ); zfµ = FTr (zdf k ), Tr = tµ+1 − tµ ; µ, k, s ∈ N0 .
© 2004 by CRC Press LLC
(5)
We use denotations xk = x(tk ); tk = kTu , ts = sTq , k, s ∈ N0 ; xfk = FTu (xs ), where Tu and Tq ≤ Tu are ﬁxed sampling periods of control and state measurement; moreover, multiplicity conditions must be satisﬁed for the periods Tq , Tu , Tr ; a variable xfk is the value of a variable xs measured with the sampling period Tq , which is ﬁltered out at time t = tk ; FTy (·) is the digital ﬁltering operator with the sampling period Ty , y = u, r. Principal problems are contained in the synthesis of: • the synchronous logic automaton (SLA) Ad for the ﬂowing structural state diagnosis: γkd = δ d (κdk , ldk ); κdk+1 = λd (κdk , ldk ), κd0 = κd (t0 ); k ∈ N0 ,
(6)
d with memory, logic vectors of state κdk , input ldk = gd (zdf k ) and output γk ;
• the SLA Ar , also with memory, for the description of damage blockkeeping and the reconﬁguration sequence depending on the automaton time µ and sampling time tµ = µTr : γµr = δ r (κrµ , lrµ ); κrµ+1 = λr (κrµ , lrµ ), κr0 = κr (t0 ); µ ∈ N0 ,
(7)
with logic vectors of state κrµ , input lrµ = gr (zfµ , γµdf ) with γµdf = FTr (γkd ) and output γµr ; • the nonlinear control law (NCL) with its reconﬁgurations due to the SLA Ar routine: xe k , yef k , yo k , γµr ); x ˆe k+1 = Fˆe (ˆ xe k , yef k , yko , uk , γkd , γµr ), x ˆe 0 = x ˆe (t0 ), uk = U(ˆ (8) e where k, µ ∈ N0 ; yef k = FTu (ψeu (ye s )); ye s = ψeo (xe s , γkd ); xe s = xe (ts ) ∈ Rnµ is the state vector of a simpliﬁed discrete object’s model: xe s+1 = Fe (xe s , uk , γkd , γµr ), xe 0 = xe (t0 ); s, k, µ ∈ N0 ,
(9)
e
ˆe (tk ) ∈ Rnµ is its estimation; neµ ≤ n = max{nν } and yko is a and x ˆe k = x programmed vector. Feedback loops (4)–(9) are intended for faulttolerant control of objects (1)–(3). Over the last two decades, the basic research on fault diagnostics, the isolation of faults, and reconﬁguration of control systems has received much attention: see survey [6]. The modelbased approach to this problem is now recognized as an important method for linear control systems; moreover there are trends in extending that methodology to nonlinear systems [7]. Therefore now we present in detail only our approach to synthesis of nonlinear control systems when their structural state is known. Along with different methods for nonlinear dynamic investigations as the primary approach, we use the method of vector Lyapunov functions [8, 9, 10, 11, 12], which has a strong mathematical basis for the analysis of stability and other dynamical properties of various nonlinear interconnected and logicdynamic systems with the discontinuous righthand side, depending on bounded control, in cooperation with exact feedback linearization technique [4]. Let there be a nonlinear controlled object with a certain sequence of the structural state γνf (t) and, for simplicity, with a complete and continuous measurement of its state: D+ x(t) = F(x(t), u, pf (t, x)); pf (t, x) ≡ {p(t, x), γνf (t)}; x(t0 ) = x0 ; x ∈ Rn ; t ∈ Tt0 ,
© 2004 by CRC Press LLC
l
l0
and also let vector norms ρ(x) ∈ R+ and ρ0 (x0 ) ∈ R+ be given. For any control law u = U(x) this model has the form D+ x(t) = F(x, U(x), pf (t, x)) ≡ X (t, x); x(t0 ) = x0 ,
(10)
where X : Tt0 × H → H is a discontinuous operator. Assuming the existence and the nonlocal continuability of the rightsided solution x(t) ≡ x(t0 , x0 ; t) of the closedloop control system (10) for its extended deﬁnition in the aspect of physics (engineering), the most important dynamic property is obtained, that is, l ρρ0 exponential invariance of the solution x(t) = 0 under the desired γ ∈ R+ : l×l0
(∃ α ∈ R+ ) (∃ B ∈ R+ ) (∃ δ ∈ Rl+0 ) (∀ ρ0 (x0 ) < δ) ρ (x(t)) ≤ γ + B ρ0 (x0 ) · exp(−α(t − t0 )) ∀t ∈ Tt0 .
(11)
k
For the vector Lyapunov function (VLF) υ : H → R+ with the components υ s (x) ≥ 0, υ s (0) = 0, s = 1 : k and the norm υ(x) = max {υ s (x), s = 1 : k}, deﬁned are the scalar function υ(x) = max{υ s (x), s = 1 : lk , 1 ≤ lk ≤ k} and the upperright derivative of the VLF υ(x) with respect to (10) as υ (x) = lim {(υ(x + δt · X (t, x)) − υ(x))/δt} . δt→0+
k
T HEOREM . Let there exist the VLF υ : H → + , such that: 10 (∃a ∈ Rl+ ) (∀x ∈ H)
ρ (x) ≤ a · υ(x);
20 (∃b ∈ Rl+0 ) (∀x0 ∈ H0 ) υ(x0 ) ≤ b , ρ0 (x0 ); k
30 (∃γc ∈ R+ ) and a function ϕγ (·) = {ϕsγ (·)} exists such that γc ≤ ϕγ (a, γ); 40 (∀(t, x) ∈ (Tt0 × H) the conditions are satisﬁed: ˙ fc (t, υγ (x)) ≡ P · υγ (x) + ˜fc (t, υγ (x)) where υγ (x) ≡ υ(x) − γc ; a) υγ (x) ≤ b) Hurwitz condition (Re λs (P) < 0, s = 1 : k) for the positive (k × k)matrix P; c) Wa˘zewski condition on the quasimonotonicity for the vectorfunction ˜fc (t, y): f˜cs (t, y ) ≤ f˜cs (t, y), s = 1 : k for ys = ys and yj ≤ yj , j = s; d) Carateodory condition for the vectorfunction ˜fc (t, y), bounded in each domain: Ωrc = (Tt0 × Scr ), where r > 0 and Scr = {y ∈ Rk : yE < r}; t∈Tt e) ˜fc (t, y)/y =⇒0 0 for y → 0 uniformly with respect to time t ∈ Tt0 . Then the system (10) is ρρ0 exponential invariant and the matrix B in the deﬁnition (11) has the form B = abt with a parameter c ∈ R+ . P ROOF. In the traditional scheme of the comparison principle [8], the basis of inequality for vector norm ρ (x(t)) in (11) is attained [13, 14, 15, 16] using the maximum rightsided solution xc (t) ≡ xc (t0 , xc0 ; t) of the comparison system (CS): . D+ xc (t) = fc (t, xc (t) − γc ) ≡ P(xc (t) − γc ) + ˜fc (t, xc (t) − γc ); xc (t0 ) ≡ xc0 = υ(x0 ). (12)
© 2004 by CRC Press LLC
For the conditions (40 ) according to the comparison principle under the condition ξc ≡ max{υ s (x0 ) − γcs , s = 1 : k} ≡ max{xc0 − γc , s = 1 : k} ≤ δc ; δc ∈ R+ , the CS (12) has the dynamic property upper exponential semiinvariance, R.I. Kozlov [17, Ch. III, §1]: ∃αc ∈ R+ and ∃βc ∈ Rk+ such that the vector inequality υ(x(t)) ≤ xc (t) ≤ γc + β c (ξc ) exp(−αc (t − t0 )) ∀t ∈ Tt0 ; 0 υ(x0 ) ≤ γc ; β c (ξc ) ≡ βc · ξc υ(x0 ) ≤  γc , is satisﬁed. Moreover, according to the conditions (10 ) – (30 ) the upper bound of the vector norm ρ (x(t)) is valid: s
ρ (x(t)) ≤ a · max{γcs + β c (ξc ) exp(−αc (t − t0 ), s = 1 : lk , 1 ≤ lk ≤ k} ≤ a · max{γcs , s = 1 : lk } + max{βcs , s = 1 : lk } a · υ(x0 ) exp(−αc (t − t0 ) ≤ γ + c · B ρ0 (x0 ) · exp(−α(t − t0 )) ∀t ∈ Tt0 , where γ = a · max{ϕsγ (a, γ), s = 1 : lk }, c = max{βcs , s = 1 : lk }, and α = αc − ε with ε ≥ 0. As is wellknown [18], a lot of similar theorems [17] are quite useless for a control engineering practice until we have solved an important problem: by what approach is it possible to create constructive techniques for constructing the vector Lyapunov function υ(x) and simultaneous synthesis of a nonlinear control law u = U(x) for the closedloop system (10) with given vector norms ρ(x) and ρ0 (x0 ) ? During the last 25 years we have persistently searched for a general approach to this basic problem and carefully varied our results with application to the gyromoment ACSs of information spacecraft in Kazan [19, 20, 21, 22], since 1984 in Irkutsk [13, 14, 23, 24, 25], and since 1997 in Moscow [15, 16, 26, 27, 28]. As a result, today a pithy technique on constructing VLF for the synthesis of nonlinear control law has been elaborated. This method is based on a nonlinear transformation of the nonlinear control system’s model to a canonical representation by solving the problem in the two stages. In stage 1, the right side F(·) in the model (10) is transformed as ˜ x(t), u) F(x(t), u, pf (t, x)) = f(x) + G(x) u + F(t, ˜ ⊂ Rn˜ ⊆ Rn with n and for some principal variables in a state vector x ∈ H ˜ ≤ n, ˜ ˜ x0 ∈ H0 ⊆ H, a simpliﬁed nonlinear model of a nonlinear controlled object is presented in the form of an afﬁne quite smooth nonlinear system (13) x˙ = F(x, u) ≡ f(x) + G(x)u = f(x) + gj (x) · uj , which is structurally synthesized by the exact feedback linearization technique. In this aspect, based on the structural analysis of given vector norms ρ(x) and ρ0 (x), and also vectorfunctions f(x) and gj (x), the output vectorfunction h(x) = {hi (x)} is carefully selected. Futhermore, the nonlinear invertible (onetoone) coordinate transformation z = ˜ with Φ(0) = 0 is analytically obtained according to Φ(x) ∀x ∈ Sh ⊆ H ˜ i }; zik (x) = Lk−1 hi (x); z(x) = {zi (x), i = 1 : r}, zi = {zik , k = 1 : n f 0 Lf hi (x) ≡ Lf hi (x) = ∂(hi (x)/∂x, hi (x); Lkf hi (x) ≡ Lf (Lk−1 h i (x)), f
