CONTROL OF CHAOS IN NONLINEAR CIRCUITS AND SYSTEMS

WORLD SCIENTIFIC SERIES ON NONLINEAR SCIENCE Editor: Leon O. Chua University of California, Berkeley
Series A.
Volume 47: Volume 48: Volume 49: Volume 50: Volume 51: Volume 52: Volume 53: Volume 54: Volume 55: Volume 56: Volume 57: Volume 58: Volume 59: Volume 59: Volume 60:

MONOGRAPHS AND TREATISES
Chaos, Bifurcations and Fractals Around Us W. Szempli´ ska-Stupnicka n Bio-Inspired Emergent Control of Locomotion Systems M. Frasca, P. Arena & L. Fortuna Nonlinear and Parametric Phenomena V. Damgov Cellular Neural Networks, Multi-Scroll Chaos and Synchronization M. E. Yalcin, J. A. K. Suykens & J. P. L. Vandewalle Symmetry and Complexity K. Mainzer Applied Nonlinear Time Series Analysis M. Small Bifurcation Theory and Applications T. Ma & S. Wang Dynamics of Crowd-Minds A. Adamatzky Control of Homoclinic Chaos by Weak Periodic Perturbations R. Chacón Strange Nonchaotic Attractors U. Feudel, S. Kuznetsov & A. Pikovsky A Nonlinear Dynamics Perspective of Wolfram's New Kind of Science L. O. Chua New Methods for Chaotic Dynamics N. A. Magnitskii & S. V. Sidorov Equations of Phase-Locked Loops J. Kudrewicz & S. Wasowicz Equations of Phase-Locked Loops J. Kudrewicz & S. Wasowicz Smooth and Nonsmooth High Dimensional Chaos and the Melnikov-Type Methods J. Awrejcewicz & M. M. Holicke A Gallery of Chua Attractors (with CD-ROM) E. Bilotta & P. Pantano Numerical Simulation of Waves and Fronts in Inhomogeneous Solids A. Berezovski, J. Engelbrecht & G. A. Maugin Advanced Topics on Cellular Self-Organizing Nets and Chaotic Nonlinear Dynamics to Model and Control Complex Systems R. Caponetto, L. Fortuna & M. Frasca

Volume 61: Volume 62: Volume 63:

NONLINEAR SCIENCE
WORLD SCIENTIFIC SERIES ON

Series A

Vol. 64

Series Editor: Leon O. Chua

CONTROL OF CHAOS IN NONLINEAR CIRCUITS AND SYSTEMS
Edited by

Bingo Wing-Kuen Ling
King’s College London, UK

Herbert Ho-Ching Iu
The University of Western Australia

Hak-Keung Lam
King’s College London, UK

World Scientific
NEW JERSEY

LONDON

SINGAPORE

BEIJING

SHANGHAI

HONG KONG

TA I P E I

CHENNAI

Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

CONTROL OF CHAOS IN NONLINEAR CIRCUITS AND SYSTEMS World Scientific Series on Nonlinear Science, Series A — Vol. 64 Copyright © 2009 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN-13 978-981-279-056-9 ISBN-10 981-279-056-X

Printed in Singapore.

However. sigma delta modulators (Chapter 9). a fuzzy model based approach written by H. The objective of this book is to review the state-of-the-art chaos control methods for some common nonlinear circuits and systems. The stability conditions in terms of linear matrix inequalities are derived v . a time-delayed approach written by H. The first part of the book consists of reviews on general chaos control methods. In Chapter 2.K. A master slave synchronization problem for chaotic Lur’e systems is considered. A switching controller is then employed to synchronize the systems. power converters (Chapter 6). Leung is reviewed. phase lock loops (Chapter 8). In particular. such as electronic circuits (Chapter 5). control theories for nonlinear circuits and systems are still very limited. different nonlinear circuits and systems could exhibit very different behaviours. The design of the time delayed feedback controller can be accomplished by means of the feasibility of linear matrix inequalities. Huang and G. Control theories for linear circuits and systems are well developed and almost complete. A chaotic system is first represented by the fuzzy model. Lam and F. and stimulate further research and development in chaos control for nonlinear circuits and systems. human brains (Chapter 7). are found almost everywhere.H. etc. it is difficult to unify a general control theory for general nonlinear circuits and systems. Understanding nonlinear behaviours as well as control of these circuits and systems are important for real practical engineering applications. Feng is reviewed in Chapter 1.PREFACE Nonlinear circuits and systems. This book consists of three parts. The synchronization of chaotic systems subject to parameter uncertainties is considered.F. Hence. Up to now. such as those listed in the above. A delay independent and delay dependent synchronization criteria are derived based on the H∞ performance.

C. Shao is reviewed. The tracking performance and parameter design of the controller are formulated as a generalized eigenvalue minimization problem which is solved numerically via some convex programming techniques. chaos controls for pulse width modulation current mode single phase H-bridge inverters written by B. Mogul are . terminal sliding mode control and non-singular terminal sliding mode control. Cvitanovic and D. as well as some practical issues. are discussed.M.X. The rough locations of all the possible jump points are identified using existing kernel methods. Tôrres. Aguirre.C. Systems with nonparametric regression with jump points are considered.M. Mendes are discussed in Chapter 5.vi Preface based on the Lyapunov stability theory. Liu. The approximation problem is formulated as an optimization problem and solved via existing optimization tools.A. M. the control energy and the model complexity is derived.M. namely to force the system states to converge to zero or to track desired trajectories. Slutzky. In Chapter 6. This method is very robust and easy to implement. K.L. In Chapter 4. A time scaling transformation is derived so as to map the undecided jump points to fixed points.B. Feki and H. Observer based chaos synchronizations for chaotic systems with single nonlinearity and multi-nonlinearities are also presented. Teo and Q. P.A. Wu. A smooth spline function is used to approximate each segment of the regression function. A time delayed feedback controller is used in conjunction with the proportional controller in its simple form as well as in its extended form to stabilize the desired periodic orbit for larger values of the proportional controller gain. The second part of the book consists of reviews on general chaos controls for continuous-time systems. An inductorless Chua’s circuit realization is presented. Yu is reviewed. chaos controls for epileptiform bursting in the brain written by M. Feng and X. In Chapter 7. The tradeoff among the control objective. R. In Chapter 3. mathematical modelling and dynamical characterization. In particular. traditional sliding mode control. L. Robert.M.J. are employed for the control of a chaotic system to realize two different control objectives. an optimal control approach written by C.W.Z.H. Palhares and E. a sliding mode control approach written by Y.A. chaos controls for Chua’s circuits written by L. Iu are discussed. Three kinds of sliding mode control methods. such as data analysis.

Prof. The editors are much indebted to the Editor of the World Scientific Series on Nonlinear Science.F. A fuzzy impulsive control approach is employed for the control of the sigma delta modulators. Ling. The local stability criterion and the condition for the occurrence of limit cycle behaviours are derived. Harb and B. the initial condition and the filter parameters are. the effect of the audio clicks and the distance between the state vectors and an invariant set are minimized supposing that the invariant set is nonempty. a fuzzy impulsive control law is formulated so that the occurrence of the limit cycle behaviours. chaos controls for phase lock loops written by A. Lakshmi Narayan for their help and congenial processing of the edition. Reiss are discussed.A. and to Senior Editor Ms. The third part of the book consists of reviews on general chaos controls for discrete-time systems. the phase lock loops will not exhibit Hopf bifurcation and chaotic behaviours.W. Harb are discussed in Chapter 8.Preface vii discussed. A nonlinear controller based on the theory of backstepping is designed so that the phase lock loops will not be out of lock. Ho and J. The techniques provide a nonlinear control pathway for terminating or potentially preventing epileptic seizures in the whole brain. C. chaos controls for sigma delta modulators written by B. Based on the derived conditions.K. Chaos analysis and chaos control algorithms for manipulating the seizure-like behaviour in a brain slice model are discussed.M.D. Leon Chua. Also. Bingo Wing Kuen Ling Herbert Ho Ching Iu Hak Keung Lam .Y. In particular. The state vectors can be bounded within any arbitrary nonempty region no matter what the input step size. In Chapter 9.

This page intentionally left blank .

M. Yu 4. Palhares and E. Feng 2.H. Mendes 6. L. Sliding Mode Control of Chaotic Systems Y. Chaos Control for a PWM H-bridge Inverter B. Wu. Robert. Feng and X.L. Synchronization of Uncertain Chaotic Systems based on Fuzzy-model-based Approach H. Tôrres.Z. Leung 3.B.K.A.M. R.F. Aguirre. Iu v 3 35 55 79 97 165 ix . A New Two-stage Method for Nonparametric Regression with Jump Points C. C.A.A. Robust Synchronization of Chaotic Systems based on Time-delayed Feedback Control H. Liu. Feki and H. Huang and G. M.M.CONTENTS Preface Section A: General Chaos Control Methods 1.X. Teo and Q.C. Chaos Control for Chua’s Circuits L. Lam and F.M.H. Shao Section B: Chaos Control for Continuous-time Systems 5. K.

W.J. Ling.M. Harb and B.Y.K. Chaos Control of Epileptiform Bursting in the Brain M.W. Control of Sigma Delta Modulators via Fuzzy Impulsive Approach B.x Contents 7.A. C. Ho and J.D. Chaos Control for Phase Lock Loop A. Cvitanovic and D.F. P. Reiss 197 227 245 . Mogul Section C: Chaos Control for Discrete-time Systems 8. Harb 9. Slutzky.

Section A General Chaos Control Methods .

This page intentionally left blank .

It basi3 . n-scroll attractors [11] and hyperchaotic attractors [12]. Feng† Department of Manufacturing Engineering and Engineering Management. 1. As is well known now..g. Kowloon. The master-slave synchronization problem for chaotic Lur’e systems has recently attracted considerable interest. such as Chua’s Circuit [10].Chapter 1 ROBUST SYNCHRONIZATION OF CHAOTIC SYSTEMS BASED ON TIME-DELAYED FEEDBACK CONTROL H. A simulation example is finally given to demonstrate the effectiveness and performance of the developed approaches. City University of Hong Kong.1. Hong Kong. Introduction Since the pioneering work of Pecora and Carroll [1].edu.edu.hk † megfeng@cityu. e. the so-called Lur’e systems are regarded as a class of feedback systems whose forward path is a linear time-invariant system and whose feedback path is a memoryless (possibly time-varying) nonlinearity satisfying a sector condition [9].hk The master-slave synchronization problem for chaotic Lur’e systems is studied in this chapter based on time-delayed feedback control. Huang∗ and G. ∗ hhuang@student. It is assumed that the master system of the synchronization scheme is subject to noise disturbances. chaotic synchronization has been an active research topic and achieved many successful applications. chemical reactions and information processing (see. and so on. such as secure communication.cityu. [2–8]). Delay-independent and delay-dependent synchronization criteria are presented such that the controlled slave system can robustly track the noise-disturbed master system with guaranteed H∞ performance. It is shown that the design of the time-delayed feedback controller can be accomplished by means of the feasibility of linear matrix inequalities. On the other hand. many nonlinear systems can be represented in this form.

Recently. The authors in [30. The absolute stability theory was utilized in the development of synchronization conditions for master-slave Lur’e systems via static and linear feedback in [4. By applying the vector modulation and either full static state feedback or linear dynamic output error feedback. the authors in [14–18] investigated the master-slave synchronization schemes for identical or nonidentical Lur’e systems.4 H. time-delayed systems have received special attention over the past few years [22]. 27] and delay-dependent synchronization criteria [25. It is also known that time delay is often encountered in various control systems. 29]. Generally. the master-slave synchronization problem via time-delayed feedback control has received much attention. 5]. From the point of view of control. The existing results related to this issue can be generally classified into two categories: delay-independent synchronization criteria [24. because it can be solved efficiently by the standard numerical packages [28. 30. as well as the relation between synchronization and global asymptotical stability. During the past few years. The propagation delay was firstly introduced into the chaotic synchronization problem in [23]. it is worth noting that in those works. and control community. and many results have been reported in the open literature. 26. the feedback gain matrices can be designed only when a nonlinear programming problem corresponding to the derived matrix inequalities conditions is solved. 31]. The authors called this problem a phase sensitivity because of the distance between two remote chaotic systems and showed that the existence of time delay may destroy synchronization. However. several delay-independent and delaydependent synchronization conditions have been proposed to ensure the error system to be globally asymptotically stable. systems. 31] have presented some LMI based synchronization criteria such that the controlled slave system can track the master system. While the delay-dependent conditions are related to the size of time delay. the delay-dependent conditions are considered to be less conser- . The delay-independent conditions are irrespective of the size of time delay. A unifying definition for synchronization between stationary finite dimensional deterministic dynamical systems was presented in [2]. Huang and G. A unified framework for synchronization of dynamical systems was presented in [13]. An excellent overview of various methods for synchronization of chaotic systems was available in [3]. In [24–27]. which is one of the main sources for instability and poor performance [19–21]. the linear matrix inequality (LMI) approach has been widely used in circuits. Feng cally relates to give sufficient conditions for master-slave synchronization for identical or nonidentical Lur’e systems.

In the synchronization scheme. manifold-based method [37]. delay-dependent synchronization conditions without/with slack variables are obtained by using the well-known Jensen’s inequality and a newly established bounding technique. The rest of this chapter is organized as follows. especially when the size of time delay is small. which includes adaptive control [32. active control [38]. Moreover. It is also shown that the design of the time-delayed feedback controller can be achieved by solving some LMIs. In Section 1. etc. A delay-independent synchronization criterion is first given such that the controlled slave system can robustly track the master system. The main objective of this chapter is to study the master-slave synchronization for chaotic Lur’e systems with guaranteed performance in H∞ sense. 33]. Section 1.6. coupling control [36].3 is dedicated to presenting a delay-independent synchronization criterion such that the existence of the feedback controller can be guaranteed by means of the feasibility of an LMI. our attention is focused on the master-slave synchronization problem for chaotic Lur’e systems by using the time-delayed feedback control method. 40] is used as an example to demonstrate the application of the proposed time-delayed feedback controller design approaches and their performance. . a wide variety of synchronization approaches have been proposed to handle the master-slave synchronization problem. fuzzy control [34. several slack variables are introduced to further reduce the conservatism of the synchronization conditions. the delay-independent conditions are often more conservative than the delay-dependent conditions.2 for chaotic Lur’e systems with time-delayed feedback control. the master system is assumed to be disturbed by the exogenous noise input. and some preliminaries are also given. delay-dependent conditions are presented to ensure the existence of the desired time-delayed feedback controller such that the master system and the slave system are synchronized.5. In recent years. which is followed by conclusions in Section 1. respectively. An example with simulation results is used to demonstrate the effectiveness of the developed approach in Section 1. In this chapter. Finally. The master-slave synchronization scheme is formulated in Section 1. the Chua’s Circuit system [10. impulsive-control method [39]. 27].Synchronization of Chaotic Systems with Delayed Feedback Control 5 vative than the delay-independent conditions. As mentioned above. such that the controlled slave system can synchronize the master system with guaranteed H∞ performance. 35]. and time-delayed feedback approach [25. Then.4.

negative semi-definite. Let L2 [0. Consider a master-slave synchronization scheme for chaotic Lur’e systems with time-delayed feedback control. For a real square matrix X. The master and slave systems are the so-called Lur’e systems with state vectors x(t). . 0]. Let R denote the set of real numbers. ∞). . Matrices. respectively. Fig. Feng 1. Problem Formulation Notations: The following notations adopted in this chapter are standard. w(t) ∈ Rm is the disturbance term belonging to L2 [0. the arguments of a function will be omitted in the analysis when no confusion can arise. are assumed to have compatible dimensions. the notation X > 0 (X ≥ 0. e. if not explicitly stated. MN . . B ∈ Rn×nh . M2 . slave system S and time-delayed feedback controller C. in which the master system is subject to noise input:  ˙  x(t) = Ax(t) + Bσ(Cx(t)) + Dw(t) M:  p(t) = Hx(t)  (1. ∞). C([−τ. ∞) be the space of square-integrable vector functions over [0. . The superscript “T” stands for matrix transposition. 0] to Rn with the norm ϕ = sup |ϕ(ϑ)|. negative definite. M2 .g. Rn the n-dimensional Euclidean space and Rn×m the set of all n×m real matrices. where | · | is the Euclidean norm −τ ≤ϑ≤0 in Rn . The notation ∗ always denotes the symmetric block in a symmetric matrix. Sometimes. 1.1) ˙  y(t) = Ay(t) + Bσ(Cy(t)) + u(t) S:  q(t) = Hy(t) C : u(t) = M (p(t − τ ) − q(t − τ )) with master system M.6 H. D ∈ Rn×m and H ∈ Rl×n are known . I is the identity matrix with appropriate dimension. .1).. Rn ) denotes the family of continuous functions ϕ from [−τ. X < 0. The shorthand notation diag{M1 . The real matrices A ∈ Rn×n . . Huang and G. X ≤ 0) means that X is real symmetric and positive definite (positive semidefinite.1 is a block diagram of the master-slave synchronization scheme (1. y(t) ∈ Rn and output vectors p(t).2. respectively). MN } denotes a block diagonal matrix with diagonal blocks being the matrices M1 . . q(t) ∈ Rl . For τ > 0. C ∈ Rnh ×n . X ∗ Y Z = X YT Y Z . .

the diagonal nonlinearity σ(·) : Rnh → Rnh satisfies a sector condition with σi (·) (i = 1. When w(t) = 0 and τ = 0.4)   ©3§¥2£%$ ( ¥¨ 01()' p(t − τ ) − u (t )  £ §" § ¥%$#!©¡ !  § ©¡ ¤ p(t ) τ  § ©¡ & q(t − τ ) τ q(t ) § ¥©§ ¥¦¤¢  £  ¨ £¡ . σi (ξ)(σi (ξ) − kξ) ≤ 0. 1. . . 15. τ ≥ 0 is a constant time delay. which may be more practical than those studied in [5. Time-delayed feedback control based master-slave synchronization scheme. Remark 1. then the error system can be obtained as follows: E : e(t) = Ae(t) − M He(t − τ ) + Bη(Ce(t). for i = 1.1) degenerates into the one discussed in [27]. That is. the master-slave synchronization scheme (1. k]. 18. (1. nh . 27]. . y(t)) + Dw(t). .3) (1. As in [9. The scheme aims at robustly synchronizing the slave system S to the noise-perturbed master system M with a guaranteed performance by utilizing the time-delayed feedback to the slave system S with control signal u(t) ∈ Rn with feedback gain matrix M ∈ Rn×l and time delay τ . The time-delayed feedback control based master-slave synchronization scheme for Lur’e systems is assumed to be subject to noise perturbation. When w(t) = 0. ∀ ξ.2) + Fig. . . It should be pointed out that based on the developed approaches.1. 26. Define the errors e(t) = x(t) − y(t) and z(t) = p(t) − q(t). 26. The purpose of this chapter is to develop several approaches to dealing with the guaranteed performance synchronization problem for chaotic Lur’e systems. some delay-independent and delay-dependent synchronization criteria can be easily obtained for Lur’e systems studied in [5. y(t)) = σ(Ce(t) + Cy(t)) − σ(Cy(t)). (1.1) reduces to the one studied in [5. 18]. 27]. 24. 27]. nh ) belonging to sector [0.1. . the master-slave synchronization scheme (1. . 18. 2. with η(Ce(t).Synchronization of Chaotic Systems with Delayed Feedback Control 7 constant matrices. ˙ z(t) = He(t). 2. 24.

8 H. . In this case. the master system M and the slave system S are said to be synchronized with guaranteed H∞ performance γ. 0≤ ηi (cT e. (1. Lemma 1. .e. 27]. . i Then. y) belongs to sector [0. The LMI S11 T S12 S12 −S22 < 0.6) (1.1 (Schur complement [28]).. . cT denotes the ith i row vector of C). find a suitable time-delayed feedback controller C such that the error system (1. Feng T Let C = c1 c2 . T T where S11 = S11 . it follows from (1. where the norm · 2 is defined as ∞ ψ 2 := 0 ψ T (t)ψ(t)dt. 18. Huang and G. −1 T S11 + S12 S22 S12 < 0. i = 1. nh (cT e = 0). nh . ∞). 2. .5) The objective of this chapter is to develop several delay-independent and delay-dependent approaches to the design of the time-delayed feedback controller.5) that ηi (cT e.3) with w(t) = 0 is globally asymptotically stable. given a prescribed level of noise attenuation γ > 0. cT e cT e i i ∀e. the following assumption is made: A: The nonlinearity η(Ce. . S22 = S22 . y)(ηi (cT e. . i i i ∀ e. . namely. y) σi (cT e + cT y) − σi (cT y) i i i i = ≤ k. . .7) under zero-initial conditions for all nonzero w(t) ∈ L2 [0.1). As in [4. y. 2. y) − kcT e) ≤ 0. and z(t) 2 < γ w(t) 2 (1. is equivalent to S22 > 0. k]. That is. cnh with ci ∈ Rn (i. i = 1. such that controlled slave system S can robustly synchronize the master system M with a guaranteed performance in the H∞ sense. . y. We end this section by recalling two lemmas. to obtain synchronization criteria for the master-slave synchronization scheme (1.

We first show that the error system (1.1. The master system M and the slave system S are synchronized with guaranteed H∞ performance γ. a delay-independent approach is proposed to deal with the master-slave synchronization problem of chaotic systems.8) ∗ −2Λ 0 0  < 0. ˙ It is clear that the LMI (1. λ2 .    ∗ ∗ ∗ −γ 2 I 0  ∗ ∗ ∗ ∗ −I where Ω1 = P A + AT P + Q.8) results in   Ω1 −P M H P B + kC T Λ  < 0. scalar δ > 0. such that the controlled slave system can robustly synchronize the noise-disturbed master system with guaranteed H∞ performance. the error system (1. . 1. which is applicable to time delay of arbitrary size. then δ δ T δ δ 0 ω T (s)X ω(s)ds ≥ 0 ω(s)ds X 0 ω(s)ds . vector function ω : [0. λnh ) > 0 such that the following LMI holds:   Ω1 −P M H P B + kC T Λ P D H T  ∗ −Q 0 0 0     ∗ (1.3) with w(t) = 0 is globally asymptotically stable under the condition of Theorem 1. . An LMI synchronization condition is derived for the design of time-delayed feedback controller. X = X T . When w(t) = 0. if there exist real matrices P > 0.10) .2 (Jensen inequality [19]). γ] → Rm such that the integrations concerned are well defined. . Proof. . For any constant matrix X ∈ Rm×m .1. Ω2 =  ∗ −Q 0 ∗ ∗ −2Λ (1. Delay-Independent Synchronization Criterion First of all.3) can be rewritten as: e(t) = Ae(t) − M He(t − τ ) + Bη(Ce(t). Theorem 1. y(t)).3. Assume that the feedback gain matrix M is given.Synchronization of Chaotic Systems with Delayed Feedback Control 9 Lemma 1.9) (1. Q > 0 and a diagonal matrix Λ = diag(λ1 .

y(t)) T . which implies ˙ V1 (t) ≤ − eT (t)e(t).11) V1 (t) = eT (t)P e(t) + t−τ eT (s)Qe(s)ds. 0} ≤ 0. (1. y(t))Λη(Ce(t). 0. y(t)) +2keT (t)C T Λη(Ce(t). one has ˙ V1 (t) = 2eT (t)P e(t) + eT (t)Qe(t) − eT (t − τ )Qe(t − τ ) ˙ = eT (t)(P A + AT P )e(t) − 2eT (t)P M He(t − τ ) +2eT (t)P Bη(Ce(t). y(t)) + eT (t)Qe(t) −eT (t − τ )Qe(t − τ ) ≤ eT (t)(P A + AT P + Q)e(t) − 2eT (t)P M He(t − τ ) +2eT (t)P Bη(Ce(t). Since Ω2 < 0. y(t)) − eT (t − τ )Qe(t − τ ) −2η T (Ce(t).13) with σ1 (t) = eT (t) eT (t − τ ) η T (Ce(t). y(t))Λη(Ce(t).12) where P > 0. Huang and G. . . Choose a Lyapunov functional candidate as t (1. Q > 0 are to be determined.11) into account. λ2 .6) Λ = diag(λ1 . y) − kcT e) i i i = −2η T (Ce(t). y(t)). Then. .9) and taking (1. Feng It follows from (1. . y)(ηi (cT e. y(t)) +2keT (t)C T Λη(Ce(t). there must exist a sufficiently small Ω2 + diag{ I. > 0 such that According to the stability theory of [20].9) is globally asymptotically stable. by directly calculating the time-derivative of V1 (t) along the trajectory of system (1. λnh ) > 0. the error system (1. y(t)) T = σ1 (t)Ω2 σ1 (t). nh that for any diagonal matrix 0 ≤ −2 i=1 λi ηi (cT e. (1. .10 H.

Theorem 1. y(t)) w T (t) T  Ω3 =   Ω1 + H T H ∗ ∗ ∗ −P M H −Q ∗ ∗ P B + kC T Λ P D 0 0  .16). which is formulated by means of the feasibility of an LMI.16) with σ2 (t) =  eT (t) eT (t − τ ) η T (Ce(t). −2Λ 0  ∗ −γ 2 I .12) one has V1 (0) = 0 and V1 (t) ≥ 0 for t > 0.14) Under the zero initial condition. (1. From (1. Let γ > 0 be a prescribed constant scalar. The following theorem presents a delay-independent synchronization controller design method to ensure that the slave system S can synchronize the master system M with guaranteed H∞ performance γ.15) = 0 Following the similar line of the proof of (1.17) which implies J(t) < 0 for w(t) = 0.13).  By employing Lemma 1. we introduce ∞ J(t) = 0 z T (t)z(t) − γ 2 wT (t)w(t) dt. (1. from (1. (1.Synchronization of Chaotic Systems with Delayed Feedback Control 11 Next.1) under zero initial condition. the LMI (1.8) is equivalent to Ω3 < 0. one thus has that for w(t) = 0 ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V1 (t) < 0. the master system M and the slave system S are synchronized with guaranteed H∞ performance γ. The master system M and the slave system S are synchronized with guaranteed H ∞ .2. z(t) 2 < γ w(t) 2 . This completes the proof.1. it is not difficult to derive that T ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V1 (t) ≤ σ2 (t)Ω3 σ2 (t) (1. to establish the H∞ performance to the master-slave synchronization scheme (1. Then ∞ J(t) ≤ 0 ∞ z T (t)z(t) − γ 2 wT (t)w(t) dt + V1 (t)|t=∞ − V1 (0) ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V1 (t) dt. Therefore. That is to say.

an LMI based synchronization condition can be immediately derived shown in the following corollary. The following section is to this purpose.  ∗ −Q 0 (1. where the obtained delay-independent synchronization criteria were expressed in terms of matrix inequalities.2. And the feedback gain matrix can be designed as M = P −1 G. Algorithm 1. . The master-slave synchronization scheme (1. . Huang and G. Remark 1. G and a diagonal matrix Λ = diag(λ1 .Q.Λ min γ 2 . From Theorems 1.   2  ∗  ∗ ∗ −γ I 0 ∗ ∗ ∗ ∗ −I where Ω4 = P A + AT P + Q. Corollary 1.2. It can be directly deduced from Theorem 1. λnh ) > 0 such that the following LMI holds:   Ω4 −GH P B + kC T Λ P D H T  ∗ −Q 0 0 0     ∗ (1. Feng performance γ.3.2. the delay-independent synchronization condition is viewed to be more conservative than the delay-dependent counterpart. P.18). λ2 . if there exist real matrices P > 0. . In general. The master-slave synchronization described in (1. 27].19) ∗ ∗ −2Λ where Ω4 is the same as the one in Theorem 1.G. It is thus worth developing delaydependent synchronization criteria. subject to the LMI (1.2 can be optimized by the convex optimization algorithm [28.1) with w(t) = 0 was studied in [26. especially when the size of time delay is small.18) ∗ −2Λ 0 0  < 0. Proof.1 and 1.1) with w(t) = 0 can be achieved with M = P −1 G.12 H. . Q > 0. G and a diagonal matrix Λ > 0 such that   Ω4 −GH P B + kC T Λ  < 0. It should be noted that the performance index γ described in Theorem 1. . Remark 1.1 by applying the change of the variable M = P −1 G. if there exist real matrices P > 0. Q > 0. 29].1.1.

21) is equivalent to Σ4 + τ 2 ΣT RΣ5 < 0. .Synchronization of Chaotic Systems with Delayed Feedback Control 13 1.4. Proof. the LMI (1. As in the proof of Theorem 1. a delay-dependent synchronization condition is first proposed in this section. It follows from the LMI (1.1. Σ3 =  (1. Q > 0.3) with w(t) = 0 is globally asymptotically stable under the condition of Theorem 1.20) that   Σ1 −P M H + R Σ2 τ AT R  ∗ −Q − R 0 −τ H T M T R   < 0. Assume that the feedback gain matrix M is given. λ2 . . . Σ2 = P B + kC T Λ. if there exist real matrices P > 0. The master system M and the slave system S are synchronized with guaranteed H∞ performance γ. there must exist a sufficiently small scalar > 0 such that Σ4 + τ 2 ΣT RΣ5 + diag{ I. (1. Delay-Dependent Synchronization Criteria By utilizing the well-known Jensen inequality [22]. 5 (1.1. 0} ≤ 0.3. .21)  ∗  ∗ −2Λ τ BT R ∗ ∗ ∗ −R By using Lemma 1. λnh ) > 0 such that the following LMI holds:   Σ1 −P M H + R Σ2 PD τ AT R HT  ∗ −Q − R 0 0 −τ H T M T R 0      ∗ −2Λ 0 τ BT R 0   ∗   < 0. 5 with   Σ1 −P M H + R P B + kC T Λ .20)  ∗ ∗ ∗ −γ 2 I τ DT R 0     ∗ ∗ ∗ ∗ −R 0  ∗ ∗ ∗ ∗ ∗ −I Σ1 = P A + AT P + Q − R. R > 0 and a diagonal matrix Λ = diag(λ1 .22) .3. Σ4 =  ∗ −Q − R 0 ∗ ∗ −2Λ Σ5 = A −M H B . 0. Theorem 1. where Then. we first prove that the error system (1.

λnh ) > 0. . Q > 0 and R > 0 are to be determined.24) +τ −τ t+θ eT (s)Re(s)dsdθ. y(t)) +2keT (t)C T Λη(Ce(t).9).25) By Lemma 1. y(t)). nh 0 ≤ −2 i=1 λi ηi (cT e. By calculating the time-derivative of V2 (t) along the solution of system (1. y)(ηi (cT e. y(t))Λη(Ce(t). one has t t T t −τ t−τ eT (s)Re(s)ds ≤ − ˙ ˙ t−τ e(s)ds ˙ R t−τ e(s)ds ˙ = −[e(t) − e(t − τ )]T R[e(t) − e(t − τ )] = −eT (t)Re(t) + 2eT (t)Re(t − τ ) −eT (t − τ )Re(t − τ ). ˙ ˙ where P > 0.26) .14 H. for any diagonal matrix Λ = diag(λ1 . Feng In addition. . . Huang and G. λ2 .2. ˙ ˙ (1. . one can derive ˙ V2 (t) = 2eT (t)P e(t) + eT (t)Qe(t) − eT (t − τ )Qe(t − τ ) ˙ t +τ 2 eT (t)Re(t) − τ ˙ ˙ t−τ eT (s)Re(s)ds ˙ ˙ = eT (t) P A + AT P + Q e(t) −2eT (t)P M He(t − τ ) + 2eT (t)P Bη(Ce(t). (1. Choose a Lyapunov functional candidate as: t (1. y) − kcT e) i i i = −2η T (Ce(t).23) V2 (t) = eT (t)P e(t) + t−τ 0 t eT (s)Qe(s)ds (1. y(t)) −eT (t − τ )Qe(t − τ ) + τ 2 eT (t)Re(t) ˙ ˙ t −τ t−τ eT (s)Re(s)ds.

y(t)) +τ 2 eT (t)Re(t) ˙ ˙ T = ξ1 (t) Σ4 + τ 2 ΣT RΣ5 ξ1 (t). y(t)) −eT (t − τ )(Q + R)e(t − τ ) − 2η T (Ce(t). from (1.1) under zero initial condition. it follows from (1.29) = 0 Following the similar line of the proof of (1.28) Under the zero initial condition. to establish the H∞ performance to the master-slave synchronization scheme (1.25) and (1. (1. it is easy to obtain that T ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V2 (t) ≤ ξ2 (t) Σ6 + τ 2 ΣT RΣ7 ξ(t) (1.26).  P B + kC T Λ P D 0 0  . y(t)) T (1.27) . with ξ1 (t) = eT (t) eT (t − τ ) η T (Ce(t). (1.30) 7 with ξ2 (t) =  eT (t) eT (t − τ ) η T (Ce(t).9) is globally asymptotically stable. y(t)) w T (t) T . we introduce ∞ J(t) = 0 z T (t)z(t) − γ 2 wT (t)w(t) dt. 5 ≤ − eT (t)e(t). (1. Then ∞ J(t) ≤ 0 ∞ z T (t)z(t) − γ 2 wT (t)w(t) dt + V2 (t)|t=∞ − V2 (0) ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V2 (t) dt. it is not difficult to derive ˙ V2 (t) ≤ eT (t) P A + AT P + Q − R e(t) +2eT (t)(−P M H + R)e(t − τ ) + 2eT (t)(P B + kC T Λ)η(Ce(t).22).23).27). It follows from [20] that the error system (1. y(t))Λη(Ce(t). −2Λ 0  ∗ −γ 2 I .Synchronization of Chaotic Systems with Delayed Feedback Control 15 Therefore. Next. Σ7 = Σ1 + H T H  ∗ Σ6 =   ∗ ∗ A −M H −P M H + R −Q − R ∗ ∗ B D .24) that V2 (0) = 0 and V2 (t) ≥ 0 for t > 0. (1.

7 It follows from (1. Feng By utilizing Lemma 1. It then follows from (1. (1. Huang and G.16 H. Q > 0.1. it is only to show that the LMI (1. for P > 0.30) that for w(t) = 0 ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V2 (t) < 0. In fact. This completes the proof.31) and thus J(t) < 0 for any w(t) = 0. Then we have the following result. Therefore.32) that   Σ8 −GH + R Σ9 PD τ AT P HT  ∗ −Q − R 0 0 −τ H T GT 0      T ∗ −2Λ 0 τB P 0   ∗  < 0. G and a diagonal matrix Λ > 0 such that the following LMI holds   Σ8 −GH + R Σ9 PD τ AT P HT  ∗ −Q − R 0 0 −τ H T GT 0      ∗ −2Λ 0 τ BT P 0   ∗  < 0. Proof. The master system M and the slave system S are synchronized with guaranteed H ∞ performance γ. Let γ > 0 be a prescribed constant scalar. (1.20) holds. By Theorem 1.4. And the feedback gain matrix can be designed as M = P −1 G.20) results in Σ6 + τ 2 ΣT RΣ7 < 0. (1. It leads to z(t) 2 < γ w(t) 2 .32)  ∗ ∗ −γ 2 I τ DT P 0   ∗    ∗ ∗ ∗ ∗ −2P + R 0  ∗ ∗ ∗ ∗ ∗ −I where Σ8 = P A + AT P + Q − R. R > 0.3. Σ9 = P B + kC T Λ. which gives −P T R−1 P ≤ −2P + R. if there exist real matrices P > 0. R > 0. Theorem 1.33)  2 T  ∗ ∗ ∗ −γ I τD P 0     ∗ ∗ ∗ ∗ −P T R−1 P 0  ∗ ∗ ∗ ∗ ∗ −I . the slave system S is synchronized to the master system M with guaranteed H∞ performance γ. P T R−1 P − 2P + R = (P − R)T R−1 (P − R) ≥ 0. the LMI (1.

The master-slave synchronization described in (1. 27] have investigated the master-slave synchronization scheme with w(t) = 0. respectively. which plays a key role in the derivation of the main results in this section.3.4. we will introduce some slack variables in our design. Therefore. we present a useful lemma to estimate the t ˙ ˙ term − t−τ eT (s)Re(s)ds. and noting the change of variable M = P −1 G. For the error system (1.4. it can be seen that the resulting LMI is the same as (1. I. Lemma 1. if there exist real matrices P > 0. the master-slave synchronization scheme (1.Synchronization of Chaotic Systems with Delayed Feedback Control 17 Now. Before proceeding further. 27] were formulated by means of matrix inequalities. N1 .2. I. N2 . I.and post multiplying the above matrix inequality (1.33) by diag{I. The proposed delay-dependent synchronization criteria in [24. P.3) and any real matrices R > 0. I}. I. an LMI-based delay-dependent synchronization condition can be obtained shown in the following corollary.G. In order to further reduce the conservatism of the synchronization condition. the following inequality . It implies that to obtain the time-delayed feedback controller. RP −1 .1) with w(t) = 0 can be achieved with M = P −1 G.1) is synchronized with guaranteed H∞ performance γ. Q > 0. By Theorem 1. pre.20). I. one needs to solve a corresponding nonlinear programming problem. Remark 1.5. I} and diag{I.34) where Σ8 is the same as the one in Theorem 1. Remark 1.4 can also be optimized by the convex optimization algorithm [28. R > 0. Corollary 1.Λ min γ 2 .Q.32).3. G and a diagonal matrix Λ > 0 such that Σ8  ∗   ∗ ∗  −GH + R −Q − R ∗ ∗  P B + kC T Λ τ AT P 0 −τ H T GT   < 0. I. N4 with appropriate dimensions. −2Λ τ BT P  ∗ −2P + R (1. N3 .R. This completes the proof. by Theorem 1. 29]. which can be facilitated readily by the Matlab LMI Control Toolbox [29]. It is noted that the performance index γ described in Theorem 1. Algorithm 1.4. subject to the LMI (1. The authors in [24. P −1 R.2.

˙ t−τ (1. (1. . one has Π2 Π1 ΠT = 2 0 0 0 R ≥ 0.36) it yields that t − t−τ eT (s)Re(s)ds ˙ ˙ t t ≤− t−τ t eT (s)Re(s)ds + ˙ ˙ t−τ t π1 (t) e(s) ˙ T Π1 π1 (t) e(s) ˙ ds =− t−τ eT (s)Re(s)ds + ˙ ˙ t−τ T π1 (t)N T R−1 N π1 (t) T +2π1 (t)N T e(s) + eT (s)Re(s) ds ˙ ˙ ˙ t T = τ π1 (t)N T R−1 N π1 (t) + 2 T = τ π1 (t)N T R−1 N π1 (t) + T π1 (t)N T e(s)ds ˙ t−τ T 2π1 (t)N T [e(t) − e(t − τ )].35) is true. the inequality (1. N T R−1 N N I 0 NT R .18 H. . (1. Let Π1 = Π2 = Then. y(t)) . eT (t) N1 eT (t − τ ) N2 N3 N4 η T (Ce(t).37) That is. Feng holds: t − t−τ T eT (s)Re(s)ds ≤ τ π1 (t)N T R−1 N π1 (t) ˙ ˙ T +2π1 (t)N T [e(t) − e(t − τ )]. −N T R−1 I which implies Π1 ≥ 0. w T (t) T . This completes the proof.35) with π1 (t) = N= Proof. Huang and G. Noting the fact that t e(s)ds = e(t) − e(t − τ ).

Θ14 = P D + N4 .3) can be given as follows: e(t) = Ae(t) − M He(t − τ ) + Bη(Ce(t). . N4 and a diagonal matrix Λ = diag(λ1 . T Θ22 = −Q − N2 − N2 .38)     Θ13 = P B + kC T Λ + N3 . . T Θ12 = −P M H − N1 + N2 . if there exist real matrices P > 0. λnh ) > 0 such that the following LMI holds:            where T Θ11 = P A + AT P + Q + N1 + N1 . (1. The master system M and the slave system S are synchronized with guaranteed H∞ performance γ. Assume that the feedback gain matrix M is given. Q > 0. When w(t) = 0. Θ11 ∗ ∗ ∗ ∗ ∗ ∗ Θ12 Θ22 ∗ ∗ ∗ ∗ ∗ Θ13 −N3 −2Λ ∗ ∗ ∗ ∗ Θ14 −N4 0 −γ 2 I ∗ ∗ ∗ τ AT R −τ H T M T R τ BT R τ DT R −τ R ∗ ∗ T τ N1 T τ N2 T τ N3 T τ N4 0 −τ R ∗ HT 0 0 0 0 0 −I        < 0. Theorem 1. N2 . we are in a position to present our main results.40) . R > 0.Synchronization of Chaotic Systems with Delayed Feedback Control 19 Now.5. N1 . ˙ From the LMI (1. λ2 . . y(t)).   (1. the error system (1. N3 .5. it is obvious that       Θ11 ∗ ∗ ∗ ∗ Θ12 Θ22 ∗ ∗ ∗ Θ13 −N3 −2Λ ∗ ∗ τ AT R −τ H T M T R τ BT R −τ R ∗ T τ N1 T τ N2 T τ N3 0 −τ R (1. we show that the error system (1. Proof.38). . Firstly.39)     < 0.3) with w(t) = 0 is globally asymptotically stable under the condition of Theorem 1.

43) V3 (t) = eT (t)P e(t) + t−τ 0 t eT (s)Qe(s)ds (1. 0} ≤ 0.39).1. N3 . the LMI (1. Q > 0. Then. 4 5 It follows from (1.45) . by directly computing the time-derivative of V3 (t) along the trajectory of system (1. 0.44) + −τ t+θ eT (s)Re(s)dsdθ. there must exist a sufficiently small scalar > 0 such that Π3 + τ ΠT RΠ4 + τ ΠT R−1 Π5 + diag{ I. λ2 . −2Λ B .40) is equivalent to Π3 + τ ΠT RΠ4 + τ ΠT R−1 Π5 < 0. R > 0 are to be determined. . nh (1. Feng Let Θ11 Π3 =  ∗ ∗ N1  Θ12 Θ22 ∗ N2  Θ13 −N3  .20 H. y(t)). . y(t)) +2keT (t)C T Λη(Ce(t). y)(ηi (cT e. ˙ ˙ where P > 0. ˙ ˙ (1. y(t)) − eT (t − τ )Qe(t − τ ) t +τ eT (t)Re(t) − ˙ ˙ t−τ eT (s)Re(s)ds. λnh ) > 0. one can derive ˙ V3 (t) = 2eT (t)P e(t) + eT (t)Qe(t) − eT (t − τ )Qe(t − τ ) ˙ t +τ eT (t)Re(t) − ˙ ˙ t−τ eT (s)Re(s)ds ˙ ˙ = eT (t)(P A + AT P + Q)e(t) − 2eT (t)P M He(t − τ ) +2eT (t)P Bη(Ce(t).42) matrix that for any diagonal 0 ≤ −2 i=1 T λi ηi (cT e. y) − kcT e) i i i = −2η (Ce(t). . Choose a Lyapunov functional candidate as t (1. 4 5 Then. Huang and G. Π4 = Π5 = A −M H By employing Lemma 1.41) (1.6) Λ = diag(λ1 . y(t))Λη(Ce(t). .

46) with π2 (t) = eT (t) eT (t − τ ) η T (Ce(t). y(t))Λη(Ce(t).48) Under the zero initial condition.1) under zero initial condition. y(t)) T . Then ∞ J(t) ≤ 0 ∞ z T (t)z(t) − γ 2 wT (t)w(t) dt + V3 (t)|t=∞ − V3 (0) ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V3 (t) dt. 7 (1.39) is globally asymptotically stable. y(t)) +2keT (t)C T Λη(Ce(t). (1.50) . one can deduce that ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V3 (t) T ≤ π1 (t) Π6 + τ ΠT RΠ7 + τ N T R−1 N π1 (t). we introduce ∞ J(t) = 0 z T (t)z(t) − γ 2 wT (t)w(t) dt.3. to establish the guaranteed H∞ performance to the master-slave synchronization scheme (1. it can be easily derived that t − t−τ T eT (s)Re(s)ds ≤ τ π2 (t)ΠT R−1 Π5 π2 (t) ˙ ˙ 5 T +2π2 (t)ΠT [e(t) − e(t − τ )].42). This together with (1. y(t)) − eT (t − τ )Qe(t − τ ) T +τ eT (t)Re(t) + τ π2 (t)ΠT R−1 Π5 π2 (t) ˙ ˙ 5 T +2π2 (t)ΠT [e(t) − e(t − τ )] 5 −2η T (Ce(t).3 and following the similar line of the proof of (1.43) and (1. 5 (1.47) According to the stability theory in [20].49) = 0 By using Lemma 1. from the definition of V3 (t). one has V3 (0) = 0 and V3 (t) ≥ 0 for t > 0. the error system (1.Synchronization of Chaotic Systems with Delayed Feedback Control 21 By following the similar line of the proof of Lemma 1. (1.45) gives ˙ V3 (t) ≤ eT (t)(P A + AT P + Q)e(t) − 2eT (t)P M He(t − τ ) +2eT (t)P Bη(Ce(t). Next. y(t)) T = π2 (t) Π3 + τ ΠT RΠ4 + τ ΠT R−1 Π5 π2 (t) 4 5 ≤ − eT (t)e(t).47). (1. (1.

6. ˙ z T (t)z(t) − γ 2 wT (t)w(t) + V3 (t) < 0. Theorem 1.22 H. Huang and G. The master system M and the slave system S are synchronized with guaranteed H∞ performance γ.  Θ11 + H T H Θ12 Θ13 Θ14 T  ∗ −Q − N2 − N2 −N3 −N4 Π6 =   ∗ ∗ −2Λ 0 ∗ ∗ ∗ −γ 2 I Π7 = A −M H B D .3. N1 .  (1.1) can be achieved with guaranteed H∞ performance γ. the LMI (1. N3 . 7 It implies that for w(t) = 0. N2 . Let γ > 0 be a prescribed constant scalar.38) can guarantee Π6 + τ ΠT RΠ7 + τ N T R−1 N < 0. Feng with π1 (t) and N being the same as those in Lemma 1. G and a diagonal matrix Λ > 0 such that the following LMI holds:   T Ξ11 Ξ12 Ξ13 Ξ14 τ AT P τ N1 H T T  ∗ Ξ22 −N3 −N4 −τ H T GT τ N2 0     ∗ T T ∗ −2Λ 0 τB P τ N3 0      T ∗ ∗ −γ 2 I τ DT P τ N4 0  < 0. R > 0. T Ξ12 = −GH − N1 + N2 . where Ξ13 = P B + kC T Λ + N3 . That is. if there exist real matrices P > 0. Ξ55 = −2τ P + τ R. z(t) 2 < γ w(t) 2 . Applying Lemma 1. (1.52)  ∗    ∗ ∗ ∗ ∗ Ξ55 0 0     ∗ ∗ ∗ ∗ ∗ −τ R 0  ∗ ∗ ∗ ∗ ∗ ∗ −I T Ξ11 = P A + AT P + Q + N1 + N1 . . Q > 0. This completes the proof.51) and thus J(t) < 0 for any w(t) = 0. T Ξ22 = −Q − N2 − N2 . Therefore. N4 . the master-slave synchronization scheme (1. Next is our result on the feedback controller design.1 again.   . Ξ14 = P D + N4 .

I.38) holds. such that the slave system is synchronized to the master system with guaranteed H∞ performance γ. respectively. N3 and N4 have been introduced into the LMI condition (1. This completes the proof. subject to the LMI (1. 29]. Proof. RP −1 .6 presents a delay-dependent synchronization condition to the design of the time-delayed feedback controller.N4 .52) implies that  Ξ11 Ξ12 Ξ13 Ξ14  ∗ Ξ −N3 −N4 22   ∗ ∗ −2Λ 0   ∗ ∗ −γ 2 I  ∗   ∗ ∗ ∗ ∗   ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ τ AT P −τ H T GT τ BT P τ DT P −τ P T R−1 P ∗ ∗ T τ N1 T τ N2 T τ N3 T τ N4 0 −τ R ∗ HT 0 0 0 0 0 −I  Pre.52). I}. The purpose of the introduction of the slack variables is to further reduce the conservatism of the derived synchronization criterion. to prove Theorem 1.6.7. Remark 1.38) in Theorem 1. N2 .Synchronization of Chaotic Systems with Delayed Feedback Control 23 And the feedback gain matrix can be designed as: M = P −1 G. Theorem 1. By Theorem 1. I. It should be pointed out that several slack variables N1 . I.N1 .5. Therefore.5. and applying the change of variable M = P −1 G. the master system M and the slave system S are synchronized with guaranteed H∞ performance γ. N2 . The matrices N1 . P −1 R. Remark 1. In view of the inequality P T R−1 P ≥ 2P − R resulting from (P − R)T R−1 (P − R) = P T R−1 P − 2P + R ≥ 0.3. The performance index γ described in Theorem 1. I. one can see that the resulting LMI is the same as (1.6.R.53) by diag{I. I.6 can be optimized by the convex optimization algorithm [28.Q. I. the LMI (1.G. P. N3 and N4 are even not required to be symmetric.3). Algorithm 1.52) by employing the developed bounding technique (Lemma 1. I} and diag{I. (1. I. it is sufficient to show that the LMI (1.Λ       < 0. I. It is thus expected that Theorem .53)     min γ 2 .and post multiplying the above matrix inequality (1.N2 .N3 .

28. N3 . Huang and G. When w(t) = 0. As can be seen in Corollary 1. N2 .55) y = x − y + z − w(t) ˙  z = −by + w(t) ˙ with nonlinear characteristic 1 h(x) = m1 x + (m0 − m1 )(|x + c| − |x − c|) 2 and parameters a = 9. Ξ12 . 1. The noise disturbance is taken as w(t) = 0. the following corollary can be directly deduced from Theorem 1. The master-slave synchronization described in (1. Remark 1.3 may be less conservative than the results reported in [26. 31].3.24 H. b = 14.5. 40] is used as an example to illustrate the application of the proposed time-delayed feedback controller design approaches and their performance. R > 0. .6.0001t sin(2π × 0. several slack variables have been used to reduce the conservatism of the synchronization condition.4 due to the increasing freedom of these slack variables. The advantage of the delay-dependent approach with slack variables over the delay-independent approach and the delaydependent approach without slack variables are also demonstrated. Corollary 1. Ξ13 . Feng 1. m0 = −(1/7). N1 .1) with w(t) = 0 can be achieved with M = P −1 G. A Simulation Example The Chua’s Circuit [10. 27.005t).6.8. m1 = (2/7). Consider the Chua’s Circuit subject to noise disturbances:  ˙  x = a(y − h(x)) + w(t) (1.3.2 and 1.   ∗ ∗ ∗ Ξ55 0  ∗ ∗ ∗ ∗ −τ R where Ξ11 .6 will be less conservative than Theorems 1.54) ∗ −2Λ τB P τ N3  < 0.01e−0. if there exist real matrices P > 0. Therefore. c = 1. Q > 0. G and a diagonal matrix Λ > 0 such that  T  Ξ11 Ξ12 Ξ13 τ AT P τ N1 T  ∗ Ξ22 −N3 −τ H T GT τ N2    T T   ∗ (1. Ξ22 and Ξ55 are the same as those in Theorem 1. Corollary 1.

3: P = Q= R= N1 = −0. 0. −0.2728 0.6346   0. Let τ = 0.7205 12.8642 0.7237 −0.0865 0.7241 5.6351 2.7378 0.3705  .6897 −0.5645 −0.2156 −0. −0.3705 0.0119   −5.2316 N2 =  −2.0233 −0. then the following results have been obtained by applying Algorithm 1.1868  −2.2322 −0.0005 0.1868 −0.1186     −0. B= 0 0 C=H= 1 0 0 T  . D= 1 −1 1 .0003   2.2491 0.2872.4542 −0.0865  .0233  −0.2156  −1.Synchronization of Chaotic Systems with Delayed Feedback Control 25 The system can be represented in Lur’e form with noise input by  −am1 a 0 A= 1 −1 1  . G =  −0.2313  2.7823 −1. 0 −b 0   −a(m0 − m1 ) .6375 −2.1055 3.0277   0.8993  . 0.8646  . N4 =  −0.0055 N3 =  0.2491  .1186  5.7241 0. This representation results in nh = 1.8635  .0097 0.8638  .7298 −0.7169 0.2742 −0.8512 Λ = 2.1. 0. −0. 1 and σ(ξ) = 2 (|ξ + c| − |ξ − c|) belonging to sector [0. k] with k = 1.2321 0.1055 0.0062  .3625 −2.2547   1.2728 1.0103  .7182 −12. 0.

