Bounded tracking controllers for the chaotic (forced) Dung equation

A.A.J. Lefeber and H. Nijmeijer

This letter deals with the design of feedback controllers for a chaotic dynamical system like the Dung equation. Lyapunov theory is used to show that the proposed bounded controllers achieve global convergence for any desired trajectory. Some simulation examples illustrate the presented ideas.

AMS Subject Classication: 93C10, 93D05, 94C05, 34C35, 34D20

Keywords: Bounded controllers, Control theory, Chaos, Feedback, Lyapunov methods, Trajectory tracking.

Introduction

Starting with [7] controlling chaotic systems has recently become an extremely exciting and interesting research area, see the bibliography [1] and the paperscollection [8]. Basically, the method advocated by Ott, Grebogi and Yorke [7] consists of making small time-dependent perturbations on available system parameters so that an initially chaotic system is steered towards any orbit contained in the chaotic attractor. Since the original OGY-paper various modications and adaptations on the forementioned method have appeared. Also a number of experimental case studies have illustrated the idea. On the other hand, the OGY-method, and more general, most literature on control of chaos, seems distinct from what is called control theory, although some clear connections exist. The purpose of the present note is to work out from a control theoretic point of view, a feedback control scheme that ensures the tracking of any desired 1

144

trajectory under given input constraints, by means of a specic example. The example we consider is the controlled forced Dung equation [2,6]: x + px + p1 x + x3 = u + q cos(t) (1)

with p 0, p1 , q and constants and u() the (physical) control. For u 0, depending on the choice of the constants, it is known that solutions of (1) exhibit periodic, almost periodic, and chaotic behaviour; see e.g. [2]. Note that, in contrast to [2] we consider (1) as the controlled form of the Dung equation, and for obvious reasons we do not allow for control actions in the x direction. Typically the control u in (1) can be understood as a force (torque) applied to the physical system that obeys the uncontrolled forced Dung equation. Clearly this also distinguishes the way controls appear from typical control of chaos work like in the OGY-method, see [7], where control actions are generated via parameter perturbations. Most notably, the dierences with the OGY-method may be summarized as: The proposed state feedback controller achieves global tracking, i.e. no matter where the original system starts and no matter whether the uncontrolled system has a chaotic attractor or not. Tracking towards any desired trajectory is guaranteed; in particular the desired trajectory may not be an uncontrolled system trajectory, the prize being that a nonzero control action is needed to stay on the desired trajectory. The tracking is achieved by means of a state feedback controller, which automatically acts as an error correction mechanism. The above dierences with the OGY-method are from one side overemphasized. It will turn out that each aspect mentioned above also has elements in common with the OGY-method, and even more, the OGY-methodology may be considered as a driving force for our present work. The results we present here are based on the analysis of one specic model (the uncontrolled Dung equation) and are motivated by earlier work, [2,6], where the systematic use of, not necessarily bounded, feedback was proposed. We do think that our results are mutatis mutandis also valid for various other systems, but certainly not for all systems. The organization of this paper is as follows. In section 2 bounded feedback controllers, which are of composite form, are given. Simulations illustrating the control performance and tracking ability of the controllers are given in section 3. Section 4 contains concluding remarks. We conclude this introduction with some terminology and notation. 2

145

Let F denote the class of nondecreasing continuous dierentiable functions f :I RI R satisfying f (0) = 0, f (0) > 0 and supxIR |f (x)| 1. This class is a subset of the class of saturating functions proposed in [3]. Examples of 2 functions f (x) F are f (x) = tanh(x) and f (x) = arctan(x). Throughout we want to discuss the tracking problem under input constraints. Therefore we assume that the desired trajectory xd we want to track is bounded and has bounded rst and second derivative, i.e. there exist positive constants B0 , B1 , and B2 such that d (t)| B1 , |x d (t)| B2 , |xd (t)| B0 , |x t0 (2)

A specic case which deserves separate attention, is formed by those desired trajectories that satisfy d + p1 xd + x3 |x d + px d q cos(t)| B (3)

0. Clearly, when B = 0, xd is a trajectory of the uncontrolled for some B system, and in this case our work is related to (but dierent from) various other control of chaos papers, see e.g. [8].