© 2004 by CRC Press LLC
where Lf hi (x) is the Lie derivative of the function hi (x) with respect to the vector ﬁeld f(x) on Rn˜ , and (˜ n1 , · · · , n ˜ r ) is the relative degree of the nonlinear system (13) with respect to the vectorfunction h(x), which ∀x ∈ Sh is deﬁned as Lgj Lkf hi (x) = 0 ∀k < n ˜ i & ∀ i, j ∈ (1 : r). For some assumptions, including that matrix Ac (x) = {[Lgj Lnf˜ i −1 hi (x), j = 1 : r ], i = 1 : r } is nonsingular ∀x ∈ Sh , the system (13) is transformed to the canonical representation: z˙ik = zik+1 , k = 1 : (˜ ni − 1), i = 1 : r;
z˙ c ≡ {z˙in˜ i } = bc (x) + Ac (x) · u,
(14)
where the vectorfunction bc (x) = {Lnf˜ i hi (x), i = 1 : r}. Now, if the control vector u = U(x) is chosen in the form u = A−1 c (x)(−bc (x) + ν ) with the canonical control ν ∈ Rr , then the system (14) can be written in the Brunovsky canonical form with the state vector z. Futhermore, for the desired spectrum S∗c = ∪S∗ci of that closedloop system, where ∗i ∗i i i S∗ci = (λ∗i ˜i ; c1 , . . . λcni ), i = 1 : r; λcs = −αcs ± jωcs , s = 1 : p i pi + 1) : k˜i λ∗i cs = −αcs , s = (˜ i with k˜i = n ˜ i − p˜i and αcs > 0, the modal synthesis of the canonical control law ν = K · z ≡ K · Φ(x) is executed, with simultaneous construction of the VLF [27]:
υ(x) = {{υis (x), s = 1 : k˜i }, i = 1 : r}; υis (x) = wis (x); w(x) ≡ Vc−1 (S∗c )Φ(x), (15) ˜ with dimension k˜ = ki , where Vc (S∗c ) is the blockdiagonal Vandermonde matrix, which is analytically formed and inverted [27]. The linear canonical control law ν = K · z provides the exponential stability for the linear system (14) in an equilibrium position z = Φ(x) = 0 ⇐⇒ x = 0 for the nonlinear system (13) with nonlinear control law u(x) = A−1 c (x)(−bc (x)+K·Φ(x)). Moreover, the matrix P in the comparison system (12) for the VLF (15) has the elementary diagonal form. Finally, bilateral componentwise inequalities for the vectors x, z, υ(x), ρ(x), and ρ0 (x0 ) are derived. It is most desirable to obtain the explicit form for the nonlinear transformation x = Ψ(z), inverted with respect to z = Φ(x). The aggregation procedure for the VLF is carried out with an analysis of the proximity for the singular of directions in the Jacobian [∂F(x, U(x))/∂x] [27]. Since the forming control uk is discrete on a period Tu and a measurement of the model (13) coordinate’s state is discrete and incomplete, we obtained a simpliﬁed discrete object’s model (9) from the continuous model (13) by means of the TaylorLie series [29], and then we form a nonlinear discrete observer in the nonlinear control law (8) by socalled k(x)duality of the nonlinear control systems [30] and carry out its modal synthesis in canonical variables with parameterization of its desired discrete spectrum Z∗s , and also simultaneously analytically [27] construct a discrete subvector Lyapunov function υs (ˆ xek ) for the estimate x ˆek of the principal variables’ vector xek . In stage 2, with the help of a VLF with the above structure for the principal variables, the problem of nonlinear control law synthesis for the complete model of the nonlinear control system (10), taking rejected coordinates and nonlinearities, and also the restriction on
© 2004 by CRC Press LLC
Figure 1 The minimumexcessive faulttolerant scheme of MGC and envelope its AM.
control, into account is solved. That is, fulﬁlled as a multistage numerical process of parametric optimization of the nonlinear comparison system (12) with a view to selecting a spectrum S∗c (or discrete spectra Z∗c and Z∗s ) [23] as well as some other parameters for the additionally considered coordinates and nonlinearities in the complete model (10), which provide the ρρ0 exponential invariance for closedloop nonlinear control system and optimize its main quality criterion.
3
MATHEMATICAL MODELS
Any model of a moving information spacecraft takes into account: • spatial angular motion of a spacecraft bogy; • movements of ﬂexible contructional units (SAPs, RTAs, etc.) about the spacecraft bogy by single or twogimbal gearstepping drives (GSDs); • the external torque’s vector Mod as the vector sum of gravitygradient Mog , magnetic Mom (both disturbing Momd and control Momc ) torques, and also that of solar pressure forces Mos for communication spacecraft and of aerodynamic forces Moa for the remotesensing spacecraft; • operation of the MGC in the form of minimumexcessive scheme “2SPEED”type [31], see Fig. 1, where hp , p = 1 : 4, is the angular momentum (AM) unit vector; • model of each gyrodine (GD), which describes – nonlinear rotation dynamics for gyrorotor with regard to its imbalance; – ﬂexibility of gyroshell’s and gyrorotor’s preloaded ball bearings;
© 2004 by CRC Press LLC
– nonlinear dynamics of the moment gearless drive (MGD) on the GD precession axis for communication spacecraft as well as ﬂexibility, a dead band, and kinematic defects in the GSD for remotesensing spacecraft; – nonlinear dynamics of stepping motor and electromagnetic damper on the GD precession axis taking into account the dry friction torque; • models of desaturation loops of the accumulated AM at the expense of lowdraft reaction thrusters (RTs) and/or a magnetic drive (MD), and also of aerodynamic forces from the ongoing atmospheric ﬂow for the LEO remotesensing spacecraft; • models of the system’s meter elements take into account the proper dynamics of these devices, nonlinearities, digital forming of output signals, discrete noise inﬂuences, and delay with respect to the main cycle of the onboard computer operation. It is very important to provide the opportunity for reconﬁguration of the gyrocomplex structure and control algorithms for 1 to 2 abrupt possible faults in any GD of the MGC, since in this case the gyrocomplex control characteristics are essentially changed — see the AM envelope for the MGC without GD4 in Fig. 1b. In communication spacecraft 4 GDs are also used, but with MGD under bounded precession angles with respect to their neutral positions, see Fig. 2a, where γ1g = π/2 and γ1g = 0; furthermore for the main mode of the spacecraft attitude control in each MGC conﬁguration only 3 GDs are used — the fourth GD is in the “old” reserve, and all GDs’ electric components are duplicated. For the MGC’s Zarrangement on the spacecraft body, when the axis Og xc is the same as the axis Oz of the body’s reference frame (BRF), for σ = π/6 and βp ∈ [−π/2, π/2], the following 4 efﬁcient (for 3axis spacecraft attitude control) MGC conﬁgurations are possible on the basis of only 3 active GDs — conﬁgurations ZI, I = 1 : 4 — the MGC without GDI, represented at the nominal state in Fig. 2b (Z4 or Z3) and in Fig. 2c (Z2 or Z1). Therefore, the MGC scheme in Fig. 2a is really faulttolerant under diagnostics of the faulted GD and reconﬁguration of the used GDs by passages between conﬁgurations ZI with special logic conditions. The BRF’s attitude with respect to the inertial reference frame (IRF) is deﬁned by the quaternion Λ = (λ0 , λ), λ = (λ1 , λ2 , λ3 ). Assume that Λp (t) is a quaternion, and ω p (t) ≡ {ωip (t)} and ω˙ p (t) are angular rate and acceleration vectors of the programmed spacecraft body’s motion in the IRF. ˜ p (t)◦Λ, the Euler paramThe spacecraft attitude error quaternion is E ≡ (e0 , e) = Λ eters’ vector is E ≡ {e0 , e}, and the attitude error’s matrix is Ce ≡ C(E) = I3 −2[e×]Qte , where Qe ≡ Q(E) = I3 e0 + [e×] with det(Qe ) = e0 . Here In is ndimensional identity matrix, symbols ·, ·, × , { · }, [ · ] for vectors and [a×], (·)t for matrixes are conventional denotations. The communication spacecraft attitude with respect to the orbital reference frame ˜ o (t)◦Λ, where (ORF) Oxo y o z o , see Fig. 2d, is deﬁned by quaternion Eo ≡ (eo0 , eo ) = Λ Λo is a known quaternion of the ORF’s attitude with respect to the IRF, by angles of yaw ψ, roll ϕ, and pitch θ for the rotational sequence {132}, and also by the matrix of direction cosines Coe ≡ [ϕ]2 [θ]3 [ψ]1 , where [α]i is the matrix of elementary rotation, and by vector of Euler’s parameters E o = {eo0 , eo }; moreover Coe = C(E o ). For a ﬁxed position of onboard equipment’s ﬂexible structures on a spacecraft body with some simplifying assumptions, the motion model of a spacecraft ACS equipped with the MGC and the moment
© 2004 by CRC Press LLC
Figure 2 The faulttolerant scheme of MGC and the spacecraft attitude angles.