The simulation results for the master-slave synchronization scheme of the Chua’s Circuits (1. The simulation results confirm the effectiveness of Theorem 1.26 H. 1.56).32 (a) and (b) represent the state x(t) of the master system and the state y(t) of the slave system.9019  . the time-delayed feedback gain matrix can be designed as   5. However.01 infeasible infeasible 0. Huang and G.1453 M = P −1 G =  1.4.6736  −7.15 0. The initial conditions are taken as x(0) = T T and y(0) = −0.1. Table 1.6 over Theorems 1.0015 . respectively.001 Algorithm 1.4 and 1.55). Furthermore. (c) is the responses of the error state e(t) of the error system. In Fig.6 for the design of the time-delayed feedback controller for the master-slave synchronization scheme of Chua’s Circuits (1.2 and 1.5363 (1. the maximum allowed time delay in Theorem 1.3. This clearly demonstrates that there still exists some extent of conservatism in the proposed design approaches and thus suggests an interesting future research topic. 1.0254 τ = 0.8349e + 003  8. Figs. The results clearly demonstrate the much better performance of the delaydependent approach with slack variables over the other two approaches. respectively. −6.55) are shown in Fig.21 −0. Methods τ = 0.15 −0.3. In order to further demonstrate the advantage of Theorem 1.5359 τ = 0. by solving the LMI (1.1 infeasible Algorithm 1. Feng with the optimal H∞ performance index γmin = 1. it has been observed throughout simulations that with the same timedelayed feedback controller (1.6 is τmax = 0.0021 minimum H∞ performances for different τ = 0. a more detailed comparison between the minimum H∞ performance indexes for different time delays are summarized in Table 1. Comparison of the time delays.3 0.14534 are shown in Fig.5 are the simulation results when τ = 0.2999 and τ = 0.14534 infeasible infeasible 3. 0.3 0.3).7490  and the simulation results for τ = 0.2. In this case. 0.52) via the Matlab LMI Control Toolbox.2.1.56) M =  0.5359.1 infeasible infeasible 1. 1.11 .14534. 1. the Chua’s Circuits can be synchronized for any τ ∈ [0.2 infeasible Algorithm 1.

2. .Synchronization of Chaotic Systems with Delayed Feedback Control 27 (a) (b) (c) Fig.1. Simulation results for master-slave synchronization of Chua’s Circuits (1.55) when τ = 0. 1.

14534.28 H. Huang and G. 1. Feng (a) (b) (c) Fig. .3. Simulation results for master-slave synchronization of Chua’s Circuits (1.55) when τ = 0.

.4. Simulation results for master-slave synchronization of Chua’s Circuits (1.Synchronization of Chaotic Systems with Delayed Feedback Control 29 (a) (b) (c) Fig. 1.55) when τ = 0.2999.

2 0 −0.4 −4 −2 x1 2 0 4 (a) SLAVE SYSTEM 15 10 5 y3 0 −5 −10 −15 4 2 0 −2 y2 −4 −10 −5 y1 5 0 10 (b) ERROR SYSTEM 8 6 4 2 e=x−y 0 −2 −4 −6 −8 0 5 10 15 20 t 25 30 35 40 (c) Fig.5. Feng MASTER SYSTEM 4 2 x3 0 −2 −4 0. 1. Huang and G.55) when τ = 0. .4 0. Simulation results for master-slave synchronization of Chua’s Circuits (1.3.30 H.2 x2 −0.

2138  .0e + 003 ×  −1.6369 N2 =  743.0e + 003 ×  −0.6136 −200.5358 Λ = 487.1355 −0.9523 R =  −170.9180 22.1636 2.0928     −485.8843 −243.1333   0. we consider the above Chua’s Circuit (1.2893 N1 =  −750.4069.2618 −0.1411 Corollary 1.55) with w(t) = 0.4339  .5308 26. from which one can clearly see that Corollary 1.8315 −0.9523 −89.6266 82. G =  63.5678 −10.039 Han [31] 0.0386 −681.3774 0.3 (Delay-dependent) 0. to show that Corollary 1.1735   5. [27] 0. Methods Maximum allowed time delay τmax Corollary 1.2536  . the maximum allowed time delay is τmax = 0.0504   132.2.2138 0.2 gives the comparisons of the maximum allowed time delay τmax .3 is less conservative than the results in [27.1636 0.2322 25.8469 −190.1551 −0.1355 P = 1. 31].2504 82.14537 31 On the other hand.6923 −170.3 is less conservative than the results in [27.4716 173. . M = P −1 G =  0.8114 −622.0424   204. 31].0021 Table 1.3 via Matlab LMI Control Toolbox.7985  .2343   359.14537 with the following feasible solution:   0.1729 12.4606 −9. 22.7996 −0.5135 0.4782 −1. Comparison of the maximum value τmax of the allowed time delay.3774  .2663  . −6.9015  . By solving the LMI (1.1551 Q = 1.4066 62.1729  .Synchronization of Chaotic Systems with Delayed Feedback Control Table 1.9180 667.6112 −89.2172 N3 =  556. 100.1 (Delay-independent) infeasible Yalcin et al. 0. −74.54) in Corollary 1. −98.5135 1.3601 258. 0.

Brown and L.32 H. L. Delayindependent and delay-dependent synchronization criteria have been derived such that the controlled slave system can robustly synchronize the noise-perturbed master system with guaranteed H∞ performance. J. 64. Chua. Conclusion The master-slave synchronization problem has been studied for a class of chaotic systems based on the time-delayed feedback control. (World Scientific. Lett. Y. 1998). J. . F. in Circuits and Systems: Tutorials IEEE-ISCAS’94. Pecora and T. and O. 821-824. Phys. Rev. [4] P. Dong. Zhang. [2] R.-II 53. Bifurcation and Chaos 7. [9] M. Phys. Absolute stability theory and the synchronization problem. 28912896. Chua. (2000). [7] M. Simple examples of partial synchronization of chaotic systems. It is shown that the time-delayed feedback controller can be determined by solving some LMIs. Vidyasagar. M. A unifying definition of synchronization for dynamical systems. (1997).6. A. Synchronization in chaotic systems. 344-349. [8] M. Curran and L. Chaos 10. (1994). (1990). (1998). (1997). and Applications. and L. Acknowledgement This work is partially supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region of China under Project CityU-113205. NJ. References [1] L. An example with simulation results is provided to demonstrate the performance of the proposed approaches and the advantages of the delay-dependent synchronization condition with slack variables over the delay-independent condition and the delay-dependent condition without slack variables. 1375-1382. Nonlinear Systems Analysis. [5] P. Int. Output regulation of discrete-time piecewise-linear systems with applications to controlling chaos. pp. From Chaos to Order-Perspectives. Int. Feng 1. E 58. 1993). Singapore. Kocarev. [3] G. Suykens. Maistrenko. [6] G. Carroll. Englewood Cliffs. Rev. Feng and T. (2006). Absolute stability theory and master-slave synchronization. Hasler. O. Methodologies. Popovich. Synchronization principles and applications. O. K. F. (Prentice-Hall. 314-326. Huang and G. Circuits Syst. J. Curran. 6843-6843. Chen and X. 249-253. Hasler. Bifurcation and Chaos 7. IEEE Trans.

Suykens and J. Vandewalle. (2000). M. F. A unified framework for synchronization and control of dynamical systems. Bifurcation and Chaos 7. Vandewalle. Master-slave synchronization of Lur’e systems. Suykens. Suykens. 1285-1298. F. Kharitonov. (1999). J. Chua. Nonlinear H∞ synchronization of Lur’e systems: dynamic output feedback case. Circuits Syst. V. and L. K. Hyperchaotic attractors of unidirectionallycoupled Chua’s circuit. Chua. Chua.. 125. 158165. O. H. Int. K. Ho. [14] J. Master-slave synchroc . K. Academic Press. Bifurcation and Chaos 4. (1997a). Experimental confirmac tion of 3. A. I. and J. IEEE J. Dyn. R. 1323-1335. [22] K. Meas. H. Liu. Gu. M. Suykens. [13] C. Fract 23. Wu and L.Synchronization of Chaotic Systems with Delayed Feedback Control 33 [10] R.and 5-scroll attractors from a generalized Chua’s Circuit. -I 44.-X. Vandewalle. Chen. Bifurcation and Chaos 13. (1994). -I 44. Int. O. W. Yang. A. and L. Hale and S. and J. P. Zhong. Quant. E. Yal¸in. (2003). 979-998. A. [25] H. New York. Survey on recent results in the stability and control of time-delay systems. (1997). T. and L. Bifurcation and Chaos 16. A. A. 1089-1092. [12] T. Chaos synchronization of general Lur’e systems via time-delay feedback control. 1993). Vandewalle. [21] V. Introduction to Functional Differential Equations. (2003). [16] J. Robust nonlinear H∞ synchronization of chaotic Lur’e systems. (1997c). J. J. -I 47. 477-482. 1986). [11] M. Yal¸in. [18] J. Madan. [20] J. Singapore. (World Scientific. 1993). Vandewalle. Chua’s Circuit: A Paradigm for Chaos. (2006).-X. (Springer-Verlag. Nonlinear H∞ synchronization of chaotic Lur’e systems. V. and J. Master-slave synchronization of general Lur’e systems with time-varying delay and parameter uncertainty. Syst. IEEE Trans. Curran. [15] J. Chua. J. Open-loop chaotic synchronization of injectionlocked semiconductor lasers with Gigahertz range modulation. IEEE Trans. O. K. New York. Curran. Stability of Functional Differential Equations. Gu and S. Nosov. Chaos Solit. Niculescu. (2000). C. Liao and G. (1997b). Chua. Li. Robust synthesis for masterslave synchronization of Lur’e systems. K. E. K. Eds. O. N. W. Bifurcation and Chaos 7. Chua. Stability of Time-Delay Systems. Vandewalle. P. [27] M. -I 46. IEEE Trans. Li. 841-850. 425-429. 665-669. [19] K. ASME J. O. IEEE Trans. Cao. Suykens. K. Int. and J. Circuits Syst. Electron. O. 207-213. Int. (1994). J. J. A. J. Lunel. Circuits Syst. J. Chen and J. 27-34. Bifurcation and Chaos 4. 891-904. (2005). B. 2003). Int. Boston. Cont. P. J. [26] X.. Suykens. and D. [17] J. F. K. [24] J. Circuits Syst. Kolmanovskii and V. (Mathematics in Science and Engineering 180. [23] H. Suykens. 281-294. Kapitaniak and L. (Birkhauser. F. J. Chen. Huang. J. Curran. Int. A. Synchronization criteria of Lur’e systems with time-delay feedback control. 36. and L.

IEEE Trans. (2005). Han. B 31. 18-24. G. C. Phys. Chiang. Feng [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] nization of Lur’e systems with time-delay.34 H. Boyd. Rev. J. (1997). Hong. (2007a). Lett. K. N. Feron. 191-197. IEEE Trans. (1986). . Inc. Yang. 1995). Liu. T. art. Yang. Q. Agiza and M. Laub. EI Ghaoui. o 1565-1577. Feng. H∞ output feedback control of discrete-time fuzzy systems with application to chaos control. (2002). J.-Y. T. H. M. (2001). 73. Switching manifold approach to chaos synchronization. Gaiza. Heagy. Chaos 15. Phys. M. 2523-2526.-M. and G.-L. Linear Matrix Inequalities in System and Control Theory.-P. L. (2005).-L. T. (1999). 563-569. Fuzzy Syst.-S. A 360. C. Hegazi.-Q. no. Guan. Y. E 59. Syst. -I 54. Circuits Syst. LMI Control Toolbox: for Use With Matlab. 1707-1722. T. Rev. J. A 278. Phys. 13. Chua. On designing time-varying delay feedback controllers for masterslave synchronization of Lur’e systems. Experimental and numerical evidence for riddled basins in coupled chaotic systems. Carroll. Yassen. E. 66-83. Huang and G. Lian. and P. S. 3528-3531. J. Bifurcation and Chaos 12. Cao. Chen. Chen. Fang. New delay-dependent synchronization criteria for Lur’e systems using time delay feedback control. 531-543. J. Man Cybern. and V. Impulsive control of Lorenz system. Komuro. Sun. -I 33. Yang. Balakrishnan. (2007b). Chiu. J. Lett. EL-Dessoky. D. Physica D 110. and L. L. Adaptive complete synchronization of two identical or different chaotic (hyperchaotic) systems with fully unknown parameters. O. Lett. Adaptive synchronization for R¨ssler and Chua’s circuit systems. IEEE Trans. Han. A. Bifurcation and Chaos 11. Lu and J.-L. N. Gahinet. Phys. (The Math Works. (1994). and M. The double scroll family. F. 043901.-B. Circuits Syst. IEEE Trans. 1072-1118. Synchronization of Rossler and Chen chaotic dynamical systems using active control. S. and C. Q. Synthesis of fuzzy modelbased designs to synchronization and secure communications for chaotic systems. P. and T. 1573-1583. A. (SIAM. Nemirovsky. L. L. Pecora. (2001).-S. PA. Int. Philadelphia. 1994). and M. Int. Chilali. Matsumoto. N. and X. A. H. (2001).

Stability conditions in terms of linear matrix inequalities are derived based on the Lyapunov stability theory. Kowloon. A switching controller is then proposed to deal with the synchronization problem of which the system states of the response chaotic system is driven to approach those of the drive chaotic system. Chaos stabilization and synchronization are 35 . Leung Centre for Multimedia Signal Processing.edu.K. The Hong Kong Polytechnic University.uk F. a fuzzy model is first employed to represent the chaotic systems. Department of Electronic and Information Engineering.H. Introduction Chaotic control is a challenging task due to the complex characteristic of the chaotic systems. Division of Engineering.hk This chapter investigates the synchronization of chaotic systems subject to parameter uncertainties.1. Simulation examples are given to show the effectiveness of the proposed approach.Chapter 2 SYNCHRONIZATION OF UNCERTAIN CHAOTIC SYSTEMS BASED ON FUZZY-MODEL-BASED APPROACH H.ac.lam@kcl. Hung Hom. Based on the fuzzy-model-based approach. Lam Department of Electronic Engineering. The tracking performance and parameter design of the proposed switching controller are formulated as a generalized eignevalue minimization problem which can be solved numerically using some convex programming techniques. 2. United Kingdom hak-keung.F. Hong Kong enfrank@polyu. WC2R 2LS. The King’s College London.

Various control approaches such as nonlinear controllers [1. Lam and F.H. This particular structure is in favour of the system analysis and controller synthesis. the synchronization of chaotic systems subject to parameter uncertainties is handled based on the fuzzy-model-based approach. In general. were reported to realize the synchronization of identical or different chaos systems. a switching controller is proposed. Comparing to the chaos stabilization problem. The fuzzy model gives a fixed framework to model some general chaotic systems. Consequently. the complex adaptive laws will always increase the computational demand and structural complexity of the fuzzy controllers.36 H. Furthermore.F. 9]. To deal with the parameter uncertainties and achieve synchronization. A switching law is then derived to guarantee the convergence of the error states between the response and the drive chaotic systems based on the Lyapunov stability theory. The conditions for system stability and tracking performance are . However. 9] of which the controller parameters were updated in an online manner according to some adaptive laws to cope with the changes of the operating conditions. Adaptability was granted to controllers [8. In such a case. a fuzzy model is derived to represent the dynamics of the chaotic system as a weighted sum of some linear sub-systems. However. Leung important topics which has drawn a great deal of attention from researchers. the synchronization problem will become more complex if the chaotic systems are subject to parameter uncertainties. A fuzzy model will first be employed to represent the chaotic system. chaos synchronization requires driving the system states of the response chaotic system to follow those of the drive chaotic system. To stabilize the chaos behaviour. If the fuzzy-model-based control approach [57] is adopted. the tracking control problem is considered. In this chapter. fuzzy controllers [5-7] and adaptive controllers [8. which is typically designed with the pole-placement or nonlinearity compensation techniques. the control objective is only to suppress the chaotic dynamics and drive the system states to the equilibrium. chaos synchronization is much harder to be achieved. 2]. active controllers [3. synchronization is achieved by modifying the chaotic behaviour of the response chaotic system with the controller [14].K. the adaptive controller offered an outstanding robustness property. 4].

This chapter is organized as follows. which can be solved readily using some convex programming technique.. 2.2. The i-th rule is of the following format. In section 2.5. 2. i = 1. . Ψ is a positive integer. In section 2. a fuzzy model [5] is employed to represent the chaotic system subject to parameter uncertainties. p. the stability conditions and the switching laws are derived based on the Lyapunov approach. α = 1.1.1) i where Mα is a fuzzy term of rule i corresponding to the function fα (x(t )) with known form. A conclusion is drawn in section 2. Based on the fuzzy model.. A i ∈ ℜ n×n and B ∈ ℜ n× n are known constant system and input matrices respectively. Stability conditions in terms of linear matrix inequalities (LMIs) are derived to guarantee the system stability based on the Lyapunov stability theory. The switching activity of the controller is able compensate the uncertain nonlinearity of chaotic systems. . x(t ) ∈ ℜn×1 is . the fuzzy model representing the dynamics of the chaotic system subject to parameter uncertainties is presented. 2..3. Fuzzy Model and Switching Controller In this section. i i Rule i: IF f1 (x(t )) is M1 AND … AND fΨ (x(t )) is MΨ THEN x(t ) = A i x(t ) + B u(t ) (2. Ψ. numerical examples on synchronizing two different chaotic systems are given to demonstrate the effectiveness of the proposed approach. A switching controller is proposed based on the fuzzy model of the chaotic system.2. It is required that B is invertible.2. a switching controller is then proposed to realize the synchronization of two chaotic systems. Fuzzy Model Let p be the number of fuzzy rules describing the uncertain response chaotic system with control input term...4.Synchronization of Uncertain Chaotic Systems 37 formulated as a generalized eigenvalue minimization problem (GEVP). 2. In section 2..

2. ˆ ˆ ˆ ˆ ˆ x(t ) = ∑ w j (x(t ))A j x(t ) j =1 q (2. µ M i ( xα (t )) . Lam and F. The proposed switching controller is defined as follows. j =1 q 2. w j (x(t )) ∈ [0 1] ˆ ˆ membership and ∑ w j (x(t )) = 1 . Consequently.2) to follow those of the drive chaotic system of (2. Switching Controller A switching controller is proposed to drive the system states of the response chaotic system of (2. ˆ A j ∈ ℜ n × n is a known is the grade of the ˆ ˆ constant system matrix.2) where i=1 ∑ wi (x(t )) = 1 . The system dynamics are described by.5).5) ˆ where x(t ) ∈ ℜn×1 is the system state vector.K.38 H.F.4) is a nonlinear function of x(t) and µ M i ( xα (t )) is the grade of α i membership corresponding to the fuzzy terms M α . are uncertain in value as the α chaotic system is subject to parameter uncertainties.2. Similarly.1) but without the control input.3) wi (x(t )) = µ M i ( f1 (x(t ))) × µ M i ( f 2 (x(t ))) × × µ Mi ( fΨ (x(t ))) 1 2 Ψ ∑ (µ k =1 p k M1 ( f1 (x(t ))) × µ M k ( f 2 (x(t ))) × 2 × µ M k ( fΨ (x(t )) Ψ ) (2. Leung the system state vector and u(t ) ∈ ℜn×1 is the input vector. the dynamics of the drive chaotic system can be represented as. x(t ) = ∑ wi (x(t ))(A i x(t ) + Bu(t ) ) i =1 p (2. wi (x(t )) ∈ [0 1] for all i p (2. the dynamics of the drive chaotic system can be represented by a fuzzy model with q fuzzy rules in the format of (2.H. It should be noted that the grades of membership. .

we have. x(t ) ) = ∑ wi ( x(t ))A i x(t ) − ∑ w j (x(t ))A j x(t ) . It should be i =1 j =1 ˆ ˆ ˆ noted that wi(x(t)). From (2.Synchronization of Uncertain Chaotic Systems 39 u(t ) = ∑ mi (e(t ))B −1G i e(t ) i =1 p (2. The scalar function of mi(e(t)) is a switching function which takes either 0 or 1 to compensate the uncertainties in the membership functions of the fuzzy models. x(t )) i =1 p q p (2. p ˆ e(t ) = x(t ) − x(t ) ˆ ˆ ˆ ˆ = ∑ wi (x(t )) A i x(t ) + Bu(t ) − ∑ w j (x(t ))A j x(t ) i =1 p j =1 p q ˆ ˆ = ∑ wi (x(t ))A i x(t ) − ∑ wi (x(t ))A i x(t ) + ∑ wi (x(t )) A i x(t ) i =1 i =1 i =1 p p ˆ ˆ ˆ ˆ + Bu(t ) − ∑ w j ( x(t ))A j x(t ) j =1 q ˆ ˆ ˆ ˆ ˆ = ∑ wi (x(t )) Ai e(t ) + Bu(t ) + ∑ wi (x(t )) Ai x(t ) − ∑ w j (x(t ))A j x(t ) i =1 i =1 j =1 p p q ˆ = ∑ wi ( x(t )) A ie(t ) + Bu(t ) + m e (x(t ).5).6) and (2. Error System The error system is formed by the response and drive chaotic systems of (2. ˆ e(t ) = ∑ wi (x(t ))A i e(t ) + B ∑ mi (e(t ))B −1G i e(t ) + m e (x(t ).2) and (2.7) ˆ ˆ ˆ ˆ ˆ ˆ where m e (x(t ). w j (x(t )) and x(t ) are bounded due to the nature of the membership functions and chaotic drive system. with the property that i =1 ∑ wi ( x(t )) = 1 .5).2. From (2. we have.2) and (2.6) ˆ where e(t ) = x(t ) − x(t ) is the error system state vector. x(t ) ) i =1 i =1 p p . 2.3. G i ∈ ℜ n×n is the feedback gain to be designed and mi(e(t)) is a scalar function to be determined later.7).

ˆ ˆ ˆ ˆ w j (x(t )) . wi(x(t)).9) where P ∈ ℜn×n is a symmetric positive definite matrix. In the following. x(t )) are denoted as wi. To investigate the stability of the error system of (2.8) ˆ + ∑ (mi (e(t )) − wi (x(t )))G ie(t ) + me (x(t ). Leung p p = ∑ wi (x(t )) A ie(t ) + ∑ mi (e(t ))G ie(t ) + ∑ wi (x(t ))G i e(t ) i =1 p i =1 i =1 ˆ − ∑ wi (x(t ))G i e(t ) + m e (x(t ). Lam and F. From (2.9).3. we have. x(t ) ) i =1 2.F. the following quadratic Lyapunov function candidate is considered. the stability of the error system of (2. w j . mi and me respectively for brevity.8).K. Stability Analysis In this section. V (t ) = e(t )T Pe(t ) + e(t )T Pe(t ) p p =  ∑ wi (A i + G i )e(t ) + ∑ (mi − wi )G i e(t ) + m e  Pe(t )   i =1  i =1  p p + e(t )T P ∑ wi (A i + G i )e(t ) + ∑ (mi − wi )G i e(t ) + m e    i =1  i =1  T = ∑ wi e(t )T (A i + G i ) P + P(A i + G i ) e(t ) T i =1 p ( ) + 2 ∑ (mi − wi )e(t ) PG i e(t ) + m e Pe(t ) + e(t ) Pm e T T T i =1 p (2. It is going to show that V (t ) ≤ 0 (equality holds when e(t) = 0) which implies the synchronization of both drive and response chaotic systems are achieved.8) is investigated based on the Lyapunov stability theory. V (t ) = e(t ) T Pe(t ) (2.10) Let the switching law be . mi (e(t )) and m e ( x(t ).40 p H. x(t ) ) i =1 = ∑ wi ( x(t ))(A i + G i )e(t ) i =1 p p (2.8) and (2.H.

p (2. Consequently. 2 2 1 1 ≤ which implies 2 2 V (t ) ≤ ∑ wie(t )T (A i + G i ) P + P(A i + G i ) e(t ) T i =1 p ( ) + me Pe(t ) + e(t )T Pme + e(t )T Pe(t ) − e(t ) T Pe(t ) + ηme Pm e − ηm e Pm e T T T .10) and (2. we have. we have wi − − 1 1 + wi − ≤ 0 .Synchronization of Uncertain Chaotic Systems 41 mi = 1 1 − sgn e(t )T PG ie(t ) 2 ( ( )) (2.3).12) V (t ) = ∑ wi e(t ) T (A i + G i ) P + P(A i + G i ) e(t ) T i =1 p 1  + 2∑  1 − sgn e(t ) T PG i e(t ) − wi e(t ) T PG i e(t ) i =1  2  ( ) ( ( )) + m e Pe(t ) + e(t ) T Pm e = ∑ wi e(t )T (A i + G i ) P + P(A i + G i ) e(t ) T i =1 p  1 1   + 2 ∑  − sgn e(t )T PG i e(t ) −  wi −  e(t )T PG ie(t )  2 i =1  2    p T ( ) ( ) + m e Pe(t ) + e(t ) T Pm e ≤ ∑ wi e(t )T (A i + G i ) P + P (A i + G i ) e(t ) T i =1 p T ( )  1 1 T + 2 ∑  − + wi −  e(t )T PG ie(t ) + m e Pe(t ) + e(t )T Pme  2  i =1  2 p (2.13) From the property of (2.11).11) where the sign function is defined as follows. we have. 1 z > 0 sgn ( z ) =  − 1 otherwize From (2.

Denote X = P −1 and G i = N i X −1 where Ni ∈ ℜn×n is an arbitrary matrix. MATLAB LMI toolbox.9) that e(t) will be increasing.K. e. we have. p (2. In the following.16) that the tracking performance is partially governed by the value of η. Hence. e(t) will be decreasing. 2. can provide a better tracking performance with a smaller tracking error of e(t).14) and (2.H. A small value of η is required to attenuate the effect of me to the tracking performance. pre-multiply and post-multiply diag{X. it can be seen from (2. T  X 0   A i + N i X −1 X −1 + X −1 A i + N i X −1 + X −1 X −1    0 X  X −1 − ηX −1      X 0  × <0  0 X ( ) ( ) .15). which gives smaller value of ηm e Pm e .16) T that V (t ) > 0 happens when the value of ηm e Pm e is sufficiently large..42 T H. Lam and F. It can be seen from (2. ..g..15) P − ηP   From (2. X} to the stability condition of (2. it can be seen that a T smaller value of η. i = 1. T V (t ) ≤ −e(t )T Pe(t ) + ηm e Pm e T (2. Leung e(t ) (A i + G i )T P + P (A i + G i ) + P P  e(t ) = ∑ wi      i =1 P − ηP   m e  (2. the problem of minimizing the value of η is formulated as a generalized eignevalue problem (GEVP). When e(t) increases to a sufficient large value which makes V (t ) < 0 happen again.16) It can be seen from (2.14)  me   p − e(t )T Pe(t ) + ηm e Pme Let (A i + G i )T P + P (A i + G i ) + P P    < 0 . we have.15). which can be solved numerically using some convex programming techniques. Under this case.F.

(2. 2. . It should be noted that the complexity of the proposed switching controller will increase . a saturation function [10] is employed to replace the sign function in the switching controller to produce smooth control signal near the origin. Hence. T 2 ( ( )) (2. To attenuate the chattering effect. z  1 for T ≥ 1  z  sat( z .  XAi T + A i X + N i T + N i − X X    < 0 ..11) is a switching function. The error system of (2. X > 0.17) Theorem 2.1.Synchronization of Uncertain Chaotic Systems 43  XAi T + A i X + N i T + N i + X X  ⇒ <0 X − ηX   The analysis results are summarized in the following Theorem. X − ηX   As the sign function in (2. ( ( )) minimize η subject to X..18) where the saturation function is defined as follows. it will introduce an undesired chattering effect to the system states and control signals.N i η > 0. mi = 1 1 − sat e(t )T PG i e(t ). we have. i = 1. p.8) is guaranteed to be input-to1 state stable if mi (e(t )) = 1 − sgn e(t )T PG ie(t ) and there exists a scalar 2 η. T ) = − 1 for ≤ −1 T  z  T otherwise  and T is a non-zero positive scalar to be determined. a symmetric matrix X and an arbitrary matrix Ni such that the following GEVP is satisfied..

K. Lam and F.1. Leung when the number of fuzzy rules of the response and drive chaotic systems is large.F. The exact fuzzy model of the response Rössler system [5] is represented by the following fuzzy rules. Simulation Examples Two simulation examples on dealing with the synchronization of chaotic systems are given in this section.4.1.5. It is assumed that x1 (t ) ∈ [cmin − d cmax + d ] and d ∈ ℜ+ = 25. 2.4.H. the Rössler system is taken as the response system while the Lorenz system is taken as the drive system. Rössler and Lorenz Systems In this simulation example. 2.4. 2. The proposed switching controller is employed to synchronize both Rössler and Lorenz systems. 1 a  x(t ) =  2  A(x(t ) ) =  0 and where   x3 (t )  b 0 − (c(t ) − x1 (t ))     1 0 0  c +c c −c B = 0 1 0 . and c2 = 7. i = 1. c1 = 4.19) −1 0 − 1   x1 (t )   x (t ) . 2. a = 0.1.44 H. x(t ) = A(x(t ) )x(t ) + Bu (t ) (2. b = 0. The Dynamics of the Response Rössler’s System The dynamics of the response Rössler’s system with input term are described as follows. Both chaotic systems are subject to uncertain parameters in this example. Rule i: IF x1(t) is Mi THEN x(t ) = A i x(t ) + Bu(t ) .20) . The inferred response Rössler system is defined as (2.7.4. c(t ) = 2 1 + 2 1 sin(t ) ∈ [cmin   2 2 0 0 1    cmax] = [c1 c2] > 0 is an uncertain value.34.

i = 1. The Dynamics of the Lorenz System The dynamics of the Lorenz system are given as follows. ˆ ˆ ˆ ˆ x(t ) = A (x(t ) )x(t ) ˆ ˆ  x1 (t )  − a  x (t ) and A(x(t ) ) =  c ˆ ˆ ˆ ˆ where x(t ) =  ˆ2    x3 (t )   0 ˆ   ˆ a 0  ˆ − 1 − x1 (t ) . 2. It can be seen that the grades of membership are unknown as the values of c(t) is uncertain.24) ˆ ˆ ˆ ˆ 0  0  − a a − a a  c −1 − d   c ˆ ˆ ˆ ˆ A1 =  ˆ and A2 =  ˆ −1 d  . b = .   b 0 − d  b 0 d      1  c(t ) − x1 (t )  w1 (x(t )) = µ M1 ( x1 (t )) = 1 + and  2 d  w2 ( x(t )) = µ M 2 ( x1 (t )) = 1 − µ M1 ( x1 (t )) . ˆ ˆ ˆ ˆ Rule i: IF x1 (t ) is Ni THEN x(t ) = A i x(t ) .2.23) (2. The inferred drive Lorenz system is defined as ˆ ˆ ˆ ˆ ˆ x(t ) = ∑ wi (x(t ))A i x(t ) i =1 2 [ ˆ d ] (2.1.21) 0 − 1 − 1  0 − 1 − 1 1 a  where A1 =  0  and A 2 = 1 a 0 .22) It was ˆ ˆ reported in [5] that the Lorenz system of (2. a = 10 . where   ˆ − b ˆ ˆ ˆ  0  0 − d − b d     ˆ 1  x (t )  ˆ 8 ˆ ˆ ˆ ˆ w1 (x(t )) = 1 + 1  and w2 ( x(t )) = 1 − w1 (x(t )) .Synchronization of Uncertain Chaotic Systems 45 x(t ) = ∑ wi (x(t ))(A i x(t ) + Bu (t ) ) i =1 2 (2.4. ˆ 2 3 d  .22) with x1 (t ) ∈ − d can be represented by the fuzzy model with the following rules [5].  ˆ ˆ x1 (t ) −b   (2. 2.

6460 − 0.0902 − 2010.1122 X= 20019. The Dynamics of the Proposed Switching Controller The proposed switching controller of (2. 2. Lam and F.1274 − 0.3418 − 0.5215×10−4 such that the stability conditions in Theorem 2.0506    0. u(t ) = ∑ mi (e(t ))B −1G i e(t ) i =1 2 (2.9381 − 0.K.0904  .1303  − 2037. Leung ˆ ˆ c = 28 and d = 25 .9385  − 2037. The .   1.6) is employed to handle the synchronization problem. we have − 0. Fig.1460 1.0506 − 2049.4778   0.1460 G1 =  − 2033.46 H. 2.1497 G2 =  − 2033.0917 20000.3414 − 0.1122 1.1919 − 0.   1.1 are satisfied.3489  − 0.1919  20023.4428  0.4.1497 0. The values of the parameters of the drive system are assumed to be unknown in this example.3.1 shows the system state responses of the drive chaotic systems T ˆ under the initial condition of x(0) = [− 1 − 1 − 1] .0775   and η = 2.H.F.3992   0. In this simulation. u(t) = 0 is employed for 0 ≤ t < 50s and the switching controller is applied for t ≥ 50s.0917  .4754  0. the saturation function with T = 10−5 is employed to replace the sign function in the switching controller. Fig. In order to alleviate the chattering effect.25) With the help of MATLAB LMI toolbox.2 shows the system state responses of the response system with the proposed switching controller subject to the initial state conditions of T x(0) = [− 1 − 1 − 1] . 2. It can be seen that the value of η is a small value to ensure a good tracking performance.1. The switching controller is defined as.

1. 2.Synchronization of Uncertain Chaotic Systems 20 x^1(t) 47 0 -20 50 x^2(t) 0 10 20 30 40 50 60 70 80 90 100 0 -50 60 40 x^3(t) 20 0 -20 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Fig. System state responses of the drive Lorenz system. System state responses of the response Rössler system with u(t) = 0 for 0 ≤ t < 50s and the switching controller applied for t ≥ 50s.2. 2. 20 x1(t) 0 -20 50 x2(t) 0 10 20 30 40 50 60 70 80 90 100 0 -50 60 40 x3(t) 20 0 -20 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Fig. .

which is applied for t ≥ 50s.48 50 e1(t) H. The Dynamics of the Response Chua’s System The dynamics of the response Chua’s system with input term are described as follows. 2. 2. 2.1. The tracking error of the response Rössler to the drive Lorenz chaotic systems.4. the Chua’s system is taken as the response system while the Lorenz system in the previous example is taken as the drive system.2. it can be seen that the proposed switching controller.26) .4. Referring to these figures. Lam and F. x(t ) = A (x(t ) )x(t ) + Bu (t ) + E (2. Leung 0 -50 50 e2(t) 0 10 20 30 40 50 60 70 80 90 100 0 -50 60 40 e3(t) 20 0 -20 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 Time (sec) 70 80 90 100 Fig.3.F. 2. tracking error is shown in Fig.H. is able to drive the system states of the uncertain Rössler system to follow those of the uncertain Lorenz system with a sufficiently small tracking error. Chua’s and Lorenz Systems In this example.3.2.K.

Synchronization of Uncertain Chaotic Systems 49  x1 (t )  − a a  x (t ) . and where  c( x1 (t ) ) for x1 (t ) > 0  φ (x1 (t ) ) =  x1 (t ) and d = 19 ≥ max φ ( x1 (t ) ) . 1 (g a (t ) − gb )( x1 (t ) + 1 − x1 (t ) − 1 ) .3428 . It is assumed that x1 (t ) ∈ [− 25 25] .27) The inferred response Chua’s system is defined as x(t ) = ∑ wi ( x(t ))(A i x(t ) + Bu (t ) ) i =1 2 (2.5285 and gb = 0. 2.  0  b 1 0 0 B = 0 1 0    0 0 1    = 14. i = 1. 2 g a + g a 2 g a1 − g a 2 g a (t ) = 1 + sin(t ) is an uncertain value. (2.    0    c(x1 (t ) ) = g b x1 (t ) + 0 1 . g a1 = −1.2850. The exact fuzzy model of the response Chua’s system7 is represented by the following fuzzy rules. 2 2 g a 2 = −1. A(x(t ) ) =  1 − 1 where x(t ) =  2    x3 (t )   0 −b    c( x1 (t ) ) E= 0  . .28) (d − 1)a a 0 where A1 =  1 − 1 1 and    0 − b 0   1  φ (x1 (t ) )  w1 (x(t )) = µ M1 ( x1 (t )) = 1 −  2 d  w2 ( x(t )) = µ M 2 ( x1 (t )) = 1 − µ M1 ( x1 (t ))  − (d − 1)a a 0 A2 =  1 − 1 1    0 − b 0   . It can be x1 ( t ) g otherwise  a seen that the grades of membership are unknown as the value of c(x1(t)) is uncertain. Rule i: IF x1(t) is Mi THEN x(t ) = A i x(t ) + Bu(t ) .7143. a = 9.

The Dynamics of the Switching Controller The switching controller in the same form of (2.8293   − 4. The dynamics and the fuzzy model of the Lorenz system are given in the previous example. 2.   − 0.5870×10−4 such that the stability conditions in Theorem 2.4415  . 2. 2.3324    − 0. In this simulation.0880  − 1984. Fig. It can be seen from these figures that the proposed switching controller applied for .5041   − 4. Leung 2. The tracking error is shown in Fig.6171 − 0.0630  − 2127.3298 − 2051.6280  19708.7217 7.9457 − 0.1.1 are satisfied. i = 1.3297  . The Dynamics of the Lorenz System The same Lorenz system of the previous example is employed as the drive system.F. The system state responses of the drive Lorenz system is shown in Fig. With the MATLAB LMI toolbox.4415 19760.6280 3.8320  − 0.9546 G1 =  − 2050.3.2.4.K.2. will be employed to handle the synchronization problem.25) with the corresponding mi(e(t)). 2.2134 X= 3. 2. u(t) = 0 is employed for 0 ≤ t < 50s and the switching controller is applied for t ≥ 50s.9454 − 0.6323 G2 =  − 2050.5.50 H.4 shows the system state responses with the proposed switching controller using a saturation function with T = 10−5 subject to T the initial state condition of x(0) = [− 1 − 1 − 1] .7885  − 4. Lam and F.   − 0.8293   and η = 2.4.2.6775 7. The value of η is small enough to ensure a good tracking performance.2310  − 4.9457 19761.H.0610 7.3324 − 2051. we have − 0.0908 7.

The guarantee of the tracking performance and the parameter design of the switching controller have been formulated as a generalized eigenvalue minimization problem. Conclusion The synchronization of chaotic systems subject to parameter uncertainties has been investigated in this chapter. which can be solved numerically using some convex programming techniques. 20 x1(t) 0 -20 50 x2(t) 0 10 20 30 40 50 60 70 80 90 100 0 -50 60 40 x3(t) 20 0 -20 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Fig. 2. Stability conditions have been derived based on the Lyapunov approach to guarantee the system stability. . 2. System state responses of the response Chua’s system with u(t) = 0 for 0 ≤ t < 50s and the switching controller applied for t ≥ 50s. Simulation examples have been given to show the effectiveness of the proposed approach.4.5. a switching controller has been proposed to handle the synchronization problem.Synchronization of Uncertain Chaotic Systems 51 t ≥ 50s is able to drive the system states of the uncertain Chua’s system to follow those of the uncertain Lorenz system with a sufficiently small tracking error. Based on the fuzzy models of the response and drive chaotic systems.

Phys. Lett. 1041-1044. Wang. Lam and F. 1021-1040. Ikeda and H. H. LMI-based fuzzy stability and synchronization of Chen's system. Controlling chaotic systems via nonlinear feedback control. Sequential synchronization of two Lorenz systems using active control. Circuits and Systems – I: Fundamental and Applications 45(10).K. [6] Y. Wang. . Leung 0 -50 50 e2(t) 0 10 20 30 40 50 60 70 80 90 100 0 -50 60 40 e3(t) 20 0 -20 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 Time (sec) 70 80 90 100 Fig. Phys. IEEE Trans. [3] E. 191-197. [4] H. Wang. (1998). W. Synchronization of chaotic systems via nonlinear control. A unified approach to controlling chaos via an LMIbased fuzzy control system design. Chaos.F. Agiza and M. T. Guan and H. Yassen. Tanaka. 271-275. Solitons and Fractals 11(7). A 320(2-3). Solitons and Fractals 23(3). T.H. R. 154-159. Lonngren. N. (2005). 1-BB9E). (2003). 2. Huang. W.5. A unified approach to controlling chaos via an LMI-based fuzzy control system design. (2000). Chaos. Acknowledgement The work described in this chapter was supported by King’s College London and the Centre for Signal Processing. H. O. O. References [1] L. [2] J. Feng and M. the Hong Kong Polytechnic University (Project No. The tracking error of the response Chua’s to the drive Lorenz chaotic systems. [5] K. A 278(4). (2004). (2001). Z. A 320. Lett. Park. E. Bai and K. Lett.52 50 e1(t) H. Phys. 1049-1054. Wang.

G. 1013-1032. Y. [10] J. Q. [9] W. Slotine and W.Synchronization of Uncertain Chaotic Systems 53 [7] K. C. International Journal of Bifurcation and Chaos 12(8). Chaos. Lin. T. J. E. Wu and W. Li. 1827-1841. Dong and D. (2002). 1991). . [8] M. (2005). Chaos. C. (2003). P. Liu. Observer-based robust adaptive variable universe fuzzy control for chaotic system. 883-890. Jiang. Feki. Bin. Lian. Solitons and Fractals 15(5). Solitons and Fractals 23(3). An adaptive feedback control of linearizable chaotic systems. Applied Nonlinear Control (Prentice Hall. Chaotic control using fuzzy modelbased methods.

This page intentionally left blank .

irregular.Chapter 3 SLIDING MODE CONTROL OF CHAOTIC SYSTEMS Y. to force it to converge to zeros or to track desired trajectories. such as chemical reactors. 55 . chaos control and chaos synchronization are discussed.edu. and predictable trajectories. Melbourne Vic 3001. observer based chaos synchronization is described.au This chapter introduces the sliding model control of chaotic systems. and random-like. and laser systems [1]. traditional sliding mode control. Yu† School of Electrical and Computer Engineering. displaying no true cyclic behaviors. Chaos has been found in many different physical systems. a time division multiplex based method is used. In order to realize chaos synchronization via one single transmission channel. Three kinds of sliding mode control methods.feng@rmit. forced oscillators. The synchronization for two kinds of chaotic systems with single nonlinearity and multinonlinearities are presented. feedback control devices.1. Australia * yong. deterministically random and hence ultimately unpredictable. and continually unstable. Introduction Chaotic systems are characterized by being extremely sensitive to initial conditions. Chaotic systems do not follow any simple. respectively. Feng* and X. In addition. The motion of chaotic systems is complex. such as an unstable limit cycle. The main two branches of chaos control.edu. terminal sliding mode control and nonsingular terminal sliding mode control are used to control a chaotic system to realize two different control objectives. regular. Royal Melbourne Institute of Technology. 3. fluid flow systems.yu@rmit.au † x.

Sliding Mode Control The basic principle of sliding mode control can be briefly summarized as follows: consider a 2nd-order uncertain nonlinear dynamical system  x1 = x2 . by using switching type of control strategies.2. and u is the scalar control . and neural network based control and sliding mode control. physicists and control engineers in the last decade. 3. Among these approaches. Feng and X. fuzzy logic control. feedback control. the sliding mode control can be used in chaos control and chaos synchronization. chaos control includes two branches of research: chaos control and chaos synchronization. sliding mode control is an important approach. Yu Generally speaking. which is well known for its robustness. f(x) and b(x)≠0 are smooth nonlinear functions of x. the closed-loop responses of the systems become totally insensitive to both the internal parameter uncertainties and the external disturbances. adaptive control. we will discuss the sliding mode control of chaotic systems. The latter is a special kind of chaos control.   x2 = f ( x ) + g ( x ) + b( x )u (3. In this chapter.1) where x=[x1. etc. Because chaos synchronization is essential in the application of secure communications. The sliding mode control is designed to drive and constrain the system states to lie within a neighborhood of the prescribed switching manifolds that exhibit the desired dynamics. Many methods and techniques for chaos control and chaos synchronization have been proposed. When in the sliding mode. parameter-dependent control.56 Y. x2]T is the system state vector. and g(x) represents the uncertainties and disturbances satisfying g ( x ) ≤ l g where lg>0. The former is to control the variables of a chaotic system to an unstable periodic orbit or an equilibrium point of the system [2]. such as open loop control. Therefore. this issue has received a lot of attentions from mathematicians. and/or to eliminate or weaken it when it is harmful. The objectives of chaos control are generally to enhance existing chaos or create chaos in a dynamical system when it is useful or beneficial. including chaos control and chaos synchronization. which is to force the trajectories of the slave chaotic system to track that of the master system starting from arbitrary initial conditions.

a commonly used control design is u = −b −1 ( x )( f ( x ) + β x2 + (l g + η ) sgn( s )) . (3. 2 dt (3.2) reaches 0 from s(0)≠0.1). a commonly used TSM control design is u = −b −1 ( x )( f ( x ) + β q q / p −1 x1 x 2 + (l g + η ) sgn( s )) p p>q s = x 2 + β x1q / p . a terminal sliding mode (TSM) controller was developed [19-21]. the sliding mode control (3.8) . (3. The sliding mode is described by the following first-order linear sliding variable s = x2 + β x1 . Assume that the initial condition of the sliding mode manifold (3.6) where β >0 is a design constant. In order to design the control law.3) where η > 0 is a constant. For system (3. which satisfy the following condition: .2) where β >0 is a design constant. integrating (3.1). and p and q are positive odd integers.Sliding Mode Control of Chaotic Systems 57 input. (3.5) can guarantee that the system is robust to its uncertainties. for system (3.4) which means from time t=0 to t= ts. a η–reachability condition guaranteeing an ideal sliding mode motion should be satisfied.1). Therefore.5) which ensures that the sliding mode occurs. (3. which is given by 1 d 2 s < −η s .7) . The TSM is described by the following first-order terminal sliding variable (3. Recently. the sliding mode manifold s (3.3) gives the time ts as follows: ts ≤ s (0) η . For system (3.2) is s(0). (3. which ensures that TSM occurs.

58 Y. TSM offers some superior properties such as fast finite time convergence. x2= 0).9). The NTSM model is described as follows: . and the case of x2≠0 while x1=0 may occur from time to time. the singularity may also occur even after the sliding mode s = 0 is reached since. 1<p/q<2. Furthermore. Compared with the linear hyperplane based sliding modes. which is able to avoid this problem completely. 13]. This underlines the importance of addressing the singularity problem in conventional TSM systems. x1 (t r ) = q/ p β ( p − q) x1 (3. However.e. This situation does not occur in the sliding mode because when s = 0.8) cannot guarantee a bounded control signal for the case of x2≠0 when x1= 0 before the system states reach the TSM s = 0 . the term x1q / p −1 x 2 is equivalent to x1( 2 q − p ) / p which is nonsingular. i. This controller is particularly useful for high precision control as it speeds up the rate of convergence near the equilibrium point.4) the system states will reach the sliding mode s=0 within the finite time tr ≤ s (0) / η . It can be seen in the TSM control (3. When the sliding mode s=0 is reached. The finite time ts that is taken to travel from x1(tr) ≠ 0 to x1(ts+tr)=0 is given by t s = − β −1 x 2 + β x1q / p = x1 + β x1q / p = 0 . the system states cannot be guaranteed to always remain in the sliding mode especially near the equilibrium point (x1= 0.9) where x1 = 0 is the terminal attractor of the system (3. x 2 = − β x1q / p . In order to overcome the singularity problem in the conventional TSM systems. hence as long as q<p<2q. ∫ x1 ( t r ) 0 dx1 p 1− q / p . in the TSM manifold (3. both the system states x1 and x2 converge to zero in finite time. due to computation errors and uncertain factors. according to equation (3. The TSM controller (3. Yu It is clear that if s(0)≠0. the system dynamics is determined by the following nonlinear differential equation: (3.9).8) that the second term containing x1q / p −1 x 2 may cause a singularity if x2≠0 when x1= 0. a nonsingular TSM (NTSM) was proposed by [12. The singularity problem may occur in the reaching phase when there is an insufficient control to ensure that x2≠0 while x1= 0.10) This means that. Feng and X. the TSM controller design methods still have a singularity problem.

14) .10). One can easily see that when s = 0. x3ref. p and q have been defined in (3.11) where β.  x = −bx + x x 3 1 2  3 (3.6).Sliding Mode Control of Chaotic Systems 59 s = x1 + 1 β x 2p q . x2ref. Theorem 3.1) with the NTSM (3.11). In this section. (3. 3. Chaos Control In order to demonstrate the principle of the sliding mode based chaos control. which are all smooth.13). For the system (3. η>0. the states x1 and x2 will converge to zero in finite time. the sliding mode control has been introduced.12) where 1<p/q<2. which provides the basis of chaos control and chaos synchronization to be described in the following sections. Furthermore.   x3 = −bx3 + x1 x2 + u (3. p (3.3.11) is equivalent to (3. the NTSM (3.13) are x1ref.11) the derivative of s along the system dynamics does not result in terms with negative (fractional) powers. Consider the Lorenz system described by the following equation:  x1 = σ ( x2 − x1 )   x2 = rx1 − x2 − x1 x3 . The control signal u is applied in the third equation in (3. This can be seen in the following theorem about the NTSM control. equation (3. Note that in using (3. then the NTSM manifold (3.1.6) so that the time taken to reach the equilibrium point x1=0 when in the sliding mode is the same as in (3.11) will be reached in finite time.13) Assume that the desired trajectories of the chaotic system (3.13) is rewritten as follows:  x1 = σ ( x2 − x1 )   x2 = rx1 − x2 − x1 x3 . that is. if the control is designed as u = −b −1 ( x )( f ( x ) + β q 2− p / q x2 + (l g + η ) sgn( s )) . a simple example is given first.

Yu The objective of chaos control is to force the variables of the chaotic system (3.14) satisfy the η–reachability condition (3.3) can be designed as follows: u = a1σ ( x2 − x1 ) − a2 ( rx1 − x2 − x1 x3 ) − bx3 − x1 x2 + a1 x1ref + a2 x2 ref + x3 ref − η sgn( s ) .n) is the hysteresis-series as shown in Fig. (3. which is particularly useful for high precision control as it speeds up the rate of convergence near the equilibrium point. 3. Feng and X.19) . the sliding mode control law (3.15) where a1 and a2 are design parameters.  2 a2 − 4a1 < 0 (3. n) = ∑ hysi ( x) .  x2 = − a1 x1 − a2 x2 + a1u  where u is the nonlinear state feedback given by (3. We now design a NTSM controller. which guarantee that the system on the sliding mode is stable. In order to guarantee the existence of the chaotic motion in system (3.1b.14).17).18) where H(x. n is the number of hystereses. which can be described as follows:  x1 = x2 .16) can realize the control of the chaotic system (3. i =1 n (3.16) Therefore.60 Y. 3. A sliding mode control law which makes the system (3.1a. A sliding mode manifold can be designed as follows: s = a1 ( x1 − x1ref ) + a2 ( x2 − x2 ref ) + ( x3 − x3ref ) .14) to track the desired trajectories.17) u = H ( x. the parameters should satisfy  a2 < 0 . (3. The sliding mode control in the above example is the conventional method which characteristic is that the convergence of the system states to the equilibrium points is usually asymptotical due to the asymptotical convergence of the linear switching manifolds that are commonly chosen. Consider a hysteresis-based PWL system constructed using a second-order linear system with a feedback of a hysteresis series as shown in Fig. Theorem 1 can be used for chaos control.

The trajectory of system (3.22). (a) (b) Fig.21) and (3.22) . if a linear sliding mode and a terminal sliding mode are chosen as equations (3.17) can be considered as being governed by piecing together trajectory parts in the subspaces. (b) A hysteresis-series (n = 4). (a) Second-order linear system with a feedback of hysteresisseries. (3.21) (3. For the chaotic system (3.17) is rewritten as follows:  x1 = x2 .17).20) If the objective of chaos control is to control the variables of a chaotic system to an equilibrium point of the system.20). l1 (t ) = γ 1s1 p1 q1 + s1 . and a control law is designed as equation (3. respectively. that is. 3. the variable states of system (3. In order to control chaotic system (3.Sliding Mode Control of Chaotic Systems 61 which gives rise to a pair of complex conjugate eigenvalues with positive real parts that result in useful oscillatory motions for generating chaos.2. equation (3.1. centered at the equilibrium points. the following theorem can be used: Theorem 3. a control signal v is applied in the system.30) can converge to zeros asymptotically: s1 (t ) = x1 + λ1 x2 .23).   x2 = − a1 x1 − a2 x2 + a1u + v (3.

According to [12] the state variables of the system . the states (x1. it can be obtained V1 < 0 . Since p1 and q1 are positive odd integers. it gets: V 1 (t ) = l1l1 p1 p1 q1 −1 1 q1 2 − p1 / q1 s1 ( s1 + s1 ) q1 γ1 p1 p 1 q1 2 − p1 / q1 = l1γ1 1 s1 p1 q1 −1 (λ1v2 + s1 ) q1 γ1 p1 p = −η1γ1 1 s1 p1 q1 −1 l1 ≤ 0 . l1 and s1 can converge to zero in finite time. , , 2) s1 = 0 s1 ≠ 0 .s1s1 .62 Y. two different cases are discussed as follows: 1) s1 ≠ 0 . v1 = ax1 − bx2 − au − and (3. λ1 0 γ1 p1 1 1 ∫ t (3. Therefore. and η 1>0 are design parameters. satisfying s1 p1 / q1 −1 >0.23c) where λ1>0. s1 ≠ 0 ) and will continue to cross the axis s1 = 0 in the phase plane 0. (3. Yu v = v1 + v2 . which satisfy the condition p > q . p1 and q1 are positive odd integers. γ 1>0. will not always stay on the points ( s1 = 0 .23b) v2 = − q1 2 − p1 / q1 s + η1 sgn(l1 )]dτ . This completes the proof. Then. The following Lyapunov function is considered: V1 (t ) = 1 2 l1 (t ) 2 Differentiating V1(t) with respect to time. Feng and X.23a) 1 λ1 x2 . Proof. q1 = l1γ1 when l1 ≠ 0 . x2) will converge to zero asymptotically.