gearless drives on the GD precession axes appears as: ˙ = Λ ◦ ω / 2; Λ
¨ Ω ˙ h } = {Fω , Fq , Fβ , Fh }, ˙ q ¨ , β, Ao · {ω,
(16)
where Fω = Mg − ω×G + Mod + Qo ;
˙ ˙ = −Ah β; Mg = −H
Fq = {−aqjj ((δ q /π) Ωqj · q˙j + (Ωqj )2 · qj ) + Qqj (ω, q˙j , qj )}; Fβ = Ath ω + Mgc + Mgd + Mgb + Mgf + Qg ; Mgc = Mg + Mgd + Mga ; ˙ ω); Mh = Mh + Mha ; Fh = Mhc + Mhd + Mhf + Qh (Ω, β, β, c o H = {Hp }; H(β) = Hp (βp ); Dq Dg Dh J Dtq Aq 0 ˙ 0 G = Go + Dq q˙ + Dg β; Ao = o o Dtg 0 Ag 0 ; G = J ω + H(β); 0 Ah Dth 0 Ah (H, β) = [ ∂H(β) / ∂β ] ; • the inertia matrices Jo , Aq , Dq , Ag , Ah and Dg are constant in the BRF, whereas the matrix Dh can be found from the relation H(β) = Dh (β)Ωh ; • ω = {ωi } is a spacecraft body angular rate vector; • q = {qj } is a coordinate vector by ﬂexible oscillations of its construction;
© 2004 by CRC Press LLC
• β = {βp }, Ωh = {Ωhp } and H ≡ {Hp } = Ah Ωh are vectors of the precession angles, the gyrorotor’s angular rates and the proper AMs of the GDs in the MGC; • G is the AM vector of a spacecraft together with the MGC and the vector Go = Jo ω + H(β) is its main component; ˙ and Qqj = Qqj (ω, q˙j ) are nonlinear continuously differentiable functions, • Qo (ω, q) ˙ ω) and Qh (Ω, β, β, ˙ ω); as well as the vectorfunctions Qg (Ω, β, β, ˙ describing the inﬂuence of limiting supports on the gyrodine • the torques Mgb (β, β), precession axes, as well as the torque’s vectors of rolling friction forces in the bear˙ ω) and in the bearings on gyrodine’s ings on the gyrodine’s rotor axes Mhf (Ω, β, β, g ˙ β, ¨ ω) are nonlinear discontinuous functions; moreover precession axes Mf (Ω, β, β, ¨ the last function vs. β for the extended deﬁnition of the gyrodine’s motion under the conditions β˙p = 0 for any p = 1 : 4; • the torques of the physically implemented damping Mgdp (β˙ p ) and Mhdp (Ωp ) are nonlinear continuous functions. The components of the gyrocomplex control vectors Mgc and Mhc with regard for the possible faults in electric circuits of the MGDs as well as electromagnetic dampers on the GD’s precession axes, and also that of the electric drives on the gyrorotor axes and arresters (cages) are described by hybrid functions: Mxp =
2
γpf xl (ν) γprxl (µ) axp ixl p , x = g, gd, ga, h, ha,
(17)
l=1
where coordinates γpyxl , y = f, r are logic variables γpyxl ∈ {0, 1}; γpyx1 ∧ γpyx2 = 0; γpyx1 ∨γpyx2 = 1, p = 1 : 4; ixl p are the control currents and currents at GD’s electromagnetic arresters in main (l = 1) and in reserve (l = 2) circuits, and axp are constants. The functions γpf xl (ν) are outputs of an ALA Af with memory used for representing fault occurrences and damage development. Functions γkrxl (µ) are outputs of an SLA Ar , also with memory, for description of damage or fault blockkeeping and the reconﬁguration sequence. The rgl hl rhl currents in GD’s control circuits igl = 1 are proportional p (t) for γp = 1 and ip (t) for γp to GD’s digital control voltages: uxp (t) = Zh[Sat(Qntr(uxpk , bxu ), Bux ), Tu ], x = g, h,
(18)
where uxpk , x = g, h are the outputs of NCLs on the GD’s precession and gyrototor’s axes, and functions Sat(x, a) and Qntr(x, a) are generalusage ones, while the holder model with the period Tu is of the type: y(t) = Zh[xk , Tu ] = xk ∀t ∈ [tk , tk+1 ).
4
THE MGC MOMENTUM DISTRIBUTION
For any gyrodine at gyrocomplex in Fig. 1 equipped with the GSDs on the GD precession axes, we investigated the inﬂuence of such factors as ﬂexibility, dead band, kinematic defects in the gear, quantization of the control signal, limitation on the stepping motor
© 2004 by CRC Press LLC
current, etc. We revealed [32] amplitude of ranges for the GD’s precession rate and acceleration, in which the command and true precession rates are close in spite of the existence of large gyroscopic torques on the GD’s precession axes by virtue of the spacecraft’s rapid SRM. This means that for selected GD parameters, the assumptions for the precession theory of the CMGs are satisﬁed for indicated ranges of the input command signals, and for Hp = hg = const ∀p = 1 : 4 the vector of the MGC’s output control torque Mg can be presented by ˙ Mg (β, u) = −H(β, u) = −hg Aγ Ah (β) · u; β˙ = u; Ah (β) = [∂h(β)/∂β] , (19) where H(β) = hg hc ; hc = {xc , yc , zc } = Aγ h; h(β) = {x, y, z} = hp (βp ); g g γ = γ1 = γ2 ; x = x12 + x34 ; y = y1 + y2 ; z = −(z3 + z4 ); xp = cβp ; yp = sβp ; zp = sβp ; x12 = x1 + x2 ; x34 = x3 + x4 ; Aγ = [{1, 0, 0}, {0, sγ , −cγ }, {0, sγ , cγ }], and cα ≡ cos α; sα ≡ sin α. The exact distribution law (DL) [33] of the normed AM h(β) = A−1 γ H(β)/hg between the GD pairs: x1 − x ˜2 ) + ρ (˜ x1 x ˜2 − 1) = 0; fρ (β) = (˜ x ˜1 ≡ x12 /qy ; x ˜2 ≡ x34 /qz ; qs ≡ 4 − s2 , s = y, z,
(20)
√ with ρ = 2 6/5, which differs from that proposed by Crenshaw [31], ensures global maximum of the Grame’s determinant det(Ah Ath ) = 64/27 and the maximum module of the warranted control torque’s vector Mg in an arbitrary direction for the MGC “park” state h(β) = 0, as well as large singularities within the central part of the MGC AM’s variation domain: S ≡ {x2 + y 2 + z 2 − 2qy qz < 8; y < 2; z < 2}, see Fig. 1a, and analytically described discrete curves in the set of smoothly passed MGC’s internal singularities Qyz (β) = Qpy ∪Qpz with subsets in a peripheral part of domain S and Qps = Q∗s ∩ S∗s ; S∗s = {s = 0; s1  = s2  = 1, s = y, z}, and subsets Q∗s are described as Q∗y = {(x34 /(2ρ))2 + (z/2)2 = 1; x34 < 0}; Q∗z = {(x12 /(2ρ))2 + (y/2)2 = 1; x12 > 0}. For the DL (20) the “rightsided differential relayhysteresis” tuning [33]: D+ fρ (β) = Φρ (fρ (β), h(β));
−Sat(φρ , µρ · fρ (β)), h(β) ∈ S \ Qyz ; Φρ (·, ·) ≡ φρ Relh(as , lρ , rs ), h(β) ∈ Qps , s = y, z,
(21)
where φρ , µρ and lρ are positive constants; nonlinear function Sat(a, x) is wellknown and relay hysteresis function Relh(a, lρ , x) = (1, if x > −lρ ) ∨ (−1, if x < lρ ), with Relh(as , lρ , rs (β(t0 ))) = as ∈ {−1; 1}, s = y, z; ry = Mπ (β1 − β2 − π); rz = Mπ (β3 − β4 − π); Mπ (α) = (α, if α ≤ π) ∨ (α − 2π Sign(α), if α > π), for MGC’s AM vector ensures its inclusion in the set Qyz (β) of singular states only at ˙ separate time moments, bijectively connects the vectors Mg and β, β.
© 2004 by CRC Press LLC
5
OPTIMIZATION OF PROGRAMMED CONTROL
During the recent 25 years we have persistently researched optimizing the programmed attitude motions of information spacecraft [28, 34, 35, 36] and control algorithms for their precise stabilization based on the methods of Lyapunov functions and vector functions. For the simpliﬁed controlled object model (16), (19) within the CMG precession theory under the spacecraft model as a free rigid body, ˙ = Λ ◦ ω/2; Λ
Jo ω˙ + [ω×] Go = Mg (β, u); β˙ = u,
(22)
the computeraided algorithm for synthesis of strict optimal control u = up (t), t ∈ Tr = [ti , tf ] for the AM’s DL of the MGC during the spacecraft SRM with general boundary conditions on the initial and ﬁnal states Λs = Λ(ts ); ω s = ω(ts ), s = i, f ; β i = β(ti ) with h(β i ) ⊆ So ⊂ S in the optimization problem of the energytype index: 1 I = 2 ◦
tf
up (τ ),up (τ )dτ =⇒ min, ti
with preassigned time Tr = tf − ti for Λi , ω i ⇒Λf , ω f ; h(β(t)) ⊆ So ∀ t ∈ Tr ⊂ Tt0 , has been created. The condition that So ⊂ S and the used AM’s DL (20) are inexplicit restrictions on the programmed control up (t). The algorithm is based on numerical solving the twopoint boundaryvalue problem by Newton’s iteration method for differential equations of Pontryagin’s maximum principle. For ensuring numerical convergence for this method, the technique of descending along the parameter from the analytically resolvable starting problem, differing from that by Junkins and Turner [2], to the initially stated problem was used. We elaborated [33, 37] the fast approximate algorithms for parametric optimization of ˙ ¨ the programmed control up (t) with restrictions to ω(t), β(t), β(t) in a class of spacecraft angular motions. Therewith we used conditions of solving Darboux’s kinematic problem; the spacecraft angular motion is represented in the form of three simultaneous rotations, each executed about the axis motionless with respect to the preceding coordinate frame. By exact DL (20) of a given normed AM h(β) there is derived a main relation δ = d(1 − (1 − 2acρ − eρ2 )1/2 )/ρ, where a = x/d; b = qy qz /d2 ; c = (qy −qz )/d; d = qy +qz ; e = 4b−a2 , and due to explicit analytical formulae xs = ((2 − rs ) (2 + rs ))1/2 /rs ; rs = (a2s + b2s )1/2 , s = y, z, where ay = (x + δ)/2; by = y; az = (x − δ)/2; bz = −z, results in x1,2 (t) = (ay ∓ by xy )/2; y1,2 (t) = (by ± ay xy )/2 for the ﬁrst GD’s pair and similarly for the second pair. 6
NONLINEAR FEEDBACK SPACECRAFT CONTROL
˜ in the angular rate vector ω is deﬁned as ω ˜ = ω − Ce ω p (t), and the If the error ω g MGC’s required control torque vector M (19) for the spacecraft model (22) is formed as ˜ then the simplest normed nonlinear Mg (·) = ω×Go +Jo (Ce ω˙ p (t)−[ω×]Ce ω p (t)+ m), model of spacecraft attitude error kinematics and dynamics is as follows: ˜ / 2; e˙ 0 = − e , ω
© 2004 by CRC Press LLC
˜ / 2; e˙ = Qe ω
˜ ω ˜˙ = m.
(23)
t By analytical transformations we derived [14, 15] the following relations: Q−1 e Qe = Ce ; −1 t t −1 −1 t Qe = Qe +e · e /e0 ; Qe e = e/e0 ; I3 −e0 Qe = Qe [e×], which are used for e0 = 0. ˜ , e20 (x) = 1 − e2 ; f (x) = {Qe ω/2, ˜ 0}. Now, On stage 1, for model (23) let x = {e, ω} 1 choosing h(x) = z = −2e we can deﬁne a nonlocal nonlinear coordinate transformation and use the EFL analytical synthesis. So, we derive
˜ ˜ z˙ c = bc (x) + Ac (x)m; z˙ 1 = z2 = zc = −Qe ω; ˜ ω ˜ +ω Ac (x) = −Qe ; bc (x) = Peω (E, ω) ˜ 2 e/2, ˜ with matrix Peω = e, ω(Q e − e0 I3 )/(2e0 ). If we choose ˜ ·ω ˜ + (ω ˜ 2 /e0 ) · e/2 − Q−1 ˜ = A−1 m c (−bc + ν) = Reω (E, ω) e ν, ˜ e [e×]/(2e0 ), and ν = −a∗0 z1 − a∗1 z2 with constants a∗0 and a∗1 , where Reω = e, ωQ which are analytically calculated by Vieta’s formulas on spectrum S∗ci = −αc ± jωc , this results in the normed nonlinear control law (NCL): ˜ · e · Sgn(e0 ) − A1 (E, ω) ˜ · ω, ˜ ˜ ˜ = −A0 (E, ω) m(E, ω)
(24)
A0 (·) = ((2a∗0 − ω ˜ 2 /2)/e0 )I3 ; A1 (·) = a∗1 I3 − Reω ; Sgn(e0 ) = (1, if e0 ≥ 0)∨(−1, if e0 < 0). Moreover, the closedloop nonlinear system (23), (24), is transformed to the simplest linear system for the spacecraft spatial attitude control. Simultaneously constructing the VLF’s components using the Vandermonde matrix is analytically executed by relation (15), 3 ˜ has become known; in this case υ(x) ∈ R+ , and for the VLF’s structure υ(x) = υ(E, ω) obtaining the inverse nonlinear transformation x = Ψ(z) in the explicit form it is sufﬁcient to use only the analytic inverse of matrix Qe [14]: o o o o q03 + q12 −q02 + q13 po11 −1 o o 0 o , −q03 + q12 po22 q01 + q23 Q−1 e = e0 · o o o o 0 p33 q02 + q13 −q01 + q23 o o o − q33 (123) with the standard chain (123); qij = ei · ej , i, j ∈ {0 : 3}. where po11 = 1 − q22 Without restrictions on the moment gyrocomplex control u ∈ R4 , the NCL u(t, Λ, ω, β) is evaluated analytically using the vector relations g {Ah (β), [∂fρ (β)/∂β]} · u(·) = {vu (·), Φρ (fρ (β), h(β))}; vu (·) = −A−1 γ M (·)/hg .