Fig. It can be seen that chaotic system (3. The control signal is activated after 240 seconds. and η 2>0 are design parameters. The control signal is activated after 270 seconds.20) tracks the limit cycle. e = [e1.25). and ′ v2 = − ∫ q2 2− p2 / q2 s2 + η 2 sgn(l2 )]dτ . 3.25) (3.Sliding Mode Control of Chaotic Systems 63 If the objective of chaos control is to force the variables of the chaotic system (3. l2 (t ) = γ 2 s2 p2 q2 (3. The chaotic behavior of system (3.e2]T=xref − x.26a) (3.20).20) and its desired trajectories. Fig. for example. For the purpose of the simplicity. which satisfy the condition p2>q2.2 – Fig.2 under the case of uncontrolled. 3. p2 and q2 are positive odd integers. 3. . the errors between the variable states of system (3. 3. Some simulation results of chaotic system (3. the following theorem can be used: Theorem 3. and a control law is designed as equation (3. if a linear sliding mode and a terminal sliding mode are chosen as equations (3. The desired trajectory is a limit cycle.3d are simulation results using Theorem 2.4d are simulation results using Theorem 3. Proof. which are all smooth. can converge to zeros asymptotically: s2 (t ) = λ2 e1 + e1 .26).20) using Theorem 2 and Theorem 3 are shown in Fig.26c) where λ2>0..26b) + s2 .11. The proof of Theorem 3 is similar to that for Theorem 2. 3.3. assuming that the desired trajectories of the chaotic system (3.3a – Fig. n is chosen as 1 in chaotic system (3. For the chaotic system (3.4a – Fig.20) can be seen in Fig. 3. It can be seen that chaotic system (3.24) (3. γ 2>0.20) converges to zeros after 240 second.20).24) and (3.20) are x1ref and x2ref. ′ v1 = −λ2 ( x2 − x1ref ) + ax1 − bx2 − au + x2 ref . 3. 0 γ p 2 2 t (3. ′ ′ v = v1 + v2 .20) to track the desired trajectories.

5 0 0 -0. 3.5 (c) Fig.5 1 0. .5 -1 -1.5 1.5 2 2.5 1 x2 x1 0. Feng and X. Yu 2. (b) state x2. (c) phase plane.5 0 50 100 150 t(s) 200 250 300 -1.5 2 1.5 -0.5 0 50 100 150 t(s) 200 250 300 (a) 1.2.5 x1 1 1. Chaotic behavior in the case of uncontrolled (a) state x1.5 -1.5 (b) 1 0.5 -1 -1 -1.5 -1 -0.64 Y.5 0 0.5 x2 0 -0.

Chaotic behavior with control as in Theorem 2 after 240 seconds (a) state x1.5 -1 -1.5 -2 -2.5 2 1. .5 0 50 100 150 t(s) 200 250 300 -1.5 1 0 -0.5 2 2.5 0. (d) control signal.5 1.5 -1 -1 -1.5 x1 1 1. 3.Sliding Mode Control of Chaotic Systems 65 2. (b) state x2.5 (b) 0.5 0 50 100 150 t(s) 200 250 300 (a) 1.3.5 50 100 150 t(s) 200 250 300 (c) (d) Fig.5 -3 0 -0.5 0 -0.5 -1 x2 0 v -1.5 -1.5 1 0.5 0 0. (c) phase plane.5 0 -0.5 1 x2 x1 0.5 -1 -0.

5 -1. (c) phase plane.5 (b) 4 1 2 0. Feng and X.5 -1 -0.5 0. 17].5 x1 1 1. many schemes for chaos synchronization have been proposed [5-8]. Since then.5 -1. It usually appears in two chaotic systems consisting of a master system and a slave system. Chaotic behavior with control as in Theorem 2 after 240 seconds (a) state x1.4. 3.5 0 0 -0.5 0 0.5 -4 -1 -6 -1.5 -0. In the first case. which are of identical structure and parameters except for different initial conditions.5 1 x1 x2 0. (d) control signal.5 0 50 100 150 t(s) 200 250 300 0 50 100 150 t(s) 200 250 300 (a) 1.5 0 x2 0 -2 v -0.4. Generally speaking. (b) state x2.5 2 1.5 1 1. sufficient .5 -1 -1 -1. Yu 2.e. The chaos synchronization means that the trajectories of the slave system track that of the master system starting from arbitrary initial conditions [15.66 Y. The pioneering work was done by [7]. Chaos Synchronization Chaos synchronization is a special kind of chaos control.5 2 2.5 -8 0 50 100 150 t(s) 200 250 300 (c) (d) Fig. 3. two main classes of chaos synchronization schemes can be identified i. based on either the transverse stability or the observability.

(3. A. y ) . Another chaotic system reconstructed is similar to the master chaotic system. (3. Although there are many methods of chaos synchronization. . named slave chaotic system: x (t ) = Ax (t ) + bf ( x ) + D + u .e. B. Rössler system. ensuring synchronization of identical chaotic systems with unidirectional coupling.28). the chaos synchronization is merely a control of the nonlinear system. is the output of system (3. Lorenz system. In this section. can be expressed one of the following form consisting of linearities and nonlinearities in [10]: x (t ) = Ax (t ) + bf ( x ) + D . 3.27) and the system (3. and D are the same as (3. such as Chua′s circuit. Considering the latter approach. (3. the chaos synchronization is a more challenging task.28b) ~ where x ∈ R n . u ~ x can be adopted two cases: u = u( x .1. and so on.27). the reconstruction of the system state from measurements of an output variable. the problem is formulated as an observation problem.27b) where x∈Rn is the state vector. x ) or u = u( ~. f(x): Rn→Rm is a nonlinear vector function. control u is used to realize the synchronization of the system (3. Synchronization of Chaotic System Using Observer Many chaotic systems. i.4. In the former. A∈Rn×n. where y∈Rp and p≥m.28a) or x (t ) = Ax (t ) + Bf ( x ) + D + u .27). b.Sliding Mode Control of Chaotic Systems 67 conditions for the negativity of all transverse or conditional Lyapunov exponents are considered. we will use the sliding mode control concept to realize chaos synchronization. (3. f(x): Rn→R is a nonlinear scale function. n≥m.27a) or x (t ) = Ax (t ) + Bf ( x ) + D . b∈Rn×1. For the later case.B∈Rn×m. sliding mode based method is an important method of chaos synchronization because of its robustness and simplicity. D∈Rn×1 are constant real matrices.

B is of full rank matrix. A robust observer is proposed in [14] and used to realize the chaos synchronization. (3. Since the method of the chaos synchronization makes the systems robust. which avoids the strict conditions. For the nonlinear function f(x). 3.3. ξ(t. x) denotes the system uncertainties. which consists of the states of the chaotic system. Because the regions of the chaotic attractors are kept in a finite space. it has significant meaning for practical applications. The second is that synchronization of chaotic systems with multi-nonlinearities was not well done. B.27). This section addresses these problems. one important method is to regard the slave system as an observer of the master system [10]. a robust sliding mode observer is proposed as follows: x (t ) = Ax (t ) + Bf ( x ) + D + ξ (t . x ) . therefore. it holds f ( x ) ≤ r1 .30) where y(t)∈Rp is the output of the chaotic system (3. a new robust sliding mode observer is designed. 3. Implementation of Chaos Synchronization Based on the Walcott-Zak observer. Chaos Synchronization Using a Robust Sliding Mode Observer Consider a class of uncertain chaotic systems: . y ) .68 Y. But the existing methods have some disadvantages. x) = Bδ(t. C∈Rn to be determined later. For the master chaotic system (3. x ) ≤ r2 + α (t . D and f(x) are the same as that in (3. The first is that the chaotic systems used for synchronization are ideal. The synchronization signal is chosen as: y (t ) = Cx (t ) .4. δ(t . where r1 is a known positive constant. where m is a positive real constant. the states of the chaotic systems are bounded.4. where r2 is a known positive constant and α is a known function.2. Feng and X.29) where A. that is x ≤ m . Yu In order to realize chaos synchronization only using the output signal y(t).29). (3. so the design can be simplified.29) and p≥m.

33) and the control strategy v is designed as follows: v = vl + v n . (3. T Define e = [e1T e2 ] . The error system is obtained from (3.34b) And rewrite s(t) (3.Sliding Mode Control of Chaotic Systems 69 ˆ ˆ ˆ ˆ ˆ x = Ax (t ) + Bf ( x ) + D − G (Cx − y ) + Bv ( x. (3. e2 (t ) = A021e1 (t ) + A022 e2 (t ) Theorem 3.32) as the decomposed form: s = M1e1 + M 2 e2 . M=FC.31) of system (3. and v ( x. B and D are the same as ˆ that of (3. 18].35a) (3. F∈R . the chaos synchronization between systems (3. x )) + B1v . . M 1∈Rm.32) where A0=A − GC.34a) and .31) and (3.31).31) is adopted as the slave system for chaos synchronization. x )) + Bv ˆ s = Me = FCe = Fe y = F (Cx − y ) ˆ e1 (t ) = A011e1 (t ) + A012 e2 (t ) + B1 ( f ( x ) − f ( x ) − δ(t .29) is guaranteed: 1) if the sliding mode manifold is chosen as (3. M2∈Rm×(n-m). where M=[M1 M2]. the following two conditions are satisfied.4. and rewrite the error system (3.31) ˆ where x (t ) ∈Rn represents the system states.31) and (3. We have the following result [16. (3. For the observer (3. A.32) as the following decomposed form: m×n m×1 ˆ e (t ) = A0 e (t ) + B( f ( x ) − f ( x ) − δ(t .29) as follows: .31). a sliding mode manifold.33) depends on the design of parameter matrix M. which is the function of the output y(t) of the master chaotic ˆ system (3. The observer (3. y ) ∈Rm is the control input. e2∈Rn-m. ˆ Define the error variable e(t) as: e (t ) = x (t ) − x(t ) .33) where M∈R . (3. is designed as the following form: .29).29) and the variable x (t ) of the slave system (3. So the design of the sliding mode manifold (3. (3. G∈Rn×1 is a constant design matrix. (3.27).35b) ˆ vl = − f ( x ) . y ) . e1∈Rm.29). the vector C is the output matrix of system (3. For the sliding mode observer (3.

33). By substituting the sliding mode control strategy (3. the matrix M makes A022 − A021 M1−1 M 2 a Hurwitz matrix. β and η are positive scalars. 2) The matrix G makes A0 in (3. Therefore.35c) where r1 and r2 are known positive constants.35) into above equation. β>0. x ) + sT MB ( r1 + r2 + α (t .32) a Hurwitz matrix. the reduced-order dynamic system of error system (3.32) will depend on the linear sliding mode manifold (3.37) Substituting (3. (3. 0<η <1. Feng and X.70 Y. (3.32) can be obtained: . Yu and vn (s =− T MB ) T sT MB ( r1 + r2 + α (t. where γ = − β MB > 0 . the dynamic behaviour of the error system (3. the above equation becomes: . y) + β ) . 2 2 The derivative of V(t) along the error system (3. the error system (3. it can be obtained: V (t ) ≤ sT MBf ( x ) + sT MBδ(t . After the error system (3.32) reaches the sliding mode manifold surface s=0. x ) + sT MBvn ˆ V (t ) ≤ sT MBf ( x ) − sT MBf ( x ) + sT MBδ(t . y ) + β ) ≤ − β sT MB ≤ − β MB s that is V (t ) ≤ −γ s < 0 for s ≠ 0 . x ) ) + Bv ) = 1 T ˆ e ( M T MA0 + A0T M T M )e + sT MBf ( x ) − sT MBf ( x ) + sT MBδ(t . Consider the following Lyapunov function: V (t ) = 1 T 1 s s= s 2 2 2 = 1 1 ( Me )T Me = eT M T Me .35a). Proof.32) will reach sliding mode manifold surface s=0 in a finite time tr ≤ s (0) γ . while the following condition is satisfied: T λmax ( M T MA0 + A0 M T M ) ≤ 0 . x ) + sT MBv ≤ sT MB f ( x ) + sT MB δ(t . x ) + sT MBv 2 From (3.37) into the error system (3.32) can be obtained: ˆ V (t ) = sT s = e T M T Me = eT M T M ( A0 e + B ( f ( x ) − f ( x ) − δ(t . (3.36). that is: e1 = − M1−1 M 2 e2 .36) where λmax(As) presents the maximum eigenvalue of As.

ecology etc [3. economics.4. The third is to add another information to the transmission signal to represent the synchronization pulse signal. the error system (3.…. synchronization can not be realized by using a single signal channel. from the transmission signal yε(t). image. There are three tasks in the transmitter.38) will be asymptotically stable. That is.5. For practical applications such as signal. 3. The section proposes a hardware structure as shown in Fig. speech processing and secure communication. εi.4. Synchronization of chaotic systems with a single nonlinearity such as Chua system and Rössler system can be well done by using an observer via a single transmission channel. The functional block diagram of the transmitter is shown in Fig. biology and medicine. The level adjustor is used to eliminate the bias level series. Multi-dimensional Signals Transmission via One Signal Transmission Channel Chaos synchronization has numerous potential applications in mechanics. i=1. after reaching the sliding mode manifold surface s=0. 9]. εi. laser and chemical technologies.p. 3. 3. to the transmission signal to distinguish each signal from the amplitudes. The transmission signal yε(t) is inputted to the two blocks: level adjustor and level comparator. it is desirable to use a smaller number of channels of signals from the master chaotic system for chaos synchronization. whereas. The challenge is how to do it via a scalar transmission channel. i=1. then the system (3.6. The second is to add the bias level series.32) will converge to the equilibrium point e = 0 asymptotically. p. The first is to combine multi dimensional signals into one signal to be transmitted in one single transmission channel. communications. One effective approach is by using observers [10. 4. a method to transmit multi dimensional signals plus a synchronization signal of the master chaotic system using only one single transmission channel is proposed in [14] based on the idea of time division multiplex. 11].Sliding Mode Control of Chaotic Systems 71 e2 (t ) = ( A022 − A021 M1−1 M 2 ) e2 (t ) = AM e2 (t ) . (3. In order to realize chaos synchronization via one signal transmission channel. While the level . for the chaotic systems with multi nonlinearities such as hyperchaotic Chua system and Lorenz system.38) Designing the sliding mode parameter matrix M to make AM a Hurwitz matrix.….

7. . . ∗ yε (t ) Level Adjustor f h1 fh2 f hp Zero-order Hold Level Comparator Fig. which frequency must satisfy the Shannon's sampling theorem. which are to be transmitted. y1(t). In the transmitter terminal. 3.72 Y.6. pluse generator + + 2 k . … . . yp(t) represent p output signals of the master chaotic system. fclk is the sampling pulse. The block diagram of the transmitter. . . The hardware structure of the receiver. y1 y2 k yε1 f1 K1 yp kp Fig. yhp(t) represent p reconstructed output signals of the master chaotic system. Feng and X.4.5. yh2. The signal fsyn is the reconstructed sampling signal of fclk. yh1. + 1 ε1 ε2 yε 2 f2 K2 ∗ yε (t ) yε p fp Kp εp * y 1 (t ) Gh(s) Zero-order Hold y h1 (t) * y 2 (t ) Gh(s) yh2 (t) y * (t ) p Gh(s) y hp (t) . 3.5. 3. . It should be . y2(t). while in the receiver terminal. Synchronization of Chaotic Systems with Multi-nonlinearities The principle of the synchronization of uncertain chaotic systems with multi-nonlinearties by using observers is shown in Fig. Yu comparator is used to extract the sampling pulse fsyn from the transmission signal yε(t). 3. … . Zero-order Hold .

yh1(t) ∗ yε yh2(t) Demux Slave chaotic system . The example is hyperchaotic system formed by a pair of coupled Chua′s circuits in [10]: x = Ax + Bf ( x ) + ξ (t . x ) + D = Ax + B ( f ( x ) + δ(t . 3. y1(t) y2(t) Mux . 0  0 0 0    0  0    f1 ( x1 )  f ( x) =  .39) where D=0.Sliding Mode Control of Chaotic Systems 73 noticed that there is only one signal transmission channel between the transmitter and the receiver. 0 1  0  and . . 0 0 α 0  −α (b + 1)  0 −α (b + 1) 0 0 10   0 0 0 −β 0 A= 1 0 1 −1 − m m   0 1 0 0 −1  0 0 0 0 −β    −α 0   0 −α    0 0  B= . Master chaotic system yp(t) yhp(t) Fig. . (3. .7. x ) ) + D .  f 2 ( x4 )  0 0  0 . . Synchronization of uncertain chaotic systems.

8 . Feng and X.8. f 2 ( x4 ) = (a − b)( x4 + 1 − x4 − 1 ) / 2 . The output matrix of the system (3. y)=0. m=1. }.2063 52. 3. ˆ x2 . α=10.005}.005. Fig. choose the spectrum of the matrix A0 as {−9. 3.7690 −0. plus a synchronization signal. 0. 3.6269 29. C =  1 0 1 0 0 0  According to Theorem 4. 3.9801 18. According to Theorem 4.27. −8.68. Fig. x ) =  . −7. −20. T F is chosen as: F=diag{0. Yu  0. a = −1.3051 30.11 presents phase trajectory of the hyperchaotic Chua′s circuit state variables x2.2535 29. T=1ms. y1 and y2. the matrix G can be calculated:  25.8 depicts the two output signals of the master chaotic system to be transmitted to the slave chaotic system. Fig. β=14. k1=k2=1. x3 and its estimates. ε2=0.41) is designed as: 0 0 0 1 0 1 .87. .6791 28.2. A0 is a Hurwitz matrix. 3. r3=0. 0. Fig.74 Y.9 shows the one channel transmission signal yε contains two signals.10 shows phase trajectory of the ˆ hyperchaotic Chua′s circuit state variables x1.9. 3. Two outputs y1 and y2. −15.Fig. The simulation results are shown in Fig.01sin(3π t )  δ(t . x2 and their estimates x1 .01sin(4π t )  where f1 ( x1 ) = (a − b)( x1 + 1 − x1 − 1 ) / 2 .1798 −14. b=−0. the parameters are designed as: r2=0.2.3055 −7. −6. 4 3 2 1 0 -1 -2 -3 -4 y1 y2 0 5 10 15 Time(sec) 20 25 30 Fig.10. y1 and y2. α(t. β=0.9225 G=   −27.7371 .2248 38. 3. ε1=8.

Sliding Mode Control of Chaotic Systems
12 10 8 6 4 2 0 -2 -4

75

0

5

10

15 Time(sec)

20

25

30

Fig. 3.9. One transmitted signal.
4 3 2 1 0 -1 -2 -3 -4 -3 master system the observer

-2

-1

0

1

2

3

4

ˆ ˆ Fig. 3.10. Phase trajectory of (x1,x2) vs.( x1 , x2 ).
5 4 3 2 1 0 -1 -2 -3 -4 -5 -4 master system the observer

-3

-2

-1

0

1

2

3

4

ˆ ˆ Fig. 3.11. Phase trajectory of (x2,x3) vs.( x2 , x3 ).

76

Y. Feng and X. Yu

3.5. Conclusions
In this chapter, the sliding mode based chaos control has been discussed. Three methods of sliding mode methods, namely, traditional linear sliding mode, terminal sliding mode and nonsingular terminal sliding mode, have been used in chaos control. The advantages of the sliding mode based chaos control can be summarized as follows: it is robust to system uncertainties, suitable for realization, and simple to calculate. Furthermore, a robust sliding mode observer based chaos synchronization method has also been described. Based on the time division multiplex technique, a transmission method of the synchronization signals from the master system to the slave system via one single transmission channel has been designed. The advantages of this method are as follows: (i) multi-dimensional signals can be transmitted using the single channel so that chaos synchronization can be realised more efficiently without incurring significant costs of additional equipment/devices for transmitting more than one-dimensional signals; (ii) it is robust to a class of chaotic systems including internal parameter uncertainties and/or the external disturbances (iii) it is a systematic design method of chaos synchronization; (iv) it can be used for highly chaotic systems.

References
[1] G. Chen and X. Dong, From Chaos to Order: Perspectives and Applications. (World Scientific, Singapore, 1998). [2] W. D. Ditto, S. N. Rauseo and M. L. Spano, Experimental control of chaos, Phys. Rev. Lett. 65, 3211-3214, (1990). [3] M. Lakshmanan and K. Murali, Chaos in Nonlinear Oscillators: Controlling and Synchronization. (World, Singapore, 1996). [4] K. S. Halle, W. C. Wu, M. Itoh and L. C. Chua, Spread spectrum communication through modulation of chaos, Int. J. Bifurcation Chaos 3, 469-477, (1993). [5] M. Itoh, W. C. Wu and L. C. Chua, Communication systems via chaotic signals from a reconstruction viewpoint, Int. J. Bifurcation Chaos 7, 275-286, (1997). [6] L. J. Kocarev and K. S. Halle, Experimental demonstration of secure communication via chaotic synchronization, Int. J. Bifurcation Chaos 2, 709-713, (1992). [7] L. M. Pecora and T. L. Carroll, Synchronization in chaotic systems, Phys. Rev. Lett. 64, 821-824, (1990). [8] T. L. Carroll and L. M. Pecora, Synchronizing chaotic circuits, IEEE Trans. Circuits Syst. 38, 453-456, (1991).

Sliding Mode Control of Chaotic Systems

77

[9] L. O. Chua, M. Itoh, L. Kocarev and K. Eckert, Chaos synchronization in Chua’s circuit, J. Circuits Syst. Computers 3, 93-108, (1993). [10] D. Cafagna and G. Grassi, Synchronizing hyperchaos using a scalar signal: a unified framework for systems with one or several nonlinearities, APCCAS 1, 575580, (2002). [11] G. Grassi and S. Mascolo, Nonlinear observer design to synchronize hyperchaotic systems with a scalar signal, IEEE Trans. Circuits Syst. 44, 1011-1014, (1997). [12] Y. Feng, X. Yu, and Z. Man, Nonsingular adaptive terminal sliding mode control of rigid manipulators, Automatica 38, 2159-2167, (2002). [13] Y. Feng, X. Han, Y. Wang and X. Yu, Second-order terminal sliding mode control of uncertain multivariable systems, International Journal of Control 80(6), 856862, (2007). [14] Y. Feng, X. Yu and L. Sun, Synchronization of uncertain chaotic systems using a single transmission channel, Chaos, Solitons and Fractals 35(4), 755-762, (2008). [15] Y. Feng, L. Sun, and X. Yu, Finite time synchronization of chaotic systems with unmatched uncertainties, The 30th Annual Conference of the IEEE Industrial Electronics Society Busan, Korea, (November 2-6, 2004). [16] Y. Feng, L. Sun and X. Yu, Synchronization of a class of uncertain chaotic systems via observer, Dynamics of Continuous, Discrete and Impulsive Systems 13(c), 7787, (2006). [17] F. Han, L. Sun, Y. Feng, P. Wang and X. Yu, Synchronization of the Duffing oscillator by using terminal sliding mode control, Proceedings of The IEEE Region 10 Technical Conference on Computer, Communication, Control and Power Engineering (IEEE TENCON) 3, 1362-1365, (2002). [18] Y. Feng and L. Sun, Chaotic system synchronization using a robust sliding mode observer, International Conference on Dynamics, Instrumentation and Control (CDIC) Nanjing, China, (August 19-20, 2004). [19] Y. Wu, L. X. Yu and Z. Man, Terminal sliding mode control design for uncertain dynamic systems, Systems and Control Letters 34, 281-288, (1998). [20] X. Yu and Z. Man, Model reference adaptive control systems with terminal sliding modes, International Journal of Control 64, 1165-1176, (1996). [21] Z. Man and X. Yu, Terminal sliding model control of MIMO linear systems, IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications 44, 1065-1070, (1997).

This page intentionally left blank

Chapter 4 A NEW TWO-STAGE METHOD FOR NONPARAMETRIC REGRESSION WITH JUMP POINTS

C.Z. Wu School of Mathematics and Computer Science, Chongqing Normal University, Chongqing, 40047, China. changzhiwu@yahoo.com C.M. Liu* and K.L. Teo† Department of Mathematics and Statistics, Curtin University of Technology, Perth, WA, 6102, Australia. * jimmy_uwa@yahoo.com.tw † k.l.teo@curtin.edu.au Q.X. Shao CSIRO Mathematical and Information Sciences, WA, 6104, Australia. Quanxi.Shao@csiro.au In this chapter, a two-staged method is presented for nonparametric regression with jump points. After the rough location of all the possible jump points are identified using the existing efficient kernel method, a smoothing spline function is used to approximate each segment of the regression function. A time scaling transformation is derived so as to map the undecided jump points into fixed points. This approximation problem is formulated as an optimization problem which can be solved by many existing techniques. The method is applied to several examples. The results obtained show that the method is highly efficient.

79

Z. such as land clearing. [18] used wavelets to provide a useful procedure for the detection of the jump points and their magnitudes. where the asymptotic properties are also studied. Shao 4. [20] used one-sided moving average to find the locations of jumps. Liu. [14] proposed a two-stage estimation scheme for the jump locations. [7] used linear splines to estimate discontinuities of the unknown regression functions.L.80 C. At the same time. [19] used the kernel method to estimate both the locations and magnitudes of the jumps. Without considering these jumps.X. C. Introduction Statistical modelling generally assumes smoothness and continuity of the phenomena of interest. For example. When we study the portfolio management. Wu. [11] proposed an algorithm to estimate regression functions when discontinuities are present. The key idea is to investigate the difference between the estimators of the left.1. Problems related to regression with jump points have been addressed and investigated for more than two decades. Teo and Q. the locations and magnitudes of the jumps can be estimated by using the maximization argument of the jump points estimation. the amount of stocks of a particular investor can be viewed as experiencing a jump when he/she purchases or sells his/her stocks. some phenomena may experience sudden or sharp change. Most of the estimators mentioned above are of the kernel type. [13] estimated the locations and magnitudes of the jumps by the boundary kernels. [8] made use of pairing pattern and linear filter to develop a discontinuity detector for the purpose of detecting the jumps and their magnitudes.M. the overall fittings obtained using these methods are not very satisfactory at around jump points and around end points. [9] used Fourier analysis for jump detection. K.and right-hand side limits for the unknown regression functions. However. we may make an serious error in drawing inference about the process under study. On this basis. . the spline method was also applied for detecting the jump points. It is clearly important to estimate both the number of jumps and their locations and magnitudes. This sudden change is reflected as a jump in visual display. However. groundwater table may undergo drastic changes in very short periods of time [15] due to sudden changes in environment. [12] applied the evolutionary algorithm to locate the optimal knots.

Nonparametric Regression with Jump Points

81

These methods are computationally expensive. In this chapter, a twostage method is developed for detecting the number of jump points, their locations and magnitudes as well as finding the spline representations for the approximation of the unknown regression function with jump points. The following notations are used in this chapter. Let X be the response variable with respect to variable t . For a given set of observations {xi ,ti } , i = 1, 2, , n , the regression function is written as: xi = m ( ti ) + ε i , i = 1, 2, , n (4.1) where m is an unknown function defined on

[ 0, T ]

,

{ε i }

,

i = 1, 2,

, n , are independent and identically distributed N ( 0, σ 2 )

normal random variables. For each i = 1, 2,

, n , {ε i } represents the

variation of xi around m ( ti ) . Without loss of generality, we can let

T = 1 . Otherwise, a re-scaling of T can be used. To form a segmented regression, we write m ( t ) as: m ( t ) = ml ( t ) , for τ l −1 < t ≤ τ l , l = 1, , N ,

(4.2)

where τ 0 = 0 and τ N = T , where τ i , i = 1, , N , are jump points. This setting covers the case of the change of mean values (step changes) as a special case by choosing ml ( t ) as constants. Note that the form of the regression ml ( t ) is usually unknown, and hence are nonparametric. It is well-known that spline functions are effective for approximating nonparametric regressions if they are smooth. However, the smoothing spline does not work well directly for regressions with jump points, since the smoothing spline is rather sensitive to the location of jump points. Some difficulties in jump point estimation have been demonstrated in [6] for linear regression with only one jump point. Intuitively, the jump points should be identified before applying the smoothing spline to each of the individual segments defined by the jump points. This is the motivation behind our proposed two-stage method. The proposed method consists of two steps. First, we should locate all potential locations of jump points. Then, we use a time scaling transformation to transform the potential jump points and spline knots into pre-fixed points. By doing this, the parameter estimation using least squares becomes an optimization problem, which can be solved by

82

C.Z. Wu, C.M. Liu, K.L. Teo and Q.X. Shao

efficiently techniques, such as the sequential quadratic approximation optimization method. Finally, the modified Akaike's information criterion is used to determine which potential jump points are real jump points. We then obtain the final model. This chapter is organized as follows. In section 4.2, we employ the kernel method proposed by [19] to locate locations of all the potential possible jump points to within respective observation points. Then, the problem formulation is given in section 4.3. section 4.4 introduces a time scaling transform, which is motivated by the one proposed in [16]. This time scaling transform is then used to derive a new time scaling transform, which maps the undecided jump points and spline knots into fixed points. The obtained optimization problem is solved by the sequential quadratic approximation optimization technique in the Optimization Toolbox within the Matlab environment. In section 4.5, we use the modified Akaike's information criterion to determine the number and locations of jump points and therefore obtain the last model. In section 4.6, several numerical examples are presented and solved by using the proposed method. Some concluding remarks are given in section 4.7. 4.2. Estimation for Potential Jump Points In statistics, the kernel method is an efficient and powerful statistical tool to detect jump points. Therefore, it is used in our two-stage method to find potential jumps. We assume that the following conditions are satisfied. 1) The number of the observation points is sufficiently large to detect all the possible jump points. That is, the number of jump points is very small in comparison with the number of observation points and the number of observation points between consecutive jump points is sufficiently large for spline fitting. 2) The jump points should be located in the interval [δ ,1 − δ ] , where δ is some small positive constant. That is, there is no jump point in the neighborhoods of the boundary points on which they become undetectable.

Nonparametric Regression with Jump Points

83

3) There is at most one jump point between ti and ti +1 , i = 0,1, , n, where t0 = 0 and tn +1 = 1 . 4) The regression errors ε i , i = 0,1, , n, are independent and identically distributed (i.i.d) random variables with mean 0 and variance σ 2 < ∞ . Let K be the kernel function with the bandwidth h . There are two popular methods to construct kernel estimators, depending on the choice of weights by either direct kernel evaluation or the convolution of the kernel with a histogram obtained from the data. Each estimator has several important advantages and disadvantages. A thorough review on this subject can be found in the paper by [2]. [4] make use of the first method to construct an estimator to detect the vicinity of the jump points. In our two-stage method, the second method is used to identify rough positions of the jump points. For the kernel function K and the bandwidth h , the Gasser-Müller estimator is defined as:

ˆ m ( t ) = ∑ xi ∫ K h ( t − τ )dτ , for t ∈ ( 0,1) ,
si i =1 −1 si−1

n

(4.3)

where K h ( • ) = h K ( • / h ) , sn = 0 , si = ( ti + ti +1 ) / 2 , i = 2, , n − 1, and sn = 1 . Let m1 ( t ) and m2 ( t ) be two Gasser-Müller estimators obtained with the kernel functions K1 and K 2 , respectively, using the same bandwidth h . Let ˆ ˆ J ( t ) = m1 ( t ) − m2 ( t ) . (4.4) To continue, we need to analyze the value of J ( t ) . If m ( t ) has no jump point, then, under the usual regularity conditions, m1 ( t ) and m2 ( t ) are uniformly strongly consistent estimators of m ( t ) . Thus, the variation of J ( t ) would not be obvious. If m ( t ) has a jump point, then the value of J ( t ) would have an obvious change around the jump point. In Fig. 4.1., we give an example where m ( t ) has a jump at t = 1/ 2 .

84

C.Z. Wu, C.M. Liu, K.L. Teo and Q.X. Shao

Fig. 4.1. The magnitudes of J ( t ) and the regression function. Write

m ( t ) = ϕ ( t ) +ψ ( t ) ,

(4.5)

where ϕ ( t ) is the continuous part of m ( t ) , while ψ ( t ) is a step function to characterize the jumps of m ( t ) . The magnitude of

J ( t ) corresponding to ϕ ( t ) is of small order. However, the magnitude
corresponding to ψ ( t ) is symmetric and convex downward in the neighborhood of the jump point if the kernel functions K1 and K 2 are chosen such that K1 ( t ) = K 2 ( −t ) , (4.6) Furthermore, if K1 and K 2 have compact support in [ −1,1] , then the widths of the neighborhoods mentioned above are no more than 2h . Based on the above discussions, we can give a numerical procedure to estimate rough locations (and hence, number) of the jump points. Algorithm 4.1. (Kernel method to detect the potential jump points) Step 1: Choose h such that h = O n −1/ 3 .

(

)

Nonparametric Regression with Jump Points

85

Step 2: Choose a nonnegative function K ( t ) with a compact support in [-1,1] and is such that

1

0

K ≠∫ K .
−1

0

Step 3: Let K1 ( t ) = K ( t ) and K 2 ( t ) = K ( −t ) . Calculate J ( t ) by using (4.3) and (4.4), where h is given in Step 1. Step 4: Find all the points, which correspond to local maxima of

J (t ) .
The points obtained by Algorithm 4.1 are considered as potential jump points due to Step 4. In the next section, we will use a time scaling transformation to find the accurate positions of these jump points. 4.3. Segmented Regression with Constraints Suppose that τ 1 ,τ 2 , ,τ N −1 , is a set of potential jump points (obtained from the kernel method crudely). Then, the interval [ 0,1] has been partitioned into N subintervals [τ l −1 ,τ l ] , l = 1, , N . For each l = 1, , N , let the observation points contained in the subinterval [τ l −1 ,τ l ] be denoted by t1,1 , , t1, N1 , and let x1,1 , x1, N1 be the corresponding observations. The regression function is denoted by ml ( t ) in the subinterval (τ l −1 ,τ l ] where l = 1, , N , τ 0 = 0, and τ N = T . Now, we use the segmented regression to fit the segment ml ( t ) , l = 1, , N . The most widely used approach to curve fitting is least squares. If we place no restrictions on the residual sum of squares, this method is in fact an intoporlant which may be caused a rapid fluctuation. To avoid this, we incorporate a smooth penalty in the cost function. In this chapter, we will introduce the cubic spline for fitting the segment ml ( t ) . A general cubic spline basis is defined as

{1, t, t , t , (t − θ )
2 3

3

1 +

,

, ( t − θ K )+

3

}.

(4.7)

where ( t − θ k ) + = max ( 0, t − θ k ) , θ k , k = 1,

, K , are the knot points.

Since the smoothness is penalized in the cost, we need the estimator ˆ ml ( t ) of ml ( t ) is continuously differentiable at the knot θ k ,

k = 1,

, K . Write

Wu. we need to compute the cost. Since the first part is only related to the coefficient vector a . 3 + al . the estimates of the jump locations and the optimal regression coefficients are obtained simultaneously. Define the cost functional as the following: ˆ J (τ.i1 + 4 . The second part is the sum of some integrals with the integral limit related to the jump points.iN + 4  .Z.τ N −1 ] . By this transformation.86 C.  T .M.L. a ) such that (4. we can choose it by the generalized cross-validation method [3].1. Thus.X.9) where τ = [τ 1 . K.1 + ∑ al . Liu. Teo and Q.1 . a1. C. We note that the cost (4. . aN .1 . . λ is the  smoothness parameter. .i ) + λ ∑ ∫ 2 l =1 i =1 i =1 N Nl N τi τ i−1 ˆ (mi′′(t )) dt T 2 (4. (4. al . Let this problem be referred to as Problem (P). Shao ˆ ml ( t ) = al . a =  a1.k ) + . here we will choose it interactively for simplicity.9) is minimized subject to the following constraints tl . there is no any need restrictions on ml ( t ) . it is easily computed.8) where θl . Nl ≤ τ l ≤ tl +1. For the penalized parameter λ . the jump points and the spline knots are all mapped into some pre-fixed points. . θl .l + 4 .9) is composed of two parts. To solve Problem (P). To overcome this difficulty.il . Since ( t − θ k ) is twice continuously differentiable. a ) = ∑ ∑ (ml (tl . are the pre-fixed knots contained in the l-th segment.k + 4 ( t − θl .10) For this optimization problem. However. our objective is to find ( τ .k +1t k + ∑ al . to compute it with its corresponding gradient is hard.i ) − xl . .1 . we just introduce a time scaling transformation. are the coefficients. 3 k =1 k =1 3 il (4. Under the least squares method with the smoothness penalty.

To achieve it. we just choose v ( s ) = ∑∑ ξ k .4. if s ∈ I . . It is defined by dt = v (s). dt = v (s). N − 1.11) t ( 0 ) = 0. χI ( s) =  ξ k . N . N . satisfy the following conditions: . . .T ] . . k = 1. ξi ] . ξ N . are transformed to pre-fixed times ξ1 .τ N −1 . .θl .τ N . k + j / ( ik +1 ) ) where χ I is the indicator function.il . j χ k + ( j −1) / ( i k =1 j =1 N ik +1  k +1 ). l = 1. . 0.. ξ 0 = 0 . i. 1. in the new time scale. ds (4. ∫ξ v ( s ) ds = τ i −1 ξi i − τ i −1 . j . which was introduced by [10] to overcome the numerical difficulties in the computation of some optimal control problems. are mapped into the fixed points 1. θl . This transformation was known as the enhancing transform. .Nonparametric Regression with Jump Points 87 4. . Now we apply this transformation to our problem such that the jump points τ 1 . . v ( s ) ≥ 0 for all s . ds with initial condition (4. (4. j = 1. otherwise. are N variable times in the time interval [ 0.e.1 .τ N . . are mapped into and the spline knots l + 1/ ( il + 1) . ik . A time scaling transformation is introduced such that the variable times τ 1 .13) where τ 0 = 0 .12) where v ( s ) satisfies the following conditions: • • v ( s ) is a piecewise constant on the interval (ξi −1 . A Time Scaling Transformation Method We suppose that τ 1 . l + il / ( il + 1) .

i1 +1 . Teo and Q. (4. Problem (P) is equivalent to Problem ( P ) in the sense that ( τ. Liu.1 − θ k . j ≥ 0. j − θ k .  .M. (4.1. Model Selection Note that the kernel method in section 4. We propose to use the . Problem ( P ) is stated as the  following.16) (4. j = 2. ik .15) ∑∑ ξ k =1 j =1 N ik +1 k.18) Let ˆ J ( ξ. ˆ Problem ( P ). j −1 ) . Several methods for choosing it have been suggested in the literature. Shao ξ k .19) where ξ = ξ1. N .15). i N l l −1 2 (4. a ) subject to (4.17) 1 1 ξ k .X. k = 1. .Z. Then. a ) is an optimal solution of Problem (P). ξ k .5. the number of jump points p is required to be chosen appropriately.14).4 to any possible combinations of potential jump points.iN +1  . j = 1. N.17).16) and (4. Suppose that the number p is the candidate of jump points.L.1 . (4. ξ N . T ˆ . K.14) (4. N .3 can be used in conjunction with the time transform introduced in section 4. j = T. k = 1. . k = 1. The spline approximation method outlined in section 4.ik . j = ( ik + 1) (θ k . C. they have the same optimal cost. Clearly.i ) − xl . ξ1. min J ( ξ.1 = θ k +1. ˆ Theorem 4. ik + 1 ik +1 + 1 By this transformation. ik . Furthermore. 4. we have ∑ ∫τ ( m '' ( t ) ) i =1 i −1 N τi 2 i dt =∑ ∫ i =1 N i i −1 ( m '' ( t ) ) v ( s ) ds. a ) is the optimal solution of Problem (P) if and only if ( ξ.ik +1 + ξ k +1.2 tries to detect all the potential jump points. (4.i l =1 i =1 N Nl ( ) 2 + λ∑ ∫ l =1 ( m '' ( t ( s ) ) ) l 2 v ( s ) ds (4. we have the following theorem. Wu. .88 C. a ) = ∑∑ ml ( tl . . .

Tong [17] applied AIC for model selection in a threshold autoregression. 3. For the real data. BIC will be used as the criterion as the data can be collected as much as we need.20) N N + p−2 where N is the number of observations. If there is no possible candidate jump points.21) In our simulation. The AICc penalizes the RSS by a function of the number of free parameters and is given by RSS N+p + . set Γ = Γ '. . where the data produced from a mathematical function. we will use BIC as a criterion instead of AIC when the number of observation AICc = ln points is large enough.9).Nonparametric Regression with Jump Points 89 Akaike Information Criterion ( AIC ) [1] for the purpose of model selection. Algorithm 4. . we will use AICc as the criterion since the observation points are limited. 2. stop. N N (4. (4. p is the number of free parameters in the model and RSS is the first part of the cost (4. 1.Choosing candidate jump points {τ i }i =1 ⊂ τ and for given spline ˆ knots θ . Γ m T is a set to store the solution and set it empty. for those examples.2. where BIC = ln RSS p ln N + . If AICc' ≤ AICc . AICc = AICc' . Hurvich and Tsai [5] derived a modified Akaike information criterion AICc in small samples and claimed that the AICc dramatically reduces the bias and improves model selection. and go to Step 2. Let Γ ' denotes the current solution which is the set of the jump points and the regression coefficients. which is a set of candidate jump points.11) to solve ( P ) and evaluate the corresponding AICc (or BIC ). AICc = 1000. Set τ = [τ 1 . use the time scaling transformation (4. Note that AIC tends to overestimate the number of parameters.τ N −1 ] . thus.

Liu. some numerical examples are presented here. Let the spline knots vector be [0.4. (4. Shao 4.3. 4.78178.2.26 3/ 5 1( t > 0. C. 0. Let n = 200. then the obtained result is depicted in Fig.26 1/ 5 1( t ≤ 0.26 ) + 1( t ≥ 0. In the next numerical examples. σ ε = 0.1. which contains the water levels (depths) in boreholes monitored irregularly over time. We use BIC as criterion and the obtained results are that there are two jump points: 0. and the data is produced from the function: m ( t ) = 2 − 2 t − 0. 0. Example 4. We rearrange the observation points with the given candidate jump points.2 . (4.4.25225. K.15.3.22) with ti = i / n and ε i sampled randomly from a normal distribution with standard deviation σ ε . Thus. From Fig. 0. we have to consider all the potential jump points and use .1 . 4. We can see that the kernel method cannot present an good jump estimator since the observation points is too little from the figure. The data are taken from the database of Agriculture of Western Australia and have been analyzed by [15]. 0. We now apply our method to a real example.M.1 to detect the rough locations of the jump points. The result is depicted in Fig.2. Numerical Example To assess the performance of our proposed method.Z. The results are depicted in Fig.1056. 4. the data were generated by the following equation: xi = m ( ti ) + ε i . the parameter λ controls the gradient change rate of the splines. 0. Teo and Q.23) First.78) .5. Example 4.90 C. For the simulated study.6.1 is depicted in Fig. then the fitting becomes linear fitting.26 ) − 2 t − 0.6].L. we use Algorithm 4. 4. we choose λ = 0. There are 49 observations.X.5. 4. The observation points as well as its jump estimator by Algorithm 4. If we choose λ=10.1. Thus. Wu. The obtained BIC is -4. we can see that if λ is enough. We do the same simulation for 100 times.

The last obtained model has 3 jump points: 0.2727. AICc = 4.0.Nonparametric Regression with Jump Points 91 AICc as criterion to choose the best model. The obtained results are depicted in Fig.1]. Example 4. 4.1.1. In the process of fitting.6.2.3. 4.0. Fig. with λ = 0.1050. Fig. with 100 times simulation.3148. 4. We re-scaling the time to the new interval [0.01 . we suppose that all of the jumps are positive since the level of the groundwater does not experience drop instantly. Example 4. .4554.

M. Observation points and jump estimator of Example 4.92 C. Example 4.X. .5. 4.1.Z.2. 4. K.L. Teo and Q. Shao Fig.4. Fig. Wu. C. with λ = 10 . Liu.

Peter and A. IEEE Transaction on Automatic Control 19. Kneip. a new two stage method has been proposed for solving spline regression problem with jump points. some numerical results are presented and which are shown that our proposed method is efficiency. Marron. 4. Wahba. H. Annales Institue of Statistics and Mathematics 51. Choosing a kernel regression estimators. (1974). Craven. 377-403. Statistical Science 6. (1999).6.7. Hurvich. References [1] H. [5] M. we detect the rough locations of jump points based on the kernel method. Biometrika 78. Tsai.2. [3] P. and C. 231-251. A new look at the statistical model identification. Then. Also. (1991). Bias of the corrected AIC criterion for underfitted regression and time series models. Conclusion In this chapter. Chu and J. [4] I. K. (1979). Numerical Mathematics 31. The fitting result of Example 4. 404-419. 499-509. we introduce a time scaling transformation reformulate our regression problem as a nonlinear optimization problem which is easily to be solved. 716-723. First. Smoothing noisy data with spline functions. . and G. Akaike.Nonparametric Regression with Jump Points 93 Fig. (1991). L. On the estimation of jump points in smooth curves. 4. [2] C. S. Gigbels.

323-335. (1997). Adaptive free-knot splines. [11] A. [7] J. [9] F. Lombard. Shen. Technomerics 30. Wu and C. W. Jennings. S. Song. Y. (1992). Campbell. Inference and estimation in a change-point regression problem. 266-284. (1993). Shao [6] S. Coping with discontinuities in computer vision: their detection. Detection of the number. Jennings. K. [17] H. Control parameterization enhancing transform technique for time optimal control problems. (1980). Rehbock and L. Journal of Royal Statistical Society B 42. 51-61.M. (1999). 305310. and measurement. Journal of Royal Statistical Society Series A 50. Jump and sharp cusp selection by wavelets. Spline estimation of discontinuous regression functions. The control parametrization enhancing transform for constrained optimal control problems.L.94 C. L. L. (1988). S. Liu. classification. G. Kernel-type estimators of jump points and values of a regression function. (1997). [20] Y. Teo and Q. Detecting change points by Fourier analysis. (1995). McDonald and A. Smoothing with split linear fits. S. 1545-1566. [13] H. Müller. Threshold autoregression. (1990). 197-213. J. [8] D. L. (1997). Journal of Computational and Graphical Statistics 6. [15] Q. 385-397. Statistics and Probability Letters 34. Müller and K. Change-points in nonparametric regression analysis. Annals Statistics 20. 321-344. K. [16] K. 445-455. 737-761. H. 243-261. Modelling trends in groundwater levels by segmented regression with constraints. [14] H. 314-335. Rehbock. . Teo. 129-141. 245-292. [19] J. [12] S. Journal of Computational and Graphical Statistics 12. Communications in Statistics-Stochastic Models 4. Owen. with discussions. Technometrics 28. Miyata and X. Koo. W. (2001). locations and magnitudes of jumps. Tong. Yin. V. limit cycles and cyclical data. Chu. (2003). C. Lee and V. Two-stage change-point estimators in smooth regression models. Australian and New Zealand Journal of Statistics 44. IEEE Transactions on Pattern Analysis and Machine Intelligence 12. (1986). 195-208. Shao and N. G. K. [10] H. (2002). J.Z. Biometrika 82. L. Wang. Lee.X. Teo. [18] Y. Annals Statistics 21. Dynamic Systems and Applications 6. (1988). Julious. Wu. Journal of Australia Mathematical Society B 40. Lee.

Section B Chaos Control for Continuous-time Systems .

This page intentionally left blank .

In order to better address this subject. Subjects of modeling [5. During the seventies some basic research took place which aimed at establishing new concepts concerning chaotic dynamics [63]. The main purpose of the present chapter is to serve as an introductory guide to the universe of Chua’s circuit control. The eighties were strongly marked by the goal of developing tools to characterize chaos [32. Mendes§ o Dept. and mathematical modelling. Palhares‡ and E. 42] of chaotic systems were mainly developed in the nineties. an inductorless Chua’s circuit realization is first presented. L. Of course.M.DELT Universidade Federal de Minas Gerais ∗ torres@cpdee.br ‡ palhares@cpdee.A.br The practical implementation of Chua’s circuit control methods is discussed in this chapter.ufmg.Chapter 5 CHAOS CONTROL FOR CHUA’S CIRCUITS L.A. As a consequence of the investigation of different control strategies applied to Chua’s circuit.M.ufmg. there is active ongoing research in these fields.br § emmendes@cpdee. Tˆrres∗ . and dynamical characterization associated to this electronic chaotic oscillator. a tradeoff among control objective. and model complexity is devised.B. R.M.ufmg.br † aguirre@cpdee. of Electronic Engineering .1. which quite naturally leads to a principle that seems to be of general nature: the Information Transmission Via Control (ITVC) for nonlinear oscillators. Aguirre† . 5. synchronization.A. mathematical modelling. followed by practical issues related to data analysis. 36] control and synchronization [30. Introduction The so-called chaos advent was rekindled after the computer simulation carried out by Lorenz in the early sixties [56]. 97 . 37].ufmg. control energy.

In the present chapter. One of the reasons for this imbalance is the obvious fact that more often than not ideas (theory) develop before actual implementation. the aim is to present practical implementations of control strategies for Chua’s circuit.M. In order to gain flexibility.A. The paper [73] described a laboratory setup. (b) the actuators can be driven by the analog outputs of a standard I/O-board. Another possible cause is the natural difficulty to assemble a suitable setup for testing new ideas as opposed to simulating solutions on a digital computer. Indeed. it is important that the real circuit should include devices— called actuators— to implement the control action. 20. experimental chaos based information transmission systems. for instance. This will require additional real-time I/O capabilities in order to drive the actuators [73]. L. the field of chaos control and synchronization of nonlinear oscillations has attracted much attention lately [13. together with data analysis and system identification techniques used to . A comprehensive account of this issue should discuss the Chua’s circuit implementation itself (Sec. In order to be able to take full advantage of the many benefits of Chua’s circuit in the investigation of control and synchronization algorithms. in the identification of new bifurcation scenarios [15]. quite understandably much less has been reported in terms of practical results [13].98 L. Such a circuit has undoubtedly become a standard benchmark in the study of nonlinear dynamics and chaos [28]. The advantages of a setup that includes actuators goes beyond implementing control and synchronization algorithms. Tˆrres. Aguirre.2). While much has been accomplished in the realm of theory.A.A. in order to be able to actuate fast enough (c) the I/O-board is driven by a real time program written for Linux and (d) an inductorless implementation of Chua’s circuit permits to slow down the original dynamics to just a few hertz. This enables implementing sophisticated control schemes without severe time restrictions in applications such as: design and verification of chaos control algorithms. with the following features: (a) composed of a Chua oscillator furnished with three actuators thus permitting mono and multi-variable control. (simulated) driven Chua’s circuits are useful in a number of different ways such as. it is usually desirable that the control law be implemented on a digital computer.M. From the mid-eighties considerable effort was devoted to build simple laboratory systems that could be used to produce chaotic data. Palhares and E. 5. Mendes o Particularly. 58]. R.B. An electronic circuit that has come to be known as Chua’s circuit is to be found among such endeavors [57. practical synchronization studies. named Chua’s Circuit Control Prototype – PCChua.M. to mention a few. 21].

1. and rL . Chua’s Circuit Implementation This section is primarily focused on presenting a robust and reliable implementation of Chua’s circuit. R is a variable resistor used to change the coupling between the active and passive parts of the circuit. dz/dt = −y/L − z(rL /L). 5.4. not shown in the figure. is composed by a network of linear passive elements connected to a nonlinear active component called Chua’s diode. In addition. 5.1. built as an autonomous oscillator. 5. 5. Chua’s circuit.1. L and R are the values of the passive linear elements shown in Fig.2.Chaos Control for Chua’s Circuits 99 assess the quality of the implemented circuit and to gain insight on the peculiarities of Chua’s circuit dynamics (Sec. 5. denotes the series internal # ) ! "¢ % & (  ¢ (5. which is one of the reasons for its popularity among scientists all over the world. a sample of control strategies applied to the problem of Chua’s circuit control and synchronization are discussed in Sec. 5. Fig. such as: dx/dt = (y − x)/(RC1 ) − id (x)/(RC1 ). As a consequence variation of R leads to bifurcation phenomena. The set of differential equations that describes the circuit behavior is obtained from Kirchhoff circuit laws. As it will be seen next.2. as seen in Fig. the circuit can be easily built. 5.1) $ % & ' 2 3 01  ¡       ©¨¦¤¢ §¥£ . dy/dt = (x − y)/(RC2 ) + z/C2 . An Overview of Chua’s Circuit The circuit in its standard form. where C1 . C2 .1.3).