This results in the solution u(·) = sop (β)−1 qg (·), where sin(β1 − β2 ) ≡ so1,2 = y1 x2 − y2 x1 ; sin(β3 − β4 ) ≡ so3,4 = z3 x4 − y4 x3 , vectors
matrix
˜ f (β) · Φρ (fρ (β), h(β)); ˜ · vu + d qg (·) = D(β) ˜ f (β) = (qy qz /dyz ) · {−x2 , x1 , x4 , −x3 }, d
−x2 pyx x1 pyx ˜ D(β) = −x4 pzx x3 pzx
© 2004 by CRC Press LLC
−y2 + x2 ryx y1 − x1 ryx −x4 ryx x3 ryx
−x2 rzx x1 rzx , z4 + x4 rzx −z3 − x3 rzx
(25)
and also dyx = qy − ρ x12 ; dzx = qz + ρ x34 ; dyz = dyx + dzx ; ryx = x12 y pzx /qy2 ; rzx = x34 z pyx /qz2 ; ps = ds /dyz , s = yx, zx; ˜ f (β) ≡ 0 is valid for any vector β. ˜ ·d moreover the identity Ah (β)D(β) Taking into account restrictions with respect to the gyrocomplex control in the form up (t) ≤ u ∀ p ∈ {1 : 4} and ∀ t ∈ Tt0 for a given constant u ∈ R+ , special nonlinear scalar and vector functions of the type “division of variables with scaling” are introduced [15, 33]. The scalar function is deﬁned as follows: x = 0; 0 s (26) d (x, s, a, m) = mx/s (x = 0) & (mx < as); a m Sign(xs) mx > as, for the scalar arguments x, s ∈ R; a, m ∈ R+ . The vector function y = Mds4 (x, s, u) has the vector arguments s, x ∈ R4 and the internal working constant u∗ u. Its components yp , p = 1 : 4, are deﬁned by the algorithm using the scalar function (26): for p = 1 : 4 do for p = 1 : 4 do for p = 1 : 4 do
rp := ds (xp , sp , u∗ , 1); r := max{rp , p = 1 : 4}; if r > 0 then mp := rp /r else mp := 1; yp := ds (xp , sp , u, mp ).
(27)
As a result, for the vector so (·) = {sop (·)} the discrete NCL obtained and applied has the form ugk ≡ {ugpk } = ug (·) = Mds4 (qg (·), so (·), u), (28) using the algorithm (27) and the vector function qg (·) (25) together with the discrete form of tuning (21) for the distribution law (20). On stage 2, the problems of this nonlinear control law synthesis are solved for the gyromoment ACS model of the LEO remotesensing spacecraft (16), (19), (24) with the GD’s GSDs [38]. Moreover, the selection of only 3 parameters αc , ωc , µρ in the nonlinear control law structure (28) and gain kdg in GDs’ electomagnetic damper, which optimize the main quality criterion and provide the ρρ0 exponential invariance for given restrictions, is fulﬁlled on the basis of multistage numerical analysis and optimization of nonlinear CS for ˜ and other VLF having derived the above structure for the principle error coordinates E, ω ˙ VLF components’ structure in the form of sublinear norms for vector variables β(t) and ˙ q(t), q(t), vs. the vector β(t).
7
NONLINEAR CONTROL OF COMMUNICATION SPACECRAFT
An arbitrary behavior of gyromoment ACS under requirements (i) and (ii) for communication spacecraft is essentially nonlinear under the gyrodine faults. The nonlinear dynamic problems of the faulttolerant ACS are quite real for the nonstoping communication sessions and spacecraft spatial angular movements:
© 2004 by CRC Press LLC
• the spacecraft body orbital stabilization, when the given state in the ORF is E oc = {1, 0}, i.e., the attitude angles ψ = ϕ = θ = 0, with reconﬁguration of moment gyrocomplex, the indirect SAPs orienting and tracing the Sun by the gearstepping drives digital control at the angles’ vector γ p = {γ1p , γ2p } on twogimbal axes for aligning the normalunit epn to the panels’ surface with respect to the Sun direction unit esn , for all that the matrixes Jo (γ p (t)), Dq (γ p (t)) and the vector ¨ p , ω) is added to the vector Fω in (16); Mop (γ p , γ˙ p , γ • the precision spacecraft stabilization in the SunEarth reference frame at ϕ = θ = 0 and ψ = ψ0 (t), with reconﬁguration of the MGC, the indirect SAPs guidance into the Sun by means of both the simultaneous attitude control of the spacecraft body with respect to the yaw angle ψ and the digital control of the singlegimbal gearstepping drive on angle γ p for the panels, see Fig. 2d; • the slow angular reorientation of the spacecraft body into a preassigned state E oc (t) in the ORF or in the SunEarth reference frame both before and after the correction of the spacecraft orbit and/or in the passage of some singular kinematic conditions, with reconﬁguration of the MGC and the indirect SAPs guidance; • the precision spacecraft angular stabilization of the given commanded state E oc (t) under correction of the spacecraft orbit and movements of the correcting RTs by digital control of the GSDs on the thruster 2gimbal axes, and also with reconﬁguration of the MGC and the indirect SAPs guidance. By VLFmethod we have also carried out research on nonlinear dynamics for faulttolerant gyromoment ACSs of communication spacecraft under their substantial parametric uncertainty, incomplete discrete measurement of the state and resource restrictions on digital control for smallmass gyrocomplexes of various types, including the mimimumexcessive scheme in Fig. 2a. For this case the kinematic relations in (16) have the form ˜ ˜ e˙ o0 = − eo , ω/2; e˙ o = Qoe ω/2,
(29)
˜ = ω − Coe ω o ; ω o = {0, 0, ωo } is the orbital rate vector, deﬁned where Qoe ≡ Q(E o ); ω in the ORF, see Fig. 2d, the matrix σ σ −Ch4 0 0 −Ch3 ∂H(β) σ σ 0 Ch1 Ch2 = 0 Ah ≡ Ah (H, β) = , ∂β σ σ σ σ −Sh1 −Sh2 Sh3 Sh4 σ σ σ σ σ σ = Xh1 + Xh2 ; X34 = Xh3 + Xh4 , X = S, C, and where X12 σ σ σ σ −C34 ; Shp ≡ Hp sinβpσ ; Chp ≡ Hp cosβpσ ; βpσ = βp +(−1)p−1 σ, /p = 1 : 4, δCgσ ≡ C12
and summary angular momentum of MGC is σ σ , S12 , δCgσ }; H(β) = Hp (βp ) = {−S34
Hp (βp ) = Hp · hp (βp ).
Due to accessible information: • the command order vector of Euler’s parameters E oc (t),
© 2004 by CRC Press LLC
• the measured attitude angles ψ(t), ϕ(t), θ(t) with the sampling period Tq and a vector E of k of Euler’s parameters, which is ﬁltered out at the time t = tk , • the measured and ﬁlteredout gyrodine vectors β fk of their precession angles and Ωhf k of the gyrorotor’s angular rates, the vector δHk of error on GDs angular momentums and the Euler’s vector E εk = {ε0k , εk } of the spacecraft attitude error are reprsented as
ε0k
f Hfk = Ah Ωhf k ; δHk = hg · I4 − Hk ; c f c f c t f = e0k e0k + ek , ek ; εk = Q (E k ) ek − ef0k eck ,
where hg = const is a nominal value of angular momentum for each gyrodine. As a result of the nonlinear control law synthesis for such ACS, the gyrocomplex control vectors ugk and uhk are computed onboard by the relations ugk
g k = −2εk ; wkg = wk−1 + kg (k + ag k−1 ); g f t f ef = −Ah (Hk , β k ) (wk + ω ok ); uhk = −kh · δHk , k ∈ N0 ,
(30)
f where ω ef ok = C(E k )ω o , and kh , kg , ag are the constant scalar parameters. The synthesized digital nonlinear control law (30) is universal for the given gyrocomplex type; it provides the precision 3axis attitude control of the spacecraft body in the process of the gyrocomplex reconﬁguration and is robust with respect to the accumulated angular momentum. Furthermore, the digital information on only the attitude of the spacecraft body and the gyrodine positions are used for forming the GDs’ control. Moreover, the ﬂexible oscillations of spacecraft structures (the SAPs, the highgain RTAs, etc.) are damped perfectly well. Operability of this digital nonlinear control law under the large torque perturbations arising in the process of the spacecraft orbit correction, as well as under slow largeangle spatial reorientations of spacecraft body has been investigated. Digital nonlinear control laws for 2gimbal gearstepping drivers of solar array panels have been developed with the aid of the Lyapunov functions method. These control laws provide asymptotic stability of indirect guidance for the unit epn to the Sun’s direction unit es under arbitrary, but measured, angular motion of the spacecraft body. Moreover, the Lyapunov function, v = 1 − es , epn ≥ 0, is a natural measure of closeness for the vectors es and epn with v = 0 iff es = epn . The nonlinear dynamics of both such indirect SAPs orienting and simultaneous spacecraft attitude control in the SunEarth reference frame with respect to yawing and by digital control of the SAPs with 1gimbal gearstepping driver, as well as tracing the Sun, has been investigated, and problems of passage through singular states in the process of panel and spacecraft attitude control have also been considered. Digital control algorithms for the gearstepping driver precision guidance of highgain antennas have also been synthesized.
8
PROVISION OF FAULT TOLERANCE
Onboard ACSs of the communication spacecraft use the twolevel logicdigital system of the fault diagnosis:
© 2004 by CRC Press LLC
• on the lower level — an integral local SLA Add with memory for automatic discrete monitoring of the status of all relevant devices with a sampling period Tq for measurement of available physical variables (currents, movements, rates, etc.) • on the higher “system” level — an SLA Ad , also with memory, bound up with the sampling period Tu for the functional diagnostics of the main control loop by comparison of outputs for normal and emergency models of the ACS operation. The high failsafe operation of gyromoment ACS has been achieved with the use of the onboard digital algorithms, in accordance with the SLAs Add , Ad , and Ar , allowing application of all the reverse complete sets of onboard devices or their electric circuits [39]. The verbal description of the provision of fault tolerance of a spacecraft ACS with a gyrocomplex in Fig. 2a, for its initial conﬁguration Z4, when Hp = hg , p = 1 : 3 and γpf xl = γprx1 = 1, x = g, gd, h; γprx1 = 0, x = ga, ha, and GD4 in the stopping state H4 = β4 = 0; γ4rxl = 0, x = g, gd, h; γ4rxl = 1, x = ga, ha, see (17), is as follows. In the normal mode, the magnetic desaturation loop ensures the condition Go ≈ 0 under formation of the magnetic control torque vector: Momc = Lm (t)×B⊕ ; Lm (t) = Zh[Lmk , Tu ], where B⊕ is a magnetic displacement vector of geomagnetic ﬁeld and o o Lmk = −lm φm (R0 , λm , bm , Rk ) emk ; emk = ck /ck ; f ck = Rk ×Bf⊕k ; Rk = Jo (γ pk )ω ef ok + H(β k ),
where a scalar relayhysteresis function: φom (a, λm , bm , x) ≡ {(1 ∀x > λm bm ) ∨ (0 ∀x < bm )}; moreover φom (a, λm , bm , a) = a, a ∈ {0, 1}, with threshold of operation bm and coefﬁcient o of return 0 < λm < 1, and lm is the modulus of the magnetic driver dipole torque. Let the fault of the torque gearless driver current circuit in GD3 occur at any time moment t = tfν ∈ [tk∗ −1 , tk∗ ); ν = 1, γ3f g1 (1) = 0. Then by SLAs AdGD−3 or Ad , and by SLA Ar as the result of circuits switching (γ3rg1 = 0; γ3rg2 = 1) is guaranteed for the discrete time k = k ∗ = k∗ or k = k ∗ = k∗ + 1, respectively. Moreover, the intensity of dynamic processes for the attitude control channels is essentially dependent not only on the time interval duration δtfk∗ = tk∗ − tfν , when there is no control, but also on the potentialities of the gyrodines, which remained operable in the aspect of compensation of disturbing inﬂuence of the angular rate vector ω o because of the spacecraft orbital motion. After such an isolation of the fault, the scheduled reconﬁguration of Z4 ⇒ Z3 process starts: • γ4rha1 = 0 and γ4rh1 = 1 with speeding up from the rest state of GD4 rotor, • the magnetic driver desaturation loop operates with Rk = H(β fk ),
© 2004 by CRC Press LLC
• at achieving of a small neighborhood for the gyrocomplex ”park” state Hp = hg ; βp = 0, p = 1 : 4, there takes place simultaneously: the GD3 caging (γ3rga1 = 1), and GD4 uncaging (γ4rga1 = 0) and switching (γ4rg1 = 1) online the control closedloop. In the ﬁnal stage of this process: the magnetic desaturation loop is returned into the nominal mode; the SLA’s Ar output γ3rh1 = 0, and the GD3 rotor is speeding down to the rest state; ﬁnally, after reaching the condition H3 ≈ 0, the GD3 is caged (γ3rha1 = 1). Thus, the moment gyrocomplex restores its excessiveness with respect to control circuits of torque gearless drivers for the online gyrodines, and it is prepared for the rapid isolation of any new gyrodine fault and for new reconﬁguration.
9
SOFTWARE AND RESULTS
The developed methods for modeling and dynamic research of the faulttolerant ACS have been implemented in the form of the software system DYNAMICS [40, 41, 42]. The general technology for usage of this software system consists of obtaining models of separate components of ACS, their automatic transforming with automatic assembling into an integrated model, and then a subsequent dialogue investigation based on both known and unique methods [43, 44, 45, 46] implemented in the software system. As discussed above, the intensity of dynamic processes is essentially dependent on the potentialities of the gyrodines that remained operable in the aspect of compensation of the spacecraft orbital motion. This fact is illustrated in Fig. 3, where the sampling period values Tq = 0.25 s and Tu = 4 s. Such processes are presented with respect to the pitch channel of ACS under the orbital stabilization for conﬁgurations Y4, Fig. 3a and Z4, Fig. 3b, when the GD3 fault takes place under t = 300.1 s. In the Y4 case, there are no operable gyrodines needed to create control torques along the axis Oz, so despite the ”fast” fault diagnostics by the SLA AdGD−3 and switching the torque gearless driver’s reserve circuit into the control closedloop online at time t = tk∗ = 304 s, there take place substantial overshoots of attitude errors. Such overshoots are absent for a similar fault in GD3 within the gyrocomplex according to the conﬁguration Z4 see Fig. 3a, since GD1 and GD2 in this case remain operable to create control torques along the axis Oz, see Fig. 3b. So, even for the “slow” GD3 fault diagnosis with the aid of SLA Ad and switching the torque gearless driver’s reserve circuit in GD3 by time t = tk∗ = 308 s, the precision angular stabilization with respect to pitch remains the same.
10
CONCLUSIONS
With the aid of the logicdynamical models, nonlinear methods, and software developed, we have conducted multilateral dynamic investigations of robust faulttolerant gyromoment ACSs on an order from the State Research and Production Association of the Russian Space Agency — TsSKBProgress (Samara) [47, 48] and NPO PM (Krasnoyarsk) [49, 50], for spacecraft both manufactured and presently under design, including those being developed for international projects: the maneuvering vehicle Ikar [51] in cooperation with
© 2004 by CRC Press LLC
Figure 3 The processes for fault in the MGD’s control current circuit of GD3.
Aerospatiale (France) and communication satellite Sesat [52, 53, 54] in cooperation with Alcatel Space Industries (France) on the Eutelsat order. Flight exploitation of designed spacecraft ACSs have proved the high efﬁciency of principles and methods employed for providing the faulttolerant operation.
© 2004 by CRC Press LLC
Acknowledgments This work was supported by the Russian Space Agency and by the Russian Foundation for Basic Research (RFBR), grant 0100293.
REFERENCES 1. B.V. Raushenbakh and Ye.N. Tokar, Spacecraft Attitude Control. Nauka, Moscow, 1974. 2. J.L. Junkins and J.D. Turner, Optimal Spacecraft Rotational Maneuvers. Elsevier, Oxford, 1986. 3. B.R. Hoelscher and S.R. Vadali, “Optimal openloop and feedback control using single gimbal control moment gyroscopes,” Journal of the Astronautical Sciences 42(2):189– 206, 1994. 4. A. Isidori, Nonlinear Control Systems, SpringerVerlag, New York, 2nd edition, 1989. 5. S.N. Singh and T. Bossart. “Exact feedback linearization and control of space station using a control moment gyros,” IEEE Trans. on Automatic Control AC38(1):184–187, 1993. 6. P.M. Frank, “Fault diagnosis in dynamic systems using analytical and knowledgebased redundancy — a survey and some new results,” Automatica 26(3):459–474, 1990. 7. P.M. Frank, “Enhancement of robustness in observerbased fault detection,” International Journal of Control 59(5):955–981, 1994. 8. V.M. Matrosov, “Vector Lyapunov functions in the analysis of nonlinear interconnected systems.” In Symposia Mathematica, vol. 6, pp. 209–242, Academic Press, London, New York, 1971. 9. V.M. Matrosov, “Comparison method in system’s dynamics.” In Equation Differentielles et Functionnelles Non Lineaires, pp. 407–445, Herman, Paris, 1973. 10. N. Rouche, P. Habets, and M. Lalya, Stability Theory by Lyapunov’s Direct Method. SpringerVerlag, Heidelberg, 1977. 11. V.M. Matrosov, “Method of vector Lyapunov functions in nonlinear mechanics,” Advances in Mechanics 6(3):59–82, 1989. 12. V. Lakshmikamtham, V.M. Matrosov, and S. Sivasundaram, Vector Lyapunov Functions and Stability Analysis for Nonlinear Systems, Kluwer Academic Publishers, Dordrecht, Boston, London, 1991. 13. Ye.I. Somov, “On synthesis of nonlinear control systems using vector Lyapunov functions.” In Differential Equations and Numerical Methods, pp. 72–89, Nauka, Novosibirsk, 1986. 14. Ye.I. Somov, “Feedback linearization and VLF techniques on the synthesis of spacecraft gyromoment attitude control systems.” In Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, vol. 4, pp. 2522–2527, Beijing, 1996. 15. Ye.I. Somov, “Nonlinear spacecraft gyromoment attitude control.” In Proceedings of 1st International Conference on “Nonlinear Problems in Aviation and Aerospace,” pp. 625–630, Daytona Beach, FL, 1997. ERAU. 16. Ye.I. Somov, “Precision nonlinear gyromoment attitude control of the remote sensing spacecraft,” Actual Problems of Aviation and Aerospace Systems 2(1):44–53, 1997.