002 0. and the continuous traces are the resulting linear regression applied to the measured data in order to reveal the piecewise affine structure of Chua’s diode curve. Aguirre. Mendes o resistance of the inductor.4) . Figure 5.2) where m0 .M. f (x) = Rid (x).004 0. Palhares and E. 0. L.1) in normalized form as: x = p [y − x − f (x)] . ˙ y = x − y + z.A.002 -0. ˜ q = −RL 2 .   t = RC2 . 5. L it is possible to rewrite Eq. ˙ ˜ ˙ = qy + r˜. (5.003 0. As a final remark on Chua’s circuit equations.003 -0.2.3) z = Rz.004 -8 -6 -4 -2 0 v (V) 2 4 6 8 Fig. Tˆrres.A.5(m1 − m0 ) {|x + Bp | − |x − Bp |} .2 shows a measured static nonlinearity of an implemented Chua’s diode. 1 2 C (5. (5.M. The static nonlinearity of Chua’s diode is the piecewise linear function given by id (x) = m0 x + 0.    r = r q. by adopting the following conventions   C t  p = C2 . R. m1 . The dots correspond to real data.M. Experimentally obtained static nonlinearity of a Chua’s diode.100 L.001 -0. z ˜ z (5. Bp ∈ R are parameters.B.A.001 i (A) 0 -0.

run sophisticated algorithms based on these data and to take some control action or to run some file or graphical output operation.1. 5. From the basic circuit topology shown in Fig. 5.Chaos Control for Chua’s Circuits 101 ˙ where x = dx/dt .3. which lead to generalizations that exhibit a richer set of dynamical behaviors [19.3. Implementation of Chua’s diode using Op. 5.2. as shown in Fig.1. ˙ ˙ ˜ z There are different ways to implement the nonlinear characteristic Eq. (5. Chua’s diode R1 R4 R2 R3 R5 R6 Fig. In order to slow down the chaotic oscillations. y = dy/dt and z = d˜/dt . many variations have been proposed in the literature. Very slow chaotic oscillations are desirable in many situations where computers are employed to measure data from the circuit. 5. 47]. A robust and reliable way is by using operational amplifiers [41]. (5. i.Amps. An Inductorless Version of the Circuit Chua’s circuit exhibits a broadband spectrum.1). but the majority of the power density is concentrated roughly around the frequency determined by the resonant sub-circuit composed by L and C2 in Fig.2) [46]. C2 and L in Eq. greater values for C1 . 5. (operational amplifiers).e. .2. it is necessary to use larger energy storage elements.

which is in general easier to measure. 65] as seen in Fig. 5.1. L{i(t)} I(s) R8 (5. This circuit is. Tˆrres.A. From Fig. and R7 through R10 and C3 are shown in Fig. Obviously Eq.4. (5.4.4. Mendes o However. One possible solution to this problem is the replacement of the inductor by a circuit that simulates the behavior of an ideal grounded inductor over several decades of frequency [71]. 62]. The voltage at z y point can be readily used to determine the “current through the realized inductor”. 35. Palhares and E. 5. Furthermore.1). which is the current through the inductor L shown in Fig. R. variable z in Eq. indeed. the same is not true for commercial inductors. the trade-off between coil size and maximum allowed inductance leads to very large and low-accuracy inductors.5) where L stands for the Laplace transform. one can see that the third state. can be readily determined by measuring the voltage zy shown in Fig. The theoretical transfer function of the simulated inductance is [65]: Z(s) = V (s) R7 R9 R10 C3 L{v(t)} = = s. (5. In this type of circuit it is mandatory that one of the terminals of the realized inductor be grounded (this corresponds to grounding the terminal of R10 ). 5.4.M.M. zy − i(t) R7 R8 R9 + − C3 R10 v(t) + Fig. where y is the state variable corresponding to the voltage across the capacitor C2 .M.A.B. Therefore the complete state vector is composed of voltages. Several other configurations of inductorless Chua circuits have been recently suggested [26. L.4 and using the expression z = (zy − y)/R7 . a gyrator circuit built as an impedance transformer [14. 5.102 L. despite the fact that off-the-shelf capacitors cover a very wide range of capacitance values.A. 5. 5. Aguirre.5) suggests that the inductor is ideal since it is a pure differentiator as indicated by the zero at the origin . Op-Amp realization of the inductor simulation circuit.

(5.1) the corresponding theoretical circuit parameters are: 0Ω ≤ R ≤ 2kΩ.2.1 is approximately 1.60 Hz.3. a realistic SPICE model was implemented and the resulting frequency response was found to be valid for almost four decades. This is a very important detail in the Chua’s circuit implementation because high internal resistance rL prevents the appearance of chaotic dynamics [11]. 5µF. rL = 0. called PCChua [73].3H.5. C1 = 23. For these components. which easily encompasses the spectral content of the measured time-series from the circuit. This is important in real-time applications because it enables the user to implement rather involved algorithms without being pressed too hard by computation time. m0 = -0. The main resonant frequency corresponding to the parameters in Table 5. in chaotic and non-chaotic regimes [73]. designed to be a testbed in the study of control and synchronization schemes.1. in Eq. . 5. PCChua – A Versatile Experimental Platform This section describes an experimental setup. L = 42. C2 = 235µF. In this case the reference point of the realized inductor is attached to the same ground reference of the rest of the circuit (see Sec. is also valid in the same frequency range. associated to the realized inductor.756mS. Chua’s diode Simulated Inductor R1 R4 R R7 C1 R2 R3 R5 R6 C2 R8 R9 C3 R10 Fig. Bp = 1. In order to investigate the limits of this configuration. m1 = -0. 5.0Ω. it was also found that the assumption of nearly zero equivalent internal resistance rL .5. 5. The component values are shown in Table 5.86V. The complete circuit is shown in Fig. Following the same procedure. A set of electronic components values that can be used in this implementation are listed in Table 5.Chaos Control for Chua’s Circuits 103 of the complex s-plane.409mS.2. 5.1.3). Implemented inductorless Chua’s circuit without actuators.

q and r depend on the components used. From a practical point of view. and u(t. R8 .1. (5. R2 220Ω R4 . x).3kΩ R7 .amps.6) can be written in detail from Eq.104 L. ˙ z = −qy + r˜ ˜ z + uz (t. (5.M.A. R5 22kΩ R3 2. Palhares and E. Mendes o Table 5.A. the addition of the control vector u(t.3) and Eq.3). Equation Eq. Components Value R1 . (5. see Eq. x). x). Possible set of electronic components that can be used in inductorless Chua’s circuit implementation. x). F(·) : R3 → R3 is the Chua’s circuit nonlinear vector field.6) where x = [x y z] ∈ R3 is the state vector. Aguirre.7) where the parameters p. L. (5.B.2kΩ R6 3. Tˆrres. A fundamental issue in any experimental platform for control or synchronization purposes is the design and construction of devices that will actually implement the decisions made by the controller(s). R10 1.6.M. x) in Eq.2kΩ R9 1. R.5µF (47µF 47µF) ICs TL08x (4 Op. (5. The PCChua setup implements the following vector differential equation ˙ x = F(x) + u(t.5kΩ R 2kΩ (multi-turn) C1 23.7) can be accomplished by including in the circuit controlled voltage and current sources (actuators).4) as x = p [y − x − f (x)] + ux (t.M. The control .A. (5. 5. In the case of the PCChua platform.5µF (47µF 47µF) C2 235µF (470µF 470µF) C3 23. as indicated in Fig. ˙ y = x−y+z ˙ ˜ + uy (t. the following describes the built-in actuators. x) = [ux uy uz ] ∈ R3 is the vector of control inputs.)1 1 Supply voltage: ±15V. Such devices are usually referred to as actuators. (5.

5. One solution to this problem is to use an independent power source to supply energy to the simulated inductor.7. In the PCChua. are necessary to minimize the effect of small nonzero voltage levels from the external reference voltages. x). inputs and the actuators are related by rx (t) = C1 ux (t. 5.2. using the electronic components listed in Table 5. .6. the realized inductor becomes a self-contained circuit with only two external terminals. RC2 1 ry (t) = uy (t. it is important to notice that there are alternative simulated floating inductor circuit topologies that can be employed. In Fig. the adjustable offset voltages. R C2 (5. x).7. x). However. With this modification. R L rz (t) = 2 uz (t. Chua’s circuit with actuators indicated. 5. The current and voltage sources have fixed gains: 1V/V for the controlled voltage sources and 1mA/V for the controlled current sources. obtained by varying R14 and R32 . 5.Chaos Control for Chua’s Circuits 105 R + Id (x) + x rx y ry + - z L Chua's Diode C1 C2 - - rz Fig.6 it is clear that the relative position of the voltage controlled voltage source rz in the circuit implies that a floating realized inductor (gyrator circuit) is necessary.8) From Fig. the actuators were implemented as indicated in Fig. An alternative is the use of a floating voltage controlled voltage source instead of a floating realized inductor. as pointed out in [45].

6). R.4). and to issue three control signals rx . 5. y and zy (Fig.M. 5. but the fastest rate at which data . 5. Mendes o R12 R17 From I/O board − R11 + R16 − + R18 rx +15V R13 R14 R15 −15V R19 (a) R21 R26 From I/O board R20 − + R25 − + R27 ry +15V R22 R23 R24 −15V R28 (b) R30 +15V From I/O board − R29 + Q1 + − Q2 rz +15V R31 R32 R33 −15V −15V (c) Fig. The multi-purpose board is used to measure the three voltages x. schematically represented in Fig. The external voltage sources are actually provided by a multi-purpose data and analog output board [73].8. L. Tˆrres. 5. Palhares and E.A. Aguirre. 5. which is driven by dedicated real-time software written for a modified version of the Linux operating system.106 L. Actuators circuits with fixed gain and offset adjustment.M. called RT-Linux. (a) Voltagecontrolled current source rx (R11 to R19 ).7.B.A. The control and synchronization algorithms are executed at fs = 1kHz.M. (b) voltage-controlled current source ry (R20 to R28 ). and (c) voltage-controlled voltage source rz (R29 to R33 ).1 and Fig.A. ry and rz to the actuators in the PCChua (Fig.

R30 . it is important to note that fs 1. R23 . Nevertheless. an example of the famous double-scroll attractor from Chua’s circuit. Chua’s Circuit Data Analysis and Modeling This section will describe a few results concerning the analysis and modeling from data of the double-scroll attractor from Chua’s circuit. R22 . R28 1kΩ R29 . 5.9. R12 . which indicates that continuoustime versions of control and synchronization strategies can be implemented with small effort. Electronic components used in the implementation of PCChua actuators.amps. due to hardware limitations. In Fig. 5. R15 4. .Chaos Control for Chua’s Circuits Table 5. R21 .)1 1 Supply 107 voltage: ±15V. can be acquired is fa = 100Hz. Components Value R11 .7kΩ R16 . is presented. Two of the DACs used to issue rx and ry have 10 bits of resolution and the third one (used for rz ) is limited to 8 bits.6Hz.3. R24 q 4.7kΩ R25 . Further details can be found in the cited references. RT Linux Data and Control signals I/O board PCChua Fig. obtained by using the PCChua experimental setup. 5. R17 . R27 . All the measured voltages in the PCChua are acquired using 12 bits of resolution.2. R13 . R33 4.7kΩ R14 . R18 .8. R31 . PCChua experimental setup overview. R19 1kΩ R20 . R32 5kΩ (multi-turn) Q1 BC547 Q2 BC557 ICs TL08x (6 Op. R26 .

h : Rm → R.A.M. a system is observable if the full state can be found based on s(t) only. with f : Rm → Rm .A. respectively. From s .5 -1 -6 -4 -2 0 x (V) 2 4 6 (a) 6 4 2 x (V) 0 -2 -4 -6 0 10 20 30 t (s) 40 50 60 (b) Fig.3. referred to as the observable and h. and s(t) = h(x). 5. (b) Time-series corresponding to the voltage across capacitor C 1 (x-variable) in Fig.108 L.M.1.9. Tˆrres. ˙ Let x = f (x). Briefly. Mendes o 1 0.M. R. as the measuring function.B.1. with h : Rm → R be a nonlinear system and a scalar observable. Data collected from PCChua during 60s.A. (a) Measured double scroll attractor from Chua’s circuit. 5. The Issue of Observability The key concept of observability is that of deciding if all the dynamical information of the full phase space Rm of an m-dimensional system can be retrieved by measuring one variable s(t) = h(x).5 y (V) 0 -0. Palhares and E. L. Aguirre. 5.

Rm (x) will be observable from s(t) if it is possible to go back from Rd (X) to Rm (x) on a one-to-one basis. Such difficulties seem to have been overcome in [4] where normalized observability indices... and we would expect to have “global” diffeomorphisms almost everywhere too. here denoted by the letter δ. s(1) . This order was confirmed by another observability test. As a consequence. Therefore. These results give the observability order x z y. This result. finally by y.Chaos Control for Chua’s Circuits 109 a reconstructed space can be formed.25 0. for instance by taking a sufficient number of successive derivatives X = (s(t).3 clearly point out that the case of the double scroll attractor is numerically badly conditioned and suggest that the best observable is x. are very small (≈ 10−15 ). Table 5. based on data. were defined. indicated by δ ∗ . that is. which is consistent with the fact that locally the system is linear almost everywhere. δz drops. some definitions of observability indices. Φs : Rm (x) → Rd (X). Basically. where s(j) is the jth time derivative of s [25].. ∗ ∗ ∗ whereas δx = δz = 1 remain.68 1 0.01 1 1 1 0. s(d−1) ). Increasing the tolerance even further. As the tolerance (a ∗ parameter used in the new definition) is increased δy quickly diminishes.02 1 0. there is Φs that maps the original phase space to the space reconstructed from the observable s.10 1 0 1 0. ∗ ∗ ∗ p δx δy δz 0. and in .3). if Φs defines a diffeomorphism.68 0 0 The results shown in Table 5. followed by z and. that is. . Another potential problem is that the application of some techniques to discontinuous systems is questionable. One of the difficulties with this system is that the great difference in the velocities of convergence and divergence along the stable and unstable manifold.3. Normalized observability ∗ indices δs for the double scroll attractor for increasing values of a tolerance. Observability of Rm (x) from s is therefore directly related to the existence of singularities in Φs . In this respect it would be interesting to answer the following question: which variable is best to observe the dynamics of Chua’s circuit? The analysis of the observability of the double scroll attractor is typically very challenging. thus giving the observability order x z y (see Table 5. The following results ∗ ∗ ∗ are obtained: δx = δy = δz = 1.

A. if they have random phase relationships such as a time series with Gaussian statistics).A.3. f3 ) = E[Af1 Af2 Af3 A∗1 +f2 +f3 ]. f2 . The Main Concepts The auto bispectrum is defined formally as the Fourier transform of the third-order correlation function of the time series. t2 (f1 . f3 . f2 ). 51]. 5. f (5.2. called the squared bicoherence. f2 . .1. Higher-order Spectral Analysis The results discussed in this section were published formerly in [33]. 5.M. Similar to the bispectrum.9) where Afi is the complex Fourier component of the time series at frequency fi .. The discrete bispectrum. f2 .9) are independent of each other (e.3. Palhares and E.g. The main objective of higher-order spectral analysis is to assess the presence and effect of nonlinearities in the frequency domain. R. appropriate for sampled data. Analogous results for other systems can be found in [50. It is convenient to recast the bispectrum into its normalized (by the power at each of the three frequencies in the triad) magnitude. f (5. f2 . L. is B(f1 .110 L. (5. f3 ). or average. Tˆrres. the auto trispectrum is defined formally as the Fourier transform of the fourth-order correlation.2. b2 (f1 . which represents the fraction of the power of the triad of Fourier components (f1 . and E[·] is the expected-value. f2 ) = E[Af1 Af2 A∗1 +f2 ]. f1 +f2 +f3 ) that is owing to cubic nonlinear interactions.M. Aguirre. the bispectrum is zero.B. and is a measure of the fraction of the power of the quartet of Fourier components (f1 . f1 + f2 ) that is owing to quadratic coupling.M. Mendes o particular the poorer observability properties of the observable y has been felt in the context of data sampling [2] and modeling [9] which is probably due to the fact that observing the system from y all symmetry information is lost. and the discrete trispectrum is T (f1 . If the three Fourier components in the triple product on the right hand side of Eq. asterisk is complex conjugation.10) The normalized magnitude of the trispectrum is called the squared tricoherence.A. operator.

Chaos Control for Chua’s Circuits

111

Higher-order spectral analysis can also be used as a means of model validation by comparing bicoherence and tricoherence functions of (measured) data and simulated data obtained by estimated models. 5.3.2.2. The Results Time series of voltage vC1 measured from a realization of Chua’s circuit for the spiral and double-scroll attractors were used to estimate nonlinear polynomial models which can be found in [33]. The circuit parameters used during the experiments were C1 = 11µF, C2 = 45µF, L = 20H, Bp = 1.1 ± 0.2 V, m0 = −0.37 ± 0.04 mS, and m1 = −0.68 ± 0.04 mS. For the spiral attractor (sampled at 12 µs) R = 1.67 kΩ and for the double scroll (sampled at 15 µs) R = 1.64 kΩ. The high inductance (Sec. 5.2.2) reduces the oscillation frequency from O(100 kHz) to O(1 kHz), thus facilitating A/D conversion and recording of the data. Polynomial model parameters were estimated from 8192 values of the voltage across C1 measured after initial transients had decayed. Both the measured and modeled spiral attractors have a relatively narrow power spectrum dominated by a primary peak fp = 2.92 Hz and its harmonics (2fp = 5.85 and 3fp = 8.78 Hz, Fig. 5.10a). The model bicoherence (Fig. 5.10c) indicates quadratic coupling between the primary and its harmonics (e.g., the horizontal band of contours for f2 ≈ 3 Hz, Fig. 5.10c), similar to that observed in the measurements (Fig. 5.10b). Quadratic interactions involving Fourier components with frequencies less than fp are significant in time series from both the data (1 < f2 < 3, f1 ≈ 3 Hz, Fig. 5.10b) and the model (Fig. 5.10c). Tricoherence spectra for the measured time series (not shown) are similar to those from the polynomial model. Cubic interactions between quartets of Fourier components are significant for the spiral attractor, suggesting higher-order coupling between motions with frequencies fp and its harmonics. The coupling between the primary and its harmonics, as indicated by the bicoherence (Fig. 5.10b,c) and tricoherence (not shown) spectra, suggest both quadratic and cubic interactions (as occur in the spiral attractor) are reproduced well by the polynomial model estimated from 8192 values of the voltage measured in Chua’s circuit. The higher-order spectral coherences from the model time series are slightly higher than those from the data, possibly because the model approximates the infinite-order piecewise linear current in the nonlinear resistor with a second- and third-order polynomial, leading to stronger quadratic

112

L.A.B. Tˆrres, L.A. Aguirre, R.M. Palhares and E.M.A.M. Mendes o

Fig. 5.10. (a) Power spectra of measured (solid curve) and modeled (dashed curve) voltage vc1 from Chua’s circuit when the system exhibits a spiral attractor. Contours of bicoherence of (b) measured and (c) modeled voltage vc1 . The units of power are arbitrary and the frequencies of the labeled power spectral peaks (panel a) are a) 2.92, b) 5.85, and c) 8.78 Hz. The minimum contour plotted is b = 0.85, with additional contours every 0.1. There are 32 degrees of freedom (from [33]).

and cubic coupling than observed in the actual circuit. Noise in the measured circuit may also reduce higher-order spectral values relative to those from the model time series. 5.3.3. Discrete-time Global Modeling The construction of discrete-time models from data has been thoroughly discussed in the literature. Specifically concerning the double-scroll attractor, [9] and [29] presented the case of NARMAX (nonlinear autoregressive moving average model with exogenous inputs) polynomial and rational models, respectively. 5.3.3.1. Monovariable Polynomial Model The polynomial structure is very attractive because of the simplicity and insight it offers of the systems properties and therefore have been extensively used for reproducing nonlinear dynamics.

Chaos Control for Chua’s Circuits

113

In order to identify dynamically valid models from data generated of Chua’s circuit, the NARMAX model [49] was chosen. The representation of such a model can be as follows y(k) = F [ y(k − 1), . . . , y(k − ny ), e(k). . .e(k − ne )] , u(k − d). . .u(k − d − nu + 1), (5.11)

where ny , nu and ne are the maximum lags considered for the output, input and noise terms, respectively and d is the delay measured in sampling intervals, Ts . Moreover, u(k) and y(k) are respectively the input and output signals. e(k) accounts for uncertainties, possible noise, unmodeled dynamics, etc. and F [·] is some nonlinear function of y(k), u(k) and e(k). The function F [·] can, of course, be a polynomial-type function with nonlinearity degree ∈ Z + . In such a case, to estimate the parameters of this map, Z Eq. (5.11) should be expressed in prediction error form as ˆ y(k) = ψ (k − 1)θ + ξ(k) , (5.12)

where ψ(k − 1) is the regressor vector which contains linear and nonlinear combinations of output, input and noise terms up to and including time k − 1. The parameters corresponding to each regressor are the elements of the ˆ vector θ. Finally, ξ(k) are the residuals which are defined as the difference between the measured data y(k) and the one-step-ahead prediction ψ (k − ˆ 1)θ. The parameter vector θ can be estimated by orthogonal least-squares techniques [76]. One of the many advantages of such algorithms is that the Error Reduction Ratio (ERR) can be easily obtained as a by-product [17, 48]. In the particular case of Chua’s circuit, one thousand data points of the output y(k), the voltage over capacitor C1 , were used in the identification. The following polynomial model was obtained:
y(k) = 3.5230y(k−1)−4.287y(k−2)−0.2588y(k−4)−1.7784y(k−1) 3 +2.0652y(k−3)+6.1761y(k−1)2 y(k−2)+0.1623y(k−1)y(k−2)y(k−4) −2.7381y(k−1)2 y(k−3)−5.5369y(k−1)y(k−2)2 +0.1031y(k−2)3 +0.4623y(k−4)3 −0.5247y(k−2)2 y(k−4)−1.8965y(k−1)y(k−3)2 +5.4255y(k−1)y(k−2)y(k−3)+0.7258y(k−2)y(k−4)2 ˆ −1.7684y(k−4)2 y(k−3)+1.1800y(k−4)y(k−3)2 + ψ ξ (k−1)θ ξ +ξ(k) ,

(5.13)

114

L.A.B. Tˆrres, L.A. Aguirre, R.M. Palhares and E.M.A.M. Mendes o

ˆ where ψ ξ (k − 1)θξ is the noise model composed of 20 linear terms of the form ξ(k − j) used to avoid bias. This part of the model was not used in the simulations shown below. The identification of model Eq. (5.13) and the cluster analysis of the double-scroll and spiral attractors have been discussed in detail in [9]. Figure 5.11 show the chaotic double-scroll attractor reconstructed directly from the original data whereas Fig. 5.12 shows the chaotic doublescroll attractor reconstructed using polynomial model Eq. (5.13). For comparison purposes the analysis of the results will be postponed to next section. 5.3.3.2. Monovariable Rational Model Another possibility for the function F [·] in Eq. (5.11) is a rational model defined as a ratio of two polynomials [16]
y(k) = a(y(k−1), . . . , y(k−ny ), u(k−1), . . . , u(k−nu ), e(k−1), . . . , e(k−ne )) + e(k), b(y(k−1), . . . , y(k−ny ), u(k−1), . . . , u(k−nu ), e(k−1), . . . , e(k−ne ))

(5.14) where u(k) and y(k) are as before, ny , nu and ne are the maximum lags of the output, input and noise, respectively. Moreover, such lags need not be the same in the numerator and denominator. a(k − 1) and b(k − 1) are polynomial functions nonlinear in the regressors taken up to time k − 1. It is convenient to define the numerator and denominator polynomials in Eq. (5.14) respectively as [16]
Nn

a(k − 1) =

j=1

pnj θnj = ψ n (k − 1)θn ,

(5.15)

Nd

b(k − 1) =

j=1

pdj θdj = ψ d (k − 1)θd ,

(5.16)

where θnj , θdj are the parameters of the regressors (of the numerator, pnj , and denominator, pdj ,) up to time k − 1. Nn + Nd is the total number of parameters to be estimated. The use of Eq. (5.14) to perform parameter estimation is not straightforward because such a function is nonlinear in the unknown parameters. An alternative solution to this problem is to multiply both sides of Eq. (5.14)

Chaos Control for Chua’s Circuits

115

by b(k − 1) and rearranging terms in order to yield [75]
Nd

y (k) = a(k − 1) − y(k)
Nn

j=2

pdj θdj + b(k − 1)e(k) pdj θdj + ζ(k)

Nd

=
j=1

pnj θnj − y(k)

j=2

= ψ n (k − 1)θn − ψ d1 (k − 1)θd + ζ(k) , where ψ d (k − 1) = [pd1 ψ d1 (k − 1)], θd1 = 1 and y ∗ (k) = y(k)pd1 = a(k − 1) pd1 + pd1 e(k) , b(k − 1) 
Nd j=1

(5.17)

(5.18)

where e(k) is white noise. Because e(k) is independent of b(k − 1) and has zero mean, it can be written E[ζ(k)] = E[b(k − 1)]E[e(k)] = 0 . (5.20)

ζ(k) = b(k − 1)e(k) = 

pdj θdj  e(k) ,

(5.19)

Equation (5.18) reveals that all the terms of the form y(k)ψ d (k − 1), because of y(k), implicitly include the noise e(k) which is correlated with ζ(k). This, of course, results in parameter bias even if the noise e(k) is white. The aforementioned correlation occurs as a consequence of multiplying Eq. (5.14) by b(k − 1) and should be interpreted as the price paid for turning a function which is nonlinear in the parameters to one which is linear in the parameters. Algorithms with intelligent structure selection can be found in [18] and [29], to mention a few. Using the same data as described above and also term clustering information [8, 9], the following rational model was obtained [29] y(k) = 1 × 2.5568y(k−1)−1.7594y(k−2)+0.2696y(k−5)+0.6192y(k−1) 3 D −1.0219y(k−2)3 −3.2455y(k−1)2y(k−5)+0.0735y(k−3)3 +0.3444y(k−1)y(k−5)2 −0.4401y(k−2)y(k−5)2
10 5

+3.4624y(k−1)y(k−2)y(k−5)+0.1986y(k−1)2y(k−2) +
i=1

ˆ θi ξ(k − i) +

j=1

ˆ θj ξ(k − j)2 + ξ(k) ,

(5.21)

(5.3350 ±0.24) 1. Bi-dimensional reconstruction (2000 points) of Chua’s double-scroll attractor.37) 1.22) (0.21) – Eq.3516 ± 0.B. (5.-2.37. (5.13) and rational model Eq. This should come as no surprise because. Aguirre.4 shows the estimated fixed points and largest Lyapunov exponents for both the original system and the identified models.21). (5.12 shows the chaotic double-scroll attractors reconstructed using polynomial model Eq. Table 5. (5.1948y(k−1)2 +1.1409y(k−5)2 −2.0563 Rational Eq.24.24) 1.2.8527y(k−2) 2 +0. Mendes o where D = 1+1.2. Dynamical invariants of data and identified models for Chua’s Circuit Data Fixed Points Largest Lyapunov Exponent (0. Table 5.M. (5.M.24.4218 ±0. . A window of one thousand points of these measured data were used in the identification. Such results also seem to indicate that the polynomial representation is slightly better as far as the modeling of attractors produced by Chua’s circuit is concerned. Palhares and E.116 L.11. L.A.4.A.0596 4 3 2 1 Vc1(k−4) 0 −1 −2 −3 −4 −4 −3 −2 −1 0 Vc1(k) 1 2 3 4 Fig. R.0343 Polynomial Eq.7662y(k−2)y(k−5)−0.5657y(k−3)y(k−5)−12.13) (0.M. the set of equations which describe Chua’s circuit can be expressed in polynomial form.5164y(k−1)y(k−2)+.22) Figure 5.-2. in fact. Tˆrres. The results suggest that both polynomial and rational models do reproduce the characteristic dynamical features of the original attractor.-2. 5.A.0470y(k−1)y(k−5) .2.

hereafter. five derivative coordinates were used to obtain a global model. which coincides with the dimension of the original phase space. Such an embedding will permit the comparison of the state portrait reconstructed from the experimental data with the portrait generated by the model. 5. such a model will be referred to as the full model.12. 5. Continuous-time Global Modeling The results of building continuous-time models from discrete-time data obtained from Chua’s circuit is briefly described in this section.4. Double-scroll attractors reconstructed from (a) simulation of the polynomial model (Eq. The x-axis is y(k) and the y-axis is y(k − 4).3.2. therefore. Conversely. These results have first appeared in [6]. (5. 5. 5. x.3.3.13)) and (b) simulation of the rational model (Eq. the models obtained implementing the structure selection procedure described in [6] and references therein will be referred to as parsimonious models in the sense that all such models will have less than 38 terms.1. Topological Characterization The chaotic attractor is embedded in a space spanned by the successive derivatives of the experimentally recorded time series.4.Chaos Control for Chua’s Circuits 117 4 4 3 3 2 2 1 Vc1(k−4) Vc1(k−4) −3 −2 −1 0 Vc1(k) 1 2 3 4 1 0 0 −1 −1 −2 −2 −3 −3 −4 −4 −4 −4 −3 −2 −1 0 Vc1(k) 1 2 3 4 (a) (b) Fig. The parsimonious models were obtained by means of structure selection techniques. The Parsimonious Models A 38 term model was obtained from the training time series without any structure selection. .21)).4. The phase space reconstructed from the recorded time series has an embedding dimension equal to 3. However. (5.

14 presents two critical points.13. The firstreturn map shown in Fig. X3 ) ∈ R2 | X1 = −1.9.0 -16. (5.0 4. 5. Palhares and E.0 Y -6. ¨ (5.M. X2 ).0 2.A. only the first three coordinates will be used to characterize the topological structure of the reconstructed state space. The analysis of the dynamics starts with the computation of a firstreturn map obtained from a Poincar´ section. The fact that this map does not exhibit any layered structure.0 14. every point in the time series may have a counterpart on the opposite side of the attractor. Mendes o Since the original dynamics can be embedded in a 3D space. X = x and Y = x.0 -3.M.118 L. R.B.0 -1.13. 5. Plane projection of the x-induced attractor generated by the experiments. Tˆrres.A. This can be viewed as a first-return map to a Poincar´ set defined as e (X2 .0 -26.0 -2. 24.M.23) The chaotic attractor is displayed in Fig. X3 ) ∈ R2 | X1 = 1.9. In the case of a symmetric e attractor. X2 < 0 ∪ (X2 . Therefore.0 3. a fundamental map is used.0 Fig. Aguirre. 5. X2 > 0 . Consequently. the reconstructed attractor should also present an inversion symmetry [53]. ˙ . the first three coordinates are X1 = x(t) x = X2 = x ˙ X3 = x .0 X 1. confirms the hypothesis of symmetry.24) built on the absolute value of one of the two variables (X1 . L. Since the system presents an order two inversion symmetry.A. This information will be used at the end of this section and will prove to be useful in enhancing model quality.0 0.

−x.0 Yn+1 C2 12. X5 ]T .5 2 10.16.0 7.3. that is for each point in the differential embedding space.5 7. .Chaos Control for Chua’s Circuits 20. please refer to [52–54]. In order to help the algorithm account for the order two inversion.4.0 Fig.0 12. 5. A further step was to compare the fundamental first-return map computed for attractors generated by the models with the one associated with the experimental data. Fundamental first-return map of the attractor reconstructed from the experimental data. if the original data set had the point x = [X1 X2 . Model Validation Figure 5. 5.5 Yn 15. The layered structure of the first return maps of models with 36 and 38 terms could be an indication of overparametrization [3]. as shown in Fig.5 10. then the mirror image of such a point. it seems fair to conclude that structure selection has been able to reduce the complexity of the full model resulting in parsimonious models with improved dynamics. For details on this procedure.5 1 0 15.3.5 20. For instance. the system becomes unsymmetrical [54]. was inserted in the data set . the first-return maps of models with 29 and 32 terms are much closer to the experimental one and despite the differences.0 17. As a result of this.14. 5. On the other hand. Models with 36 and 38 terms are characterized by fundamental first-return maps that have a layered structure which is not observed in the experimental data. a symmetrical counterpart was included in the data used to estimate the model. . a new model was obtained using a measure which respects the symmetry of the attractor.0 119 C! 17.15 shows the projections of the attractors of some estimated model.

120 L. the identified model still reveals difficulties. the topology of part of the attractor is different. 5.15.A. as seen in the experimental map (Fig. 5. reveal some differences between the model and the dynamics underlying the ex- . the four ˙ parsimonious models and the full model (Np = 38).17. Palhares and E. Aguirre.A. An Np = 20 term model was found.16). R. this has important consequences in the resulting population of periodic orbits. 5. This is a signature of perfect symmetry as pointed out in [54].M.17 does not coincide with the x = y line as observed in other models (Fig. 5.M. L.M. that is. for the Np = 20 model the location of this maximum in relation to the x = y line is very similar to that of the experimental dynamics (Fig. It should be mentioned that this model possesses an important property of the experimental flow that is not shared by the other models since all the models mentioned in Fig. Secondly. the half sum of oriented crossings are computed.16). The first return map of which is shown in Fig. The first-return map does not present the layered (ghostly) structure as before. Plane (x × x) ≡ (X1 × X2 ) projections of the original attractor.16). the maximum of the first return map shown in Fig. 5. Tˆrres. The linking numbers. before choosing the centers.16 have symmetric fixed points but the related flow is not symmetric due to the presence of quadratic terms of the form Xi Xj (∀i.B. As a matter of fact. rather than decreasing.A. j = 1). 5. Because of the third branch which is increasing. 5. Although improvement in some sense can be observed. Mendes o Experimental Np=18 Np=29 Np=32 Np=36 Np=38 Fig.

Interestingly. it visits the two scrolls in the same way. This enabled collecting the data with a sampling time of Ts = 30 ms.3.18(a)) and for the 20 term model (Fig. considers data collected from PCChua. Such an example confirms that the experimental and the 20 term model attractors are not topologically equivalent although both attractors are rather difficult to distinguish by simple visual inspection.4. 5. . 5. 5. Experimental Np=18 Np=29 Np=32 Np=36 Np=38 Fig. 5. 10) is equal to -2 in both cases but it is noticed that the period-2 orbit of the model is a symmetric orbit. which is not observed for the experimental dynamics. 5.16.6 Hz [73].18(b)). the first-return map of a Chua-type circuit with a smooth nonlinearity [44] also presents a similar increasing branch. A likely reason for such a difference is the third increasing branch in the first-return map of the model. first presented in [10].17(b). with the frequency of the main spectral peak around 1. however. First-return maps for the five different models. the pair of periodic orbits encoded by (200) and (10) are displayed for the experimental dynamics (Fig. Multivariable Polynomial Modeling The following example. For instance. that is. The linking number lk(200.4. whereas it is asymmetric and confined in to a single scroll for the dynamics underlying the experimental data. as shown in Fig.Chaos Control for Chua’s Circuits 121 perimental data.

2 5) As before (Sec.35 −0.3 9 8 09 7 7 03 6 7 5 ×1 0+5 y(k − 1 )z 2 (k − 1 )     + 0.25) cannot produce a mathematically symmetrical attractor because it includes even parity terms [54].8 9 5 2 5 2 009 8 6 6 y(k − 1 ) − 0. The automatic structure selection scheme followed by an orthogonal extended least squares estimation routine yielded the following model  +2 z(k − 1 )  x(k) = 1 .M. Palhares and E.1).50 0.00 0.3 1 5 3 6 3 09 3 5 7 5 ×1 0−5 x2 (k − 1 ) + 0.2 103 6 2 03 2 06 x(k − 1 ) + 0.1 8 7 6 1 9 3 2 1 5 3 7 ×1 0−2 x(k − 1 )z(k − 1 ).6 3 2 8 6 8 3 4 5 4 4 1 ×1 0−5 y 2 (k − 1 )     −0.25 0.5 0 1 Yn+1 0. Phase portrait (a) and the corresponding first-return map (b) generated by the symmetrical model.9 3 8 2 6 5 05 7 18 6 z(k − 1 ) − 0.25 0.5 0.75 −3. only the deterministic part of the model is shown.25) multilayer perceptron (MLP) .5 1.55 0. Np = 20.75 0.5 3.3 0.2 04 05 04 4 6 8 1 2 ×1 0+1 x(k − 1 )y(k − 1 )z(k − 1 )    z(k) = 0 .M. 5.6 0.A.1 7 1 3 1 5 9 2 4 8 5 8 ×1 0−1 x3 (k − 1 )      + 0. It is important to notice that model Eq. (5.9 7 2 04 7 1 8 3 7 3 3 ×1 0   + 0. Tˆrres.4 0.5 −1. The data used to build the model are shown in Fig.6 7 7 2 1 4 8 1 7 4 3 2 ×1 0−3 x3 (k − 1 )     + 0.B.3 6 2 1 3 4 02 01 8 5 ×1 0−4 x(k − 1 )     + 0.A.5 −0.65 X Yn (a) (b) Fig. 5.35 0.25 0.6 3 7 4 8 5 6 8 4 2 2 0×1 0−3 y(k − 1 )      −0. R.1 1 1 07 2 5 6 3 5 09 ×1 0+3 y 2 (k − 1 )z(k − 1 )     y(k) = 0 .4 Y −0.5 2. Besides the polynomial model Eq.3.50 0. This model has twenty terms. Aguirre.65 0. (5.2 2 3 3 03 6 8 1 1 1 9 ×1 0+6 z 3 (k − 1 )    + 0.5 0.55 0. L.1 2 9 3 2 5 1 8 03 3 3 ×1 0+3 z(k − 1 ) − 0.5 0.122 L.19.3. When symmetry conditions were imposed during modeling. the new model did not settle unto the double scroll attractor and state estimation results did not improve.A.17.6 4 3 3 1 8 6 8 2 3 1 2 ×1 0−5 x(k − 1 )y(k − 1 )   + 0.08 3 7 1 4 3 4 9 2 9 4 7 x(k − 1 ) + 0.M.5 0.6 0.25 2 −2. (5.1 8 1 08 5 2 6 8 7 8 3 ×1 0−3 x2 (k − 1 )      −0.3 −0.5 9 8 2 2 1 2 8 3 6 7 5 y(k − 1 )     −0.8 1 9 5 9 3 3 5 4 8 5 6 ×1 0+4 x(k − 1 )z 2 (k − 1 )     −0. Only 1000 observations of each state variable were used to build the model. Mendes o 0. 5.45 0.45 0.

0 (10) (200) negative crossings 0.18.5 3.75 0.0 -15.0 0. and (b) model dynamics. The linking numbers between periodic orbits encoded by (200) and (10) is equal to -2 for both the experimental dynamics and the 20 term model dynamics.0 x 1.0 -0.0 2. 5. The first set was composed of three fully connected networks with two nodes in the hidden layer and the second was composed of three fully connected networks with seven nodes in the hidden layer. The activation functions of the nodes in the single hidden layer were hyperbolic tangent and the single output node was linear in all three networks.0 1.0 y -5.0 5. The input layer in each case was the same.0 15. one for each state variable.25 −0.0 (10) (200) negative crossings 0.Chaos Control for Chua’s Circuits 123 25.0 -25.00 −0.50 0.5 1.75 0. (a) Experimental dynamics.50 −0. Such neural models will be referred to as MLPs2 and . Three networks were trained. Nevertheless. z(k − 1)] .0 2. the period-2 orbit does not present the same symmetry properties.25 Y 0.5 (a) 0. neural networks were trained from the same set of data.5 3.0 X (b) Fig.5 1.5 2. y(k − 1). In this section we discuss the results obtained by two set of networks. namely [x(k − 1).5 2. See text for a possible explanation.

as described in the following paragraph. although the model might display improved global features. Briefly. Palhares and E.92427441 is used. . For instance.25) in order to obtain the double-scroll attractor.19. The measured double scroll attractor of the implemented electronic oscillator. R.M. MLPs2 and MLPs7— could be competitive in the implementation of a Unscented Kalman Filter – UKF [40].20(a). (5. it was desired to verify if models with improved global dynamics would outperform locally optimal models.20(b)a . Tˆrres.A.M.25) settles to the attractor shown in Fig. Mendes o x 10 5 4 3 2 −3 z(k) (mA) 1 0 −1 −2 −3 −4 −5 1 0.25) settles to the double-scroll attractor shown in Fig. 5.5 1 2 3 4 5 y(k) (V) −1 −5 −4 −3 −2 −1 0 x(k) (V) Fig. Aguirre. the farther from the optimal set of parameters) the worse the performance of the UKF in estimating the states. MLPs7. In the present case this is achieved by slightly increasing the underlined parameter.B.e. (5.124 L. model Eq. Nevertheless. (5.A. if 0. L. it has been argued that the model can be perturbed (this can be done in different ways) in order to become chaotic [1].M. a Other parameter values were tested for perturbing model Eq. Such an attractor seems to be close to a genuine solution of the system in the sense that it could be the “stable version” of one of the unstable periodic orbits that compose the original attractor.25). 5. The free-run of model Eq.A. It was believed that all the estimated models— Eq. (5. respectively.5 0 −0. (5. In order to investigate this issue slight changes were implemented in model Eq. 5. (5.25).25) a posteriori. There are the data used to build model Eq. When this is the case. In what follows we shall refer to this as the perturbed double-scroll model. the greater the disturbance (i. MLPs2 also settled to limit cycle and MLPs7 showed a very long chaotic transient— consistent with the double-scroll attractor— before settling to a spiral strange attractor. However.

with initial conditions x(0) = 3.009.5 0 −0. iii) using Eq. (5. The following situations were tested: i) using Eq.25) as obtained from the modeling step. (5.Chaos Control for Chua’s Circuits 125 x 10 5 4 3 2 1 0 −3 z(k) −1 −2 −3 −4 −5 1 0. The UKF with the identified model Eq.37 and −0. (5.5 1 2 3 4 5 y(k) −1 −5 −4 −3 −2 −1 0 x(k) (b) Fig. ii) slightly perturbing the underlined parameter of Eq. y(0) = 0.3.25) in order to approximate the resulting attractor to the original double scroll attractor. 5. iv) the same as in (i) but using a null process noise covariance .25) was used to estimate the full state vector of the electronic circuit using only the x component to drive the filter.5 1 2 3 4 5 y(k) −1 −5 −4 −3 −2 −1 0 x(k) (a) x 10 5 4 3 2 1 0 −3 z(k) −1 −2 −3 −4 −5 1 0.5.2. From [10]. State Estimation The results in this section were first presented in [10]. Attractors produced by identified model (a) without and (b) with parameter perturbation. (5.25) but jointly estimating the three parameters in italic (which correspond to the terms of each equation with higher ERR index).5 0 −0. 5.20.

86% (vi) MLPs7 The very low values of RMS for the estimation errors of the x component are a consequence of low measurement noise in PCChua experimental setup (Sec. It is important to declare that in the aforementioned situations. R.M. Tˆrres. Aguirre. situation investigated x y z (i) model 2.3) and due to the fact that the x variable is used to drive the filter.25) — in several contexts (see text)— and with the neural models MLPs2 and MLPs7. 5. This problem can be circumvented by setting a lower bound for the covariance matrix P by properly setting the process noise covariance matrix Q.85% 2. see Fig. Q was taken as a diagonal matrix with elements equal to the onestep-ahead prediction error variance of the respective modelb . Normalized root mean square (RMS) error of estimated states for the electronic system. In so doing the filter remains more active and the most recent measurements gain more weight compared to the model and as a result the UKF settles to a doublescroll attractor.03% 1. (5.33×10−7 % 1. Table 5. closer to the original system (for instance. L.06% 1. In other words. In case (iii). From [10].M. 5.46% 76. When the UKF was implemented using Eq. in a sense.2.70×10−7 % 3. quite naturally.A.A. This explains the very high values of RMS for situation (iv) shown in Table 5. the UKF estimates also settled to the same attractor (possibly just a limit cycle) shown in Fig.75×10−7 % 5.B. both have a positive Lyapunov exponent).87% (ii) perturbed model 3.5. .86% 3. An interesting remark that can be made based on the results of Table 5.5 is that although the perturbed model is. except in (iv). when convergence was achieved. where three parameters were also estimated.20a. The results are summarized in Table 5.5.27% 0.18% (v) MLPs2 9.A.21(a).24% (iii) joint parameter estimation (iv) model with Q = 0 88. (5. Otherwise.58% 0. the trace of matrix P would tend to zero. the unperturbed model— which has a lower one-step-ahead prediction error— is better suited for use within the UKF. 5. Palhares and E.M. the elements of Qθ were taken as the corresponding parameter variances obtained as a byproduct from the extended least squares estimator used during modeling step.46% 2. there are two important b Obtained by one-step-ahead simulation using each model.25) then.126 L. The UKF was implemented with the identified model Eq.5.46% 85.80×10−7 % 2.57×10−7 % 2. Mendes o matrix Q.

22.08 3. that is.93 12.42 MLPs2 10% 3. Table 5. the trace of R ought to be directly related to the degree of uncertainty in the measurements. Consequently situations (i) and (iii) have close performance indices in Table 5.82 17. 5.6.66 5.63 2.37 Polyn 25% 4.25) and with the neural models MLPs2 and MLPs7.75 .13 MLPs7 50% 2.20 18. Similarly. the UKF implemented with the neural model with 7 hidden nodes is significantly more robust to noise than the filters implemented with the other models.84 Polyn 50% 17. in which case the predictions are all we have. model/noise x (%) y (%) z (%) Polyn 10% 1.6 summarizes the UKF performance when extra white Gaussian noise is added to the x component in the case of polynomial and neural models. (5. on the other hand.Chaos Control for Chua’s Circuits 127 sources of information in the filter: on the one hand.37 1. there are the measurements. the trace of Q should be increased.08 MLPs2 25% 8. was corrupted with extra additive white Gaussian noise.48 5. poor performance in predicting the driving signal is not as serious as poor performance in predicting the remainder of the state vector. As can be seen.84 MLPs2 50% 25.01 2.6 because the filters are driven by the x variable. there is the model that propagates the sigma points and. It should be noticed however that this weakness of MLPs7 does not appear in Table 5. The UKF was implemented with the identified model Eq. the estimated parameters converge to the values obtained during the modeling stage as indicated in Fig.20 12.88 1.55 MLPs7 10% 1.97 3.35 6. Suitably defining Q and R is a way of weighting differently these two sources. The x-component.57 3.5.5 where MLPs7 has the highest error for the x variable.54 MLPs7 50% 1. At first sight this could seem inconsistent with the results shown in Table 5. So as to express lack of confidence in the model. Normalized root mean square (RMS) error of estimated states for the electronic oscillator. that drives the filter. When parameters are estimated in addition to the states.09 2.56 27.95 5.53 1. Table 5.

This is explained by the simultaneous conjugation. The estimated double scroll attractor of the implemented electronic oscillator.M. z(k) (mA) 1 0 −1 −2 −3 −4 −5 0 5 10 15 t = kTs (s) (b) Fig. In the plot (a) the transient at the beginning ˆ was also included to give an idea of how quickly the filter settles. 5. Chua’s Circuit Control and Synchronization The control (and synchronization) of Chua’s circuit has been chosen as a benchmark in nonlinear control of dynamical systems by many researchers around the world.B.A. Aguirre. 5. on Chua’s circuit. Tˆrres.128 L. L.5 1 2 3 4 5 y(k) −1 −5 −4 −3 −2 −1 0 x(k) (a) 5 4 3 2 x 10 −3 z(t). (5. Palhares and E.A. simple construction. (b) detail of the z component (-·-) measured data. From [10]. of complex behavior. Estimation was performed by the UKF with model Eq.4. Mendes o x 10 5 4 3 2 1 −3 z(k) 0 −1 −2 −3 −4 −5 1 0.5 0 −0. (a) Full state vector.M.M. R.25) and the x component driving the filter. and (—) estimated component z (k).A. and manageable mathematical .21.

. it is possible to register: feedback linearization [55].1 1 0 5 10 15 20 25 30 (b) 1 0. sliding mode [39]. From [10]. As a consequence.8 0. As a sample of modern nonlinear control strategies used to control Chua circuits.-) indicate the values for these parameters estimated by extended least squares during the modeling step. or one of the many variations of this famous chaotic oscillator. description composed by a set of only three piecewise affine differential equations.8 0 5 10 15 20 25 30 t = kT (s) s Fig.6 0. to control and/or to synchronize Chua’ circuits has become a paradigm in nonlinear control systems theory. The dotted lines (·) indicate plus and minus one standard deviation (± diag{Pθ }) of the parameters estimated by joint UKF.4 2 θ 0 5 10 15 20 25 30 (c) 1.2 1 θ 3 0.25). among others.2 1.3 1 θ 1. (5. The dashed lines (. The solid lines (—) show the results obtained from the joint estimation of the three parameters in italic in Eq. a sample of results from the application of different control and synchronization strategies to the circuit of Chua will be presented. backstepping [74]. and fuzzy control [23].Chaos Control for Chua’s Circuits 129 (a) 1.4 1. 5. The aim is to show how Chua’s circuit can be employed to draw deep conclusions on the more general problem of controlling nonlinear systems. In this section.22.

General Considerations – Control Objective. which are the fundamental bricks in the system chaotic attractor.g.130 L. R.A. This is indeed verified in the famous OGY chaos control method [61]. with which one would be capable of explaining the plethora of phenomena found in these systems. e. when the control objective is to force one or more state variables of the chaotic oscillator to follow an arbitrary reference. On the other hand.4. in the problem of synchronization of unidirectionally coupled identical nonlinear oscillators. The lack of a universal theory. simple and/or low energy controllers could be employed. almost prevents the achievement of general results. one expects that simple and low energy controllers would suffice. Mendes o 5. one expects that complex and/or high energy controllers would be necessary.B.M. such that a slave system has to be controlled in order to follow it. 66] and references therein). 64.e. as it will be seen in Sec.2.4. as it will be presented in Sec. if one considers that this is indeed a control problem where the reference signal is the trajectory produced by the so-called master system. and (iii) the controller structure complexity itself. as powerful as the well known Linear System Theory [27]. although not obvious. In this case. Palhares and E. Another example is the stabilization of unstable periodic orbits (UPO). by taking into account a number of aspects such as: (i) maximum control effort allowed (energy limits). and therefore are close to the natural behavior of the uncontrolled system. whose variations have been applied to Chua’s circuit in [23] and [34]. (ii) the dissimilarities of the control objective (goal dynamics) when compared to the natural behavior of the uncontrolled system.A. a sinusoidal signal. The tradeoff among these aspects can be clearly illustrated on the control and synchronization of Chua circuits. the demanded energy could be reduced.4. 5. Despite this fact. This is a consequence of the fact that the control .A. which has been approached from many different points of view (see [43. in [12] it was possible to elaborate broader conclusions on the performance of controllers for nonlinear dynamics. if the reference signal is constant at a specific value such that the problem becomes one of stabilization of unstable fixed points of Chua’s circuit.4. L.1. Aguirre. For example. Available Energy and Mathematical Modelling The control of nonlinear dynamical systems is undoubtedly a vast field of research. the necessity or not of accurate models of the system to be controlled in the design of the control law. i. Similarly.M. and the controller could be made simpler than the one used in the former case.M. Tˆrres. 5.

27) (5. as shown in Fig. to the perturbation (or information) p(t) signal injected at the master oscillator. after synchronization has been reached. as long as y(t) → x(t). This will be discussed in the next section.23.4.1. f : R → R .28) where x ∈ R is the master oscillator state vector. p(t) Master x sx (t) η (t) Controller u(t) Slave y sy (t) Fig. . pn (t)] is a vector of information signals. 5. p(t) and u(t) are the perturbation and the control signals. the control effort should vanish in the absence of noise and perturbations. n n n (5.Chaos Control for Chua’s Circuits 131 objective is to force the slave system follow one of its natural trajectories after synchronization has been attained.23. sx (t) and sy (t) are measured signals from the oscillators. p(t) = [p1 (t) p2 (t) . . 5. and in [69] for discrete-time systems. x(t) and y(t) are the state vectors of the master and the slave nonlinear oscillators.1. in the absence of noise η(t). ˙ sx = h(x) + η(t). 5. a general principle called Information Transmission Via Control – ITVC was presented and proved. ˙ ˆ y = f (y) + u(t). where it is allowed . The Information Transmission Via Control principle [72] states that u(t) → p(t). Consider the following information transmission system based on additively perturbed nonlinear oscillators: x = f (x) + p(t).26) (5. The ITVC Principle Interesting considerations naturally arise from the close relationship between control and synchronization of nonlinear oscillators. which tend to synchronize by means of the control action u(t). Unidirectionally coupled systems or master-slave configuration. which relates the control effort u(t). demanded to keep the systems synchronized. In [72] for continuous-time systems. If the goal trajectory belongs to the library of natural behaviors of the slave then. respectively.

A. L. i.A. In addition. high-energy controllers should not be used to recover p(t). .30) where · denotes the Euclidean norm. .32) (5.e. When this occurs. (5.B.M. . respectively. 2. . In addition. ∀x ∈ Rn . one has the following result: t→∞ lim e = 0 ⇒ lim p(t) − u(t) = 0.1.27) ˆ oscillators (f (x) = f(x). An important fact. Aguirre. R. (Information recovery when both systems are not identically synchronized). ˙ as long as E → 0. Palhares and E. Define the synchronization error vector as e = x − y.26) and Eq. .M. f : Rn → Rn . 0 (5. it is possible to write that ed (t) ≤ e(t) + ρ(y)E + φ(E 2 ) ˙ ed (t) ≥ e(t) − ρ(y)E − φ(E 2 ) ˙ (5. suppose that 0 ≤ e ≤ E < ∞.29) Suppose that the systems Eq. y ∈ Rn is the slave ˆ oscillator state vector. was also proved in [72]. n}. (5. the following can be established: Theorem 5. ∀x ∈ Rn ). More specifically. h : Rn → Rq is a general mapping. then ed (t) → e . f (x) = f(x). as can be deduced from the following result: Theorem 5. ηq (t)] is a vector of noise signals. and u(t) = [u1 (t) u2 (t) . t→∞ (5. identical synchronization ( e → 0) is not possible due to unavoidable noise and parameters mismatch. (ITVC – Information Transmission Via Control Principle). . Mendes o to have pk (t) = 0 for some values of k ∈ {1. Tˆrres. un (t)] is a vector of control signals.M. .132 L. sx ∈ Rq is the vector of transmitted signals.2.27) have C (continuous) ˆ ˆ vector fields f (x) and f(y). (5.31) . In practice. (5. Considering the recovering error defined as ed = p(t) − u(t). but instead it remains limited: 0 ≤ e ≤ E < ∞. η(t) = [η1 (t) η2 (t) . also represented as continuous functions of time. such that the synchronization error does not go to zero. In this case. such that q ≤ n. concerning the general result of the ITVC principle.26) and slave Eq.A. the two autonomous oscillators are identical. Under these assumptions. Suppose that f : Rn → Rn is a C ω (analytic) function representing the vector fields of both master Eq. .

Similar remarks concerning the fact that low-energy controllers should be used in this validation approach (due to the impossibility of achieving perfect synchronization between real data and identified model) were also proved in [69].4. the theorem clearly establishes that it is possible to achieve a dynamic condition arbitrarily close to the identical synchronization ( e ≈ 0).Chaos Control for Chua’s Circuits 133 y y where ρ(y) = Df is the norm of the Jacobian matrix Df of f (e + y) evaluated around e = e0 = 0. In Sec. Linear State Feedback In this section. i. The formulation presented here can be viewed as an extension of the work in [70] where the focus was the controller structure design. and the perturbation signal is generally kept small ( p(t) ≈ e ) in order to not destabilize the master oscillator. The objective is to provide a more general method to determine the non zero elements in the controller gain matrix in order to issue stable controlled systems.2. has also been used as the basis for model validation procedures that rely on the control effort spent to synchronize identified discrete-time models with the corresponding acquired data set from a real system [7]. 5.e.2 an example of this situation will be discussed. The function φ(E 2 ) : R → R+ represents high order terms in E that vanish when E → 0. if the controller structure is such that it results in e ˙ 0. The ITVC principle. the control and synchronization of Chua’s chaotic oscillator is investigated based on the design of linear state feedback controllers. one sees that the error ed (t) in recovering the perturbation (information) approaches the derivative of the synchronization error e. limE→0 φ(E 2 )/E = 0. as stated in [69]. because usually it is observed negligible mismatch between master and slave oscillators ˆ ( f (x) − f (y) ≈ 0). Therefore. A method for selecting the controller gain matrix is proposed based on the use of the optimization technique known as the ellipsoid algorithm [31].4. Such controllers.4. the control laws that lead to the condition e = ˙ ˆ f (x) − f (y) + [p(t) − u(t)] 0. i. which depends on the value of y. By applying the above theorem. Under these assumptions. it follows that e ˙ 0 ⇒ u(t) 0. 5. . and still do not recover accurately the original perturbation signal p(t). are called in this context “high-energy controllers”.e.

Note that the derivatives were taken relative to the normalized time variable t = t /RC2 . Aguirre. 5. +Bp ].1).2.36) (5. (5.M. but substituting z ˜ by z for clarity purposes: x = p [y − x − f (x)] + ux (t. Palhares and E. The control inputs ux (t).7). if x ∈ [−Bp . (5. Consider the controlled circuit of Chua (Fig. +Bp ].34).2). It is shown that the synchronization problem is similar to the equilibrium point stabilization problem concerning the linear controller design requirements. Mendes o The connection between the control and synchronization problems is explored in detail.A.B. the matrix Ax is the Jacobian matrix of the nonlinear vector field of the autonomous chaotic oscillator. x).37) .A. z = −qy + rz ˙ + uz (t. (5. x).34) where x = [x y z] is the state vector. The nonlinear function f (x) = Rid (x) in Eq. if x < −Bp . (5. (5. where t represents the actual time (Sec. if x < −Bp or x > +Bp .33) is proportional to the current through the Chua’s diode. uy (t) and uz (t) in Eq. Tˆrres. 5. the reference vector is denoted by xref (t) = [xref (t) yref (t) zref (t)] and K ∈ R3×3 is the controller gain matrix. A2 ∈ R3×3 . Using the Chua’s diode characteristic Eq. assuming linear state feedback control.   b1 = pR[Bp (m0 − m1) 0 0] . which is given by Eq. page 100. repeated here. if x > +B . R. (5. page 105) described by the set of differential equations Eq.33) in matrix form as ˙ x = Ax x + bx + K (x − xref ) where Ax = A1 ∈ R3×3 .6. (5.M. it is possible to rewrite Eq. ˙ y = x−y+z ˙ + uy (t. (5. Defining the error vector as e = x − xref . (5. (5.33).2) together with Eq. x). can be rewritten in vector form as u(t) = [ux uy uz ] = K(x − xref (t)).33) where p = C2 /C1 and q = −R2 C2 /L.A.  bx = b2 = [0 0 0] .   b = pR[B (m − m ) 0 0] .35) Actually.134 L.M. 3 p 1 0 p (5. if x ∈ [−Bp . L.

which in turn become smaller for greater values of |λc |. (5. the smaller the maximum steady-state error. h(t) = exp [Ak (t − t0 )] . Note that. . This qualitative analysis was carried out to stress that the larger the absolute values of the real parts of the instantaneous eigenvalues of Ak in Eq.39) (5. This effect can be obtained by means of gain augmentation.39). δ(t) . (5. (5. ·] denotes the linear convolution operator. (5.43). the time evolution of the error vector inside each linear region is given by [27] e(t) = exp [Ak (t − t0 )] e(t0 ) + conv [η(t) . in such a way that the expression Eq. (5. which is a well known result from linear control theory. Ak = Ax + K. through the use of large values in a diagonal matrix K. ˙ η(t) = Ax xref (t) + bx − xref (t). the greater the value of |λc | the faster the decay of the components in the vector function h(t). the error behavior can be approximated by e t |λc |−1 ≈ conv η(t) . (5.40) Assuming that it is possible to adequately choose the gain matrix K in Eq.41).g.Chaos Control for Chua’s Circuits 135 it is possible to derive the following differential equation describing the error dynamics as a linear time-variant non-autonomous system represented by de dxref = Ax (e + xref ) + bx + Ke − .41) where t0 corresponds to the time instant when the system enters the linear region. λc = min |real{λ(Ak )}|.38) (5. λc < 0. dt dt ˙ e = Ak e + η(t). from the above assumption of asymptotic stability. as can be seen from Eq. exp [Ak (t − t0 )]] (5. (5.42) becomes similar to e(t > T ) ≈ conv η(t) . (5.39) so that the instantaneous eigenvalues of Ak have negative real parts. and conv [· . e.43) where δ(t) is a vector of impulsive functions whose magnitudes are proportional to the areas under the graphs of the h(t) components. despite the complexity in Eq. h(t) .42) Moreover.

M. (5. (5.44) constitutes the very nonlinear vector field of an autonomous Chua’s circuit whose state vector is xref . The most important requirement concerning the matrix K selection is to guarantee an asymptotically stable matrix Ak in Eq.47) .39) can be softened due to the nature of the control objective. (5. Applying the Routh criterion [27] on the coefficients of the characteristic polynomials det [(A1 + K) − λI] = 0 ⇒ λ3 + α2 λ2 + α1 λ + α0 = 0. Suppose that the control objective is to stabilize an equilibrium point peq of the chaotic oscillator.39). R.B. (5.44) becomes  xref = peq = [xeq yeq zeq ] .e. In this case.46) The above result reveals that there exists a strong similarity between the Chua’s circuit equilibrium point stabilization and synchronization problems. (5.136 L.A. Palhares and E.44) Note that the first two terms in Eq. (5.40) can be rewritten as ˙ η(t > T ) = Axref xref (t) + bxref − xref (t). the Eq. (5. where I ∈ R3×3 is the identity matrix and det[·] denotes matrix determi- (5.   Axeq peq + bxeq = 0. Tˆrres. (5. xref = 0. reducing the error vector norm in order to guarantee that the state vector x and the reference state xref will be in the same affine region after a time interval T .38) for all t ≥ 0. L.36) and Eq.   ˙ ⇒ η(t > T ) = 0. only one of two possible characteristic polynomials will be relevant in the stability analysis.M. det [(A2 + K) − λI] = 0 ⇒ λ3 + β2 λ2 + β1 λ + β0 = 0. i. at any instant.A. the requirement on providing larger absolute values for the instantaneous eigenvalues of Ak in Eq.45) A similar result can be derived when the reference trajectory is generated by another Chua’s oscillator with the same parameters. Aguirre. Clearly. (5. ˙ xref = Axref xref + bxref ⇒ η(t > T ) = 0. a masterslave synchronization problem.M.A. In both cases. from Eq. (5. Mendes o Assuming that the matrix K was appropriately chosen. Eq.

Clearly. Note that the coefficients αn and βn . nonlinear functions of the elements kij . arbitrary reference signal xref (t). and the set of restrictions Eq. in matrix K. where m is the number of non zero elements in the matrix K.50) and using the centercut ellipsoid algorithm described in [31].39). 0. 2. the circuit is controlled only by the voltage source in series with the inductor L shown in Fig.48). as discussed previously on the similarity between the synchronization and the equilibrium point stabilization problems. n = 0. On the other hand. 0. 0.6.47) are the negative sum of the stable eigenvalues of (A1 +K) and (A2 +K). Considering only the restrictions gi in Eq. The great complexity exhibited in Eq. issues the following set of restrictions   g1 : α0   g : α α − α  2 1 2 0    g3 : α2 Ω:  g4 : β0   g : β β − β  5 1 2 0    g6 : β2 > > > > > > 0. 3. (5. 0.49) k1 k2 k3 nant.33) so that. other control structures can be tested based on the same approach. (5. (5. an additional requirement is to maximize the minimum absolute eigenvalue of Ak in Eq. (5. i. Concerning the control problem.39) is asymptotically stable. 2. (5.Chaos Control for Chua’s Circuits 137 and there will be only two nonlinear restrictions in Eq. note that the coefficients α2 and β2 in Eq.e. it is possible to select a viable matrix K that leads to asymptotic stability in the synchronization problem. (5. 5. (5. respectively. are. (5. i. when dealing with the synchronization problem. namely g2 and g5 .48) bounds a region Ω in the m-dimensional controller parameter space for which the matrix Ak in Eq. examining the restrictions in a cyclical order. In the case of the general control problem. This means that there exists an optimal set of control parameters k ∗ inside the region Ω. Hence. (5. in order to achieve a better tracking performance. j = 1. 0. assuming that ux (t) = uy (t) = 0 in Eq. page 105. For example.48) . (5. the requirement on the minimum absolute eigenvalue of Ak is not critical. the matrix K becomes   0 0 0 K =  0 0 0 . 1. in general.48) can be reduced using simpler control structures.

138

L.A.B. Tˆrres, L.A. Aguirre, R.M. Palhares and E.M.A.M. Mendes o

a possible way to maximize the minimum absolute eigenvalue of Ak is to solve the following optimization problem max{α2 + β2 } s. t. gi > 0; i = 1, 2, . . . , 6; (5.50)

where the restrictions gi were defined in Eq. (5.48). The use of the ellipsoid algorithm on Eq. (5.50), issues a gain matrix K well suited for control purposes. A simulation procedure was carried out using the following set of Chua’s circuit parameters: C1 = 23.5µF, C2 = 235µF, L = 43.43H, R = 1.8kΩ (see Fig. 5.6, page 105); and m0 = −0.36mS, m1 = −0.68mS and Bp = 1.0V as Chua’s diode parameters; Eq. (5.2), page 100. The results are shown in Fig. 5.24 to Fig. 5.26 for the control structure represented by the gain matrix K shown in Eq. (5.49) with k1 = −19.00, k2 = −46.47 and k3 = −3.40. The objective is to demonstrate that a badly designed linear controller, exhibiting large error signal as shown in Fig. 5.24 and Fig. 5.25, can still efficiently synchronize two circuits of Chua as shown in Fig. 5.26. Note that the gain and offset distortions for a reference signal completely inside one affine region of Eq. (5.35) is different from the signal distortion observed for a reference signal inside another affine region, as shown in Fig. 5.24 and Fig. 5.25 respectively. This is a result of different sets of eigenvalues for (A1 + K) and (A2 + K) in Eq. (5.38) that leads to different temporal behaviors for the state vector x in each corresponding region described by Eq. (5.35). The above results indicate that the synchronization problem is quite similar to the equilibrium point stabilization problem concerning the stability requirements on the gain matrix controller selection, and that the nature of the control objective is indeed crucial in the controller performance evaluation. 5.4.3. Robust Control and Synchronization via Linear Matrix Inequalities The synchronization problem of two Chua’s circuits via state feedback can also be addressed using optimization techniques based on the so called linear matrix inequalities (LMIs). The main advantage of this kind of strategy is that the control law can be derived in a systematic way if the system model is available. As shown in the previous section, the nonlinear Chua’s circuit can be modelled as a piecewise linear framework. Roughly the idea is that Chua’s

Chaos Control for Chua’s Circuits

139

2.5

2

1.5 x, x ref (V) 1 0.5 0 0

2

4

6

8 t (s)

10

12

14

16

(a)
0.8

0.6

0.4

0.2

0 rz (V) −0.2 −0.4 −0.6 −0.8 −1 0

2

4

6

8 t (s)

10

12

14

16

(b)
Fig. 5.24. Chua’s circuit linear control using only a voltage source in series with inductor L (see Fig. 5.6). Control activated at t = 2.5s. (a) (—) Voltage across capacitor C 1 ; (− −) Reference signal: xref (t) = 1.5+0.5 sin(0.4πt), yref (t) = 0, zref (t) = 0. (b) Applied control signal.

circuit can be represented by a set of “local” systems. With such modelling procedure, diffused techniques from the robust H∞ control theory can be easily applied to these systems. Section 5.4.3.1 presents the LMI based formulation to obtain the control gain to synchronize a master-slave scheme. To evaluate the synchronization control performance, the information transmission problem is addressed in Sec. 5.4.3.2 using the PCChua

140

L.A.B. Tˆrres, L.A. Aguirre, R.M. Palhares and E.M.A.M. Mendes o

2.5

2

1.5

1 x, x ref (V)

0.5

0

−0.5

−1

−1.5

0

2

4

6

8 t (s)

10

12

14

16

(a)
1.5

1

0.5

0

rz (V)

−0.5

−1

−1.5

−2

−2.5

0

2

4

6

8 t (s)

10

12

14

16

(b)
Fig. 5.25. Chua’s circuit linear control using only a voltage source in series with inductor L. Control activated at t = 2.5s. (a) (—) Voltage across capacitor C 1 ; (− −) Reference signal: xref (t) = 0.5 sin(0.4πt), yref (t) = 0, zref (t) = 0. (b) Applied control signal.

(Sec. 5.2.3); i.e. the information signal can be recovered in the slave system using a coherent demodulation technique, if the robust synchronization is established. Notice that other robust H∞ synchronization strategies for coupled Chua’s circuits based on LMIs have been reported in the recent literature; e.g. a fuzzy control based approach [60] and a time-delay feedback control [67].

Chaos Control for Chua’s Circuits

141

3

2

1

ref

(V)
0

x, x
−1 −2 −3 0

2

4

6

8 t (s)

10

12

14

16

(a)
0.6

0.4

0.2

0

rz (V)
−0.2 −0.4 −0.6 −0.8 0

2

4

6

8 t (s)

10

12

14

16

(b)
Fig. 5.26. Chua’s circuit synchronization using only a voltage source in series with inductor L. Control activated at t = 5.0s. (a) (—) Voltage across capacitor C 1 ; (− −) ˙ Genuine chaotic trajectory: xref = Axref xref + bxref . (b) Applied control signal.

5.4.3.1. Robust H∞ Synchronization Consider a synchronization scheme of nonlinear discrete time systems with state and output transitions described as:   xm = A(ρk )xm + B(ρk ) k k+1 (5.51) Master  y m = Cxm k k   xs = A(ρk )xs + B(ρk ) + uk k+1 k (5.52) Slave  y s = Cxs k k

142

L.A.B. Tˆrres, L.A. Aguirre, R.M. Palhares and E.M.A.M. Mendes o

where xm,s ∈ Rn and xm,s ∈ Rn denote the state vectors in the discrete k k+1 m,s time instants k and k + 1, respectivelyc , yk ∈ Rm denotes the mean sured output vectors, uk ∈ R denotes the synchronization control vector, A(ρk ) ∈ Rn×n represents the dynamic matrix of the system, B(ρk ) ∈ Rn is the affine term, C ∈ Rm×n is the output matrix and ρk ∈ Rp is a parameter vector that depends on time k. The synchronism is achieved when dynamical systems simultaneously describe a common trajectory on the state space. By defining the synchronization error as the difference between the master and slave states, εk = xm − xs , and assuming a control law k k m s uk = L(ρk )(yk − yk ), where L(ρk ) ∈ Rn×m is the matrix of synchronization control gains, the synchronism of the master and slave systems can be investigated by checking the synchronization system error given by: εk+1 = (A(ρk ) − L(ρk )C)εk (5.53)

Global synchronization of the master and slave systems means that, for any initial state xs of the slave, and any initial state xm of the master, the 0 0 convergence condition presented below is verified
k→∞

lim

εk → 0,

∀ε0 ∈ Rn

(5.54)

where · is the Euclidean norm. The difficulty to establish the general conditions for the synchronization of global chaotic systems has led many researchers to consider particular classes of nonlinear systems. In the literature, a class of systems commonly investigated is the so-called Lur’e systems. Numerous approaches such as those ones given in [22, 38, 68] deal with the synchronization of this class of systems. This section focus on the synchronization of a class of Lur’e discrete time systems represented by local descriptions in the state space. The nonlinear systems are then modelled in a piecewise linear framework with the state N space partitioned into N regions Ri with i=1 Ri ⊆ Rn . At each discrete time instant k, the state vector xk visits a unique region Ri of the state space associated to a set of constant matrices Ai , Bi e Li . Notice that such a modelling procedure allows to apply some techniques borrowed from the robust control theory.

c ‘m’

and ‘s’ denote the systems master and slave, respectively.

The main result is based on the concept of robust H∞ stability of the error system Eq. Notice that the induced 2 gain is the rate between the energy of the weighting synchronization error signal ‘z’ and the energy of the exogenous disturbing signal ‘w’.57) where · 2 represents the 2 -norm. (5. the synchronization could be completely lost.Chaos Control for Chua’s Circuits 143 Consider a scenario where the master system is disturbed by exogenous signals wk ∈ Rq in their state and measured output: Master   xm = A i xm + B i + E i w k k k+1  y m = C xm + D w i k i k k (5. The results stated in the sequel deal with the problem of chaotic system synchronization in a noise environment.53) rewritten as: Error εk+1 = (Ai − Li C)εk + (Ei − Li Di )wk ˜ ˜ zk = Cεk + Dwk (5.55) where Ei ∈ Rn×q and Di ∈ Rm×q are weighting matrices. The determination of a control law uk that guarantees a robust (insensible to the presence of noise) synchronization of the master and slave systems becomes fundamental. In this noisy context the synchronization of the master and slave systems could suffer a performance degradation and.56) where zk ∈ Rr denotes a weighing vector of the synchronization error and ˜ ˜ C ∈ Rr×n and D ∈ Rr×q are weighting matrices. To solve this synchronization problem. it is sufficient to find a feasible solution to the following LMI optimization problem: . Therefore the synchronization problem can be formulated as a H∞ robust stability problem of the synchronization error system given in Eq. namely to determine the synchronization control gain matrices Li . in the worst case. (5.56). The robust H∞ -norm performance index can be guaranteed to obtain the upper bound γ to the induced 2 gain defined as: sup w 2 =0 z w 2 2 <γ (5.

27. j) ∈ {1 · · · N }. Experimental Results with Information Transmission A practical implementation of the Chua oscillator to achieve the robust H∞ placementssynchronization is considered. the minimum H∞ disturbance attenuation level from the input wk to the weighting synchronization error output ‘zk ’ corresponds to the optimal solution of the minimization problem.A. N . see Sec. . 5. N . . . . 5. Tˆrres.27.M.2. 5.58) ∀(i..A.1. Consider the information transmission scheme exhibited in Fig. η(t) + y m (t) - y s (t) R id + i(t) R id Transmitter + C1 Controller + C2 L u1 (t) u2 (t) u3 (t) Receiver C1 + C2 + L - Fig. L. 5.B. The synchronization gain is given and Fi ∈ Rn×r . which is the x-variable. It is interesting to note that the use of the voltage across C1 .2. i. Unidirectional communication system.A. i = 1.Fi .1.4.3.3. the information transmission problem is investigated in the light of the information transmission via control principle – ITVC discussed in Sec. Further.144 L. Sj ∈ Rn×n .M. and the scalar γ −1 by Li = V Fi .3. has been found to be the best scalar variable through which the dynamics of Chua’s circuit can be observed. This scheme shows the Chua’s oscillator circuits on a unidirectional coupling. .t.4. any matrices V ∈ Rn×n √ δ > 0. Palhares and E.1. Mendes o (5. Moreover. 5.V δ       S j − V T − V V A i − F i C i V E i − F i Di  T  ˜ ˜ ˜ ˜   AT V T − Ci FiT −Si + C T C CT D s. where the unknowns are the symmetric definite positive matrices Si .Sj . . Proof details for this result can be found in [24]. R.  i   T T T T T ˜ T ˜ ˜ ˜ E i V − D i Fi D C −δI + D D 0. 5. Aguirre. γ = √ δ. All the experiments were performed in the laboratory setup PCCHUA described in Sec.Si . .M.e. . j = 1.   minδ. The information transmission test is used as a performance index to validate the proposed approach for robust synchronization. .

with Fig.9(a).9. 5. Mathematically this perturbation corresponds to a modification on the differential equations that govern the motion of the oscillator in the direction of vC1 (t). Then a scalar signal y m (t) is used to carry the information and to serve as a reference signal to allow the synchronization of the slave oscillator (receiver ). This figure should be compared to Fig. 5.28. 5. 5. Typical information perturbed double scroll attractor exhibited by the master oscillator – real data. This signal can be corrupted by noise interferences η(t) on the transmission channel. The information signal is injected using a voltage controlled current source in parallel with capacitor C1 . page 108. The transmitter system can be represented by the set of differential equations . the oscillator unperturbed dynamical behavior is considerably altered.28. 1 0.Chaos Control for Chua’s Circuits 145 The information to be transmitted i(t) is injected on the master Chua’s chaotic oscillator (transmitter ) as a perturbation only in the direction of vC1 (t). page 105. Despite the fact that the information signal must be of very low amplitude in order to keep the master oscillator in the chaotic regime. shown in Fig.5 -1 -6 -4 -2 0 x (V) 2 4 6 Fig. 5. as can be seen by comparing Fig. page 108.5 y (V) 0 -0.6.

1.60) ∗ ∗ = −vC2 (t) − rL i∗ (t) + u3 (t) L ∗ = m ∗ vC1 + 0 ∗ = vC1 (t) ∗ ∗ (m∗ −m∗ )(|vC ∗ +Bp |−|vC ∗ −Bp |) 1 0 1 1 2 ∗ ∗ ∗ where C1 . L. rL∗ . together with the Kalman filter like algorithm known as UKF [40]. This was done by using real data acquired from PCChua.M. R∗ .B. y m (t) is the transmitted signal and η(t) is the noise signal on transmission channel. R. m∗ e Bp are the receiver circuit’s parame0 1 ters. u2 (t) and u3 (t) are scalar control signals and y s (t) is the receiver output measured signal. rL. Mendes o  dvC (t) 1  C1 dt    dvC (t)   C2 2   dt   diL (t) L dt     i (v )  d C1      m y (t) = = vC2 (t)−vC1 (t) R vC1 (t)−vC2 (t) R − id (vC1 (t)) + i(t) + iL (t) (5.4. . L∗ . C2 . to estimate the parameters shown in Table 5. L.A. u1 (t).A. Aguirre.A. 5. i(t) is the scalar information signal.59) = −vC2 (t) − rL iL (t) = m 0 vC1 + (m1 −m0 )(|vC1 +Bp |−|vC1 −Bp |) 2 = vC1 (t) + η(t) where C1 . R. Palhares and E. or quasi-identical. The ITVC principle (Sec.M. m0 .1) states that any controller that guarantees an identical.27 have to be estimated. 5.7.146 L. C2 . master-slave synchronization can actually perform as a demodulator and thus recovering the transmitted information. Tˆrres. To carry out the proposed approach the parameters of the Chua oscillator (or the communication system) shown in Fig. m1 e Bp are the circuit’s parameters. m∗ . By construction. the control signal u1 (t) corresponds to the demodulated information signal i(t). Analogous to the transmitter the receiver system has its motion governed by the set of differential equations  dv ∗ (t)  C 1 C1  ∗ dt    ∗   C ∗ dvC2 (t)  2 dt    ∗ di∗ (t) L L dt      id (vC ∗ )   1     s y (t) vC ∗ (t)−vC ∗ (t) 2 1 = = R∗ vC ∗ (t)−vC ∗ (t) 1 2 ∗ − id (vC1 (t)) + u1 (t) R∗ + i∗ (t) + u2 (t) L (5. In this way.M. the receiver oscillator is fully controllable by the controller signals u(t).

the transmitter Eq. a piecewise linear representation for the transmitter and receiver systems was obtained that describes the motion of Chua oscillator circuits by means of three linear local models:   xm = A i xm + B i + E i w k  k+1 k   xs = A i xs + B i + u k k+1 k  y s = C xs i k k  m  y = C xm + D w i k i k k ι(k) Transmitter (5.801mS 1. Parameter Value Parameter R 1673Ω rL C1 30. 5.60) were discretized and the resulting communication scheme similar to the one shown in (Fig.61) Receiver with the following matrices describing the local motion  1 − T /(RC1 ) − T m0 /C1 T /(RC1 ) 0 A1 =  T /(RC2 ) 1 − T /(RC2 ) T /C2  0 −T /L 1 − T rL /L   1 − T /(RC1 ) − T m1 /C1 T /(RC1 ) 0 A2 =  T /(RC2 ) 1 − T /(RC2 ) T /C2  0 −T /L 1 − T rL /L   1 − T /(RC1 ) − T m0 /C1 T /(RC1 ) 0 A3 =  T /(RC2 ) 1 − T /(RC2 ) T /C2  0 −T /L 1 − T rL /L  . as the PCChua is implemented in a digital platform and the approach has been proposed in discrete-time setting. (5.28H Bp PCChua osValue 0Ω −0.Chaos Control for Chua’s Circuits Table 5. Estimated parameters of the cillator circuit.27) (now in the discrete domain) is represented as a robust master-slave synchronization scheme.74V 147 Further.365mS −0. (5.66µF L 52.59) and receiver Eq.14µF m0 m1 C2 185.7. By applying the discretization approach presented in [59].

the information signal i(t) has its influence minimized over the transmitted signal y m (t). 3} in Eq.45 × 10−3 2. L. . sampling time T = 10 ms.e. represented by piecewise linear models. R. the following result can be obtained: G = 99. Di =  0  001 0 By solving the optimization problem in Eq.001 1 Ei =  0  . (5. the performance index H∞ -norm appears as a criterium that ensures a better information reconstruction.M.148 L. The entry term ι(k) = Ei wk assumes an equivalence of the information T signal at discrete time instants k.7.A. was simplified and the transmission of index model ‘i’ to the receiver system is no longer necessary. consequently. Di = 0.58). Ci =  0  . i. B3 =   0 0 0   T 0.0001 0 0    100 0 ˜ ˜ C i =  0 1 0  .93 × 10−3 51. 2. (5.56) tends to minimize the effect of the exogenous entry ‘wk ’ over the weighting synchronization error vector ‘εk ’ and.A. and a unique synchronization gain (Li = G. B2 =  0  . (5.B.58) and considering the discretized master-slave synchronization scheme with the parameters shown in Table 5.M. Aguirre. Tˆrres. 2. respectively. 3}). The weighting matrices Ei and Di were defined based on the amplitudes of the information signal i(t) and the noise signal η(t). By doing Li = G for i ∈ {1. Palhares and E. the implementation of the proposed approach for robust H∞ synchronization of the chaotic systems. This argument is based on that the robust H∞ stabilization of the error system Eq..M.70 × 10−4 of disturbance attenuation H∞ .50 × 10−9 T  (5. ˜ ˜ The weighting matrices C and D were conveniently defined in order to establish the directions and intensities of the weighting synchronization error vector. ι(k) ≡ C1 i(kT ). Mendes o  B1 =   +T Bp (m1 −m0 ) C1 0 0       −T Bp (m1 −m0 ) 0 C1     . In the context of communication.A. ∀i ∈ {1.62) that guarantee a level γ = 9.

as shown in Fig. 5. The transmission occurs in the time interval from 15 sec to 105 sec. The synchronization error signal ε(t) corresponding to the difference between the transmitted signal y m (t) and the measured slave signal y s (t) is shown in Fig. The information signal starts to perturb the natural behavior of the transmitter oscillator at 30 sec and stop at 90 sec.63) In order to emphasize the effectiveness of the proposed methodology and to validate its implementation two information transmission experiments were performed in the PCChua platform.99 × 10−3 =  0. .3t) mA. 5.29.Chaos Control for Chua’s Circuits 149 Notice that synchronization gain G in Eq.62) must be transformed in order to include the effect of the zero-order hold actuators and to generate compensation signals compatible to the receiver system Eq. In the first experiment. 5. a sinusoidal signal i(t) of amplitude of 0.1 placements −0. Figure 5.12 mA and frequency of 0.12sin(2π0.1 i(t) (mA) 0 −0. (5.2 0.30 depicts the signal y m (t) transmitted by the master oscillator (transmitter) that serves as a reference signal to the synchronization of the slave oscillator (receiver).60).69 × 10−3    C1 G1 /T  C2 G2 /T  LG3 /T GP Cchua ≡ (5. The synchronization gain considered in the PCChua implementation becomes  2.54 × 10−3  2. 0.2 0 30 Time (sec) 90 120 Fig. Outside the time span from 15 sec to 105 sec the controller is inactive and the synchronization error signal corresponds to the natural behavior of the free running receiver oscillator.29. [E1] Information signal: i(t) = 0.3 Hz is used as the information signal to be transmitted. (5.31.

5.31. 0.0 Hz) and a sawtooth signal (1. Aguirre.A. it can be readily noticed that the information from 75 sec to 95 sec is recovered via ITVC principle. 6 4 e(t) (V) 2 0 −2 −4 −6 0 15 30 90 105 120 placements Time (sec) Fig. 5. L.12 mA.30.B. . that is. [E1] Transmitted signal.4 Hz). In the second experiment an information signal i(t) with a more complex behavior is considered.33 shows a close comparison between the transmitted information i(t) and the recovered information represented by the control action u1 (t). The synchronization error ε(t) signal and the control signal u1 (t) are depicted in Fig. Palhares and E. Figure 5. 5.M.35 and Fig. 5. Mendes o 8 6 4 y m (t) (V) 2 0 −2 −4 −6 −8 0 15 105 120 placements Time (sec) Fig. Tˆrres. [E1] Error signal: e(t) = y m (t) − y s (t) V.150 L. a square signal (1.37.32 shows the control action u1 (t) and Fig. All signals have the same amplitude.36. 5.M.A.5 Hz). In Fig. R. Figure 5. This information signal is the summation of a sinusoidal signal (0.M.34 shows the transmitted signal y m (t). 5. respectively.A.

as briefly presented in Sec.5 −1 −1.4. Nonlinear State Feedback In this section.59).1.e. (5. in this section the control action will be restricted to the first equation in Eq. and controller complexity.2. and y (t) is the receiver output measured signal. However.4.Chaos Control for Chua’s Circuits 151 1. similarly to what was presented in Sec. (5. u2 (t) = u3 (t) = 0. The design of the following control structures will highlight the tradeoff between control energy. and detailed in [12]. 0 1 s u(t) is a scalar control signal. page 146. .60). in order to send information from the master to the slave oscillator. C2 . m∗ . (5. [E1] Control signal: component u1 (t).5 0 −0.64). 5.64) 1 1 ∗ ∗ = −vC2 (t) − rL i∗ (t) L ∗ = m ∗ vC1 + 0 ∗ = vC1 (t) ∗ ∗ (m∗ −m∗ )(|vC ∗ +Bp |−|vC ∗ −Bp |) 0 1 2 ∗ ∗ ∗ where C1 .32. and the slave Eq. by means of the scalar control signal u(t) at the receiver. The objective is to synchronize both oscillators. and to recover the original information signal i(t) as a byproduct of the synchronization process. control objective.4. two different control structures will be used to control and synchronize connected Chua circuits. m∗ e Bp are the slave system parameters.5 1 u1 (t) (mV) 0.5 0 15 30 90 105 120 placements Time (sec) Fig. 5. L∗ . such that  dv ∗ (t)  C 1 C1  ∗ dt    ∗  ∗ dvC2 (t) C  2 dt    ∗ di∗ (t) L L dt      id (vC ∗ )   1     s y (t) = = vC ∗ (t)−vC ∗ (t) 2 1 ∗ − id (vC1 (t)) + u(t) R∗ vC ∗ (t)−vC ∗ (t) 1 2 R∗ + i∗ (t) L (5. 5.4. rL∗ . the master Eq.3. 5. R∗ . i.

The only consideration is that.B.33.4.4.1 0 −0. 5.4.4. 5.1 placements −0.1 0 −0.2 and Sec.4. Aguirre. Control signal u1 (t) . [E1] Recovered information without filtering (a) and with filtering (b) (Butterworth filter: 2nd order.1 placements −0.3.2 i(t).2 i(t). Tˆrres.red Fig. The proposed controller structure is quite general.M.A.152 L.u1 (t) (mV) 0. it is known that a low energy controller should be implemented (see also Sec.blue 0. as can be verified by the following set of equations: ∗ eT (t) = y s (t) − vC1 (t). Low Energy Adaptive Proportional Controller In this section a nonlinear controller designed with no prior knowledge of the system dynamics is presented.A.u1 (t) (mV) 0.65) u(t) = kp (eT ) eT (t). Palhares and E. . from the results obtained in the previous Sec. 5.1). L. 1 2 kp (t) = kp + kp 1 − 1 1 + δp e2 (t) T .1. 5. 5. fc = 2 Hz).M.2 75 Time (sec) 90 95 (b) Information signal i(t) . (5.2 75 Time (sec) 90 95 (a) Information signal i(t) . R.black.A. Control signal filtered u1 (t) .M. Mendes o 0.black.

kp and δp are parameters of this adaptive proportional controller. [E2] Error signal: e(t) = y m (t) − y s (t) V.38). in order to fulfill this requirement.65) there are no assumptions about the master oscillator structure and. 6 4 e(t) (V) 2 0 −2 −4 −6 0 15 30 90 105 120 placements Time (sec) Fig.65) is that. for which the static characteristic curve is shown in Fig. (5. It is important to note that in Eq. and therefore a minimum control effort is necessary to keep the synchronization condition. [E2] Transmitted signal. 5. a low energy controller should be used as demodulator for master-slave connected nonlinear oscillators. as discussed in Sec. and at . 5.1. 1 2 where kp .34. 5. 5. at a first sight.1. it is far from obvious that the control action can actually synchronize both Chua circuits. the slave oscillator is following a trajectory that is very close to its natural behavior.38. At the same time. 5.4. (5.35.Chaos Control for Chua’s Circuits 153 8 6 4 y m (t) (V) 2 0 −2 −4 −6 −8 0 15 105 120 placements Time (sec) Fig. The rationale for the control law Eq. when the circuits are almost synchronized. and therefore the proportional gain is reduced when |eT | ≈ 0 (Fig.

as it can be seen by comparison with the result shown in Fig. R. is mainly related to the strong nonlinear character of Eq. 5.5 1 u1 (t) (mV) 0.5 0 −0.A. . i.2). This shows again that the knowledge of the dynamical system model is useful in designing the controllers.B.e.40(b). L. is due to the phase lag introduced by the low-pass Butterworth filter. High Energy Sliding Mode Controller Following the approach of minimal controller complexity of the previous section.5 0 15 30 90 105 120 placements Time (sec) Fig. 5. The same phenomenon does not occur with information signals that have smaller amplitudes. In addition.4.2.40 one can see that the filtered recovered information signal is worse than those obtained with the H∞ state feedback controller (Sec. designing a nonlinear controller for Chua’s circuit without relying on accurate knowledge of system model. 5. page 152. 5. Measured results using control Eq. 5.M.40(b).65). 5. The slight attenuation of the recovered signal. 5.154 L. it is worth noting that the delay time between the original and the recovered information signal observed in Fig. (5. the same time it will be able to approximate the original information signal after the synchronization takes place.M.3.65) are shown in Fig. Tˆrres.A.4. Such attenuation becomes more evident during the peaks of the recovered signal. but its importance is weighted by the nature of the problem under consideration. In Fig. Palhares and E.M. [E2] Control signal: component u1 (t).5 −1 −1.33. 5. Aguirre. depicted in Fig.A. a simple sliding mode controller will be proposed that is capable of synchronizing two Chua circuits to an arbitrary precision.40.4.39 and Fig. (5. 5.36. Mendes o 1.

Consider the following set of equations representing a sliding mode controller applied to the slave oscillator Eq.64): ∗ eT (t) = y s (t) − vC1 (t).4). [E2] Recovered information without filtering (a) and with filtering (b) (Butterworth filter: 2nd order.black.2. u(t) = where umax = −umin = 0.66) .5mA and δs > 0. Moreover. Control signal u1 (t) . 5. What is relevant in this example is the fact that it is possible to have both oscillators in a quasi-identical synchronization condition and umax . fc = 10 Hz). The case δs = 0. (5.red Fig.black.u1 (t) (mV) 0 placements −0. the control action is switched according to the hysteresis behavior depicted in Fig. simulated using normalized equations (Sec. In this case.42. eT (t) < −δs .blue 0. (5. (5. 5.41. it is possible synchronize both Chua circuits with arbitrarily small synchronization error. page 100. by choosing appropriately the value of δs in Eq.5 75 90 95 Time (sec) (b) Information signal i(t) .37.66).5 i(t). 5.u1 (t) (mV) 0 placements −0. Control signal filtered u1 (t) .5 i(t).4. Eq. 5. umin .Chaos Control for Chua’s Circuits 155 0. eT (t) > δs .1. (5.5 75 90 95 Time (sec) (a) Information signal i(t) . is shown in Fig.

it remains as one of the most robust and easily built chaotic circuits. 5.2. L.A.01. mathematical modeling. the control action will still be composed only by positive and negative saturation levels umax and umin . despite the fact that the circuit can be made to follow arbitrary reference signals.38.5. which comprises a complete com- . Palhares and E.A.M.004 0. In order to properly address this involved theme.0007. Chua’s circuit is undoubtedly a paradigmatic nonlinear oscillator that is capable of exhibiting many interesting dynamical behaviors.012 0.M. Mendes o 0. data analysis. 5.2.156 L. kp = 0. dynamical characterization. specially when realized by using simulated inductors (Sec. 5. which includes the synchronization of Chua circuits in master-slave configuration as a special case. And this is true even in the case when the strong assumptions of identical vector fields and absence of noise in the transmission channel are considered to be valid. In addition. 5. Therefore.006 0. δp = 1.B.0.008 kp 0. Tˆrres. Conclusions This chapter has focused on chaos control for Chua circuits.M. Aguirre. This path has been followed in the implementation of the so-called PCChua experimental setup (Sec. one is able to obtain Chua circuits that exhibit very slow oscillations appropriate to digital analysis and control [71]. an account of various published works on Chua’s circuit implementation.3). With this peculiar construction.002 0 −5 e [V] T 0 5 1 2 Fig. Adaptive proportional controller characteristic curve for k p = 0. R.A. the ITVC principle cannot be realized due to the high energy nature of the sliding mode controller.2). no matter the waveform of the information/perturbation signal injected at the transmitter.01 0. control and synchronization was presented.

page 104. (a) cillator. Controller activated at t = 15s.65) with kp = 0. 5. (b) Control action at the receiver.39. Control law as in Eq. One of the most interesting Chua’s circuit dynamical behaviors – the chaotic regime – which is usually represented by the so-called double scroll attractor. Synchronization between the perturbed master oscillator and the slave osmin max = 0. kp p Transmitted signal s(t). 5. As a byproduct of such effort. (c) Synchronization error [s(t) − y1 (t)]. (5.x] (V) 2 0 -2 -4 -6 0 10 20 30 t (s) 40 50 60 70 Fig. so that the system bifurcation diagram can be easily obtained from real data. δ = 1. Information signal perturbation starts at t = 30s. puter based system flexible enough to investigate chaos control through the realization of any control method that can be represented mathematically by Eq. many dynamical properties associated to the circuit of Chua.Chaos Control for Chua’s Circuits 157 8 6 4 2 s (V) 0 -2 -4 -6 -8 0 10 20 30 t (s) 40 50 60 70 5 4 3 2 u (mA) 1 0 -1 -2 -3 -4 0 10 20 30 t (s) 40 50 60 70 6 4 e = [s . the PCChua can also be used to simulate parametric variations in Chua’s circuit. was thoroughly investigated in Sec. As a consequence of this flexibility. as shown in [73].0007.01.6).0. were re- .3 through different attempts of data analysis and mathematical modeling. (5. and also topological characteristics of the double scroll attractor.

One of the key results of such endeavor is the principle called Information Transmission Via Control – ITVC (Sec. Recovered information signal. Aguirre.65) with k p max = 0.1 -0.4).2 -0. Sec.3 i(mA) u(mA) (filtered) 0.4 0. Linear and nonlinear state feedback based control techniques. 5.2 0.3 -0.0.1.3 and Sec.M.2 u (mA) 0. 5.158 L.0007. (a) Control signal at the receiver.2.4 a broad discussion took place in order to show how intertwined are the control of nonlinear dynamical systems with specific properties of control methods such as control energy. higher-order spectral coherence.M.4 25 30 35 40 45 t (s) 50 55 60 65 min = Fig. Tˆrres. that was extensively used on various control and synchronization strategies discussed in this chapter.1). 5. together with robust control and synchronization design methods. etc.4 0.1 0 -0. usually viewed as a trajectory tracking control problem.01. L. control objective and control law complexity (Sec. 5.1). (· · · ) Original information signal: i(t) = [0. Concerning the problem of controlling Chua’s circuit. were applied to the master-slave configuration of two unidirectionally connected Chua circuits (Sec.5Hz). (b) (−) Control 0. R. 5.A.4.23 sin(2π(0.4.1)t)]mA.4.4. vealed such as observability indices.1 -0. δ = 1. that is similar to the signal produced . Control law as in Eq. one should consider the interesting fact that to follow a chaotic reference. kp p signal after applying an order 4 Butterworth low-pass filter (fc = 0.40. These examples were aimed to highlight that in the synchronization of chaotic oscillators.4 25 30 35 40 45 t (s) 50 55 60 65 0.3 -0. 5. 5.A. symmetry.M.1 0 -0.2 -0. in Sec. Palhares and E.A.4. Mendes o 0. (5.B.3 0.

6 0. control and synchronization. Switching law Eq. by the uncontrolled nonlinear system.15 0. data analysis. (5. ||e|| .3 0. Synchronization error norm for the oscillators coupled via switching control law Eq. Control activated at t = 35s. 5.Chaos Control for Chua’s Circuits 159 0.4 0. sinusoidal signal. The long list of cited references reflects the authors desire to provide an adequate introduction to these subjects.41. is easier to accomplish than following a simple.66).42.1 0.25 0. It is hoped that this chapter can serve as a starting point for investigating Chua’s circuit dynamical characterization.4 2δ s −3 −2 −1 e [V] T 0 1 2 3 Fig.35 0. mathematical modeling.4 0. (5.2 u(t) [mA] 0 −0. but unnatural. 25 20 15 ||e|| 10 0.66). 5.2 −0.05 0 50 55 60 t (s) 65 70 5 0 0 10 20 30 40 50 t (s) 60 70 80 90 100 Fig.2 0.

Aguirre.A. Aguirre. B. Fixed point stability analysis of Chua’s o circuit: A case study with a real circuit. and E. Furtado.O. 1-18. Aguirre. Lett.A. (1995). Rev. 145. 2. 7.A.M. pp.G. Bifurcation and Chaos 5.M. (2008). M. pp.A. (2005). 3. Rodrigues. Structure-selection techniques applied to continuous-time nonlinear models.B. Mendes. 135-145. C. 1-4. Bifurcation and Chaos 10. Teixeira.A. S. Letellier. 1. 1.M. Phys.A. 026226. Nonlinear identification and cluster analysis of chaotic attractors from a real implementation of Chua’s circuit.A.A. Physical D 100.B. A nonlinear correlation function for selecting the delay time in dynamical reconstructions. (2002).B. References [1] L. Aguirre and C. (2001).A. J.F. [2] L. [5] L.A. Tˆrres. J. [3] L. (1995). 6. J. pp. Aguirre. 1411-1423. E 72. 88-94. pp. 409414. Fradkov. Physical D 158. pp. J. [7] L. [13] B. Phys. Letellier. [6] L.A.A. 203A. Aguirre and L.160 L.A. Control of nonlinear dynamics: Where do o models fit in?. Control Theory Appl. Aguirre.A. Tˆrres. Recovering map static nonlinearities from chaotic data using dynamical models. Alves. (1998). Mendes o Acknowledgments The authors are grateful to financial support from Brazilian governmental research-funding agencies Conselho Nacional de Desenvolvimento Cientfico e Tecnol´gico – CNPq and Funda¸ao de Amparo a Pesquisa do Estado de o c˜ ` Minas Gerais – FAPEMIG. Observability of nonlinear dynamics: normalized indices and a time series approach. Aguirre. pp. Retrieving dynamical invariants from chaotic data using NARMAX models. and Letellier C.A. Maquet. [12] L. Int. Andrievskii and A.M. Tˆrres. U. [8] L. Evaluation of dynamical o models: Dissipative synchronization and other techniques.S. International Journal of Circuits. pp. [10] L. Submitted.R.A.A. 449474. Aguirre. 2. Int. and C. J. (2000).3. Cluster analysis of NARMAX models for a signal-dependent systems. [11] L.B. and L.A.2.R. IEE Proc. 4. (2006). Using data-driven o discrete-time models and the unscented Kalman filter to estimate unobserved variables of nonlinear systems. Systems and Computers 7. 2. (1997). Billings. R.S.B. 41-57. Maquet. L. Bastos. Aguirre.L. C. Tˆrres. pp. E. (1997).A. Control of chaos: Methods and appli- . Aguirre. [9] L.A. Rev.A. E 74. J´come. 1-12. Phys.A. Aguirre and L. Bifurcation and Chaos 6. Palhares and E. Aguirre and S. Freitas. Tˆrres. Int. 667-681. pp.M.C. and L. Int. 111-115. pp. (1997). [4] L. Induced one-parameter bifurcations in identified nonlinear dynamical models. Bifurcation and Chaos 12.B. and J.

J. 1-60. Zhou. The control of chaos: theory and applications. Corrˆa. State space reconstruction in the presence of noise . Bifurcation and Chaos 8. Maza. Nonlinear Dynamics 17. Ecker. and D. Palhares. and M. P. M. (2007). Part I. J. pp. J. Press. Chen. 1. International Journal of Control 50. Cartwright. Billings and Q.A. pp. S. Int. S. 3. I. 4. Bowong. Campos.-T. pp. Bifurcation and Chaos 14. Phase-locking and bifurcations of the sinusoidally-driven double scroll circuit. (1994). Systems and Computers 14. Mendes. (1969). 3rd edn. Journal of Circuits. Ditto and K. Aguirre. (1989). and R. (1997). Physical D 51. pp. 103-197. H. o Mozelli. global data and threshold model identification of severaly nonlinear systems. S. Gibson. 99-107. J. Proceedings IEE 116. and C. Chen. (2000). Chua Circuit: An Overview 10 years later. 529-546. Pantano. Chaos 7. Guest Editors. Realisation of gyrators using OpAmps and their use in RCactive network synthesis. E. Systems and Computers 4.B. U. 2. Physics Reports 329.M. (1998). and F. S. G. W.H. and M. (1998). S. E.M. F. Methods.A.L. Bifurcation and Chaos 17. (2000). L. Kupferschmid. pp. (2004). Physics Reports 366. S. Using deep cuts in an ellipsoid algorithm for nonlinear programming. Lai. (2003). Farmer. Identification of mimo nonlinear systems using a forward-regression orthogonal estimator. (2005). L. pp. pp. Chen. Modeling chaotic dye namics with discrete nonlinear rational models. International Journal of Control 54. (1991). 1897-1923. Automation and Remote Control 64. R. D.V. 119139. Kili¸. Casdagli.M. 5.M. 117-159. 2. 3199-3209. J. and J.A. Calvo and J. Y. J. 5. Journal of Circuits. S. Int. Int.Chaos Control for Chua’s Circuits 161 [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] cations. Int.A. Kakmeni. Grebogi. E. 1019-1032. Osipov. pp. (Oxford Univ.C. A.S. Bifurcation and Chaos 17. Experimental Results on Chua’s Circuit Robust Synchronization via LMIs. Extended model set. Int. Focus Issue: Control and synchronization of chaos. 7. 1838-1850. 1. pp.M. S.A.A. 5. 3. Inductorless realization of nonautonomous MLC ¸ c chaotic circuit using current-feedback operational amplifiers. J. Mendes. Fuzzy Control of Chaos. Showalter. J.S. Billings. A new synchronization principle for a class of Lur’e systems with applications in secure communication. S. 6.G. 2157-2189. Kurths. Stranges. 9. Linear System Theory and Design. Zhu. M. (1989). 2477-2491. The synchronization of chaotic systems.A.D.M.A. (2002). Dziuban. pp. 673-713.M. Mathematical Programming . Rational model identification using an extended least-squares algorithm. Baptista and I. pp. Antoniou.L. pp. Caldas. and L. Billings and S.M.J. pp. 1743-1747. Cam and R. International Journal of Control 49. Tˆrres. pp. and L.D. Mancini. Korenberg. 8. Eubank. Koina. Bocaletti. 11. (2007).L. Bilotta. C. pp. Valadares.O. C. 52-98. M. Bifurcation and Chaos 10. pp. 1-101.E. A gallery of Chua attractors. 1998).T. O. Chua. C. (1991). Bocaletti.