© 2004 by CRC Press LLC
17. V.M. Matrosov, S.N. Vassiljev, R.I. Kozlov, et al., Algorithms for Theorem Derivation of Vector Lyapunov Functions Method, Nauka, Novosibirsk, 1981. 18. Lj.T. Gruji´c, “Coordinated Lyapunov’s methodology for stationary nonlinear systems,” Automation and Remote Control (12):35–73, 1997. 19. A.S. Zemlyakov, Ye.I. Somov, and R.A. Sabitov, “Computeraided synthesis of the hierarchical motion control systems on the basis of vector Lyapunov functions.” In Stability Theory and Its Applications pp. 264–274, Nauka, Novosibirsk, 1979. 20. Ye.I. Somov, “On the synthesis of spacecraft attitude control systems.” In Computing and Controlling Systems of Flying Vehicles, pp. 13–22, KAI, Kazan, 1981. 21. Ye.I. Somov, “Synthesis of invariant systems for spacecraft spatial gyromoment stabilization using the vector Lyapunov functions method.” In Invariance Theory, Sensitivity Theory and their Applications, pp. 184–185, ICS, the USSR Academy of Sciences, Moscow, 1982. 22. Ye.I. Somov, “Synthesis of nonlinear control law for spacecraft spatial rotational maneuver.” In Optimizing Processes in Aviation Engineering, pp. 66–70, KAI, Kazan, 1982. 23. V.M. Matrosov, R.I. Kozlov, Ye.I. Somov, et al., “Methods and software for computeraided design of spacecraft attitude control systems.” In Spacecraft Dymamics and Control, pp. 163–179, Nauka, Novosibirsk, 1992. 24. G.P. Anshakov, V.M. Matrosov, Ye.I. Somov, et al., “Gyromoment attitude control systems dynamics of rapid manoeuvring remote sensing spacecraft.” In Proceedings of 1st International Aerospace Congress, vol. 2, pp. 129–132, Petrovka, Moscow, 1995. 25. Ye.I. Somov, “Nonlinear methods for synthesis of spacecraft failsafe gyromoment attitude control systems.” In Proceedings of the 3rd International Workshop “New Computer Technologies in Control Systems,” pp. 68–70, IPS of RAS, PereslavlZalessky, 1996. 26. V.M. Matrosov, M.F. Reshetnev, V.A. Rayevsky, and Ye.I. Somov, “Nonlinear methods for dynamic synthesis of faulttolerant spacecraft attitude control systems,” Journal of Computer and Systems Sciences International (6):120–130, 1997. 27. Ye.I. Somov, “Constructing the vector Lyapunov functions at synthesis of linear control systems with a partial measurement of state,” Journal of Computer and Systems Sciences International (3):73–86, 1997. 28. Ye.I. Somov, “Nonlinear dynamics and optimization of the spacecraft precision gyromoment attitude control systems.” In Proceedings of 2nd IMACS/IEEE Conference “Computational Engineering in Systems Application,” vol. 2, pp. 72–77, Ecole Centrale, Lille, 1998. 29. N. Kazantzis and C. Kravaris, “Systemtheoretic properties of sampleddata repersentations of nonlinear systems obtained via TaylorLie series,” Intern. Journal of Control 67(6):997–1020, 1997. 30. A.P. Krishenko and S.B. Tkachev, “Nonlinear k(x)dual systems,” Automation and Remote Control (2):21–34, 1995. 31. J.W. Crenshaw, “2SPEED, a singlegimbal control moment gyro attitude control systems,” AIAA Paper (73895):1–10, 1973. 32. Ye.I. Somov, A.V. Sorokin, and O.A. Kondrat’yev, “Precession models of control moment gyroscopes on optimization of spacecraft singularly perturbed control systems.” In Proceedings of 3rd International Workshop, “Singular Solutions and Perturbations in Control Systems,” pp. 99–102, IPS of RAS, PereslavlZalessky, 1997.
© 2004 by CRC Press LLC
33. Ye.I. Somov, Ye.A. Bondarenko, and N.B. Kapitonova, “Synthesis of a gyromoment spatial stabilization system on the basis of vector Lyapunov functions and parametric optimization.” In Problems of Analytic Mechanics, Stability and Control, pp. 257–264, Nauka, Novosibirsk, 1991. 34. Ye.I. Somov and E.F. Fatkhullin, “Optimal control of a gyrostat programmed motion.” In Memoirs of Kazan Aviation Institute, vol. 180, pp. 37–40, KAI, Kazan, 1975. 35. Ye.I. Somov, “Optimization of extensive control for spacecraft reorientation by control moment gyroscopes.” In Optimizing Processes in Aviation Engineering, pp. 78–82, KAI, Kazan, 1981. 36. Ye.I. Somov, “Stabilization of the extensive reorientation optimal behavior for a spacecraft with control moment gyroscopes.” In Stability and Control, pp. 55–59, KAI, Kazan, 1983. 37. Ye.I. Somov and I.A. Gerasin, “The rotational maneuver realizability estimation for spacecraft controlled by an excessive gyrodine system.” In Proceedings of the Tsiolkovsky Cosmonautics Academy, pp. 138–143, Russian Academy of Cosmonautics, Samara, 1998. 38. Ye.I. Somov, V.M. Matrosov, G.P. Anshakov, et al., “Ultraprecision attitude control of the loworbital remote sensing spacecraft.” In Proceedings of the 14th IFAC Symposium on Automatic Control in Aerospace, pp. 45–50, Elsevier Science, Oxford, 1999. 39. V.A. Rayevsky, V.V. Monakhov, and A.F. Glazunov, “Digital orientation and stabilizaton system for geostationary satelliteretransmitter Lutch.” In Spacecraft Dynamics and Control, pp. 40–46, Nauka, Novosibirsk, 1992. 40. Ye.I. Somov, S.A. Butyrin, et al., “The software system dynamics for computeraided design of spacecraft attitude control systems.” In Proceedings of the AllRussian Scientiﬁc School “Computer Logic, Algebra and Intelligent Control,” vol. 3, pp 401– 431, ICC, the RAS, Irkutsk, 1994. 41. Ye.I. Somov, S.A. Butyrin, et al., “Software tool DYNAMICS in simulation of faulttolerant spacecraft attitude control systems,” Gyroscopy and Navigation 2(25):92–107, 1999. 42. Ye.I. Somov, S.A. Butyrin, V.A. Rayevsky, et al., “Computer aided design and ﬂight support of spacecraft control systems.” In Proceedings of the 14th IFAC Congress, vol. L, pp. 445–450, Elsevier Science, Oxford, 1999. 43. Ye.I. Somov and Ye.A. Bondarenko, “Stability analysis of multiply discrete stabilization system of the ﬂexible ﬂying vehicle,” Soviet Aeronautics (4):26–30, 1988. 44. Ye.I. Somov, “Dynamics of the multiply digital system for spatial gyromoment stabilization of ﬂexible spacecraft.” In Spacecraft Dymamics and Control, pp. 46–76, Nauka, Novosibirsk, 1992. 45. Ye.I. Somov and I.A. Gerasin, “Stability of linear continuousdiscrete multiply control systems with a delay.” In Proceedings of the Tsiolkovsky Cosmonautics Academy, pp. 133–137, Russian Academy of Cosmonautics, Samara, 1998. 46. Ye.I. Somov, “Robust stabilization of a ﬂexible spacecraft at partial discrete measurement and a delay in forming control,” Journal of Computer and Systems Sciences International 40(2):287–307, 2001. 47. D.I. Kozlov, G.P. Anshakov, Yu.G. Antonov, Ye.I. Somov, et al., “Precision control systems of motion on Russian spacecraft for ecological remote sensing.” In Proceedings of the 14th IFAC Symposium on Automatic Control in Aerospace, pp. 27–38, Elsevier Science, Oxford, 1999.
© 2004 by CRC Press LLC
48. Ye.I. Somov, “Nonlinear synthesis, optimization and design of the spacecraft gyromoment attitude control systems.” In Proceedings of 11th IFAC Workshop “Control Applications of Optimization,” vol. 1, pp. 327–332, Elsevier Science, Oxford, 2000. 49. V.A. Bartenev, V.A. Rayevsky, Ye.I. Somov, et al., “Orbit and attitude control systems of Russian communication, geogetic and navigation spacecraft.” In Proceedings of the 14th IFAC Symposium on Automatic Control in Aerospace, pp. 15–26, Elsevier Science, Oxford, 1999. 50. Ye.I. Somov, V.M. Matrosov, V.A. Rayevsky, et al., “Faulttolerant attitude control systems of the communication and navigation spacecraft.” In Proceedings of the 14th IFAC Symposium on Automatic Control in Aerospace, pp. 128–133, Elsevier Science, Oxford, 1999. 51. Ye.I. Somov, S.A. Butyrin, G.P. Anshakov, Yu.G. Antonov, et al., “Dynamics of the maneuvering vehicle Ikar control system by orbital placement of Globalstar satellites.” In Preprints of 15th IFAC Symposium on Automatic Control in Aerospace, pp. 372– 377, University of Bologna, Bologna, 2001. 52. V.A. Bartenev, G.P. Titov, V.A. Rayevsky, Ye.I. Somov, et al., “The attitudeorbit control subsystem and the antenna pointing perfermance of the Sesat spacecraft.” In Proceedings of 4th ESA International Conference on Spacecraft Guidance, Navigation and Control Systems, pp. 35–48, ESTEC, Noordwjik, the Netherlands, 2000. 53. G.P. Titov, Ye.I. Somov, V.A. Rayevsky, et al., “Nonlinear dynamics of gyromoment attitude control and precise antenna pointing by the Sesat spacecraft.” In Preprints of 15th IFAC Symposium on Automatic Control in Aerospace, pp. 73–78, University of Bologna, Bologna, 2001. 54. Ye.I. Somov, S.A. Butyrin, V.A. Rayevsky, G.P. Titov, et al., “Nonlinear dynamics of gyromoment attitude control system at communication satellite Sesat.” In Preprints of 5th IFAC Symposium “Nonlinear Control Systems,” vol. 5, pp. 1481–1486, IPME of RAS, St. Petersburg, 2001.
© 2004 by CRC Press LLC
EBook Information

Series: Nonlinear systems in aviation, aerospace, aeronautics, and astronautics 2

Year: 2,004

Edition: 1

City: Boca Raton, FL

Pages: 328

Pages In File: 328

Language: English

Topic: 245

Identifier: 9780415308526,0415308526

Org File Size: 3,665,012

Extension: pdf

Tags:
Транспорт
Аэрокосмическая техника

Toc: Advances In Dynamics And Control......Page 1
Contributors......Page 5
Preface......Page 7
Contents......Page 8
Introduction......Page 10
Table Of Contents......Page 0
The Model......Page 11
State Feedback Stabilization......Page 13
Stabilization Without Rate Feedback......Page 15
References......Page 21
Introduction......Page 23
System Model......Page 24
Adaptive Learning Controller......Page 28
Stability Analysis......Page 29
Simulation Results......Page 31
Conclusion......Page 32
Acknowledgments......Page 33
References......Page 35
Introduction......Page 36
Statement Of The Problem......Page 38
Operator Reformulation Of The Problem......Page 40
Asymptotic And Spectral Properties Of The Matrix Differential Operator......Page 42
Asymptotical Distribution Of Aeroelastic Modes......Page 44
Structure And Properties Of Matrix Integral Operator......Page 45
Properties Of The Adjoint Operators And Operatorvalued Functions......Page 47
References......Page 48
Introduction......Page 51
Equations Of Motion......Page 52
Continuation Method......Page 55
Numerical Results......Page 56
Recovery From A High Rollrate Motion......Page 59
Conclusion......Page 60
References......Page 61
Introduction......Page 62
Autopilot Model And Performance Objective......Page 63
Fuzzy Dynamic Gainscheduled Autopilot......Page 66
Results And Analysis......Page 72
Conclusions......Page 74
References......Page 75
Introduction......Page 77
Model Predictive Control Of Nonlinear Systems......Page 78
Control Problem......Page 79
Linear Parametervarying Approximation......Page 80
Mpc Implementation......Page 81
Application To Nonlinear Vibration Control......Page 83
Mpc Architectures For Fcs......Page 85
Maneuvers For Evaluation......Page 86
Piloted Simulations......Page 87
Conclusions......Page 90
References......Page 91
Introduction......Page 92
Semide.nite And De.nite Timeindependent Vector Functions......Page 93
Semide.nite And De.nite Timedependent Vector Functions......Page 94
Kinds Of Vector Lyapunov Functions......Page 96
System Description......Page 97
Stabilizing Control Problem Statement......Page 98
Stabilizing Control Problem Solution......Page 99
Tracking And Stabilizing Control Problem Statement......Page 100
Output Tracking And Stabilizing Control Problem Solution......Page 101
Conclusion......Page 103
References......Page 104
Introduction......Page 106
Some Basic Properties Of Unstable Linear Systems With Control Constraints......Page 107
Maximum Stabilizer Design For One Unstable Eigenvalue......Page 110
Controllability Region Computation......Page 113
Stability Suboptimal Linear Controller......Page 114
References......Page 118
Introduction......Page 120
Quest Algorithms For Gpsbased Attitude Determination......Page 122
Attitude Propagation Equations Of Gyrostat Satellites......Page 125
Stability Analysis Of A Kelvin Gyrostat Satellite......Page 128
Suppression Of Attitude Tumbling Of The Gyrostat Satellite......Page 129
Simulation Procedures And Results......Page 130
References......Page 137
Introduction......Page 140
Preliminaries......Page 141
Main Results......Page 144
References......Page 156
Description Of The Problem......Page 157
The Pde Model......Page 158
The Main Results......Page 159
About The Problem And The Literature......Page 161
The Necessary Inequality......Page 163
Step 1 (the Conservation Relation)......Page 164
Step 2 (the Acoustic Wave Estimates)......Page 165
Step 4 (final Statement Of Derived Estimate)......Page 170
Supporting Results......Page 171
The Main Estimate For......Page 172
Completion Of The Proof Of Theorem 1(a)......Page 173
References......Page 174
Introduction......Page 176
The Continuum Mechanical Model......Page 178
Temperature And Dopant Concentration Evolution In The Melt In Ampoules With A Smalldiameter Crosssection In A Space Experiment In Strictly Zero Gravity......Page 183
Modelbased Simulation Of The Flow Evolution, Heat Transport, And Ga Dispersion In The Melt In A Lowgravity Experiment......Page 184
Modelbased Prediction Of The Axial And Radial Segregation In Lowgravity Experiments......Page 185
References......Page 188
Introduction......Page 189
System Model......Page 190
A. Numerical Example......Page 193
B. Numerical Example (continued)......Page 197
C. Numerical Example Of Damage Detection......Page 200
Conclusions And Discussion......Page 201
References......Page 204
Introduction......Page 205
Optimization......Page 207
Building And Parameterizing A Shell Model......Page 210
Design Of Carillon Bells......Page 211
Design Of Loudspeaker Diaphragm......Page 212
Design Of Sedan Body Panels......Page 213
References......Page 215
Introduction......Page 220
Fuzzy Modeling Of Dynamical Systems......Page 221
Neural Networks For Control......Page 223
Adaptive Control Of Aircraft Systems......Page 224
Experimental Results......Page 225
References......Page 227
Introduction......Page 229
General Analysis Of 3d Relative Motion......Page 231
3d Ipn With Nonmaneuvering Targets......Page 233
Numerical Results......Page 240
References......Page 242
Introduction......Page 245
Dissipation And The Nonlinear H2/ H8 Control Problem......Page 246
Formulation Of H2/ H8 Guidance Problems......Page 250
Solutions Of The Hamiltonjacobi Pdi......Page 251
Numerical Results......Page 254
Conclusions......Page 255
References......Page 257
The Problem Motivation......Page 259
Autonomous Technical Controllability Of An Object......Page 261
Working Domain Of The Object’s Motion......Page 262
The Solving Of The Formulated Problem......Page 263
Conclusion......Page 266
References......Page 267
Mathematical Setting And Formulation Of The Control Problem......Page 269
Existence Of A Unique Optimal Pair, Characterization, And Regularity Properties: Proof Of (a1)......Page 274
The Evolution Operator .........Page 276
Singular Estimate For .........Page 279
The Operator .........Page 282
The Function Rw And Its Differential Equation. Proof Of (a9)......Page 283
Two Classes Of Structural Acoustic Problems......Page 287
A Class Of Composite Beams Models......Page 293
Composite Beams Models Under Homogeneous Hinged Mechanical B.c.......Page 296
A Structural Acoustic 2d Model With A Composite (sandwich) Beam As .exible......Page 299
Appendix: Two Results On Singular Estimates......Page 300
References......Page 304
Introduction......Page 306
Nonlinear Faulttolerant Feedback Control......Page 307
Mathematical Models......Page 312
The Mgc Momentum Distribution......Page 315
Nonlinear Feedback Spacecraft Control......Page 317
Nonlinear Control Of Communication Spacecraft......Page 319
Provision Of Fault Tolerance......Page 321
Conclusions......Page 323
References......Page 325