1475-1493.G. Freitas. pp. Physica D 9. Huang. pp.M. pp. [42] M. Chandran. and C. New Method for Stabilization of Unstable Periodic Orbits in Chaotic Systems. Rev. 401-422. Galias. 8. 4955-4972. (2006). Ogorzalek. 3-4.P.K. Billings.A. J. R. 2. Int. (2004). Uhlmann. (1995). A comparative study on realization of Chua’s circuit: hybrid rec alizations of Chua’s circuit combining the circuit topologies proposed for Chua’s diode and inductor elements. [37] P. Billings. and P. Unscented filtering and nonlinear estimation. 281-294. [48] M. 10. Aguirre. 191-203. [38] H.A. [39] M. pp. Master-slave synchronization of general Lur’e systems with time-varying delay and parameter uncertainty. Zhong. pp. Y. C. pp. [40] S. pp. 189-208.M. (2003).K. Guest Editors. 363-384. Bifurcation and Chaos 13. Tˆrres.M. [45] R. J.A. Li.P. 1. Sliding mode control of chaos in the cubic Chua’s circuit system. Kennedy and M. Int. International Journal of Control 48. Yildirim. Chen.O. of the Franklin Institute – Engineering and Applied Mathematics 343. 173-178.X. [46] R. Bifurcation and Chaos 12.162 L. Roose. 3. Aguirre. 6. Ruelle.J. Gouesbet and C. pp. Reviews of Modern Physics 57. Ghandi. Chua. Elgar. 2002). and F. 6. pp. and L. Leontaritis and S. 281-295. pp. D. Mendes o Study 25.J. Kili¸. (1985) [32] J. 475-491. [47] R. J. Nonlinear Systems. Orthogonal parameter estimation algorithm for nonlinear stochastic systems.J. (1994). J. pp. L.S. Proceedings of the IEEE 92.-J.-L. 5. 2. pp. (1983). An improved Chua’s circuit and its use in hyperchaotic circuit.J. Int. B. Kili¸. [35] G. Khalil. 2. pp. Circuits Systems and Signal Processing 22. and V. and Signal Processing 46. 66-80. 193-210.A. Palhares and E. Saracoglu. [41] M. On current feedback operational amplifier-based realizations of c Chua’s circuit.A. 2. Analog Integrated Circ. 2425-2431. Kennedy. J. Int. Jang. [33] S. Grassberger and I. Vanhoff. pp.-K.P. Korenberg. Int. On periodic orbits and homoclinic bifurcations in Chua’s circuit with a smooth nonlinearity. Eckmann and D. 1437-1449. J. Phys. (2003). S. [49] I. Julier and J. J. 1. Bifurcation and Chaos 16. Bifurcation and Chaos 5. (1992).A. Robust OP Amp realization of Chua’s circuit. H. (1993). [34] Z. (1989). (1985). 93-107. Input-output parametric models for non- . (1997). HigherOrder Spectra of Nonlinear Polynomial Models for Chua’s Circuit. Int. [36] G. Liu. Bifurcation and Chaos 12. (1998). (Prentice Hall. Ergodic-Theory of Chaos and Strange Attractors. Procaccia. Letellier. (2006). Khibnik. Special issue on chaos synchronization and control: Theory and applications. McIlroy. and J. O. L. [44] A. Frequenz 46. 1-2. IEEE Transactions on Circuits and Systems Part I 44.I. Kili¸. (2002). pp. 6. A new nonautonomous version c of Chua’s circuit: Experimental observations. U. pp. [43] H. E 49. Global vector field reconstruction by using a multivariate polynomial L2 approximation on nets. Bifurcation and Chaos 3.J. Measuring the Strangeness of Strange Attractors. Chen. 61-656. (2006). 3.B.

Smith.M. 2003). E. and J. Circuits. (Oxford University Press. Topological characterization of reconstructed attractors modding out symmetries. Tˆrres. Phys. Palhares. Applied Nonlinear Control. International Journal of Control 41. (1971). USA. 1-8.M.D. C. Yorke. Chaotic synchronization and information transmission experiments: A fuzzy relaxed H∞ control approach. 11. 427-449. Rev. Gilmore. L.A. 131-140. IEEE Transactions on Circuits and Systems Part I 33. Li.A. Nonlinear Systems: analysis. E. Letellier and R. An inductorless CMOS realization of Chua’s circuit. Mozelli.A.A. (1996). Ott.A. 3. Radwan. Relation between observability and differential embeddings for nonlinear dynamics. Campos. Chaos 12. Soliman. and E. J. T. o Mendes. Control of Chua’s circuit with a cubic nonlinearity via nonlinear linearization technique. pp.A. Chaos. pp. 4. pp. Aguirre. Palhares. Letellier and G. Rev. (2002). Tˆrres. Solitons and Fractals 18.C. (2005). D. Bifurcation and Chaos 6. R.. A note on discretization of nonlinear differential equations. (1984). 1. 549-558. 1. Mendes. . 4. Grebogi. E 71.A. Souza.L. Liao and F. (1998). Takens. A. L. (1996). 2. Robust o H∞ control for master-slave synchronization of Lur’e systems with timedelay feedback control. 1990). L. E.Chaos Control for Chua’s Circuits 163 [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] linear systems part II: stochastic nonlinear systems. Special issue on Chua’s circuit: a paradigm for chaos. J.A. (1990). A. Letellier. 16151638. 25312555. F. 12. Topological analysis of chaos in equivariant electronic circuits. 1055-1058. G. Ruelle and F. R. stability and control. Inc. Aguirre. Lorenz.B. Billings. (SpringerVerlag New York. Systems and Computers 3.A. Journal of Atmospheric Sciences 20. pp. A. Matsumoto. A chaotic attractor from Chua’s circuit. (2000). R. (1985). C. Controlling Chaos. and L. Sastry. J. Circuits. (2007). 6. 329-344.-W. C. S. (2008). pp. Int. 12. Systems and Signal Processing 26. 149-158. Letellier and L. Chaos 12. Rulkov. Microelectronic Circuits. Communications in Mathematical Physics 23. Mendes and S. 1196-1199.A. 719-731. (Prentice Hall. Gouesbet. Slotine and W. Phys.M. Investigating nonlinear dynamics from time series: the influence of symmetries and the choice of observables. Journal of Circuits. (2003). C. Covering dynamical systems: Two-fold covers . E.M. El-Sedeek.A.M. Deterministic Nonperiodic Flows. pp.M. 066213. pp.M. 1999). Systems and Signal Processing 17. and J.M.-J. (2002). pp. Int. Madan. Nature of Turbulence. (1993) T.B.O. pp.G. Physical Review Letters 64.M. C. (1963). Letellier. C. C. Journal de Physique II 6. E (in press). Bifurcation and Chaos To appear. Gouesbet and N. 66-71. Chen. Maquet.-L. and A. pp.S. pp. pp. Sedra and K.

Billings. Physica D 228. Physica D 164. pp. Aguirre.B. Curran. Aguirre.M.A. 871-886. 2349-2360. (1999). 1.A. pp. . pp. Zhu and S. 53-58. Palhares and E. 1915-1916. pp.A.A. 841-850.-S. (2005). Chen.A.K. Wu and M. Zhu and S.B. 8.M.M. (1999). pp.B.O. [76] Q. An extended chaos control method applied o to the Chua’s circuit. Electronics Letters 35. R. 10. [70] L. Aguirre. [74] T. International Journal of Control 64.A. Tˆrres and L. and L. 768-770. Aguirre. J. Tˆrres. pp. 5. (1996).A. IEEE Transactions on Circuits and Systems Part I 46.A.A. 3-4. Recursive parameter estimation for nonlinear rational models. Tˆrres.B. (2004). [75] Q. PCCHUA – A laboratory setup for relatime o control and synchronization of chaotic oscillations.A. pp. Inductorless Chua’s circuit.A. Chua.M. Int. Suykens. P.B. 387-406. Robust synthesis for masterslave synchronization of Lur’e systems. pp.A. (2002). L. [73] L. Tˆrres and L. Tˆrres and L.A. Discrete-time dynamic systems synchronization: Information o transmission and model matching. Bifurcation and Chaos 15. [72] L. 31-39. (2000). Aguirre. [71] L. Tˆrres and L. Electronics o Letters 36.M. Chaos control of modified Chua’s System. Fast ortogonal identification of nonlinear stochastic models and radial basis functions. (1991). 7.A. J. Eng. 63-76. pp. 23.A.F.164 L. [69] L. Mendes o [68] J. (2007). Syst. 1.B. Billings. Physica D 196. Transmitting information by controlling o nonlinear oscillators.

a time-delayed feedback controller (TDFC) is used in conjunction with the proportional controller in its simple form as well as in its extended form (ETDFC). Iu School of Electrical.edu. using control theory and numerical simulations.fr H.C. France.uwa. how to apply a method to stabilize the desired periodic orbit for larger values of the proportional gain. residential and also aerospace environments. Electronic and Computer Engineering. Perth.1. To accomplish this aim. Its field of application is wide and concerns industrial. The main advantages of those methods are the robustness and ease of construction because they do not require the knowledge of an accurate model but only the period of the target unstable periodic orbit (UPO). Australia. Electrotechnics and Automatic. to improve the dynamical performances. herbert@ee.feki@univ-reims. The University of Western Australia. Our aim in this chapter is to show. Robert* and M. Moreover. 6.au Pulse width modulation (PWM) current-mode single phase inverters are known to exhibit bifurcations and chaos when parameters vary or if the gain of the proportional controller is arbitrarily increased. Introduction Power electronics is a discipline that has emerged from the need to convert electrical energy.H.Chapter 6 CHAOS CONTROL FOR A PWM H-BRIDGE INVERTER B. commercial. * bruno. University of Reims-Champagne-Ardenne.robert@univ-reims. 165 . an optimal criterion and an adaptive law are defined to determine the control parameters. Feki† Department of Electronics.fr † moez.

TDFC is known for its robustness and simplicity of construction. the TDFC has been proposed as an alternative method [16. This method has been successfully applied to the control of . 19]. Recently it has been shown by [19] that control parameters themselves may lead to bifurcation and chaos if they are not properly selected. Once the desired UPO is stabilized the control signal vanishes. 7]. In fact. much work has focused on bifurcation and chaos due to parameter variations [1. Knowing that a chaotic attractor contains infinitely many UPO. 17]. hence engineers usually attempted to avoid chaos. Besides. 11] have shown that switching power converters can exhibit several nonlinear phenomena. tools of analyzing bifurcations and chaos have been well developed. 11. the overall operation is compared to a piecewise smooth nonlinear dynamical system. bifurcations and chaos. Due to the existence of various operating modes and control saturations.166 B. Iu Power converters are basic switching circuits that are modeled by a number of linear differential equations corresponding to different topologies. The control signal is proportional to the difference between the current state and the τ -delayed state. Feki and H. Robert. control of chaos has been the focus of a growing literature. Since the seminal paper [15] (OGY). quasi-periodic operations. the OGY methods take advantage of the great sensitivity of chaotic orbits to stabilize a UPO by appropriately perturbing an accessible parameter [9. During the last two decades. but more often than not this phenomenon was considered as strange and undesirable. 4. To overcome this deficiency. 10]. These include subharmonic oscillations. Chaos in power electronics have been observed by many researchers [8. The toggling between different topologies can either be done naturally due to switching characteristics or done under the force of a feedback control system. the system model need not be known but only the length τ of the UPO to be stabilized is essential. Early studies by Hamill and coworkers [6. M. Therefore. Actually. the investigation of nonlinear dynamics in power electronics has become popular.C. These methods suffer from lack of robustness to imprecise measurements and to uncontrolled parameter variations.H. engineers have frequently encountered chaos in power electronics systems.

Section 6. respectively. A discrete model describing the behavior of the converter is developed and digital controllers are designed.5 will be devoted to the presentation of the TDFC and ETDFC controllers and their results. In section 6.2. 20]. Sections 6. 6. Using the Jury criteria a stability zone in the parameter space is defined.Chaos Control for a PWM H-Bridge Inverter 167 duffing equations [2]. In this section. However. With the goal to increase the efficiency of the power stages. In section 6.7. Finally. H-Bridge Model In order to increase dynamical performances. In this study we show the effect of the TDFC and ETDFC on the bifurcation diagrams and we present two-dimensional bifurcation diagrams. There is a wide range of conversion structures.4 and 6. 12. a growing number of applications in the field of electrical engineering require that they be fed by a precise current generator.3 gives an overview on the effect of the variation of the proportional gain and boosting of the chaotic behavior. Indeed.4 we show a two-degree of freedom design procedure and the resulting two-dimensional bifurcation diagram. It therefore follows that converters are necessary to adapt voltage sources to loads. A convenient way to adapt sources is to add a current control to a voltage converter. our conclusions and remarks are stated in section 6. Section 6. all industrial power sources are voltage generators. The ETDFC controller presented in section 6.6 includes results on sinusoidal output tracking. discrete chaotic systems [14] and to the control of power converters [3. In this chapter we propose to control a PWM current-mode H-bridge inverter using TDFC [13] and ETDFC methods in conjunction with the proportional controller. static converters operate by switching the load between several voltage sources. most of them are naturally inductive and are deemed to be current sources.5 leads to a three-degrees of freedom design and improves the results obtained in the preceding sections. owing to the presence of many windings in electrical machines. we describe the converter structure and its running mode before setting up its sampled data model. more or less .2 we present the discrete model of an H-bridge converter circuit. Our work will be outlined as follows.

The PWM modulator defines the switching period T. centered in the period. is very widespread. M. S2) and (S3. S2 off and S3.C. is the HBridge. States of the switches define two distinct topologies of the inverter. Fig. whose duration is defined by the current controller acting on the duty ratio d. For example. used in most variable speed drives. Feki and H. A well-known structure.H. a DC/AC converter. S2. and yield to two opposite voltages across the load.1 illustrates the circuit of this voltage inverter i. S4 off) and T2 (S1. The low state (C = 0) is split into two parts at the beginning and the end of the period. S3 and S4. 6. So the normal running mode . it is possible to vary the average output voltage. The output current is controlled by a current loop. whose choice depends on many parameters. Pairs are controlled by the PWM modulator in a complementary way. This technique. Iu complex. The four equivalent switches. It generates a high state (C = 1). dn denotes the duty cycle on the nth functioning period. Robert. By controlling the switching pattern over the operating period.e. T1 (S1. 6. H-bridge. are shared among two pairs (S1. topology T1 implies that the voltage across the load v = E.1. S4 on). R).168 B. called Pulse Width Modulation (PWM). S4). named S1. S2 on and S3. The inverter is fed by a voltage source E and it supplies a resistive and inductive load (L. S1 L E i v S4 R S3 S2 Fig. A switch is realized in this bridge by combining a bipolar transistor and an anti-parallel diode.

it is possible to track a sinusoidal current reference (see section 6.6). As explain before. For the sake of generality. a nonlinear map will be required as a model of the converter instead of a usual linear averaged state space model. At that point the sampling frequency is equal to the fundamental PWM frequency. Due to the sampled nature of the current controller. 6.2 also shows different possible waveforms of the output current under a normal periodic operating mode. Fig. In fact.2. For that aim (see the dashed curve).Chaos Control for a PWM H-Bridge Inverter 169 of the converter involves a sequence of three topologies: T2 −T1 −T2. it is important to reduce the current ripple within a period. This pattern avoids switchings and interferences at the sampling times because the digital controller samples current at the beginning of the switching period. Fig. the inverter period is chosen much smaller than the electric time constant L/R.2 and v = −(−1)CE. 6. the current is . This figure assumes also that the reference current In is constant. 6. Centered PWM voltage and current waveform. Then the voltage v evolves as depicted in Fig.

2ms. in +1 = αin + β (2 sinh(δd n ) − sinh(δ )). M.6375.32 × 10−3.e. (6. . the constants are chosen in order to make it easier with a power that does not exceed a few kW and to satisfy the frequency condition presented above. α = 670. 1]. For the sake of clarity. To maintain the ripple current at a low level.C.2). Robert. the model becomes a nonlinear discrete time map.1) where δ = RT/2L << 1 in order to reduce the ripple current. On each interval. we focus on the standard model (6. Saturated running modes are characterized by a unique topology over T namely T1 or T2.2) where α = e−2δ and β = 2e−δ. β = 1. it is important to note that due to boundedness of the duty cycle d. These phenomena have been thoroughly investigated in previous works [18. Feki and H. Iu scaled with respect to the maximum output current E/R and the voltage is scaled to E.H. the one-dimensional model is described by a single linear differential equation. Saturation of the modulator requires in fact a piecewise model and can be studied in an analytical way by the normal form theory. E = 400V. we set L/R = 0. With the objective of the design of an experimental prototype. 2δ dt (6. Because our goal in this chapter is to define a new controller stabilizing the normal running mode over a wide range. By integrating on the three intervals and stacking up the solutions.5ms and T = 0. hence parameters were chosen as follows: R = 40Ω.2. leading to the following constants: δ = 0. to keep the duty cycle d inside the interval [0. the PWM modulator introduces saturations to limit too much wandering i. Saturations lead to the emergence of border collision bifurcations when controller parameters vary.170 B. − 1 di = i + (−1) C . 19]. L = 20mH.

The proportional corrector is the main controller used to control the switching process and is given by: γ n = γ Pn = k ( I n − in ). 2 2 (6. Using the discrete-time map given by (6. the constants are chosen in order to make it easier with a power that does not exceed a few kW and to satisfy the frequency condition presented above. To maintain the ripple current at a low level. 6. the T-periodic output corresponds to a fixed point of order one. we set L/R = 0. The local stability of the T-periodic mode is analyzed by deriving the Jacobian of the discrete-time map in the neighborhood of the .5ms and T = 0. Current-programmed single phase inverter. In this case.2) and (6. the PWM modulator generates the duty cycle: dn = 1 1 + γ Pn = + k ( I n − in ).3) where k is a proportional gain and In is the reference current.3. 6. (6.3. it is important to have a T-periodic output current.3 depicts an H-Bridge current-programmed single phase inverter. Current-programmed Inverter Fig.4).Chaos Control for a PWM H-Bridge Inverter 171 6.2ms.4) Clock:T S1 v E L R S3 In in Controller γn PWM Modulator Current sensor i Digital Current Control S4 S2 Fig. when static converters are driven by a T-periodic clock. With the objective of the design of an experimental prototype. hence parameters were chosen as follows: As a matter of fact.

5) ⎧(α − 1)i * + β (2 sinh(δd * ) − sinh(δ )) = 0. fixed points. 6.6) . and thereby we can determine the loci of the eigenvalues when k varies.2) and the proportional controller (6.4) we obtain a first-order closed loop iteration with a unique eigenvalue λP given by: λ P = α − 2βδk cosh(δd * ). Feki and H. Robert.H. Iu Fig.C.4. Bifurcation diagram with proportional corrector.172 B. ⎪ 1 ⎨ d * = k ( I n − i * ). ⎪ 2 ⎩ (6. M. By solving in+1 = in we obtain the fixed point denoted by i∗ as a function of k. where d∗ is the solution of: (6. By combining the inverter model (6.

4. 2 (6. The aim of this section is to stabilize the UPO of length T.7) Therefore.8) where η is the TDFC gain to be adjusted. The bifurcation diagram depicted in Fig. 2 βδ (6. Controller Design The TDFC controller is similar to a proportional corrector but referenced to the same state delayed by a time τ equal to the length of the UPO to be stabilized. 6. Using the fact that δd∗ <<1 it follows that cosh(δd∗ ) ≈ 1 and k0 ≈ α +1 = 2. Time-delayed Feedback Controller 6.2) and (6.52 is obtained.Chaos Control for a PWM H-Bridge Inverter 173 When k varies the eigenvalue crosses the unit circle at λ0 = −1 with k = k0. Fig. 6. dn = 1 + k ( I n − in ) + η (in − in −1 ). The whole control signal and the duty cycle become: γ n = γ Pn + γ Dn .1.4. It has been shown in [19] that the fixed point of order one (corresponding to T-periodic orbit) continues to exist but becomes unstable.55.4 was obtained by iterating (6. 6.5 shows the block diagram of the TDFC controller. for k > k0 the T-periodic mode becomes unstable. Using the discrete-time notation.4) and the exact value of k0 = 2. this corresponds to stabilizing the fixed point of order one for higher values of k thus the TDFC expression is given by: γ Dn = η (in − in −1 ).9) . (6.

H. the discrete system defined by (6. 6.T Sampled Digital Control I in γDn Delay in-1 η εn γPn γn k PWM Fig. Robert.2) and (6. Let us define xn = in−1 − i∗ and yn = in − i∗ . 6. Fig. Stability zone in the k-η space. using xn and yn as the new state variables and combining (6.8) becomes second order since it involves in and in−1. Stability Analysis When TDFC is applied. M. 6.5.C.174 B. Iu Clock.2. Feki and H.2) and (6. TDFC controller.4.8) yields .6.

J (d * ) = ⎢ * * ⎥ J 1 (η . 1 − J 2 − J 1 > 0. d )⎦ ⎣ (6. d * ) = −2βηδ cosh(δd * ).12) The Jury criterion is used to evaluate the stability domain in terms of k and η. k . 1 + J 2 − J 1 > 0. (6.14) (6.16) . k . the system is stable if the following conditions are satisfied: − J 1 < 1. J 2 (η . (6.6) and it follows that J 1 (η .13) (6.10) We investigate the local stability by deriving the Jacobian evaluated at the origin: 0 1 ⎡ ⎤ .11) Since at equilibrium we have in = in−1 then TDFC signal is null. According to the Jury criterion. 2βδ cosh(δd * ) (6.15) η< 1 . Thus d∗ is again the solution of (6. * (6.13) yields to an upper limit on η (6. y n +1 = αy n + (α − 1)i + β (2 sinh(δd n ) − sinh(δ )). d ) J 2 (η .Chaos Control for a PWM H-Bridge Inverter 175 x n +1 = y n . The characteristic equation is given by: z 2 − J 2 z − J 1 = 0. d * ) = α + 2βδ (η − k ) cosh(δd * ).

then the control signal is similar to a Proportional-Derivative (P. 6.H.) controller: γ n = K P ε n + K D (ε n − ε n −1 ).15) leads to an affine relation between η and k: η> k α +1 − . Ordinarily we have KP > 0 and KD < 0. Iu (6.C. (6. The shaded zone has a special interest if we assume that the reference is constant. 2 4βδ cosh(δd * ) (6.16)–(6. (6. Feki and H. then the conjunction of the proportional and the TDFC controllers can be assimilated to a P. Robert. The last of Jury’s conditions (6. Denote εn = In−in the steady state error.176 B. in order to satisfy condition (6. In view of Fig.14) yields to a lower limit on k k> α −1 2βδ cosh(δd * ) . It is known from linear control theory that the larger the proportional gain k the less is the steady state error ε.D. 6.18). M. controller inside the shaded zone of Fig. a stability zone in the parameters space k − η is defined and depicted in Fig.18) Using conditions (6. then it is sufficient to choose k > 0.17) as well as the positiveness of the proportional corrector gain. 6. where KP = k and KD = −η.6.9) can be written in the following form: γ n = k ( I n − in ) − η (( I n − in ) − ( I n − in −1 )). k is maximized when η is maximized and the limit of stability is the point M which is the intersection of boundary lines L1 : η = 1 2βδ and L2 : η = k 1+α − 2 4βδ .17) We should note that α − 1 < 0.6.D.6. Indeed.

we show that a drawback of TDFC is to degrade the dynamical response by increasing the settling time.8 shows that the fixed point of order one that was embedded in a zone of period two mode (2. 6. 2βδ (6. Fig.4. 6. when η = 0 we observe that as k increases the output current is of period one then becomes period two and for a short interval of k becomes period four before it goes into chaotic mode.4. Zones of higher periods as well as chaotic zones are also shown as parameters k and η vary. We . We also notice that as η increases and a TDFC controller is acting then the period one mode extends for higher values of k as expected (see also Fig.3. Particularly.0) has been stabilized by the application of the TDFC in conjunction with the proportional controller. Optimality Criterion The main concern of the previous design was to obtain stability of the fixed point of order one for larger values of k which reduces the steady state error εn.4.5 with k = 2 and different values of η.16) and (6.7 indicates in brown the zone of the period one mode and this matches the stability zone shown in Fig. Results 1 2βδ = 1.Chaos Control for a PWM H-Bridge Inverter 177 obtained from (6.18) with the approximation cosh(δd∗ )≈ 1.20) The two-dimensional bifurcation diagram depicted in Fig.3 ≤ k ≤ 5.6.5.6. 6.9 shows the sampled and hold output current resulting from a step reference from In = 0 to In = 0.5 ≤ k ≤ 3. As k increases further the output becomes period six and finally arrives at period three mode and this matches very well with the one-dimensional bifurcation diagram shown in Fig.19) 3+α = 5. 6. 6. In this section. 6. (6.8).3) and in a chaotic attractor (3. 6. Fig. The coordinates of M are: η max ≈ k max ≈ 6.4.

10) gives ⎡ x n +1 ⎤ ⎢y ⎥ = ⎣ n +1 ⎦ This implies that [5] ⎡x ⎤ ⎡x ⎤ J ⎢ n ⎥ = J n +1 ⎢ 0 ⎥. thus η should be as small as possible. that is when the discriminant is equal to zero. When η tends to ηmax. 2 βδ (6. M.H. Robert.10 shows the loci of the eigenvalues when η varies for different values of k. the transient time becomes too large. To obtain a fast response we should choose k and η that minimize σ(J).4. η ≤ ηc).21) where π(n) ≥ 0 is a positive polynomial in n and σ(J) is the spectral radius of J.e. ⎣ yn ⎦ ⎣ y0 ⎦ x n +1 y n +1 = J n +1 x0 y0 ≤ π (n)σ ( J ) n +1 x0 y0 . 6. .11 depicts the effect of adapting the TDFC gain η according to the value of the proportional gain k.178 B. d * ) + 4 J 1 (η c . Indeed.C. Iu note that the transient time is increased when the TDFC is applied. 6. Solving for ηc we obtain ηc = k + 2 − α − 2 2βδk + 1 − α . In case of complex eigenvalues σ(J) = √ −J1. d * ) = 0. To overcome this problem we present an adaptive law to calculate a value of η that leads to a fast response for each value of k. Feki and H. The linearization of system (6. When η is less than or equal to a critical value ηc (i. the eigenvalues become real and the spectral radius is minimum if both eigenvalues are equal. (6.22) Fig. Fig. 6. for k = 4 the proportional gain on its own leads to a chaotic output as shown in Fig. 2 J 2 (η c . k .

8.7. 1-D bifurcation diagram with TDFC controller. 2-D bifurcation diagram.Chaos Control for a PWM H-Bridge Inverter 179 Fig. Fig. . 6. 6.

6 0.4 in 0. M.4 0 5 10 15 20 With TDFC η = 1 25 30 35 40 45 50 in 0.2 0 0.10. Eigenvalues loci for different values of η and k.2 0 0 5 10 15 20 Without TDFC 25 30 35 40 45 50 0. 6. Robert. 6.H.180 B.4 25 n 30 35 40 45 50 Fig. Dynamic responses to step with k=2.C. .9. Fig. Iu 0.4 in 0. Feki and H.2 0 0 5 10 15 20 With TDFC η = 1.

3) the fixed point of order one is stabilized and the step reference is tracked with a certain static error.2 0 0 5 10 15 20 With adaptive TDFC η = 0. The ETDFC extends the effect of earlier states to the present output with a decaying weight as we go further in the past. Eventually.86) yields to the stability of the fixed point with a considerably shorter settling time equal to 7 cycles (1. The output signal of the ETDFC is given by: .3 25 30 35 40 45 50 in 0.11. Dynamic responses to the step with k=4. Controller Design The ETDFC is a generalization of the TDFC and provides the designer with a third parameter r hence one more degree of freedom.5. However. Extended Time-delayed Feedback Controller 6.4ms). if the TDFC is arbitrarily added (η = 1.1.2 0 0 0.4 in 0.4 in 0.5.2 0 0 5 10 15 20 Without TDFC 25 30 35 40 45 50 0. Nevertheless.86 25 n 30 35 40 45 50 Fig. the settling time equal to 50 cycles (10ms) is significantly long. the adapted TDFC (η = 0.4 5 10 15 20 With TDFC η = 1. 6.Chaos Control for a PWM H-Bridge Inverter 181 0.6 0. 6.

The block diagram realization of the ETDFC is shown in Fig. Stability Analysis To derive necessary conditions for local stability of the fixed point of order one.23) where k.H. 6. Robert.C.24) Clock T I + k γPn r Delay ++ γn γEn PWM in Delay in−1 + η + + Fig.23) is equivalent to (6. η and 0 ≤ r < 1 are control parameters to be fixed to guarantee stability of the fixed point of order one. 6. Iu γ En = η ⎜ in − (1 − r )∑ r m −1in − m ⎟. ETDFC controller. m =1 ⎝ ⎠ .5.2. dn = n 1 ⎛ ⎞ + k ( I n − in ) + η ⎜ in − (1 − r )∑ r m −1in − m ⎟. Let us first denote n ⎛ ⎞ z n +1 = η ⎜ in − (1 − r )∑ r m −1in − m ⎟. 6.12. ⎝ m =1 ⎛ n ⎞ ⎠ (6. we start by finding the Jacobian matrix of the discrete-time map.8) if r = 0. 2 m =1 ⎝ ⎠ (6. The overall control signal and the duty cycle are expressed as: γ n = γ Pn + γ En . We clearly notice that (6. Feki and H.182 B.12. M.

25) Since at equilibrium ETDFC vanishes then again we can define xn = in−1−i∗ and yn = in−i∗ where i∗ and d∗ are the solutions of (6. yn and zn as the new state variables. (6. z n +1 = η ( y n − x n ) + rz n . system (6. J 3 = 2βrδ cosh(δd * ). .26) ⎡ 0 J = ⎢ J1 ⎢ ⎢− η ⎣ where 1 J2 η 0⎤ J3 ⎥ ⎥ r⎥ ⎦ J 1 = −2βηδ cosh(δd * ). Using xn. dn = 1 + k ( I n − in ) + η (in − in −1 ) + rz n . Hence the Jacobian matrix is (6. 2 z n +1 = η (in − in −1 ) + rz n . y n +1 = αy n + (α − 1)i * + β (2 sinh(δd n ) − sinh(δ )). J 2 = α + 2β (η − k )δ cosh(δd * ).25) becomes x n +1 = y n .Chaos Control for a PWM H-Bridge Inverter 183 and notice that zn+1 = η(in − in−1) + rzn. Thus the discrete-time inverter model can be described as follows: in +1 = αin + β (2 sinh(δd n ) − sinh(δ )).6).

2βδ cosh(δd * ) (6. The Jury stability criteria yields to conditions on control parameters k.31) We recall that since δ = 0.29) η < rk + η > rk − 1 − rα . the aim is to stabilize the origin of the system hence we obtain in = in−1 = i∗ and ETDFC becomes zero.27) Using the system description (6. * ⎟ 4 βδ cosh(δd ) ⎠ (6.28) η > (1 + r )⎜ − ⎜ ⎞ 1+ α ⎟.28) reduces to k > 0. Robert. M.30) (6. (6. ⎟ ⎝ ⎠ L4 : η = rk + 1 − rα . (6. 2βδ . η and r: k> α −1 2βδ cosh(δd * ) ⎛k ⎝2 .C. 2 βδ cosh(δd * ) 1 + rα . Feki and H.184 B.26). thus condition (6. Iu The characteristic equation of the linearized system is: z 3 − ( J 2 + r ) z 2 + (rJ 2 − ηJ 3 − J 1 ) z + ηJ 3 + rJ 1 = 0.2 and 0 ≤ d∗ ≤ 1 then we have cosh(δd∗ )≈ 1 and α < 1. Let us now define three lines (η = L(k)) ⎛ k 1+α ⎞ L3 : η = (1 + r )⎜ − ⎜ 2 4βδ ⎟.H. A proportional gain is usually considered positive.

6. it follows that condition (6. r) therefore they are parallel. ⎛ k 1+α ⎞ ⎛ 1 + rα ⎞ 1 − r ⎛ 1−α ⎞ (1 + r )⎜ − ⎜ 2 4 βδ ⎟ − ⎜ rk − 2βδ ⎟ = 2 ⎜ k + 2βδ ⎟ > 0.3. 2βδ We notice that L4 and L5 have the same slopes (i. Fig. ⎟ ⎜ ⎟ ⎜ ⎟ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ Thus. Stability zone in the parameter space. 6.14–6. Results Figs. It is worth noting that the period one zone shown in brown in Fig. We further see that since r < 1 then L3 is above L5 when k > 0.29) is satisfied and that as r tends to 1 condition (6. 6.e. 6. Moreover. we notice that when r tends to zero L4 ≡ L1 and L3 ≡ L2.29).13 sketches the stability zone in the parameters space for a fixed value of r.16 show two-dimensional bifurcation diagrams.5. Fig.31) tends to condition (6.31) is obviously satisfied when condition (6.Chaos Control for a PWM H-Bridge Inverter 185 L5 : η = rk − 1 + rα .13.14 matches the . 6.

186 B.H. 6. Feki and H. M. . 2-D bifurcation diagram (k-r).15. Fig.14. 2-D bifurcation diagram (k-η).C. Iu Fig. 6. Robert.

1-D bifurcation diagram with ETDFC controller. . 6. 2-D bifurcation diagram (η-r).Chaos Control for a PWM H-Bridge Inverter 187 Fig.17.16. 6. Fig.

Fig. Optimality Criterion Similarly to the case of TDFC. Robert. However. 6.188 B. the ETDFC stabilizes the fixed point of order one.4.C.5. 6. 6. small values of k no longer lead to stability but to different other modes.4. the ETDFC may induce significantly slow dynamic response if the parameters are not properly chosen.21).14 as well as in Fig. we should think of a sinusoidal mode functioning. M. For different values of k and r.20 sketches the sampled and hold output current for k = 6. . We notice that as we move to the boundary of the stability zone the settling time gets longer.18 shows the optimal value of η necessary to minimize the settling time. whereas for large values of r the settling time increases gradually. Fig. we can notice that for small values of r the settling time increases rapidly as we move towards the stability boundary. a new frequency condition has to be met: the reference frequency must be much smaller than the inverter’s in order to guarantee a variation of the current close to a sinusoid (as shown in Fig. 6.H. Moreover. When TDFC with adaptive value η is applied the response is again period 6.13. In the foregoing sections we have considered a constant reference current for simplicity of the analysis. In that case.17. 6. 6. Iu zone depicted in Fig.6. Feki and H. Moreover. we see that the fixed point of order one can be stabilized for higher values of k when η and r increase. In this section we present numerical simulations results to find optimal parameters values that lead to a fast response. this is clearly depicted in Fig. 6. 6. The criterion considered herein is to find the shortest settling time to be within an interval of 5% around the fixed point. 6.19 depicts the settling time in the number of periods. Fig. Results on Sinusoidal Output Tracking When considering an industrial application. The optimum value ηopt is coded on the colorbar and we see that it increases when both k and r increase. However. 6. When no TDFC is applied the response is period 6 as shown in Fig. we notice that for large values of η and r.

Optimal values of η for different values of k and r. 6. 6.Chaos Control for a PWM H-Bridge Inverter 189 Fig. Settling time shown in number of periods T.19.18. . Fig.

C.5 0 1 With TDFC η = 1. 6.21.5 0 0 5 10 15 20 25 n 30 35 40 45 50 Fig. Iu Without TDFC in 0. Input reference current and controlled current in the load. Dynamic responses to the step with k=6. Feki and H.73 0 5 10 15 20 25 30 35 40 45 50 in 0. M.H. r = 0.5 0 1 With TDFC η = 4 . Robert. .55 0 5 10 15 20 25 30 35 40 45 50 in 0.20. Fig.190 1 B. 6.

The sampled and hold current is sketched. 6. Fig. In Fig. In Fig.2ms. we apply adaptive TDFC.5.4 that when the proportional gain is raised to k = 4. besides T has to be sufficiently large to allow the application of a digital control and to limit the losses to an acceptable level. Considering a reference frequency ranging from zero to 100Hz and a minimal switching frequency to reference frequency ratio equal to 50. Fig. the switching period is upper limited to T = 0.22.24 we present sinusoidal . Sinusoidal reference tracking with k=2. 6.23 shows that the settling time can be further minimized with the action of ETDFC. 6. 6. and we see that when the adaptive TDFC or ETDFC is applied the performance is almost unchanged. Although.22 we compare in the case of k = 2 the performance of the proportional. TDFC should be applied to obtain stability of the period one mode.Chaos Control for a PWM H-Bridge Inverter 191 T has to be small enough for obtaining an output current close to a sinusoid. We have seen in section 6. TDFC and ETDFC controllers.

Robert.H.24. Sinusoidal reference tracking with k=7. Feki and H.192 B. Iu Fig.2. 6.C. Fig. .23. Sinusoidal reference tracking with k=4. M.5. 6.

6. We note that a suitable choice of the parameters leads to faster settling time. we have determined the stability domain of .2 which can only be obtained with ETDFC. it is possible to predict and control chaotic behaviors.25. In this work TDFC presented in section 6.7.Chaos Control for a PWM H-Bridge Inverter 193 Fig.25 the continuous time-domain waveform of the output current is depicted to show how the ripple is minimized in the periodic mode compared to the chaotic mode.4 and ETDFC presented in section 6. The Extended Time-Delayed Feedback Control is an efficient method to stabilize unstable periodic orbits.5 have been applied to stabilize a current-programmed PWM single phase inverter. By analyzing the Jacobian matrix of the map. Conclusion By modeling power electronic converters using nonlinear discrete time maps.2. Time-domain waveform for sinusoidal reference tracking with k=7. 6. In Fig. reference tracking when k = 7. 6.

M. Callier and C.-I 47.C.22)] and numerical analysis of ETDFC has been presented to give the optimal parameters that lead to minimum settling time (see Fig.22–6. El Aroudi. Toribio and E. [2] M. subharmonics and chaos in power electronic systems. Gomez. Bifurcation and Chaos 13. E. 6. DP0559109). Acknowledgement This work is supported in part by ARC Discovery Projects Grant (No. an adaptive TDFC has been proposed [see (6. Benadero. Power Electron. Chaos in power electronics: An overview. G. Feki and H. It has been found that the TDFC and ETDFC induce longer settling time if parameters are not appropriately chosen (see Figs. Appl.9 and 6. 27. Linear System Theory (Springer-Verlag 1991). IEEE Trans. [8] M. (1990). Circuits Syst. [7] M. Desoer. K. Genesio and A. (1999). Circuits Syst. Results on tracking sinusoidal reference have been presented (see Figs. Stabilization of periodic orbits of the buck converter by time-delayed feedback. 617–631.-I 44. Batle. Hamill. Int. [3] C. Circuit Th. Chen and T. 5. We have also presented the stability zones in the parameters space (see Figs. 6. Fossas and G. 16. Chap. E. 260–268. Vasca. Olivar. 13.24) with the perspective of an experimental realization of the presented control methods. Basso. Chakrabarty. IEEE Trans. (1998). On discrete time maps for the analysis of bifurcations and chaos in dc/dc converters. (1997). Iu the T-periodic running mode. Singapore. Deane and D. Tesi.13). The range of the stable T-periodic mode of the inverter is widened and the dynamical performances are improved. A. 317–340.7. R. Di Bernardo and F. Tse. 130–143. Instability. 6. J. Olivar. Nonlinear modeling and bifurcations in the boost converter. eds. G. 6. Int. (2003). (World Scientific. Robert. 6. 2002). 252–260. Ueta.19). [4] L.16) corresponding to different control parameters. Power Electron.14–6. To tackle this problem. we have presented 2-D bifurcation diagrams (see Figs. Stabilizing periodic orbits of forced systems via generalized Pyragas controllers. For each controller. Two-dimensional bifurcation diagrams background pattern of fundamental DC-DC converters with PWM control. (2000). 6. 1023–1027.194 B. J. . [6] J.18). 427–451. Di Bernardo and C.6 and 6. References [1] S.H. Chaos in Circuits and Systems 11. It may be worthwhile to study the effect of the system parameters variations on these stability zones. Banerjee and K. IEEE Trans. IEEE Trans. [5] F.

(2003). Konishi and H.-I 50. Lett. Croatia. Ditto. (1992). 323– 330. of Shanghai Int. Int. A. Spano and J. IEEE Trans. Rev. Robert. A 206. (2003. Robert and C. Nonlinear Sciences and Applications Shanghai. Robert. Jefferies. IEEE Trans. Iu and B. Control of chaos in a PWM current-mode H-bridge inverter using time-delayed feedback. [18] B. (2002). Lett. 1125–1129. Controlling chaos in high dimension. 75. Techniques for the control of chaos. Circuits Syst. Feki. Phys. “Observer-based delayed-feedback control for discretetime chaotic systems. Pyragas. [20] B. C. [12] H. 1196– 1199. [19] B. Power Electronics and Motion Control Conf. A 248. [13] H. Circuits Syst. China. Circuits Syst. IEEE Trans. A 170. 519–534. Yorke. Pyragas. Robert. (1988). Border collision bifurcations in a one-dimensional piecewise smooth map for a PWM current-programmed H-bridge inverter. Hamill and D. Subharmonics and chaos in a controlled switched-mode power converter. 1059–1061. Physica D 86. Control of chaos in a PWM inverter using time-delayed feedback. 971–975.Chaos Control for a PWM H-Bridge Inverter 195 [9] W. Iu and M. . Grebogi and Y. J. C. Kokame. [11] D. (1998). Adaptive time-delayed feedback for chaos control in a PWM single phase inverter.-I 44. Symp. Robert. Proc. Lett. 421–428. 1356–1367. 359–368. (1990). 64. [10] C. Continuous control of chaos by self-controlling feedback. Lai. 198–211. Phys. 10th Int. (1995). Comput. M. [15] E. Iu and B. Circuits and Syst. H. Phys.-I 35. Grebogi and J. Ott. (1997). [16] K. EPE-PEMC Dubrovnik. Robert. [14] K. [17] K. J. Phys. Lett. Robert and C. (1995). Contr. Proc. Linder. 13. (2002). Controlling chaos. Control of chaos via extended delay feedback. (2004). Border collision bifurcations in a chaotic PWM H-bridge single phase inverter.

This page intentionally left blank .

Under certain conditions. is a pathological state of the brain that afflicts over one percent of the world’s population. Chicago.Chapter 7 CHAOS CONTROL OF EPILEPTIFORM BURSTING IN THE BRAIN M. 837 State Street. One experimental model used to test these algorithms is the brain slice where a select region of the brain is cut and kept viable in a well-oxygenated artificial cerebrospinal fluid. defined as recurrent seizures. IL 60616. Several groups are exploring the use of electrical stimulation to terminate or prevent epileptic seizures.cvitanovic@physics.W.edu Epilepsy. Although pharmacological agents are the primary treatment for preventing or reducing the incidence of these seizures. Mogul Department of Biomedical Engineering. USA Predrag.northwestern. 251 East Huron Street. 10 West 32nd Street. Room 116. thereby providing an in vitro 197 .edu P. USA mogul@iit. Cvitanovic School of Physics.gatech. Northwestern Memorial Hospital. IL 60611. USA mslutzky@md. Slutzky Department of Neurology. over 30% of epilepsy cases are not adequately helped by standard medical therapies. GA 30332. Chicago.edu D. such slices may be made to spontaneously and repetitively burst. Illinois Institute of Technology. Atlanta.J. Georgia Institute of Technology. Seizures occur as populations of neurons in the brain become overly synchronized.

Some of these patients have the option of surgically removing the seizure-generating part of the brain (the focus). so their long-term effects are unclear.W. thus there is a great deal of room for improvement. not to mention adverse events such as infections and hemorrhage [3]. but do not often prevent all seizures. anxiety and depression) that come from having seizures. and over 30% of these people are not sufficiently helped by medications [1]. we discuss our efforts at applying chaos analysis and chaos control algorithms for manipulating this seizure-like behavior in a brain slice model. These techniques possess the advantage of requiring relatively infrequent stimulation of the tissue. Mogul model of epilepsy. This would reduce the . P. many groups are working to develop less-invasive alternatives to surgery for treating drug-resistant epilepsy [4]. Some currently available therapies use electrical stimulation of either the vagus nerve [5]. Moreover. it can have serious side effects such as memory loss –in up to 35% of patients [2] – or speech deficits. their mechanism of action is uncertain and they have several adverse effects [5]. Introduction More than 60 million people worldwide are afflicted with epilepsy. In this chapter.J. Therefore. These stimulators have only been available for a few years. since studies have shown that a reduction in seizure frequency alone does not remove their psychosocial issues (i. However. While surgery is often successful in preventing seizures. Cvitanovic and D.198 M. These techniques may provide a nonlinear control pathway for terminating or potentially preventing epileptic seizures in the whole brain. Long-term stimulation of the brain at a tonic high frequency could potentially create new epileptic foci in those areas. Schiff et al. Slutzky.1. Vagus nerve stimulators reduce the number of seizures by 50% or more in approximately 35-40% of patients [9].e. the centromedian thalamic nucleus [6]. Existing stimulators most often succeed in reducing seizure frequency. [10] suggested that chaos control techniques might be used to prevent or stop seizures with intermittent electrical stimuli. 7. the anterior thalamic nucleus [7]. the goal of invasive therapies is now to render patients seizure-free. or the hippocampus [8] to stop seizures with modest success. not just reduce seizure frequency.

15] which can be misleading in practice because of the need for many assumptions and the prevalence of false positive results. synchronous discharge of one or several groups of neurons in the brain. thus. the period during which a seizure takes place is referred to as the ictal period. recorded from the scalp) during this interictal period. one should first demonstrate that the system is indeed chaotic. the hippocampal slice is a frequently used in vitro model for investigating seizures.Chaos Control of Epileptiform Bursting in the Brain 199 likelihood of inducing new epileptic seizures and decrease power requirements. 17]. 7. Searching for Evidence of Chaos in the Brain A seizure is an abnormal. However. 18] and contradicting [17] the existence of nonlinear determinism. On an EEG. Other measures have shown evidence either . Bursts are the in vitro analogues of these spikes and can be induced to occur spontaneously in the transverse rat hippocampal slice providing an in vitro model of epilepsy. Several studies of EEG voltage have focused on calculating the correlation dimension [14. 19]. Complex partial seizures are the most common type of seizures in adults and are the most likely to be refractory to medical treatment. Most complex partial seizures arise in the temporal lobe. One of the hallmarks of epilepsy is the presence of spikes in the electroencephalogram (EEG. and found evidence both supporting [16. the period between seizures is thus the interictal period.2. prior to using chaos control techniques. However. While the precise role of interictal spikes in epileptogenesis is not currently certain [11-13]. excessive. the algorithms used perform poorly in systems with a substantial stochastic component in addition to determinism. and thus have not proven very reliable in characterizing biological systems [17. Significant efforts have been made to try to characterize the system behavior in epilepsy. both important considerations for an implanted device. The hippocampus is buried deep within the temporal lobe of the brain and makes up part of the limbic system. Some have searched for chaos by calculating Lyapunov exponents of EEG seizure activity [16. it is plausible that developing a method to control them could provide a way to control seizure activity as well. particularly in the hippocampus or amygdala.

However. a) b) c) Fig. Cvitanovic and D. Fixed points lie along the identity (45o) line. a) Schematic of hippocampal slice organization. 7. it is difficult to achieve either of these conditions in a neurophysiological system. Our laboratory used a combination of new [27] and old techniques to characterize epileptiform behavior [28].200 M. stationary time series or low noise. Mogul for the presence of determinism in experimental data. b) Example of a spontaneous burst recorded extracellularly in the CA3 pyramidal layer.1. P. The recording electrode was placed in the pyramidal cell body layer (P) of the CA3 region. Bursting in the hippocampal slice. The stimulating electrode was placed in the Schaffer collateral (SC) axons to the CA1 region. . c) A return map of 1000 interburst intervals (IBIs) recorded during spontaneous bursting. most of these methods rely on either very long. [20-23] or against it [24-26]. Slutzky.J.W.

1b. zero magnesium.1. Fig.g. thus producing a one-dimensional time series. Using time-delay embedding.. IBIs were recorded at one point in space. Global Measure of Chaos: Lyapunov Exponent Estimates and Short-time Expansion Exponential divergence of two nearly identical points in state space can be quantified by the Lyapunov exponents of a system. see Fig. For small and noisy data sets. The system state variable was interburst intervals (IBIs). 7.Chaos Control of Epileptiform Bursting in the Brain 201 7. Such IBI embeddings have also been shown to be sufficient to characterize most of the original system’s dynamical behavior [30]. 7. Most algorithms. the computation of Lyapunov exponents can be very difficult. In the case of short data sets recorded from in vitro bursting experiments in rat hippocampi. 7. If the largest exponent in a deterministic system is positive. Epileptiform bursting was induced by bathing the slices in artificial cerebrospinal fluid (ACSF) containing either relatively high potassium concentration ([K+]o=10.1a shows the setup used to record and manipulate bursts. Both morphology and burst duration were consistent for all three induction protocols. Transforming Brain Signals into State Space Our laboratory investigated epileptiform behavior in the in vitro hippocampal slice. scalar data can be converted into vectors which form a more complete representation of the system in state space and preserve the geometric and dynamical properties of the system [29].5 mM). or the neurotransmitter gamma-aminobutyric acid (GABAA) antagonists bicuculline and picrotoxin. 7. it was necessary to use larger initial neighborhoods to . this method did not provide meaningful information for short time series with a great deal of noise or extremely rapid expansion [28].2.1c). A typical burst is shown in Fig. this signifies exponential expansion and thus chaos. IBIs were first embedded into two-dimensional vectors consisting of the current and next IBIs (e.2.2. in our previous experiments. look at the expansion rates of small neighborhoods of points localized in state space and average them out as time evolves. However. in particular Kantz’ method [31].

short-time expansion (STE).. It averaged this rate of spread over the entire “attractor” (i. The global expansion rate (Lave) was then obtained by averaging the local estimates. ntotal was the total number of points.J. this comparison enabled us to identify the presence of determinism in the data.W.e. These larger neighborhoods then expanded to the size of the entire attractor so quickly that it was impossible to calculate an accurate average expansion rate. Cvitanovic and D. Slutzky. For each point in the data set. we developed a new method of detecting determinism. Lave = ∑ i =1 B λi ⋅ ni . That is. P. and the square root of the ratio of the two largest principal components then provided a measure of the one-step expansion rate. To overcome these obstacles. Thus. it was a global measure of the mean expansion rate over the entire state space. Mogul obtain enough initial points for the method to work correctly. The square root of the largest principal component component (i.. which measured the rate of spread among nearby points in state space after only one time step [27]. a given number of nearest neighbors was found and was then fit to an ellipse using principal components analysis (PCA).e. The points in the neighborhood were then evolved one iterate into the future. PCA was again performed on the evolved neighborhood. it would converge to the largest Lyapunov exponent in a noiseless system.  p0  (7.202 M. all the points in the data set) which reduced the effects of local noise on the computations.1) where p0 and p1 were the largest principal components of the initial and iterated clouds of points. the standard deviation) was used as a measure of the initial spread between the points. ni was the number of points in each box. The global expansion rate of the data was then compared with that of randomized surrogates of the data. While the resulting measure of expansion was not a Lyapunov exponent. respectively. B was the number of . The local Lyapunov exponent could be estimated by the natural logarithm of this rate. ntotal  p  λi = ln  1  . Since a deterministic system should expand at a slower rate than a stochastic systems (if the amplitude of the noise was comparable to the size of the attractor).

If the system contained a large stochastic component. Lave should be smaller for a deterministic system than for a stochastic system. nearest neighbor (NN) curve for the original data was significantly different from the average curve computed for the surrogates. 7. Also. then the likelihood of determinism being present would be strengthened. Curves of Lave for one set of high-[K+]o experimental data (M) and corresponding surrogates (F) both decline logarithmically to zero with increasing NN. and λi was the local expansion rate. the points would likely spread out over most if not all of the attractor after only one time step. we generated surrogate data by randomly shuffling the order of the IBIs [19]. No plateaus are evident which indicates that the data are stochastic.2. The surrogate data represented the null hypothesis that the experimental data could be explained by a linear stochastic process.Chaos Control of Epileptiform Bursting in the Brain 203 boxes. In order to test for determinism in the short-time expansion rate analysis. . If the Lave vs. additive (extrinsic) noise should average out in this calculation leaving primarily the deterministic component. Therefore. Fig. It was expected that after only one time step small neighborhoods of points would not spread out as quickly in deterministic systems as they would in strongly stochastic systems.

This technique uses a transformation to concentrate data around UPOs thus creating peaks in the histogram of the transformed data that correspond to the locations of the UPOs.3. especially if the amplitude of the noise were larger than the region in which control was desired. The significance of these peaks was assessed using surrogate data. Cvitanovic and D. since UPOs are the points around which control can be applied [10]. there was no noticeable flattening of the data curve. [13] to search for low-period orbits. Local Measure of Chaos: Unstable Periodic Orbits A chaotic attractor can be described by a skeleton of unstable periodic orbits (UPOs) [32]. So et al. Other data curves displayed even greater similarity to their surrogates’ curves.204 M. This suggested that chaos control might be difficult to achieve in practice. Furthermore. another group found period-1 orbits in guinea pig hippocampal slices induced to burst using soman and 4-aminopyridine [35]. it was almost parallel with the surrogate mean curve. The curve in Fig. 7. Le Van Quyen et al. Our group used the method of So et al.2 shows the experiment with the biggest difference between data and corresponding surrogate average. 7. [34] found several period-1 orbits in both the high-[K+]o in vitro model and in human epileptic EEG recordings as well as a few orbits of periods 2 and 3 in intracellular recordings from CA1 neurons in normal ACSF. However. the curve in this case was displaced from the surrogates. More recently.W. which may mean that it is not quite as disordered as the surrogates. These UPOs are periodic paths in state space to and from which the system recurrently approaches and recedes. Thus. the results of this analysis suggest that globally the bursting data contain a great deal of stochastic noise with no determinism detectable by our measure of the global expansion rate.2.J. Mogul This method was used to analyze 12 sets of IBI data from high-[K+]o and zero-[Mg2+]o experiments and 5 surrogates of each data set. Even in this case. The presence of UPOs in a system implies the presence of determinism and suggests chaos. . [33] found evidence of a few period-1 orbits in epileptic EEG recordings from three human subjects. Slutzky. the presence of UPOs strengthens the rationale for using chaos control techniques to manipulate bursting. P.

The transformation shifted all points in the linear region of a fixed point x* even closer to x*. Note that the peak for the surrogates is much lower than that for the transformed IBI data.. (7. This transform was applied to every point x in the set and summarized in a spatial distribution function of the experimental data approximated by a histogram with a bin size of 0. Since the peak in Fig. F(x) was a vector of the next iterate of x.3c is above the 96% significance line.02 seconds. drifting of the mean IBI length) due to factors such as fatigue of the neurons.e. it marks the location of a . R ) ⋅ z]. After the transform is performed on the data (Fig. A histogram of one window of the original (untransformed) IBIs from a high-[K+]o experiment is shown in Fig. network plasticity. or fluctuations in temperature and pressure. R ) = ∇ F( z ) + R ⋅ [F( z ) − z].3a.98 s).R) was a d×d matrix function of x and a d×d×d random tensor R given by S( z. Fifty surrogate data sets ^ were generated and transformed. and S(x. 7.3b). A cumulative histogram of the maximum deviation of each surrogate from the surrogate mean at each point was used to estimate the probability that ZW a peak in the transformed data was due to the presence of a true UPO at that point. 7. 7.2) where x was the d-dimensional time-delay vector of IBIs. and their corresponding probability distributions were calculated and averaged together. I was the identity matrix. thus creating a peak in the distribution function (x).Chaos Control of Epileptiform Bursting in the Brain 205 For period-1 orbits. the transformation was defined as z ≡ [I − S( z. (7. a sharp peak in the data (solid line) is seen at the corresponding location of the period-1 orbit (1.3) Here ∇F(x) was the d×d Jacobian matrix of F(x). each set of data was divided into windows of 256 IBIs and then analyzed. A UPO was declared statistically significant if the peak in the distribution function of the data (minus the surrogate mean) was greater than 96% of the maximal peaks of the transformed surrogates (minus the surrogate mean). Note that no discernible peak is shown at the location of the period-1 orbit. R )]−1 ⋅ [F( z ) − S( z. which was calculated using a least-squares fit of 3 nearest neighbors in state space. To account for nonstationarity of the data (i. The dashed line shows the histogram of the mean of the transformed surrogates.

and 75% of the eight GABAA blockade experiments. b) Histogram of transformed data (solid) and mean of transformed surrogates (dotted). 7. 7.J.4b) were found in more than 35% of the high-[K+]o experiments. Period-1 orbits were found in 71% of the 17 high-[K+]o experiments. 7. and disappeared over the course of an experiment.98 s. and 50% of the GABAA antagonist .98 s) with p<0. Detection of a period-1 orbit using the UPO transform. Slutzky.4a) and period-3 (Fig.W. period-1 orbits emerged. P.206 M.04 (dashed line represents 96% significance line). a) Histogram of the raw IBI data. Cvitanovic and D. The case for chaos is made even stronger when orbits of higher periods are considered. UPOs of period 2 (Fig. Due to nonstationarity. drifted. Mogul a) b) c) Fig. 50% of the zero-[Mg2+]o experiments.3. 25% of the eight zero-[Mg2+]o preparations. true period-1 orbit at 1. c) A fixed point is found at the location of the peak (1.

The significance in this three-dimensional representation of a two-dimensional histogram is coded by color and height. Chaos Control of Epileptiform Bursting The property of sensitive dependence on initial conditions would lead one to believe that chaotic systems cannot be controlled.” Thus our laboratory set out to control bursting using a more rigorous approach.. they might suggest that in vitro epileptiform bursting may contain local islands of determinism (UPO detection) within a globally stochastic sea (Lyapunov analysis). It is possible that the noise level of the system was so high that it drowned out the determinism even using global averages. the interburst interval) instead of a system parameter. [10] were the first to control epileptiform bursting in the hippocampal slice using PPF.. However. their results did not reveal how many of the bursts were “stimulated” vs. We used a modification of PPF called stable manifold placement (SMP) [46] which is simpler and more robust than PPF because it requires less assumptions to be made about system parameters. A variant of OGY called proportional perturbation feedback (PPF) [43] modified the system state point (viz. OGY control). and these manifolds can be linearly approximated within a small radius (hereafter referred to as the control radius).Chaos Control of Epileptiform Bursting in the Brain 207 experiments. While these results seem to present conflicting evidence. how many were “spontaneous. and Yorke [37] showed that chaos can be controlled using relatively small perturbations to system parameters (a.3. The basic principle is that all saddle fixed points have at least one stable and one unstable manifold associated with them. Thus strong evidence of determinism was found using UPO analysis.a. Ott. Grebogi. However. 7. it has been shown [36] that it is indeed possible to control chaotic behavior – i. Another. simpler method was developed by Hunt [38] and successfully applied to a nonlinear diode circuit.k. Schiff et al. Modifications to these two techniques have subsequently been developed. .e. Other control schemes using artificial neural networks have also been proposed [39-41]. Since then. to move the system from a chaotic orbit to a periodic orbit. many chaos control algorithms have been advanced [42-45].

In this section. we explore in detail the modification of bursting behavior using techniques from chaos control. that helps determine the validity of fixed point estimates and assesses the feasibility of chaos control. The two sets of three red peaks (circled in green and blue) manifest the three points along two period-3 orbits. .W. Accurate fixed point estimation is crucial to the success of chaos control. developed by Christini and Kaplan [47]. we implemented for the first time in a biological system a method of continuously refining the fixed point and stable manifold estimates.4. P. in addition to SMP control. b) Two period-3 orbits from a zero[Mg2+]o model. Significance is shown by the color bar on the right and by the height. A fixed point is circled in orange. Therefore. a) Significance plot of a period-2 orbit (red peaks) in a high-[ K+]o. We also describe a novel protocol.J. state-point forcing. Cvitanovic and D.208 M. Period-2 and period-3 orbits in 3-D representations of 2-D state space. Mogul a) b) Fig. Slutzky. 7.

calculated when a stimulus needed to be delivered. when the state point fell outside of the control radius. 7. and sent out the stimulus signal. The host computer ran the control software that performed fixed point detection (using the algorithm described above [34. it was obtained through simple algebra.Chaos Control of Epileptiform Bursting in the Brain 209 7.1a). The control algorithm also had to account for the delay between the stimulus and the recorded . instead of moving the stable manifold to the state point. Thus.5). A bipolar tungsten stimulating electrode was placed in the Schaffer collaterals (Fig. 80 µs square-wave current pulses with amplitudes 0.4) where xn was the current IBI. since there was no notable effect of changing it.3.3 mA.1. the variance of IBIs was reduced compared to uncontrolled bursting. Stimuli consisted of single. 38] with 10 surrogates used for significance testing).2. 7. adaptive control techniques. This assisted our assessment of the quality of control attained. data storage and display.1-0. the same number for the Hénon map. the desired IBI value was given by xn+1= λs (xn . The displayed IBIs were marked as stimulated or unstimulated. Using the SMP algorithm. x* was the fixed point. and enabled changes to system parameters and control techniques in real time. Slice Electrophysiology A real-time data acquisition processor recorded the IBIs.3. and λs was the eigenvalue (slope) of the stable manifold. 7. The number of nearest neighbors used to fit the Jacobian matrix for the UPO transform was optimized to 4.x*) + x*. instead. (7. The advantage of using SMP was that the calculation of the desired IBI did not require estimation of the next “natural” IBI (as PPF does). There were several parameters that could affect the performance of this algorithm. That is. but tight control was not achieved (Fig. the state point was moved onto the stable manifold. including several parameters in the UPO detection algorithm. a stimulus was sent to the slice to trigger a burst at the exact time needed to shift the state point onto the stable manifold. SMP Control In SMP control.

5. It was kept constant during each experiment. Beginning of control and control radius marked by dashed lines. When RUPO was small (0.J. P.5-1 s) and there were several long IBIs. but there was no noticeable improvement in control quality. 7. we also tried applying control only when the previous state point was within a set radius (RUPO) of the fixed point.3.3.210 M. IBIs were stimulated (●) vs. SMP control in a bursting experiment. sometimes it would take a long time before the system state would get close enough to the fixed point for control to resume. 7. Variance of IBIs did decrease modestly with control. natural (○). Cvitanovic and D. Mogul Fig. Effect of Control Radius (Rc) and Synaptic Plasticity on Control Efficacy Previous work [49] suggested that stimulating the Schaffer collaterals at low frequencies (~1 Hz) could sometimes cause a form of synaptic plasticity called long term depression (LTD). burst due to finite conduction velocity in the tissue. RUPO was varied from 0.W. Since the linear approximation of the stable manifold was only accurate within a small distance of the fixed point. Slutzky.5-10 s. which could cause IBIs to . This stimulus-burst delay was varied from 15-50 ms depending on the properties of each slice.

as well as Christini and Collins [44]. it could be possible that their control algorithm was constantly stimulating due to a very small control radius. That is. the interval from when the stimulus was applied to when the burst was detected fluctuated by a few milliseconds from burst to burst. This occurred because Rc was smaller (2 ms) than the precision of the stimulus-burst interval. 7. [10]. Rc was again decreased in stepwise fashion. We found that the control radius that optimally balanced variance vs. and then in high-[K+]o ACSF containing 50 µM AP-5.6 shows examples of experiments in which Rc was varied using high-[K+]o only (Fig. 7. At extremely small control radii. This conclusion was true for both the standard high-[K+]o experiments and for experiments using AP-5. the system performed almost identically to demand pacing. and therefore the estimate could be off by as much as 5 ms. As Rc decreased. To assess whether some of the nonstationarity and impediments to control were due to this form of synaptic plasticity. Fig. the NMDA-receptor antagonist AP-5 was used. and four were done using high-[K+]o plus AP-5. While LTD might occur using other receptors besides the NMDA type. Slices were bathed in high-[K+]o ACSF for 20 minutes.6b).6a) or high-[K+]o with AP-5 (Fig. This could explain why our control of bursting did not seem as striking as those reported earlier. the control algorithm kept stimulating but could not get the IBI within the control region. 7. This similarity of results suggested that synaptic plasticity does not significantly influence the quality of control obtained. Thus. Six experiments were done with high-[K+]o only. number of stimuli was usually 50-100 ms depending on the initial variance of the bursting. variance decreased but the proportion of IBIs that were stimulated increased. the results looked very similar to those obtained by Schiff et al. In this case.Chaos Control of Epileptiform Bursting in the Brain 211 lengthen over time. AP-5 should block the majority of LTD via this pathway. The effect of the size of Rc on control efficacy was investigated both with and without AP-5. especially if the IBIs were not designated as stimulated or natural. The AP-5 solution was washed in for 5 minutes before starting the same control sequence used above with high-[K+]o. Since their work did not make this distinction.6a). nearly every IBI was stimulated (see inset in Fig. Interestingly. . 7.

One way of overcoming this obstacle is to dynamically refine the approximation of these parameters. 7. Assuming that the data were in the neighborhood of a UPO. but the ratio of stimulated (●) to unstimulated (○) IBIs increased.4. solid lines). Cvitanovic and D.6. Thus.212 M. Adaptive Control Techniques Accurate estimates of the fixed point location and stable manifold slope were the two key elements needed to achieve good control. Experiments varying Rc using basic SMP control. the variance also decreased. Nonstationarity is a significant obstacle in both estimates. there did not appear to be any effects from LTD. adapted from Christini and Kaplan [47] fitted data to linear approximations of the dynamics in the neighborhood of the fixed point to re-estimate the fixed point and stable manifold slope. The magnification (right) of the boxed area shows the demand pacing phenomenon: all of the IBIs after IBI 1370 are stimulated.W. since the system can never get within the tiny Rc. especially in fixed point location.J. Slutzky. the unperturbed system . 7. Mogul a) b) Fig. P. b) Similar results were found in experiments using AP-5. This adaptive tracking method. As Rc decreased.3. a) A high-[K+]o experiment varying control radius (Rc.

Also. xn). but the preceding two points could be stimulated or unstimulated – were used. Too few triplets could cause poor fits and result in volatile fluctuations of the parameter values. The least-squares fit itself had several notable caveats. The algorithm successfully adjusted for the drift of the system (Fig. Our refinements of the tracking algorithm and its parameters produced small but incremental improvements in control quality. the stable manifold could not be estimated accurately since the state point would not approach the fixed . If most of the IBIs were stimulated. However. This regression was performed after every new natural triplet (or. In this case. respectively. The parameters x*. equivalently. this equation became x n +1 = ( λs + λu ) x n − λs λu x n −1 + x *(1 + λs λu − λs − λu ) (7. The number of natural triplets used for the fit (NT) varied from 4-20. Only “natural triplets” – those where xn+1 was an unstimulated IBI. xn.Chaos Control of Epileptiform Bursting in the Brain 213 dynamics could be approximated by the linear equation xn+1 = axn + bxn-1 + c. Adaptive tracking noticeably improved control over basic SMP alone [48]. the state point tended to jump out from the fixed point along the unstable manifold. where the current state point was (xn-1. when unstimulated bursts occurred. When a control stimulus was applied. Several parameters had to be carefully adjusted. Rewritten in terms of the fixed point parameters. the system would behave according to (7. That is.7a).5). Too many would not allow the algorithm to track the system quickly enough. they circumvented the normal fixed point dynamics. natural IBI). and λu could then be estimated by a least-squares fit of the data triplets (xn+1. λs. 7. the system dynamics were described by x n + 1 = λs ( x n − x * ) + x * ^ * (7.6) where signifies that λs and x were estimates. and then control was turned off. xn-1). when control stimuli were applied. Therefore. the fit would sometimes not accurately represent the natural fixed point dynamics. the natural values of λs and x* could not be estimated from stimulated bursts.5) where x* was the fixed point and λs and λu were the stable and unstable eigenvalues.

While the results using this technique were not dramatic. depending on the variance of all the data. the parameter estimates would not be adjusted. as seen in Fig. We therefore required that the new estimate move in the same direction as most of the natural IBIs are located relative to the current fixed point. NT was set to 10. singular value decomposition (SVD) was used to perform the least-squares fit. but if FAM was set too small. Even with these modifications. but the fixed point estimates still often fluctuated. the algorithm wouldn’t adapt well. the fixed point estimate tracked with the system smoothly over time. 7. the new fixed point estimate would still sometimes be set smaller than the current estimate. the tracking algorithm produced some fluctuations in the fixed point estimates which inhibited control. Therefore. we limited the distance by which the fixed point estimate could change from the current estimate in any one fit. Slutzky. in proportion to the initial variance of the system. On several occasions in tracking experiments. Improving the fit also reduced fluctuations in the fixed point estimates. If the ratio between the fit’s largest and smallest singular values were very large (>106). Cvitanovic and D. The optimal range for RNT was about 0. In that case. For example. Therefore. they were a definite improvement over basic SMP control.J. Mogul point at all. This modification helped stabilize control.7b. triplet regression would still occasionally make “illogical” adjustments to the fixed point location.4-1 s. If the majority of the NT triplets behaved this way. We also observed a few shorter encounters where the . when all the natural IBIs were greater than the fixed point. the fit would be very poor.214 M. This adjustment reduced variability substantially. then the fit would be poor (ill-conditioned).5-1s. the system briefly remained in a period-2 or period-3 orbit for approximately 6-12 IBIs.W. called the fixed point adjustment maximum (FAM). After adding regulating parameters such as the FAM and RNT. P. was optimized and normally set in the range of 0. At first. In all of the later experiments. This parameter. Another modification allowed the algorithm to remove outliers by including only natural triplets in the fit if the natural IBIs were within a certain radius (RNT) of the current fixed point location.

Adaptive tracking control of bursting.7.Chaos Control of Epileptiform Bursting in the Brain 215 a) b) Fig. . The fixed point estimate and control region (solid lines) tracks along well with the system. but a large amount of stimulated IBIs (●) is still needed. 7. shown in the dashed boxes near IBIs 375 and 920. Variance is relatively low. b) Two brief encounters with possible period-2 orbits in one experiment. a) An example of a tracking experiment.

we used the change in the center of mass (∆Xcm) of the cluster of points around the forced point and after an iterate If the system state were forced onto a true fixed point. extremely close to the fixed point) for a few IBIs. the subsequent iterates should in general stay closer to the fixed point than to the arbitrary point. then ∆Xcm should be small. 7. P. If the fixed point estimate were accurate.W. However. the new approach kept stimulating until the state point was on (or very close to) the fixed point in two dimensions. We therefore hypothesized that if we first forced the state point onto the fixed point and then forced it onto some other arbitrary point in the system attractor. The fact that higher-period orbits were attained more often than period-1 orbits suggests that the fixed point estimate was close but still not accurate enough. if the fixed point estimate were not accurate (or no fixed point existed at that time) then the state point could end up anywhere on the next iterate.) The stable manifold slope estimate was set to zero to minimize the amount of stimuli needed. so it merely had to be within 40 ms of the fixed point.216 M. In contrast. we developed a novel control protocol called state point forcing. These encounters provided tantalizing hints that control was indeed possible. at least for short periods of time. the control algorithm only applied a single stimulus that placed the state point on the stable manifold estimate and let the system drift in to the fixed point on its own.J. To quantify these differences. the state point could not always be forced exactly onto the fixed point. If there were a significant difference between the two cases. but recurrence to the same period-2 orbit later in the experiment was not usually seen probably due to nonstationarity. it would imply that the fixed point estimate was relatively accurate. (Due to limitations in stimulus-burst interval precision. while if it were forced onto an arbitrary point (or . These “close encounters” were seen in different experiments. Mogul system stayed in a period-1 orbit (i. Previously.e. Feasibility of Control and Fixed Point Detection by State-point Forcing To help ascertain exactly what were the obstacles to control.4. Cvitanovic and D.. Slutzky. then the state point should remain close to the fixed point on the next iterate.

8. Forcing alternated between the fixed point and arbitrary points 4 times (counted as 4 fixed point trials in analysis). tracking seemed to be working and had stabilized at one location after a while). When a significant fixed point was found.. until the fixed point seemed to be drifting due to nonstationarity. the state point was alternately forced onto the fixed point or the arbitrary point for 30-40 IBIs each.e. then ∆Xcm should be relatively big. and state-point forcing resumed.. State point forcing in a bursting hippocampal slice experiment. λs was set to 0. Tracking was then turned off and forcing turned back on. The forcing point switched between the fixed point and the arbitrary point several times. 7. then tracking was turned off. If a suitable fixed point estimate was found (i. Then tracking was turned on (at n~700) until a new fixed point was found at n=820. . Natural IBIs stayed closer to fixed points than to arbitrary points. An example of a state-point forcing experiment is shown in Fig. an inaccurate fixed point estimate). 7. At this point. state-point forcing was turned off and the fixed point was relocated using adaptive tracking.Chaos Control of Epileptiform Bursting in the Brain 217 Fig.8.

004. the difference in ∆Xcm was strongly significant (P<0. we sought to verify that our fixed point estimates were valid. We designed the state-point forcing protocol to ascertain whether control was at all feasible and to help identify the obstacles in detail. The results of the forcing experiments suggested that the fixed point estimates were indeed valid.W.J. Another notable result was that when the arbitrary forcing point was shorter than the fixed point.258 s vs.218 M.0001). only 23% of the fixed points detected online (using 10 surrogates) were found to be statistically significant offline. When the UPO transform was applied offline (using 50 surrogates to test significance) to state-point forcing data. whereas when the arbitrary forcing point was longer than the fixed point. Slutzky. Also. Mogul The value of ∆Xcm was computed for each fixed point and its corresponding arbitrary forcing point. The data were first compared for both UPO types combined. it is possible that much of the difficulty we had with control was due to inaccurate or false fixed point estimates. Moreover. arbitrary point. More detailed analysis revealed that forcing to fixed points found with adaptive tracking (AT) produced significant results whereas those found with the UPO transform (UPOT) did not. signed rank test). This suggested that perhaps the adaptive tracking algorithm provided better estimates of the fixed point than did the UPO transform algorithm. if the system state would not stay close to the fixed point for a few iterates even when placed directly onto the fixed point. but the AT fixed points did have significantly smaller ∆Xcm when forcing to the fixed vs. This could be the reason that the forcing protocol did not produce significant results for UPO-transformed fixed points. Specifically. Cvitanovic and D. Statistical comparisons of ∆Xcm were made with the paired t-test or the Wilcoxon signed rank test. the difference . P<0. then no control short of pacing (tonic stimulation) would have much chance of working. The fixed point trials were evenly split (51 each) between those detected with the UPO transform (UPOT) and those found with the adaptive tracking algorithm (AT). and ∆Xcm was significantly smaller when forcing to the fixed points than to the arbitrary points (median values 0. P.404 s. 0. The UPOT fixed points alone did not show a significant difference in ∆Xcm. since forcing onto the fixed points produced significantly less divergence in the subsequent iterate than did forcing onto the arbitrary points.

so we could not determine whether noise was the biggest impediment to control. Alternatively. It is possible (due to small sample size) that the four neighbors used for the Jacobian were often outside of this linear region. A high level of noise certainly would have hindered attempts to control the system – especially if the amplitude were on the order of the control radius – since the system state would often bounce out of the control region as soon as it was placed within it. Estimation of the stable manifold was itself a very difficult task. The next IBIs (after a stimulated burst) stayed relatively close to the fixed points.4 s. This may have been due to the high levels of stochastic dynamical noise that were also present in the system. With the addition of high-level noise. the method of fitting local neighbors of the fixed point using a linear least-squares algorithm was a simplification. fitting the stable manifold could become problematic because there are too few natural triplets that approach the fixed point. in a range from 0. It is far from certain that the manifolds were linear. although not as close as would be required to obtain tight control.8). Even with a low-noise system. We tried to avert this second possibility by shifting the arbitrary point by different amounts. when the arbitrary point was shifted down.Chaos Control of Epileptiform Bursting in the Brain 219 in ∆Xcm was not significant (P=0. Obstacles to Chaos Control The implications of these results for the feasibility of control were not as clear. This result could be explained in two ways. It would have also impaired the fixed point and stable manifold estimates. the neurons often burst spontaneously before stimuli could be applied. the stimuli would be occurring more quickly. there was no way of empirically determining the level of noise in the system. Unfortunately.1 s to 0. 7.5. and thus there were fewer points that could be used in the calculation of ∆Xcm. The first is artifactual: when the arbitrary point was shifted up. . and therefore the next spontaneous bursts could have been longer simply because the neurons were refractory. so any linear approximation would only be accurate within a very small radius of the fixed point. thus giving the neurons less time to recover.

The process of slicing the hippocampus itself severs many regulatory connections.W. it would have been extremely difficult to precisely control extracellular voltage levels using a point-source electrode. We used it because it greatly simplified the calculations and because it was the same dimension used as previously reported by Schiff et al. There were many potential sources for drift of IBI values. Slutzky. 48] have used only one spatial dimension. P. The use of IBIs may have contributed to noise and nonstationarity since they vary on a long time scale and thus many factors can change in between each measurement.[10]. Synaptic plasticity did not appear to have a significant effect. Also. Fluctuations in flow rate and temperature could easily have altered the system behavior by system metabolism. Finally. the phenomenon of nonstationarity was a large obstacle to control.6. both extrinsic and intrinsic. Adaptive tracking noticeably improved control over non-adaptive methods and seemed to counter nonstationarity. mainly made to stay consistent with prior studies [10. all of these studies [10. 7. We attempted to control for fatigue by ensuring that burst amplitudes remained constant throughout the experiment. it is possible that increasing the spatial dimensionality of the state variable may help control the system. 43. There also could have been intrinsic nonstationarity in the bursting itself. but intrinsic randomness may have prevented us from obtaining precise control.220 M. The use of IBIs as the state variable was somewhat arbitrary. All of these assumptions were necessary to make but also could have compromised control efficacy if they were inaccurate. However. Perhaps using a variable such as the raw extracellular voltage could have improved the results.J. 44. Mogul Several necessary assumptions were made that may have had an effect on the quality of control achieved. a two-dimensional embedding may have been too small to capture the entire system’s dynamics. but it is possible that the slice became “fatigued” or refractory. since burst amplitude is an indicator of neurotransmitter supply [49]. Cvitanovic and D. It is . Moreover. Conclusion Chaos control techniques showed modest success at controlling spontaneous epileptiform bursting. 44].

Brodie. Morris and W. (2000). Kwan and M. J.e. Nguyen and S. Achten. [3] E. Jerger and D. 342-347. Once the ability to manipulate the system is established. i. Zentner et al. 1731-1735. (2000). J. Improving quality of life with epilepsy surgery: The seizure outcome is the key to success. Vagus nerve stimulation therapy for epilepsy in older adults. H. L. Neurology 68(23). Behrens. 556-565.Chaos Control of Epileptiform Bursting in the Brain 221 possible that the intact brain could be less noisy or more stationary than the in vitro hippocampal slice. Caemaert. Controlling chaos in the brain. [2] G. 314-319. S. Nature 370(6491). [4] B. M. (1994). Long-term amygdalohippocampal stimulation for refractory temporal lobe epilepsy. Neurology 53(8).. N. 63-71. [6] F. Electrical stimulation of the centromedian thalamic nucleus in control of seizures: long-term studies. Lee and Y. L. [10] S. D. Annals of Neurology 52(5). and hence easier to control. should help reveal whether chaos control is a practical solution for preventing epileptic seizures. Epilepsia 48(2). (2007).. Velasco. Surgical and neurological complications in a series of 708 epilepsy surgery procedures. Neurology 54(5). if one could maneuver the system from a chaotic to a periodic state. Epilepsia 36(1). (1999). P. (2007). J. .. Velasco and A. [7] S. L. Naritoku et al. 590-600. Schiff. M. Cascino. (2001). References [1] P. 1967-1968. J Neurosci 21(2). M. Long-term treatment with vagus nerve stimulation in patients with refractory epilepsy. [9] J.. E.. 1179-1182. Gluckman. [8] K. However. B. Schramm and J. Neurosurgery 41(1). Sirven. Thus. Adaptive electric field control of epileptic seizures. Tsai. continued exploration. (1995). Nor is it known how tightly the spikes would have to be controlled to be successful. de Reuck and J. and extension of analysis algorithms into multiple spatiotemporal dimensions. Lim. It is still not known whether controlling interictal bursts or spikes (or anticontrolling. then it is likely that one could also convert it from a periodic to a chaotic one. 1-9. Weinstein et al. [5] G. Velasco et al. (2002). (1997). Mueller. Sperling and D. K. Duong et al. making them more disordered [50]) could prevent a seizure. Early identification of refractory epilepsy. N Engl J Med 342(5). J. Vonck. Electrical stimulation of the anterior nucleus of the thalamus for intractable epilepsy: A long-term follow-up study. 615-620. T. H. T. I. one could then determine whether a chaotic or periodic rhythm would be desirable to stop or prevent seizure activity.. J.

Vanneerven and A. Marciani. Chaos Solitons and Fractals 8(3). 38-48. 335-341. Abarbanel and H. Sauer. [26] S. J. T.. Longtin et al. (1994). (1994). F. (1991). (1995). 77-94. 327-334. J. (1988). Slutzky. Pueyo. S. Theiler. (1985). Chang. Mogul. Mogul. Yaylali. Mogul [11] J. Biophys J 67(2). (1995). (1996). Looking for chaos in brain slices. Hippocampal slice. Avoli. Theiler. D. P. Physica D 106(1-2). Nonlinear analysis of epileptic activity in rabbit neocortex. Detecting Strange Attractors in Turbulence. Physics Letters A 196(5-6). 671-683. 381-399. Cvitanovic and D. Physica D 58(1-4). [21] P. K. [15] I.222 M. H. Identification of determinism in noisy neuronal systems. Ann Neurol 17(6).. J Neurosci Methods 59(1). 684-691. Petracchi. Physica D 99(2-3). P. Schaffrath. 37-44. 521-547. On the Evidence for low-dimensional chaos in an epileptic electroencephalogram. Jayakar. (Springer-Verlag. [13] M. Cvitanovic and D. Schiff. Kocak and P. C. (1997). Lecture Notes in Mathematics 898. Barbarosie and M. [29] F Takens. P. J Neurophysiol 59(1). Cvitanovic and D. C. H. L. and S. Eubank and A. 259-276. Biological Cybernetics 78(1). Noest et al. . Iasemidis and J. Pockberger.. Interval embedding of chaotic signals. The study of chaotic dynamics by means of very short time series. [22] S. [30] T. 597-603. [20] D. Slutzky. Detection of seizures from small samples using nonlinear dynamic system theory. Testing for nonlinearity in time-series The method of surrogate data. and C. [19] J. Sauer. Biophys J 67(2). Gotman and M. CA3-driven hippocampal-entorhinal loop controls rather than sustains in vitro limbic seizures. [16] J. Aitken. G. S. Kaplan. W. Exceptional events as evidence for determinism.. Grassberger. 41-48. Physica D 73(1-2). Potassium-induced spontaneous electrographic seizures in the rat hippocampal slice. Characterizing nonlinearity in invasive EEG recordings from temporal lobe epilepsy. drug levels. IEEE Transactions on Biomedical Engineering 43(7).J. T. Deterministic chaos and noise in three in vitro hippocampal models of epilepsy. Monosynaptic spinal cord reflexes. (1997). Stochastic versus deterministic variability in simple neuronal circuits: I. Casdagli. Jerger and T. (2002). J. 153-161. [14] M. Schreiber. [18] J. [12] S.. J. Stochastic versus deterministic variability in simple neuronal circuits: II. Chang et al. International Journal of Bifurcations and Chaos 1. T. (1997). [23] P. Electroencephalographic spiking activity. (1994). (1991). Pijn. and seizure occurrence in epileptic patients. J. 607-618. [24] T. Traynelis and R. Ann Biomed Eng 29(7). (1995). 743-751. I. Schiff and T. [17] J. Chaos 5(1).W. J. New York. P. 371-381. Sackellares et al. Schiff. 57-65. Slutzky. Sarnthein. (2001). Electroencephalography and Clinical Neurophysiology 79(5). J Neurosci 17(23). [27] M. [28] M. (1992). 1981). Characterizing nonlinearity in invasive EEG recordings from temporal lobe epilepsy. (1996). Dingledine. W. The search for periodic unstable orbits in periodically driven spike trains. 127-132. [25] D. Sauer et al. (1998). Nonlinear time sequence analysis. G. 9308-9314. D. Chaos or noise in EEG signals Dependence on state and brain site. J Neurosci Methods 118(2).

720-726.The diode resonator. 183-186. Exploring chaotic motion through periodic orbits. M. P. Biophys J 74(6). Collins. Christini and J. Chaos 7(4). Kaplan. Christini and D. M. 1953-1955. Controlling nonchaotic neuronal noise using chaos control techniques. 489-499. J. (2000). Physical Review Letters 67(15). Francis and T. [33] J. 846-859. [38] E. 421-428. S. Spano and W. Continuous control of chaos by self-controlling feedback. 1196-1199. Tattersall and R. 73-111. Periodic orbits: a new language for neuronal dynamics. Phys Rev Lett 58(23). Adaptive estimation and control method for unstable periodic dynamics in spike trains. [37] E. 559-570. C. [42] K. Controlling low-dimensional chaos: Determination and stabilization of unstable periodic orbits by Kohonen neural nets. Collins. IEEE Trans Biomed Eng 50(5). (1991). [40] M. Biswal et al. Longacher and K. (1997). Control of chaos in excitable physiological systems: A geometric analysis. A. Chaos 16(2). W. J. Nature Medicine 4(10). [43] A. Physics Letters A 185(1). IEEE Transactions on Neural Networks 10(4). Cvitanovic and D. Auerbach. Ott. [46] D. C. Nature Neuroscience 2(8). Z. (1998). [32] D. Physical Review E 53(5).. Lin and C. Pyragas. 4334-4344. (2004). Herrmann and R. Mogul. Advances in Physics 44(2). Alsing. 1230-1235. (1992). A. P. (1990). Harrison. Singh and B. Hunt. Controlling chaos by GA-based reinforcement learning neural network. (2003). Adam and M. Ding. Eckmann et al. M. [50] K. Phys Rev Lett 75(14). J. Using neural networks for controlling chaos. [47] D. 1225-1231. Funke. Grebogi and J. (1998). Physics Letters A 170(6). In et al. Shinbrot. H. (1999). Neuroscience Letters 357(3).Chaos Control of Epileptiform Bursting in the Brain 223 [31] H.. (1997). 2776-2785.. [49] J. Physical Review E 49(2). Clement. [36] T. J. Netoff et al. Physical Review E 61(5). K. Martinerie. (1992). R. J. 544-549. [34] P. International Journal of Adaptive Control and Signal Processing 11(6). H. Adaptive targeting of chaotic response in periodically stimulated neural systems.. Cvitanovic and J. Epileptic seizures can be anticipated by non-linear analysis. (1987). [48] M. Slutzky. L. L. J. Controlling chaos in high dimensions: Theory and experiment. J. [41] C. M. Jou. Physical Review Letters 64(11). 77-87. Yang and V. M. Der. Gavrielides and V. (1995). Le Van Quyen et al. Bains. J. J. I. [35] P. (1996). (1994). Manipulating epileptiform bursting in the rat hippocampus using chaos control and adaptive techniques. E. Kantz. [44] D. J. Reciprocal interactions between CA3 network activity and strength of recurrent collateral synapses.. Ditto et al. Stabilizing high-period orbits in a chaotic system . T. Yorke. (1994). Science 257(5074). A robust method to estimate the maximal Lyapunov exponent of a timeseries. [39] P. Staley. T. 2387-2389. 1173-1176. Kovanis. (1999). P. P.. W. Progress in the control of chaos. 5149-5153. T. Garfinkel. (2006). P. . 2782-2785. Controlling chaos. Christini and J. So. A. [45] M. Controlling cardiac chaos. (1995). Periodic orbit analysis reveals subtle effects of atropine on epileptiform activity in the guinea-pig hippocampal slice. Gupta.

This page intentionally left blank .

Section C Chaos Control for Discrete-time Systems .

This page intentionally left blank .

Harb Department of Electrical Engineering.Chapter 8 CHAOS CONTROL FOR PHASE LOCK LOOP A. Jordan aharb@just. As a result. Introduction 8.edu. Yarmouk University. 8. The study showed the effectiveness of the designed nonlinear controller in controlling the undesirable unstable behavior and pulling the PLL back to the in-lock state.jo Previous study showed that a third-order phase locked loop (PLL) with sinusoidal phase detector characteristics experienced a Hopf bifurcation point as well as chaotic behavior. Jordan University of Science and Technology. The main goal of this chapter is to control this chaotic behavior. Its ability on tracking phase-varying signal makes it very attractive to be used in systems where 227 .1.1. They play a major role in electronic synchronization circuits. Irbid. this behavior drives the PLL to the out-of-lock (unstable) state. 2]. and capability of providing high power and frequency [1.A. controllability.edu.1. Irbid.M. Harb Department of Electrical Engineering. The analysis was based on a modern nonlinear theory such as bifurcation and chaos. Jordan bharb@yu. The Concept of Phased Lock Loop Phase locked loops (PLLs) are an important part of modern electronic communication and control systems due to its accuracy.jo B. A nonlinear controller based on the theory of backstepping is designed.

It is known that the PLL under consideration exhibits a chaotic behavior preceded by a series of period doubling for the periodic orbit associated with false lock state of the PLL. This behavior drives the PLL to the out-oflock state [12]. φo is an asymptotically stable solution of the autonomous. 8. This kind of response is undesirable for other communication systems. nonlinear differential equation describing the closed loop phase error. The idea of chaos synchronization was utilized to build communication systems to ensure the security of information been transmitted [11].M. This means that small perturbation from φo. in the phase error will eventually dampen out as a result of the closed loop dynamic. Also. The PLL can be false locked [4-7] where a periodic orbit exists for a certain loop parameter. In such systems. a time-invariant loop filter and a voltage controlled oscillator (VCO) [3]. and it is necessary to produce a second sinusoidal voltage having a frequency which is an integer multiple of the reference’s frequency. chaos prediction and controlling (when it is undesirable) are of a great importance for the designer. For example. Often. . one is supplied with a sinusoidal reference voltage. 5]. Josephson junctions [9] and Van der Pol oscillator [10] have been studied extensively in the literature. a phase detector. Equivalently. In both cases. The PLL is considered to be phase locked when the loop's phase error is constant φo and the loop is in stable equilibrium state [4. research on chaos in phase-locked loop has extended beyond analysis of chaotic behavior.A. it is often necessary to estimate the instantaneous phase of a received signal which has been contaminated by random noise/or other type of interference. Harb and B. PLLs offer a practical method of performing these tasks.228 A. Phase locked loops like many chaotic systems such as Chua’s circuit family [8]. Chaos in phase-locked loops has been investigated by many researchers for at least two decades. Recently. The PLL under consideration is considered as a chaos generator for such systems. when the PLL is used to demodulate an FM signals. the output of the VCO may become chaotic for certain loop parameters.1. they are used extensively in applications requiring the synthesis of highly stable sinusoidal signals. Harb it is necessary to estimate the phase of a received signal. The basic configuration of a PLL is shown in Fig. It consists of three building blocks.

A . al. [24-27] used the backstepping recursive nonlinear controller. limit cycles. Harb et. al. the PLL is unlocked after the Hopf bifurcation point. The objective can be stabilization and reduction of the amplitude of bifurcation orbital solutions.[17]. [23]. It was shown that this kind of PLL has a chaotic attractor. al. [21] and Nayfeh and Balachandran [15] designed state feedback nonlinear controllers. Calvo and Cartwright [28]. In other word. or open loop control. Ott [16] and Ott et. perturbation method. al. Many researchers are proposed other methods to control chaos.Chaos Control for Phase Lock Loop 229 Endo and Watada [13] determined the bifurcation sets of the Shilinkov-type homoclinic orbits for a third-order autonomous PLL with a symmetric periodic triangular phase detector having a second order loop filter. and Jackson [20] the methods are based on classical control. As the control parameter was increased. The analysis was based on modern nonlinear theory such as bifurcation and chaos theory. Hubler and Luscher [19]. and Zaher et. a controlling scheme is proposed to stabilize the chaotic behavior and drive the PLL under consideration to the steady state (in-lock-state). The control can thus be static or dynamic feedback control. [29] introduced the use of fuzzy theory control in chaotic systems. Ikhouane [22]. and Mann et. That means. the limit cycles deformed and at the end culminating to chaos. a small periodic solutions. Bifurcation control deals with using a control input to modify the characteristics of a parameterized nonlinear system. In Hubler [18]. reshaping of the bifurcation diagram or a combination of these. The point was found to be supercritical one. Harb and Harb [14] showed that a third order PLL with sinusoidal phase detector characteristics experienced chaos through Hopf bifurcation. Bifurcation control are used in [15]. has been used to find the normal form at the vicinity of the Hopf bifurcation point. In this chapter. great attention is been paid to chaos and bifurcation control. The method of multiple scales. Harb and Abdel-Jabbar [30] introduced a model-based control strategy based on global state feedback linearization (GLC) to control the chaotic behavior in a power system. optimization of a performance index near bifurcation. these controllers are designed to control the bifurcation route that leads to chaos. al. Abed et. are borned at the Hopf bifurcation point. Recently.

0 t (8. F ( s) = . Equations (8. and eo(t) is the filter's zero-input response which depends only on the initial conditions existing in the filter at t = 0. (8.4) where F1(s) and F2(s) are two cascade lag-lead filters.3) can be combined to obtain dθ v = ωo + k v ∫ x(u ) f (t − u ) du = ωo + k1k v ∫ f (t − u )sin (θ i (u ) − θ v (u ) )du . Then. a timeinvariant loop filter and a voltage controlled oscillator (VCO). dt 0 0 (8. Mathematical Model After Harb and Harb [14]. It consists of three building blocks. Harb nonlinear controller based on backstepping recursive theory is designed to control the chaotic oscillation in a third order PLL. eo(t) → 0 as t → ∞ for any set of initial conditions. t ≥ 0 . the basic configuration of a PLL is shown in Fig. (8.3) where f(t) is the impulse response of the filter.1) where ωo is called the VCO quiescent or center frequency and kv is the VCO gain which has units of (rad/sec-volt).230 A.1).5) t t . If the filter is stable. Harb and B. 8. The loop filter which will be considered in this chapter has the form 1+τ z1 s 1+τ z 2 s = F1(s) .2).1. The VCO's instantaneous frequency is given by dθ v = ω o + k v e(t ) rad/sec. 8. the output of the loop filter becomes e(t ) = eo (t ) + ∫ x(t − u ) f (u ) du . and (8.2) The term sin(θi+θv) contains higher fundamental frequency components which are eliminated by the loop filter. F2(s). a phase detector.A. The output of the phase detector is given by x(t ) = k1 [sin(θ i − θ v ) + sin(θ i + θ v )] . dt (8. 1+τ p1 s 1+τ p 2 s (8.M.2.

and τz2 are the loop filters time constants. dt dt 0 (8. and (8. After simplifications. (8.8) θ 2 = θ v − ωot . The results which follow are simplified by defining (8. = 2 + +   + 2 τ p1τ p 2  dt  τ p1τ p 2 dt τ p1τ p 2 dt τ p1τ p 2 dt (8.6) k = k1 k v . the above equation becomes d 3φ d 2φ  τ p1 + τ p 2 kτ z1τ z 2 cos φ  dφ  1 k (τ + τ )cos φ  +   + 2 + + z1 z 2 3  τ τ  dt  τ τ  dt dt  p1 p 2 τ p1τ p 2  τ p1τ p 2  p1 p 2  − 2 kτ z1τ z 2 sin φ  dφ  k sin φ d 2θ1 τ p1 + τ p 2 d 2θ1 1 dθ1 . (8.5).Chaos Control for Phase Lock Loop 231 Define the closed loop phase error as φ ≡ θi −θv .12) If the input frequency is constant. and the closed loop gain as (8. τz1. t (8.7) θ1 = θ i − ω o t . = − k d 1+τ d dt dt 1+τ p1 p2 dt dt [ ] (8.9) These quantities can be used with equations (8. dt sin(φ ) .7) to write dφ dθ1 = − Ak ∫ f (t − u ) sin (φ (u ) )du for t ≥ 0 .11) where τp1 . τp2 .13) .6) and (8. then θ1(t) = (ωi-ωo) t + θo.10) The differential equation that describes the closed loop phase error in the PLL is given by 1+τ z1 d 1+τ z 2 d dφ dθ dt .

e = τz1 τz2 s2 . (8. we show the bifurcation diagram. There is only one Hopf bifurcation point H at Ko = 7299. δ = ωos / k and ωos = ωi . as K was increased above .15a) (8. b = τz1τz2 s2. Equilibrium Solution The equilibrium solution of the system of equations (8.15b) y = − ay − b cos( φ ) y − cx − d cos( φ ) x + e sin( φ ) x 2 − sin( φ ) +δ .15a)-(8.15a)-(8. and (8. which is the variation of the control parameter K with the state variable φ . while the dashed line denotes unstable foci. =d/dt' . Harb and by normalizing the time variable using t' = (k/τp1τp2)1/3 t.15c) evaluated at the equilibrium point. such as BIFOR2 and AUTO94. Endo and Seishi [13] and consider the case of the sinusoidal phase detector. Harb and B. The solid line denotes stable nodes. In Fig. To this end.15a)-(8.M. Using the normal form near the Hopf bifurcation point H.2. The solution of this equation as a function of one of the control parameters are defined by using a continuation scheme. we end up with a nonlinear algebraic equation. we find that H is a supercritical point. d = (τz1 + τz2 ) s . we follow Watada.15c) corresponding to φ = φ = φ = 0 . The above equation can be written as: ( ) φ = x. c = s/k . The stability of an equilibrium solution depends on the eigenvalues of the Jacobian matrix of equations (8.01. a = τ p1 +τ p 2 s 2 / k .14) where . equation (8.3. So.ωo.A.232 A. In this chapter we write our own program for calculating the fixed points and their bifurcations rather than use an available bifurcation software package. 8. Setting the right hand sides of equations (8. (8.15c) 8.12) becomes φ + aφ + b cos(φ ) φ + cφ + d cos(φ ) φ − e sin(φ ) φ 2 + sin(φ ) = δ . x = y. s = (k/τp1τp2)1/3 .15c) equal to zero.

and 2 (8.18) Design Methods. a sequence of deformed (asymmetric) periodic solutions was observed. Let e1 = x1 + sin −1 (δ ) . Thus. 8. and (8.19a) (8. (8. let x1 = φ .17) x3 = − Ax3 − Bx3 cos( x1 ) − Cx 2 − Dx 2 cos( x1 ) + Ex 2 sin( x1 ) − sin( x1 ) + δ + u . a designed control signal u is added to equation (8.20b) e3 = − A(e2 + c1e1 ) − B(e3 + c 2 e1 + c3 e2 ) cos(e1 − sin −1 (δ )) − C (e2 + c1e1 ) − D(e 2 + c1e1 ) cos(e1 − sin −1 (δ )) + E (e2 + c1e1 ) 2 sin(e1 − sin −1 (δ )) − sin( e1 − sin −1 (δ )) + δ + u − c 2 (e2 + c1e1 ) − c3 (e3 + c 2 e1 + c3 e2 − c1(e 2 + c1 e1 )) .Chaos Control for Phase Lock Loop 233 7299.19b) e3 = x3 − x3des = x3 − c 2 e1 − c3 e2 . x 2 = x3 .4. (8. e2 = x2 − x 2des = x 2 − c1e1 . and (8. Backstepping Recursive Nonlinear Controller Appendix 8A gives more details about the methodology of backstepping control theory. 8. e2 = e3 + c 2 e1 + c3 e2 − c1 (e2 + c1e1 ) . leading to chaos at the control parameter K= 85299 as shown in Fig. First. Second.15c). Therefore.16) (8.01.3. one can rewrite the system as follows x1 = x 2 .20a) (8.20c) . (8.19c) e1 = c1e1 + e2 . x 2 = x. and x3 = y .

choose ci and ki such that the first and second terms in V are equal to zero.4a and 8. 8. the PLL is driven to the out-of-lock state and it will not achieve phase lock (stable equilibrium state). In order to see the effectiveness of the designed controller. So.22) Substituting equations (8. i=1.4b). the main objective of this chapter is control and to get red of the chaotic behavior.21) V = ∑ k i ei ei =k1e1e1 + k 2 e2 e2 + k 3 e3 e3 .5b) show the error signals e2 and e3. 8.2. 8.3 (Figs. This means that the phase locked-loop is (in-lock-state).3b) to the controlled system shown in Fig. on can see that the system is recovered from its chaotic behavior and exhibits a stable equilibrium solution.5a and 8. and integrate the new system. But the system without any control action is experiencing chaotic behavior via Hopf bifurcation. we apply the designed control signal after 50 second of the integration time. 8.15a. (8. 2 i =1 (8. for the value of control parameter Ko = 85299. 2.For c1 = c 2 = 0 = c3 = 0. .16)-(8.6 (Figs. 8. by comparing the uncontrolled system shown in Fig.23) to make sure that the system is stable. The results were achieved by using one control signal.23) into the original system of equations (8.4 (Figs. Substituting the control signal equation (8. 8.234 A. k i > 0 . It so clear that.A. 8. and 8. 8. and k 3 = 1 .3 shows the uncontrolled system. 8.15b. and 1. As a result. Harb Let V = then 1 3 ∑ k i ei2 .21) into equation (8.3. 8. Fig. choose u such that: u = −e3 + Be3 cos(e1 − sin −1 (δ )) + Ce2 + De2 cos(e1 − sin −1 (δ )) 2 − Ee2 sin(e1 − sin −1 (δ )) + sin(e1 − sin −1 (δ )) − δ .4.6a and 8. As we mentioned. Harb and B. one obtain the simulation results as shown in Fig.3a and 8. 8. i =1 3 (8. k1 = 0 = k 2 = 0.22). Fig. Fig. the chaotic oscillations appear via period doubling rout to chaos.M.15c).5 (Figs.6b) show how the controller brings the system to the equilibrium solution or in-lock-state.

(8. 8. i. This behavior drives the PLL to the out-of-lock (unstable) state. Appendix 8A―Methodology of Backstepping To explain the methodology of the backstepping controller. The study showed the effectiveness of the designed nonlinear controller in controlling the undesirable unstable (chaotic) behavior and pulling the PLL back to the in-lock state.4.A2) where η and ξ are the state variables and u is the control signal. Conclusions In the existing literature. it was showed that a third-order phase locked loop (PLL) with sinusoidal phase detector characteristics experienced a Hopf bifurcation point as well as chaotic behavior. is asymptotically stable.Chaos Control for Phase Lock Loop 235 8. (8. and (8. it drives the chaotic oscillation into an equilibrium constant behavior (phase lock state). one obtain the equivalent representation shown in Fig. a nonlinear controller based on the theory of backstepping has been designed to control this chaotic behavior.7a. If the component of equation (8. ∂η (8. Adding and subtracting g(η)φ(η) on the right hand side of equation (8. as shown in Fig. It is the integrator backstepping. 8.A4) . with φ (0) = 0 . Let η = f (η ) + g (η )ξ .7c and mathematically as follows: η = [ f (η ) + g (η )φ (η )] + g (η )[ξ − φ (η )] . then.A1) ξ =u.7b and Fig. In this chapter. Suppose that we know a smooth and positive definite Lyapunov function V(η) that satisfies the inequality ∂V [ f (η ) + g (η )φ (η )] ≤ −W (η ) . let me start from a special case that has been taken from Khalil’s book [31]. 8. the origin of η = f (η ) + g (η )φ (η ) .A1).e.A3) where W(η) is positive definite.A1) can be stabilized by a smooth state feedback control law ξ = φ (η ) .

z=0) is asymptotically stable. ∂η ∂η (8.A2).A5) are similar to equations (8. Using the following Lyapunov function as Vc (η .236 A.M. (8. Substituting for v.A5) η = [ f (η ) + g (η )φ (η )] + g (η ) z . and V (η ) is radial unbounded.A. This feature will be exploited in the design of v to stabilize the overall system. one obtain the state feedback control law u= ∂φ ∂V [ f (η ) + g (η )ξ ] − g (η ) − k[ξ − φ (η )] .A8) ∂V ∂V ∂V [ f (η ) + g (η )φ (η )] + g (η ) z + zv ≤ −W (η ) + g (η ) z + zv .A6) z =v. and (8. so. or (8. z. 2 then Vc = (8. ∂η where k>0. Since φ (0) = 0. except that now the first component has an asymptotically stable origin when the input is zero. the origin (η=0. Harb and B. which shows that the origin (η=0. Harb and ξ =u.A4) and (8.A9) Choosing v=− ∂V g (η ) − kz . one obtain. ∂η ∂η ∂η (8. Equations (8.A7) where z = ξ − φ (η ) and v = u − φ . ξ ) = V (η ) + 1 z 2 . Vc ≤ −W (η ) − kz 2 .A1) and (8. ξ=0) is also asymptotically stable. we can conclude that the origin is globally asymptotically stable. . and φ .A10) If all the assumptions hold globally.

. 8. Bifurcation diagram (variation of state variable versus the control parameter K).2. Block diagram of a phase-locked loop with second-order loop filter.01 6000 8000 k 10000 12000 14000 H Fig. 35 30 25 20 x 15 10 5 0 2000 4000 ko=7299.1. 8.Chaos Control for Phase Lock Loop 237 2 A sin(ωο t ) Ak1 sin φ F1(s) F2(s) 2 K v cosθ v (t ) VCO Fig.

A. 5 -1 -0 . a) The chaotic time response and b) state-plane of the state variables X2 and X3 at Ko = 85299. Harb 1 0.M. 5 -1 -1 .5 b) Fig.238 1. 5 0 X 2 0. 8. 5 -1 -1 . Harb and B.5 X 0 3 -0 . Uncontrolled simulations.5 2 2.3. 5 -2 -1 .5 1200 1400 1600 1800 2000 a) 1 0. .5 X 0 3 -0 .5 A. 5 0 200 400 600 800 1000 Tim e 1. 5 -2 .5 1 1.

15 b) Fig.03 0. 1 -0 . 0 5 X 2 0 0. 2 -0 . . 0 3 a) -0 .03 239 0.02 0. 0 1 -0 . 1 5 -0 . 0 2 -0 .1 0. Controlled simulations: a) Time history and b) state-plane of the state variables X2 and X3 .01 0 X 3 -0 . 8. 0 1 -0 . 0 3 -0 .05 0. 0 2 -0 . 0 4 -0 .4.02 0.01 0 X 3 -0 . 0 4 300 400 500 600 Tim e 700 800 900 1000 0. at Ko = 85299.Chaos Control for Phase Lock Loop 0.

0 5 -0 .15 A. 0 3 -0 .M.A. 1 -0 . . 0 2 -0 . 2 3 -0 . 8. 0 4 -0 . 2 300 0. Controlled simulations: a) Time history and b) state-plane of the error signals e2 and e3.02 0.1 0.240 0.1 0.05 E rror S ignals 0 -0 .5.05 0. 1 5 a) -0 .01 0 e -0 . 0 5 e 2 0 0. at Ko = 85299. 1 5 -0 . Harb and B.15 b) Fig. Harb 0. 0 1 -0 .03 400 500 600 Tim e 700 800 900 1000 0. 1 -0 .

8.5 1 X 0.5 241 2 1.5 2 0 -0 . 5 -1 -1 .Chaos Control for Phase Lock Loop 2.5 1 and X X 3 0.6. 5 -1 -1 .5 2 1.5 2 0 -0 . Applying the control signal after 50 seconds. 5 -2 0 2 4 X 1 6 8 10 b) Fig. a) the time history and b) the state-plane of state variables X2 and X3 . at Ko = 85299. . 5 0 10 20 30 40 50 T im e 60 70 80 90 100 a) 2.

242 A. .)φ(. J.) a) u ∫ b) ξ + g(η) + ∫ ƒ(. c) backstepping -φ(η) through the integrator. [2] A. M. References [1] A. Viterbi.M. Blanchard.7. [3] F. Phase-Locked Loops (New York: John Wiley 1976).) -φ` Fig.)+g(. b) Introducing φ(η).) η -φ(η) u Ζ η g(η) + + ∫ c) ∫ ƒ(. 8.A. 1966). Harb and B. Harb u ∫ ξ g(η) + η ∫ ƒ(. Principle of Coherent Communications (New York: McGraw Hill. a) The block diagram of the system. Phase Lock Techniques (New York: John Wiley 1979). Gardner.

Phys. [8] S. Campbell. Proc. Chaos. Harb and J. 153172. (1998). Endo and L. Jackson. Chua. Acta 62. [15] A. [11] T. Zohdy. [17] E. Harb. 974-980. 431437. Fu. Harb. Stensby and B.Chaos Control for Phase Lock Loop 243 [4] J. Endo and H. I. I 41. USA. ed. O. 1994). Salam and S. Solitons and Fractals 19(3).locked loops. Zaher and M. Ikhouane and M. of ACC Alaska. IEEE Transactions on Circuits and Systems 38. Controlling chaotic dynamical systems. Kennedy and L. O. Piccardi. on Aut. 191-199. (1994). Naturwissenschaften 76. (2000). Seishi. Nayfeh and B. L. 48394853. Sastry. 1993. Hubler. (1996). Hubler and E. New York. H. . Circuits Syst. IEE Electronics Letters 1995. [20] E. [21] E. IEEE Transactions on Circuits and Systems 35. 1580-1588. Nonlinear recursive chaos control. Robustness of the tuning functions adaptive backstepping design for linear systems. Applied Nonlinear Dynamics (John Wiley. T. (2004). (1998). A 44. Harb. Phys. Chaos in Dynamical Systems (Cambridge University Press. A. A. [10] M. (1991). (1989). Ott. Yorke. [7] J.locked loops. Chua. Rev. 1022-1032. Helv. Proceedings of the IEEE 75. [16] E. [13] K. Journal of Franklin Institute 330(5). [14] B. Chua. H. (1988). IEEE Transactions on Circuits and Systems–I: Fundamental Theory and Applications 45. 987-1003. H. Dynamics of the forced Josephson junction circuit: The region of chaos. (1986). L. Harb and A. The half-plane pull-in range of a second-order phase locked loop. Shilnikov orbits in an autonomous third-order chaotic phase-locked loop. K. Systems and control letters 7. 979-983. Cambridge. of ACC 1473-1478. D. Control of dynamic flows with attractors. Stensby. CHAOS (Soviet-American Perspectives on Nonlinear Science. [6] C. 1992a). 1993). Endo and L. Wu. Zaher and M. 315-320. Chaos from phase. Luscher. Abed and J. Chua’s circuit family. IEEE Transactions on Circuit and Systems 32. 775-786. An adaptive control of chaotic systems. (2002). (1989). [18] A. Local feedback stabilization and bifurcation control. (1981). Hopf bifurcation. 11-17. Ott. (1987). Synchronization of chaos in phase. [12] T. (1991). England. [24] A.. Robust control of biped robots. [19] A. 784-795. Proc. Watada. USA. Resonant stimulation and control of nonlinear oscillators. Balachandran. O. [9] F. IEEE Transactions on Circuits and Systems 33. L. (1980). [5] B. Stensby. [23] A. C. Control 43(3). Computing the half-plane pull-in range in a second-order phase locked loop. IEEE Trans. 67-69. Saddle node bifurcation at a nonhyperbolic limit cycle in a phase locked loop. 343-346. Chicago IL. IEEE Trans. Van der Pol and chaos. [22] F. Zohdy. Chaos and bifurcation in third-order phase locked loop. Krstic. Journal of Franklin Institute 333B(2). P. A. 667-672. Grebogi and J. Bifurcations of limit cycles in periodically forced nonlinear systems: The harmonic balance approach.

1055-1063. . (2001). IEEE Trans. Zaher. M. Abedl-Jabbar. Zohdy. of ACC 4514-4519. Areed and K. G. Harb and B. F. M. USA. Calvo and J.244 A. M. Controlling Hopf bifurcation and chaos in a small power system. (2001). Arlington VA. Khalil. (1998). Hu and R. Nonlinear Systems (Prentice Hall. B. Zaher. [31] H.Soliman. Soliman. (1999). Harb [25] A. Analysis of direct action fuzzy PID controller structures. Conf. Areed and K. [27] A. I. Areed and K. K. Zohdy. [30] A. Robust estimation-based design for uncertain plants. (2003). System and Cybernetics 29(3). Zohdy. AMSE conference Dearborn. Real-time model-reference control of non-linear processes. Cartwright.A. Fuzzy control of chaos. Solitons and Fractals 18(5).M. 2nd Edition. Robust model-reference control for a class of non-linear and piece-wise linear systems. Harb and N. Proc. F. Soliman. (2000). International Journal of Bifurcation and Chaos 8(8). [28] O. [26] A. G. F. Mann. MI. 1996). 2nd Int. USA. on Computers in Industry Manama Bahrain. Zaher. Gosine. [29] G. Chaos.

a fuzzy impulsive control strategy is applied to high order interpolative lowpass sigma delta modulators.ho@qmul.reiss@elec. Reiss Department of Electronic Engineering. United Kingdom c.ac. Mile End Road. the local stability criterion and the condition for the occurrence of limit cycle behaviors are derived. Ling Department of Electronic Engineering. the initial condition and the filter parameters are.Y. first.uk C.Chapter 9 CONTROL OF SIGMA DELTA MODULATORS VIA FUZZY IMPULSIVE APPROACH B. King’s College London.uk J. Ho School of Mathematical Sciences. Division of Engineering.ling@kcl.ac. WC2R 2LS. E1 4NS. The state vectors that the impulsive controller resets to are determined so that the state vectors of interpolative lowpass sigma delta modulators (SDMs) are bounded within any arbitrary nonempty region no matter what the input step size. Second. To work on this problem.K.ac. London. E1 4NS. the occurrence of limit cycle behaviors and the effect of audio clicks are minimized. Mile End Road.qmul.W. Queen Mary College London. United Kingdom wing-kuen. London. United Kingdom josh. as well as a practical consideration based on the boundedness of the state variables and a 245 . Queen Mary College London.uk In this chapter. based on the derived conditions.F. as well as the state vectors are close to the invariant set if it exists.D.

employ SDMs in their A/D and D/A conversions.Y. Ho and J. respectively.246 B. the signals are sampled with the frequencies greater than twice of their bandwidths. . such as in the consumer and professional audio processing systems [2]. It is well known that when the input signals are critically sampled. 9.W. C. By applying simple lowpass filtering.1. the initial condition and the filter parameters. noises outside the signal bands are filtered and good signal-to-noise ratios (SNR) could be achieved. The working principles of sigma delta modulation are based on the oversampling and the noise shaping techniques. signals and noises are separated and better SNR could be further achieved. By a proper design of a loop filter. Many systems.K. Hence. communication systems [3]. that is. the magnitudes of the signal transfer function in the signal band and in the noise band are approximately equal to one and zero. the implementation cost of sigma delta modulators (SDMs) could be very low [2] and nowadays SDMs are very popular in analog-to-digital (A/D) and digital-to-analog (D/A) conversions [1]. and that of the noise transfer function in the signal band and in the noise band are approximately equal to zero and one. then the signal bands are decreased and the bandwidths of the signals are inversely proportional to the oversampling ratios. etc.F. The controlled state vectors are then determined by solving the fuzzy impulsive control law. Reiss heuristic measure on the strength of audio clicks. respectively. Introduction Because of the advance in electronic technology. Ling. Noise shaping technique is further to shape the noise in the signal band to the noise band by applying a negative feedback. and precision measurement devices [4]. that is.D. then the signals can be perfectly reconstructed and the signals are spread over the whole frequency band. fuzzy membership functions and a fuzzy impulsive control law are formulated. and that over the existing nonlinear control strategy are the efficiency and the effectiveness in terms of lower frequency of applying the control force and higher signal-to-noise ratio (SNR) performance. the signals are sampled with the frequencies equal to twice of their bandwidths. One of the advantages of the fuzzy impulsive control strategy over the existing linear control strategies is the robustness to the input signal. When the input signals are oversampled.

Besides. Although there are many existing linear control strategies for stabilizing interpolative SDMs. so it is not guaranteed that a particular class of controllers can stabilize all types of interpolative SDMs. Hence. there may be a large jump between the unclipped and clipped state levels. For audio applications [2]. SDMs are suffered from instability problems. these linear control strategies stabilize the loop filter by changing the effective poles of the loop filter. such as variable structure compensation (sliding mode control strategy) [5] and time delay feedback control strategy [6]. as the set of the state vectors under the clipping control strategy is usually not the same as the invariant set. This situation is found very frequently when the input signal is very slow time varying or the clipped level is set at very low value. etc. In order to solve these problems. in which it is to reset the state vectors to different positions in the state space whenever the control force is applied. which should also be avoided. was employed [2]. audio clicks may be observed. as the state variables are always reset to the same values. Moreover. control is required. it is not guaranteed that the SNR of the controlled SDMs is still maintained or even improved if the effective poles of the loop filter are changed. nonlinear control strategy. Consequently. it cannot be guaranteed that the controlled SDMs are stable for all initial conditions in the state space. This is because the magnitude of the loop filter response at the signal band has to be tended to an infinity in order to achieve the magnitude of the signal transfer function equal to one and that of the noise transfer function equal to zero. In order to control the SDMs without changing the effective poles of the loop filter. an impulsive control strategy is discussed in this chapter. periodic output sequences may result and this periodic behavior is known as limit cycle behavior. such as the clipping control strategy. In addition. which should be avoided. Furthermore. the controlled SDMs may still be unstable when the magnitude of the input signal is increased. the clipping force may be applied very frequently. the occurrence of limit cycle behaviors results to the annoying audio tones. This implies that the loop filter is unstable or marginally stable. Since the loop filter is usually designed to have a very noise shaping characteristics. As a result.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 247 However. the occurrence of limit cycle behaviors and the effect of . the parameters in the controller depend on the loop filter parameters. For the clipping control strategy. in order to achieve good noise shaping characteristics. Furthermore.

6. In Section 9. The input to the SDM and the output of the loop filter are denoted as.4. Finally. which are used for the formulation of fuzzy membership functions and fuzzy impulsive control law. Reiss audio clicks can be minimized with the guarantee of the bounded state variables. u (k ) and y (k ) .1. Moreover.2. The outline of this chapter is as follows. We make those assumptions because this type of SDMs is commonly used in the industry [2]. respectively. cause a difficulty to solve the state vectors analytically. 9. We assume that the loop filter is a single input single output real system and the input is also real. if the invariant set exists. then we only need to reset the state variables of the loop filter once and the state vectors of the SDMs are guaranteed to be within the invariant set forever if the effects of limit cycle behaviors and audio clicks do not consider.3. To solve this problem. Ling.K.2. such as fractal and chaotic behaviors. Ho and J.5. a fuzzy approach is employed because employing fuzzy approach can simplify the complicated problems and can capture heuristic knowledge in the system in an easy manner. C. nonlinear behaviors. Denote the coefficients in the denominator and numerator of . some simulation results are presented to illustrate the effectiveness of the fuzzy impulsive control strategy. In Section 9. combined with the practical consideration on the boundedness of the state variables and a heuristic measure on the strength of audio clicks. Notations The block diagram of an interpolative SDM is shown in Fig. 9. Since the SDMs consist of a quantizer. The transfer function of the loop filter is denoted as F ( z ) .W. u (k ) ∈ ℜ . F ( z ) is assumed to be causal. a fuzzy impulsive control strategy is discussed. rational and proper with the order of the polynomial of z −1 in the numerator being equal to that in the denominator and there is a delay in the numerator.F. that is. a conclusion is summarized in Section 9.248 B. the conditions for the occurrence of limit cycle behaviors and the local stability criterion of the SDMs are derived.D. so y(k ) ∈ ℜ .Y. In Section 9. As there are usually an infinite number of state vectors in the invariant set. we introduce the notations which appear throughout this chapter. this chapter suggests a method to determine the state vectors that the impulsive controller resets to. In Section 9.

2) x(k +1) = Ax(k ) + B(u(k ) − s(k )) for k ≥ 0 . −i (9. they can be converted to the direct form realization using simple transformations.1.4) T is the state vector of the SDMs. Since this chapter is based on the feedforward structure of the SDMs. N .1. without loss of generality. . N and b j for j = 1.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 249 F ( z ) as. where ∑b z j −j F (z ) = ∑a z i i =0 j =1 N .  0 b1  a0   (9. (9. 0  0     0  and B ≡    1  0 a1   bN −  a0 a0    . x N (k )] ≡ [ y (k − N ). Then N . ai for i = 0. u (k ) ≡ [u (k − N ).1) u (k ) + _ F (z ) y (k ) Q Fig. s(k ) ≡ [s1 (k ). N is the order of the loop filter. respectively. Q( y (k − 1))] . u (k − 1)] . y (k − 1)] T (9.5) 0   . we assume that the loop filter is realized via the direct form because the expressions will be much simplified. For other minimal realizations. s N (k )] ≡ [Q( y (k − N )).6) 0 . The block diagram of an interpolative SDM.  0   A≡   0 − a N  a0  1 0 T T . .3) (9. the SDMs can be described by the following state space equation: (9. . T . where x(k ) ≡ [x1 (k ). 9. Hence.

limit cycle behaviors should be avoided.2. respectively. Ling. nd .Y. N and k ≥ 0 . Let nd be the number of eigenvalues of matrix A on the unit circle with their phases are integer multiples of k i ∈ Z and i = 1. Hence.3. (9. N }. we further assume that u(k ) = u for k ≥ 0 . i 2π . the input can be approximated as a step signal. xi (k ) < Vcc for i = 1.2. N be the i row of j 2πk ∑A j =0 P −1 P −1− j B(u(k0 + j ) − s(k0 + j )) . C.W. Conditions for Occurrence of Limit Cycle Behaviors and Local Stability Criterion As discussed in Section 9. . Suppose the eigen decomposition of matrix A exists. such that A = TDT−1 . For the direct form realization.250 B.7) Q( y) ≡   −1 otherwise Since the oversampling ratio of the SDM is usually very high. Denote λi and ξ i for i = 1.2.F. This is essential for formulating a fuzzy membership function for avoiding the occurrence of limit cycle behavior. that is. we denote the desired bound on the state variables as Vcc . Ho and J. (9. 9. . since all the state variables are the delay versions of the output of the loop filter.K. That is. N be the eigenvalues and the corresponding eigenvectors of the matrix A . Denote Bo as the set of the desired state vectors. the magnitude of the state variables of the SDM should not be larger than certain values. λi + N − nd = e P for P th .2.2.8) .D. We make this assumption because it is satisfied for most of SDMs employed in the industry [2].1. In many practical situations. Otherwise. That is. Reiss in which Q is a one bit quantizer defined as follows: y≥0 1 . . the conditions for exhibiting limit cycle behavior are discussed below. . the SDM is guaranteed to yield an unwanted behavior. there exists a full rank matrix T and a diagonal matrix D which consist of the eigenvectors and eigenvalues of matrix A . before we suggest the fuzzy impulsive control strategy. Hence. Denote Li for i = 1. That is. Bo ≡ {x : xi < Vcc for i = 1.

nd .n i r = rN − nd + n for . (9.11) Since T is a full rank matrix.1. 0 ] (9. ] rank (Q ) ≤ N − nd .Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 251 where P ∈ Z + and k0 ≥ 0 . Lemma 9. nd such that N − nd i =1 N − nd i =1 ∑c i. ([( ) . P N − nd N − nd 1 N − nd 1− λ ≠ 0 i = 1. rank (Q ) ≥ N − nd . nd . nd .n i r x(k 0 ) = LN − nd + n for n = 1.2. . . Proof. rank (Q ) = N − nd . Denote Q ≡ I − A P . nd .n Li = LN − nd + n for n = 1. N be the j row of I − A P . and ∑c i. If ΨP = Ø or ∃n ∈ {1.2. where Ø denotes the empty set.9) ΨP ≡ { x(0 ) : ri x(k0 ) = Li for i = 1.2. ∃x(0) ∈ ℜ N such that ri x(k0 ) = Li for i = 1. .10) and (1 − λ )ξ .2.n Li ≠ LN − nd + n . As a result. (1 − λ P N − nd )ξ N − nd . N − nd . then the SDMs exhibit limit cycle behavior with period P . As N − nd i =1 . rank (QT) = N − nd .n Li = LN − nd + n for n = 1. . Since ∑c i. The number of linearly independent rows in the matrix I − A P is N − nd . that is. Hence.2. nd } such that N − nd i =1 ∑c i. . Let r j for j = 1. 0. Since P i for . Hence. 0 T −1 . If ΨP ≠ Ø. [ . . ∑c i. nd . N − nd . 0]) . . th . Since ΨP ≠ Ø. . . N − nd }. 0.2.2.2. N − nd and n = 1. However.n i r = rN − nd + n . (9. N − nd i =1 . 0. Since A = TDT−1 and λi + N − nd = e ki ∈ Z and i = 1. where I is an N × N identity matrix.2.2. ξ } are linearly independent. (1 − λ )ξ P N − nd N − nd . As P rank (QT) = rank 1 − λ1 ξ1 .2. . . {ξ . rank (QT) ≤ rank (Q ) . then there will not exist any fixed point or periodic j 2πki P state sequence. N − nd i =1 ∑c i. Denote (9. . the number of linearly independent rows in the matrix I − A P is N − nd . ∃ci . for [( ) . and ΨP is the corresponding nonempty set of initial condition. n ∈ ℜ for i = 1. n = 1.12) P Q = (1 − λ1 )ξ1 .2. we have: P QT = 1 − λ1 ξ1 .

then the P matrix Q will also drop rank. but if there exist some poles on the 2π unit circle with their phases are nonzero integer multiple of . ΨP is the corresponding nonempty set of initial condition. Although the global stability of the SDMs is usually preferred because global stability implies local stability. Ho and J. Hence. sometimes global stability cannot be achieved.2. we reveal that even though there is not DC pole. Hence. This set of initial condition will be used for the formulation of fuzzy rules shown in Section 9. P for n = 1.Y.n Li ≠ LN − nd + n . C.2. . This result is a generalization of [2]. it mainly considers the DC pole cases. This implies that (I − A )x(k ) = ∑ A 0 P −1− j B(u(k 0 + j ) − s(k0 + j )) . Define the forward dynamics and one of the possible backward dynamics of the system as ℵf : ℜ N → ℜ N and ℵb : ℜ N → ℜ N . nd . In [2]. 2. there will not exist any fixed point or periodic state sequence. Obviously. When ΨP = Ø or ∃n ∈ {1. The importance of this Lemma is to characterize the set of initial condition that corresponds to the limit cycle behaviors with period P for k ≥ k0 . (9. and this completes the proof. The local stability is discussed as follows. that is ki = 0 for i = 1. the stability analysis is performed before the fuzzy impulsive control strategy is suggested. nd .4. Reiss rN − nd + n x(k 0 ) = LN − nd + n n = 1. respectively. Hence.13) As a result.252 B. then the eigen decomposition of matrix A exists and Lemma 9.D.W. the SDMs exist limit cycle behaviors with period P for k ≥ k0 . nd } such that N − nd i =1 ∑c i. N . That is: . . Ling. if the degeneracy is equal to the multiplicity of the eigenvalues of matrix A . P −1 j =0 . Only local stability can be achieved and local stability may be enough for some applications. ri x(k0 ) = Li for .1 and 9.F. Hence. As discussed in Section 9. when there are more than one DC poles in the loop filter transfer function. However.2. we have x(k0 ) = x(k0 + P ) . stability is an important issue.1 is still applied.K. Besides. then there does not exist x(0) such that x(k0 + P ) = x(k0 ) . such as for audio applications [2].2.

15) respectively.24) . (9.   (9.19) Hence.21) (9.  x N −1 (k ) .  x′(k ) − Q( x′(k )a N )bN Q(x′(k )aN ) − Q  aN  and ˆ ˆ Ax(k ) + B(u − Q(x(k ))) = [x1 (k ).18) If x′(k ) > bN . x1 (k ). aN  . Denote x′(k ) ≡ bN u + ∑ bN −i (u − Q( xi (k ))) − ∑ aN −i xi (k ) i =1 i =1 T N −1 N (9. If x′(k ) < bN . then Q( x′(k ) − Q(x′(k )aN )bN ) = Q(x′(k )) .16) and  x′(k ) − Q (x′(k )a N )bN ˆ x(k ) ≡  . then  x′(k ) − Q( x′(k )aN )bN Q( x′(k )aN ) − Q  aN   =0   (9. x N (k ) + N a0   ′ ′  Q (x′(k )a N ) − Q x (k ) − Q( x (k )a N )bN   aN       . (9. then Q( x′(k ) − Q(x′(k )a N )bN ) = −Q(x′(k )a N )Q(bN )   = Q( x′(k )aN ) − Q(x′(k )aN ) = 0   (9.  (9.20) . xN (k )] = x(k ) . T (9.     T   x1 (k ).23) If Q(a N bN ) = −1 .Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 253 x(k + 1) ≡ ℵf (x(k )) in which x(k + 1) = Ax(k ) + B(u − Q(x(k ))) (9.22) and  x′(k ) − Q( x′(k )a N )bN Q( x′(k )a N ) − Q  aN    = Q( x′(k )aN ) + Q( x′(k )a N )Q(a N bN ) .   (9. x N −1 (k ).14) and x(k − 1) ≡ ℵb (x(k )) in which x(k ) = Ax(k − 1) + B(u − Q(x(k − 1))) .17) Then ˆ ˆ Ax(k ) + B(u − Q(x(k ))) = b .

then the local stability is guaranteed.4. As discussed in Section 9.26) aN   Suppose the above conditions are satisfied ∀k ∈ Z . then x(k ) ∈℘ ∀k ∈ Z .2 shows the block diagram of how the fuzzy impulsive controller influenced the SDMs. . Lemma 9. Although it was reported in [7] that if the invariant set exists and there exists an initial condition in the invariant set. if ∃k0 ∈ Z such that x(k0 ) ∈℘ . it is not sufficient to conclude the global stability of the SDMs.2.2. That is. two step procedures are employed. However. Fuzzy Impulsive Control Strategy 9. (9. C. Hence.K. x1 (k ).F.25) Ax(k ) + B(u − Q(x(k ))) = x(k ) . then x(k ) ∈ ℜ N \℘ ∀k ∈ Z . the conditions on the existence of the invariant map are not explored and this relationship is explored in Lemma 9. Hence. x N −1 (k ) .D.Y.4. if x′(k ) > bN the possible backward dynamics of the SDMs can be defined as ˆ ˆ (9. the fuzzy impulsive controller determines the controlled state vectors and reset the state variables of the loop filter to the controlled state variables via a reset circuit.W.254 B. If x′(k ) > bN or x′(k ) < bN and Q(a N bN ) = −1 . Proof.1.28) ℑ(x) ≡ Ax + B(u − Q(x)) . Ling.27) and a map ℑ :℘ → ℘ such that (9. Fuzzy Impulsive Control Strategy Fig. and ℵb (x(k )) ∈℘ for k ≤ 0 } (9. Denote ℘ ≡ { x(0 ) : ℵf (x(k )) ∈℘ for k ≥ 0.2 is that it provides information for formulating a fuzzy membership function to achieve local stability. Reiss and Hence. 9. To determine the controlled state vectors. Ho and J. 9. The result follows directly from the definition. ℑ(℘) = ℘ . It is worth noting that if ∃k0 ∈ Z such that x(k0 ) ∈ ℜ N \℘ . The first step of the procedure is the training phase in which the invariant set and the set of state vectors that exhibits limit cycle . and x(k ) may diverge.1. or x′(k ) < bN and Q(a N bN ) = −1 . then ℘ is an invariant set under ℑ . The importance of Lemma 9. then the one of T  x′(k ) − Q ( x′(k )a N )bN  ℵb (x(k )) =  .

Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 255 behaviors are learnt through training. so we define a vector x r ∈ B0 such that x(k0 + 1) − x r 2 is minimum and our goal is to minimize the distance between x c (k0 + 1) and x r via a triangular fuzzy membership function as follows: µ continuous ( where  N x c (k 0 + 1) ≡  ∏ f i x c (k 0 + 1). As discussed in Section 9.    1 (9.1. The details are discussed in below. 9. By generating a set of DC signals inputted to the system with different initial condition.29)  xic (k 0 + 1) − Vcc  r  c xi − Vcc  x (k + 1) + Vcc f i x c (k0 + 1). (9. we want to minimize the effect of audio clicks. The second step of the procedure is the control phase in which the controlled state vectors are determined and the state variables are reset to the corresponding values. the state vectors are tested if they form an invariant set and exhibit limit cycle behaviors or not. To achieve this goal. The block diagram of the interpolative SDM under the fuzzy impulsive control strategy.2. Fuzzy impulsive controller u (k ) + _ F (z ) y (k ) Q Fig. x(k 0 + 1) may be outside the desired bounded region B0 . x r ≡  i 0r  xi + Vcc   0  xir ≤ xic (k0 + 1) ≤ Vcc ( ) − Vcc ≤ xic (k0 + 1) ≤ xir . x r   i =1 ) ( ) N .30) otherwise . However. we want to minimize the distance between the original state vectors x(k 0 + 1) and the controlled state vectors x c (k0 + 1) .

µ stable (x c (k 0 + 1)) = 1 when x c (k 0 +1) = x p and 0 ≤ µ stable (x c (k 0 + 1)) ≤ 1 ∀x c (k0 + 1) ∈ ℜ N . Ho and J. C. µ continuous (x c (k 0 + 1)) = 0 when c x c (k 0 + 1) ∈ ℜ N \ B0 . Denote x p ∈℘∩ B0 such that x(k 0 + 1) − x p is 2 minimum. and the effect of audio clicks is minimized if x c (k0 + 1) is closed to x r . then x p = x(k0 + 1) . which also x(k0 + 1) ∈℘∩ B0 . the trajectory may not be inside B0 because ℘ is usually not equal to B0 . then x(k ) ∈℘ ∀k ∈ Z if ∃k0 ∈ Z such that x(k0 ) ∈℘ . By setting x c (k 0 +1) = x p .2. If ℘∩ B0 ≠ Ø.F. Note that if x(k0 + 1) ∈ B0 . Ling. Hence. this fuzzy membership function can minimize the effect of audio clicks.K.2. µcontinuous (x c (k0 + 1)) = 1 c N when x c (k 0 +1) = x r . Supposing that ℘∩ B0 ≠ Ø. x p  . Hence.    i =1  Since a triangular fuzzy membership function is employed and x p ∈ B0 . the local .1 and 9. As discussed in Section 9.31) µstable x (k0 + 1) ≡  ∏ f i x c (k 0 + 1). 1 ( c ) ( ) µstable (x c (k0 + 1)) = 0 when x c (k 0 + 1) ∈ ℜ N \ B0 . In order to guarantee that the trajectory is bounded within B0 .256 B. if x′(k ) > bN or x′(k ) < bN and Q(a N bN ) = −1 . According to Lemma 9. we want the controlled state vectors to be inside ℘∩ B0 . then we define the following triangular fuzzy membership function:  N N (9. If stability criterion is satisfied. x′(k ) > bN or x′(k ) < bN and Q(a N bN ) = −1 .Y. x c (k 0 + 1) ∈℘ ∩ B0 . and 0 ≤ µcontinuous x c (k 0 + 1) ≤ 1 ( ) ∀x (k 0 + 1) ∈ ℜ . the local stability criterion is an important issue.W. This implies that there exist state vectors that achieve local stability within the set of the desired bounded state variables.D. µstable (x c (k 0 + 1)) force the new state vectors x c (k0 + 1) to be within B0 . Reiss Since a triangular fuzzy membership function is employed and x r ∈ B0 . then x r = x(k0 + 1) and there will be no audio click effect by setting x c (k 0 +1) = x r . Since µcontinuous (x c (k0 + 1)) captures the knowledge on the closeness between x c (k0 + 1) and x r . µ continuous x (k 0 + 1) force the new state vectors c ( ) x (k0 + 1) to be within B0 . Since µ stable (x c (k 0 + 1)) captures the knowledge on the closeness between x c (k0 + 1) and x p . that is. However.

1.    ∀P >0   ∀k ≤ k0  If PER(k0 ) ∩ B0 = B0 .Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 257 reflects the closeness between x c (k0 + 1) and the set of state vectors that achieved local stability within the desired bounded region. then the local stability criterion is not guaranteed. if PER(k0 ) ∩ B0 = Ø. we do not want to move x c (k0 + 1) into ∀P >0 ∪ {x(k )} too. we do not want to move the new state vectors x c (k0 + 1) into ∪ ΨP . (9. this fuzzy membership function can capture the local stability criterion into the system. In order to avoid this case to be happened. then all the state vectors in B0 may result limit cycle behaviors and this situation should be avoided. Hence. The reasons why small value of δ stable can avoid the instability problem are discussed in Section 9. Moreover. Define     (9.32) x c (k 0 + 1) ∈ ℜ N \ B0  0 where 1 ≥ δ stable > 0 and δ stable is very closed to zero. if PER(k0 ) ∩ B0 = B0 or PER(k0 ) ∩ B0 = Ø. This is because after a certain number of iterations. In this case.4. we define the fuzzy membership function as . then we define x c (k0 + 1) ∈ B0 . then we cannot find a state vector x q ∈ B0 ∩ PER (k0 ) such that x(k 0 + 1) − x q 2 is minimum.2. if ℘∩ B0 = Ø.33)   PER(k0 ) ≡  ∪ ΨP  ∪  ∪ {x(k )} . this fuzzy membership function will δ µstable (x c (k0 + 1)) ≡  stable force the new state vectors x c (k0 + 1) to be within B0 . Or if ∃k ′ ∈ Z such that x′(k ) < bN and Q(a N bN ) = 1 . However. the ∀k ≤ k 0 state vectors may go to the same points in the state space and cause limit cycle behaviors to occur. the occurrence of limit cycle behaviors should be avoided. then x p does not exist. As discussing in Section 9. the SDM may suffer from an instability problem. On the other hand. Since the fuzzy membership value of the state vectors outside B0 is exactly equal to zero. Since ∪ ΨP is the set of state vectors that exhibiting ∀P >0 limit cycle behavior. or if ∃k ′ ∈ Z such that x′(k ) < bN and Q (a N bN ) = 1 . if ℘∩ B0 = Ø.

Once the fuzzy membership functions are defined.W. C. we define the fuzzy membership function as µaperiodic (  N c q   c  x (k0 + 1) ≡ 1 −  ∏ f i x (k0 + 1).258 B.34) x c (k 0 + 1) ∈ ℜ N \ B0  0 > 0 and δ aperiodic is also very closed to zero.2. Hence. the reason why small value of δ aperiodic can avoid the occurrence of limit cycle behaviors is discussed in Section 9. Similarly. Ho and J. Reiss δ µaperiodic (x c (k 0 + 1)) ≡  aperiodic where 1 ≥ δ aperiodic x c (k 0 + 1) ∈ B0 . Otherwise. Since f i q µaperiodic (x c (k 0 + 1)) = 0 x c (k0 + 1) ∈ ℜ N \ B0 (9. (9. when and x(k0 + 1) ∈ B0 ∩ PER(k0 ) because x = x(k 0 + 1) when x(k0 + 1) ∈ B0 ∩ PER(k0 ) .Y. Since µ aperiodic (x c (k 0 + 1)) captures the knowledge on the separation between x c (k0 + 1) and B0 ∩ PER(k 0 ) . (9.35) is a triangular fuzzy membership function and x q ∈ B0 . µ x c (k +1) (x c (k0 + 1)) ≡ 0 . .K.4.D. Ling. we can define the fuzzy impulsive control law as follows: If Ax(k 0 ) + B(u − Q (x(k 0 ))) ∈ ℜ N \ B0 . no control force is applied to the SDMs. x  i =1  0  ) ( ) N    1 x c (k0 + 1) ∈ B0 . which also reflects the separation between x c (k 0 + 1) and the set of state vectors within the desired bounded region that exhibits limit cycle behaviors.F. then the fuzzy impulsive controller will reset the state variables of the loop filter to x c (k0 + 1) where x c (k0 + 1) is the state vector such that the following function is maximized. µ aperiodic (x c (k0 + 1)) force the new state vectors x c (k 0 + 1) to be within B0 .36) 1 c c c 3 max N (µ stable (x (k 0 + 1))µ aperiodic (x (k0 + 1))µ continuous (x (k 0 + 1))) c x ( k 0 +1)∈ℜ Otherwise. µaperiodic (x c (k0 + 1)) can be used to avoid the occurrence of limit cycle behaviors. µaperiodic x (k0 + 1) = 0 when c ( ) x c (k 0 + 1) ∈ ℜ N \ B0 0 ≤ µ aperiodic x c (k 0 + 1) ≤ 1 ( ) ∀x c (k0 + 1) ∈ ℜ N .

if ) Ax(k 0 ) + B(u − Q (x(k 0 ))) ∈ ℜ N \ B0 . As a result. N . 0 +1 PER(k0 ) ∩ B0 ≠ B0 . x (k ) ∈ B0 for k > 0 . ∃x (k 0 + 1) ∈ B0 \ PER (k 0 ) c c if PER(k0 ) ∩ B0 ≠ B0 . µ continuous (x c (k 0 + 1)) > 0 and µstable (x c (k0 + 1)) > 0 . and ∪ ΨP = Ø or ∪ ΨP = B0 . .N and Proof. N . Hence. ∀k0 ≥ 0 . ai for i = 0. Another advantage of this fuzzy impulsive control strategy is that we can alter the maximum bound of the state variables easily by setting the value of Vcc appropriately. ∀ai ∈ ℜ for i = 0. ∀P >0 However.1. . then the fuzzy impulsive controller will reset the state vector of the loop filter to x c (k0 + 1) where x c (k0 + 1) ∈ B0 \ PER (k 0 ) .3 guarantees that the ∀P >0 ∀P > 0 controlled trajectory is bounded within B0 no matter what the input step size. x c (k 0 + 1) = x(k0 + 1) ∈ B0 . Thus.3 is still applied even though ℘ = Ø or ℘ = B0 . ∀x (0 ) ∈ ℜ N . ∀b j ∈ ℜ for j = 1. ∀x (0 ) ∈ ℜ N . It can be seen that ∀u ∈ ℜ . the initial condition and the filter parameters. . the initial condition and the filter parameters are. ∀k0 ≥ 0 and ∀x c (k 0 + 1) ∈ B0 . N . . Hence. x c (k ) ∈ B0 for k > 0 . PER(k0 ) ∩ B0 ≠ Ø c and since such that µaperiodic (x (k 0 + 1)) > 0 .1. ∃x (k 0 + 1) ∈ B0 \ PER (k 0 ) such that µ x c ( k ) (x (k c 0 + 1) > 0 .1. Lemma 9. which is independent of the input step size. If PER(k0 ) ∩ B0 = B0 or then Although PER(k0 ) ∩ B0 = Ø. since no control force is applied to the SDM. N . N and b j for j = 1. Hence. µ aperiodic (x c (k 0 + 1)) > 0 . x c (k ) ∈ B0 for k > k0 . And this completes the proof. ∀b j ∈ ℜ for j = 1. This is very important because we do not want the trajectory of the SDM to be unbounded if the input step size is increased. will affect the existence of ℘ and ∪ ΨP . µ aperiodic (x c (k 0 + 1)) = 0 x(k0 + 1) ∈ B0 ∩ PER(k0 ) .Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 259 Lemma 9. c . Lemma 9. If Ax(k0 ) + B(u − Q(x(k0 ))) ∈ B0 . . ∀u ∈ ℜ . It is worth noting that different values of u .3. ∀ai ∈ ℜ for i = 0. x(0) . . or the initial condition or the loop filter of the SDMs are changed.

Ling. x c (k ) ∈ B0 for k > 0 . no matter what the input step size. x (k + 1) − x c r 2 .N and ≤ 2Vcc N for k > 0 . As discussed in above. N and ∀b j ∈ ℜ for j = 1. etc. For the parameters δ aperiodic and δ stable . if the hardware operates normally in a safe condition only when the state variables are bounded by 20V . ∀x(0 ) ∈ ℜ N .1] . c c then ∃M > 0 / such that x (k ) = x (k + M ) for k > k0 .Y. the fuzzy impulsive controller works properly ∀δ aperiodic ∈ (0. . Reiss Lemma 9. ∀ai ∈ ℜ for i = 0.2.W.5. since δ aperiodic represents the fuzzy membership value of how to avoid the occurrence of limit cycle at x c (k 0 + 1) when PER(k0 ) ∩ B0 = B0 or . Vcc is the maximum allowable bound on each state variable and this value is determined based on the real situations.F. For example.1] and ∀δ stable ∈ (0. N . Proof. while our approach will minimize the occurrence of limit cycle behaviors. However. the initial condition and the filter parameters are. They are Vcc . The proof follows directly from Lemma 9.4. ∀ai ∈ ℜ for i = 0.1.4. Lemma 9. We will show the contrast in Section 9.3. N . ∀u ∈ ℜ . Since ∀u ∈ ℜ . . δ aperiodic and δ stable . The importance of this Lemma is that it guarantees the norm of the difference between x r and x c (k + 1) being bounded by 2Vcc N . The importance of this Lemma is that it states the condition that limit cycle behaviors do not occur when the fuzzy impulsive control strategy is applied at once. C. then Vcc may be set accordingly. ∀x(0 ) ∈ ℜ N . .260 B.D.1. and Ax(k 0 ) + B(u − Q (x(k 0 ))) ∈ ℜ N \ B0 . the result follows directly. ∀b j ∈ ℜ for j = 1. Parameters in the Fuzzy Impulsive Controller There are only three parameters in the fuzzy impulsive control strategy.K. 9. Ho and J. Proof. If ∃k0 ∈ Z such that PER(k ) ∩ B0 ≠ B0 for k ≥ k0 . such as the hardware constraints and the safety specifications.5 that the clipping control strategy usually results in the limit cycle behaviors. we do not want the norm of the difference between x r and x c (k + 1) to be too large because the effect of audio clicks may be too large for these situations.

the SDM may exhibit divergent behavior if the fuzzy impulsive control strategy is not applied.36). since the input signals are typically oversampled at 64 or 128 [2]. 9.36). since it represents the fuzzy membership value of the local stability of the SDM at x c (k0 + 1) when ℘∩ B0 = Ø. the number of samples inputted to the SDM per second is 2. such as 10−3 . these three basic rules and a simple fuzzy engine is enough for achieving the objectives. 9. 10−3 . Because several megasamples are needed to process per second.4.6448M.5. we recommend the SDM control designers to set this value as a small number closed to zero too. The Nyquist rate for audio signal is 44. many existing reset circuits can be employed [8]. 9. and all the state vectors in B0 may cause the trajectory to exhibit limit cycle behaviors if PER(k0 ) ∩ B0 = B0 . this will increase the complexity of the system and may cause real time processing problems. Implementation of the Fuzzy Impulsive Controller As discussed in the above. can be employed for solving equation (9. such as MATLAB or MATCAD.4. we suggest the SDM control designers to set this value as a small number closed to zero.4. To reset the state variables of the loop filter. or if ∃k ′ ∈ Z such that x′(k ) < bN and Q (a N bN ) = 1 . For δ stable .8224M or 5. Simulation Results To illustrate our results. the fuzzy impulsive controller resets the state variables of the loop filter to the controlled state variables of x c (k0 + 1) if Ax(k 0 ) + B(u − Q (x(k 0 ))) ∈ ℜ N \ B0 . According to the simulation results shown in Section 9.1kHz [2]. particular for audio applications [2]. and in this case. and x c (k0 + 1) is calculated based on equation (9.3.5. a fifth order SDM with loop filter transfer function . Complexity Issue Although more fuzzy rules and sophisticated fuzzy engine will improve the performance of the SDMs. only three basic fuzzy rules are captured and only a simple fuzzy engine is used to reduce the complexity for processing. for example.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 261 PER(k0 ) ∩ B0 = Ø. Numerical solvers.

0.39) y (k ) = Q C~ (k ) .37) 1 − 5 z −1 + 10. Ho and J.9793]T when u = 0.09375  0  0. the maximum absolute value of the state variables of the controlled SDM is bounded by Vcc for k > 0 and 0 0 0 0 ∀u ∈ ℜ .0025 z −5 is illustrated. which may be too large for some practical applications [2]. we can guarantee that the state 1 1 ~  A ≡ 0  0 0  T variables are bounded by 20 . 9. Ling.00589 0 0 1 1      (9. Fig. 0]T .F. B ≡ 0 and C ≡  1  . the maximum absolute value of the state variables is between 20.0497 z −3 − 64.0015 z −4 + 14. C. This fifth order SDM is commonly employed in the industry [2].D.40) Assume that the initial condition of this SDM is zero. that is. By using a simple transformation.262 B. Reiss 20 z −1 − 74 z −2 + 103 .0025 z − 2 − 10. 0]T ) if the input step size is approximately x between − 0.0075 z − 4 − 1.3. 0. where ~ (9. 0. and may diverge if the input step size is outside this range.0523 and 59. From the simulation result. 1 1 0 0      0 1 1 − 0.25. 32. .0018 0 0  ~    ~   .9584 z −5 (9.4633 . − 5.5. 0. 35. 9. According to Lemma 9.K. We can check that the trajectory of this SDM is bounded for this initial condition ( ~ (0 ) = [0. we can see that even though the trajectory is bounded for this range of input step size. this SDM x can be realized by the direct form and the corresponding initial condition is x(0 ) = [0.Y. 28.3 also shows the plot of the maximum absolute value of the state variables (also realized in the direct form) for k > 0 versus the input step size when the fuzzy impulsive control strategy is applied at Vcc = 20 . even though u ≥ Vcc .3.71 and 0. x ( )  1  20   0   6  1 − 0.75 . The relationship between the maximum absolute value of the state variables (realized in the direct form) and the input step size is plotted in Fig.38) x x for k ≥ 0 .W. The SDM can be implemented via the Jordan form [2] and can be realized as the following state space equation ~ ~ ~ (k + 1) = A~ (k ) + B(u (k ) − y (k )) (9. Hence. ~ (0 ) = [0.75 . 0.000685 0  0.0075 z −3 + 5. 0.

x This SDM is not globally stable. 0. we see that the . On the other hand. 9.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 263 Fig.4b that even though the SDM exhibits acceptable behavior when ~ (0 ) = [0. 9. 9. the maximum absolute value of the state variables is always bounded by Vcc for k > 0 and ∀x(0 ) ∈ ℜ N if the fuzzy impulsive control strategy is applied.001.3. the SDM exhibits divergent behavior when ~ (0 ) = [0. 0.4a and Fig. 0. It can be seen from x Fig.4c and Fig. 0]T x T and ~ (0 ) = [0. Fig.4a and Fig. 0. 0. 0.75 . 0] . 9. 0. 0. when u = 0. 0]T and the behaviors of the x SDM for these two different initial conditions are very different. 0]T and the difference between x these two initial conditions is very small.4d show the corresponding state responses when the fuzzy impulsive control strategy is applied at Vcc = 40 . 0]T . 9. That means. 9. 0.4b show the responses of x1 (k ) with ~ (0 ) = [0. Plot of the maximum absolute value of the state variables (realized in direct form) against the input step size when ~ (0 ) = [0.3. ∃~ (0 ) ∈ ℜ N such that x the trajectory is unbounded. 0. For example. 0. respectively. according to Lemma 9. 0. Fig. 0.001. 0. From the simulation result. 0. 9.

0. 0. (d) Initial condition ~ (0 ) = [0.2. Fig. 0]T when no control strategy is x applied. Denote λi for i = 1. 0. 0.264 B. . 0. . 6 max λi occurs at K c = 0 . Ho and J. (c) Initial condition ~ (0 ) = [0. (b) Initial x condition ~ (0 ) = [0. For comparison with other control strategies.001.K. 0. in which the controller is in the form of − K c (1 − z −1 ) . 2 .D. Ling. C.001.2.F. 0.W. When K c = 0 . consider the time delay feedback control strategy suggested in [6]. . 0. 0. . 0. it reduces to the .75 and (a) initial condition ~ (0 ) = [0. Since λi for i = 1. 0]T when the fuzzy x impulsive control strategy with Vcc = 40 is applied. 0. it can be shown that max λi > 1 ∀K c ∈ ℜ and the minimum value i =1. 2 .Y. The response of x1 (k ) when u = 0. 6 of i =1.6 be the poles of the effective loop filter. Reiss SDM exhibits acceptable behavior with the state variables bounded by Vcc for both of these two initial conditions. 0]T when no control strategy is applied. 0]T when the fuzzy impulsive control strategy x with Vcc = 40 is applied. 9.6 depends on the value of K c .4. 0.

and setting the initial condition and the input step size as the previous values.25. 39. that is.25. Fig. 9. .5. Fig.5. Hence. − 5. T x(0 ) = [0.9793. 0. that is. By selecting a value of Kc which is very closed to zero. N .75 and initial condition x(0 ) = [0. 0. it is found that the trajectory diverges as shown in Fig. the time delay feedback control strategy fails to stabilize this SDM. for example K c = 2 × 10 −5 . 9. 28.5. 28. 35.5. 9.5612] and u = 0. 32. The response of x1 (k ) with input step size u = 0.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 265 uncontrolled case.9793.2. 0]T and the x . 39. 32.75 . set xi (k ) = VccQ( xi (k )) if xi (k ) ≥ Vcc for i = 1. it is found that limit cycle behaviors may occur if the clipping control strategy is applied.5612 ]T when the time delay feedback control strategy with K c = 2 × 10 −5 is applied. 0. To compare the fuzzy impulsive control strategy to the clipping control strategy. ~ (0 ) = [0. 35.6 shows the magnitude response of s(k ) when u = 0.75 (the initial condition is determined based on zero initial condition of the Jordan form). − 5.

W.7 that the SNR x .266 B. 0. On the other hand. the spectrum is quite flat for the SDM when the fuzzy impulsive control strategy is applied with Vcc = 40 . C.K. Ho and J. SNR is calculated using [9]. 9. The oversampling ratio is 64. and initial conditions are given by ~ (0 ) = [0. Magnitude response of the output sequence when u = 0.F.6. where the frequency of the input sinusoidal signals is 2 of the passband 3 bandwidth.6 that there is an π if the clipping control strategy is applied. 0.7 shows the SNR of SDMs under the clipping control strategy with the clipped level set at 28 .75 and initial condition ~ (0 ) = [0. 0]T for both the clipping and fuzzy x impulsive control strategies are applied with the state variables bounded by 40 . 0. It can be seen from Fig. 9.Y. Fig. 9. Ling. 0]T . 9. 0.D. which impulse located at 2 demonstrates that the SDM exhibits a limit cycle with period 2. 0. Reiss clipped level is set at 40 . Fig. It can be seen from Fig. 0. which demonstrates that the SDM exhibits acceptable behavior and the limit cycle behavior is avoided.

8281dB improvement compared to the clipping control strategy for outside this input magnitude range. Hence. On the other hand. 0.7. if the input magnitude exceeds this range.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 267 of both the clipping and fuzzy impulsive control strategies with the state variables bounded by 28 are the same when the input magnitude is less than 0. However. the SDM performs normally under the fuzzy impulsive control strategy.0135 for the input magnitude range greater than or equal to 0.52. 0]T and the x state variables are bounded by 28. SNR of SDMs when input sinusoidal frequency is 2 of the 3 passband bandwidth. the SNR of the SDM under the fuzzy impulsive control strategy has an average of 41. It can be seen from Fig. 0. 9. initial condition ~ (0 ) = [0. 0.8 that the probability of the control force to be applied by the fuzzy impulsive control strategy is 0. This is because both the maximum absolute value of the state variables (realized in the direct form) do not exceed 28 in this input magnitude range.2562dB because of the occurrence of limit cycle behaviors. as opposed to a . Fig. the SNR corresponding to the clipping control strategy may drop to less than 1. 9.52 .

This demonstrates that the fuzzy impulsive control strategy is more efficient than the clipping control strategy. Hence.6926 for the clipping control strategy.0069 z −3 + 5. (9. Ho and J. 0. 9. Ling. 0.0023 z −5 .9094 z −3 − 2. Fig.W. initial condition 3 ~ (0 ) = [0.0023 z − 2 − 10.268 B.D. x To verify the independence of the filter parameters on the fuzzy impulsive control strategy. Reiss probability of 0.Y. the number of reset action on the state variables of the loop filter is much reduced when applying fuzzy impulsive control strategy.5498 z −5 .8.7919 z −1 − 2. Probability of control force applied to the SDM when the input sinusoidal frequency is 2 of the passband bandwidth. 0]T and the state variables are bounded by 28. C. This is because the fuzzy impulsive control strategy tends to reset the state vectors inside the invariant set if it exists and the state vectors will tend to stay inside the invariant set without applying control force again soon afterwards.0069 z − 4 − 1. 0.8630 z −2 + 3. consider another fifth order SDM with the following transfer function [2] 0.41) 1 − 5 z −1 + 10.F.3873 z −4 + 0.K.

9b. 0. The response of with initial condition x1 (k ) ~ (0 ) = [0.9.9a that the trajectory diverges. 9. Fig. 9.Control of Sigma Delta Modulators via Fuzzy Impulsive Approach 269 This SDM is also widely used in the industry [2]. the maximum absolute value of the state variables (realized in the direct form) is always bounded by Vcc for k > 0 .6. 0]T is shown in x Fig. when the fuzzy impulsive control strategy is applied with Vcc = 2 . (b) when the fuzzy impulsive control strategy with Vcc = 2 is applied. . Conclusion 9. N .9a. we have suggested the fuzzy impulsive control strategy for the stabilization of high order interpolative SDMs in which the occurrence of limit cycle behaviors and the effect of audio clicks are . according to Lemma 9. N and ∀b j ∈ ℜ for j = 1. The trajectory of this SDM with u = 0. 0. 9. 0]T and input step size u = 0. On the other hand.59 and ~ (0 ) = [0. and it can be seen from Fig. as shown in Fig.3. strategy is applied. 0. 0. 0.59 (a) when no control x In this chapter. .1. ∀ai ∈ ℜ for i = 0. 0. 9.

Reiss minimized. Cerman. 38-44. Chattopadhyay and V. V. Candy. Uçar. (2002). Ramanarayaanan. (2003). Philips Research.0 Analog Devices Inc. (1997). the SNR performance of the SDMs is maintained or improved after control. IEEE Transactions on Circuits and SystemsI: Fundamental Theory and Applications 49. Moreover.W. C. IEEE Transactions on Communications 52. 1010-1018. A use of limit cycle oscillations to obtain robust analog-to-digital converters. (2003).Y.F. IEEE Transactions on Circuits and SystemsI: Fundamental Theory and Applications 44. 41-53. R. A. IEE ProceedingsCircuits. (1974). Janssen. Ling. G. A single-reset-integrator-based implementation of line-current-shaping controller for high-power-factor operation of flyback rectifier. Ho and J. Ginis and J. (2003). [4] [5] [6] [7] [8] [9] . (2004). 490-499. Devices and Systems 150. (2002). IEEE Transactions on Instrumentation and Measurement 52. Variable-structure compensation of delta-sigma modulators: stability and performance. Comparisons between the fuzzy impulsive control strategy and some existing control strategies show that the fuzzy impulsive control strategy is much effective in terms of producing much higher SNR and efficient in terms of requiring less control force applied to the system. T. An algorithm for computing convex positively invariant sets for delta-sigma modulators. Goodson and B. IEEE Transactions on Communications COM-22.270 B. M. Signal processing for direct stream digital: A tutorial for digital sigma delta modulation and 1-bit digital audio processing. A weak magnetic field measurement system using micro-fluxgate sensors and delta-sigma interface. IEEE Transactions on Industrial Applications 38. S. Bounding integrator output of sigma-delta modulator by time delay feedback control. Schreier. Kawahito.. Schreier. Cioffi. A. K. The delta-sigma modulators toolbox version 6. the initial condition and the filter parameters are.K. Reefman and E. Aramaki and Y. Johns. Eindhoven. R. 31-37.D. White Paper (2002). A. M. Zourntos and D. Tadokoro. Optimum bandwidth partitioning with analog-to-digital converter constraints. References [1] [2] [3] J. 103-110. Since the effective poles of the loop filter are not affected by the control strategy. S. 298-305. D. the controlled trajectory is guaranteed to be bounded no matter what the input step size. Zhang. C.

Sign up to vote on this title
UsefulNot useful