State feedback control of the Dung equation

Assume we want the system (1) to follow any smooth desired trajectory xd (t), satisfying (2). For this purpose, we select the control via a state feedback law u=x d + px d + p1 xd + x3 d q cos(t) Kp e Kd e where e x xd , Kd and Kp are positive constants, and f1 , f2 F . The controller (4) consists of two components, namely d + p1 xd + x3 (i) a feedforward part: x d + px d q cos(t), which can be computed o-line, (ii) a P(roportional) D(ierential) feedback part: Kp e + Kd e . The feedback terms in (ii) are required to guarantee that the system converges towards xd (t). Once on this trajectory, the feedforward component keeps the system moving along it; note that this part equals zero if xd is a trajectory of the uncontrolled system. 3 (4)

146

The closed-loop system consisting of (1) and (4) is described by the timevarying second-order dynamics + (p1 + Kp )e + e3 + 3e2 xd + 3ex2 e + (p + Kd )e d = 0 (5)

Proposition 1 Let xd (t) be any smooth bounded reference trajectory, satisfying (2). Then the state feedback controller (4) guarantees that (1) asymptotically converges towards xd (t), i.e.
t

lim (e(t), e (t)) = (0, 0)

provided that Kp > p1 and Kd p + 2B0 3. PROOF. We will use Lyapunovs direct method for obtaining our results, see for instance [4,5,9]. Dene = (Kd + p)/2 and consider the candidate Lyapunov function 1 1 1 + e)2 + (2 + p1 + Kp )e2 + e4 ; V (e, e ) = (e 2 2 4 (6)

which is positive denite. Along the closed-loop error dynamics (5), its timederivative becomes:
2 (e, e V ) = e 2 (3xd e + 3x2 (p1 + Kp + e2 + 3xd e + 3x2 (7) d )ee d )e

which is negative denite on provided that

3 2 3 2 Since |xd | 3 and (p1 + Kp + e2 +3xd e +3x2 d ) = p1 + Kp +(e + 2 xd ) + 4 xd > 0 we obtain:

(e, e Therefore, V ) is negative denite which completes the proof.

When we know the initial error-state, we are able to determine an upperbound for the control law. Since (6) is a decreasing function along solutions of (5), it 4

147

follows that for all t 0: |e(t)| and |e (t)| resulting in + Kp D0 + Kd D1 |u(t)| B (10) 4V (e(0), e (0)) D1 (9) (2 + p1 + Kp )2 + 4V (e(0), e (0)) (2 + p1 + Kp ) D0 (8)

This upperbound on the controller (4) depends on V (e(0), e (0)), and the larger the initial errors are, the larger the upperbound will be. To overcome this problem, we use a composite control law. Apart from the tracking controller (4) we develop a bounded controller that globally controls the system into a specic area containing the origin. As soon as the system is in that area, we have bounds on both e and e , such that we can switch to the tracking controller (4). Since the rst phase controller will be globally, i.e. it controls the system towards the specic area from any initial error state, the composite controller will be a bounded globally asymptotically stable tracking controller. Proposition 2 Consider the system (1) (with p 0), together with the control law u = q cos(t) Kp f1 (x) Kd f2 (x ) (11)

with Kp 0, Kd > 0 and f1 , f2 F . Then for all C0 > 0 and C1 > 0 there (t)| C1 for all t ts , provided exists a ts 0 such that |x(t)| C0 and |x that Kp is chosen large enough, i.e. Kp > 0 for p1 > 0 and Kp > p1 x max x[ p1 , p1 ] f1 (x)

otherwise.

PROOF. The controller (11) results in the time-invariant closed-loop system ) = 0. x + px + p1 x + x3 + Kp f1 (x) + Kd f2 (x 5 (12)

148

Consider the radially unbounded candidate Lyapunov function 1 2 1 1 V1 (x, x + p1 x2 + x4 + Kp F1 (x) ) = x 2 2 4 (13)

where F1 (x) = 0x f1 ( )d . This Lyapunov function is positive denite. Dierentiating (13) along solutions of (12) yields 1 (x, x ) = px 2 Kd xf 2 (x ), V which is negative semi-denite in the state (x, x ). Therefore, we have to de 1 (x, x ) = 0}, which is the termine the largest invariant set in {(x, x ) I R 2 |V origin. Application of LaSalles theorem completes the proof. 2 Corollary 3 Consider the system (1). For all C0 > and C1 > 0, there exists a switching time ts 0 such that the composite control law

u=

q cos(t) Kp,1 f1 (x) Kd,1 f2 (x ) t < ts

x d + px d + p1 xd + x3 t ts d q cos(t) Kp,2 e Kd,2 e

(14)

results in a globally asymptotically stable closed-loop system, provided that 1x Kp,2 > p1 , Kd,2 > p + 2B0 3 and Kp,1 > maxx[p1 ,p1 ] fp 1 (x) Furthermore, there exists a > 0 such that the composite controller (14) satises |u(t)| for all t 0.

(ts )| C1 . From PROOF. Let ts be a moment both |x(ts )| C0 and |x Proposition 2 the existence of ts follows and from Proposition 1 it follows that the second phase controller is asymptotically stable. Therefore the composite controller (14) is globally asymptotically stable. d (t)| B1 for all t 0, we have |e(ts )| B0 + C0 Since |xd (t)| B0 and |x and |e (ts )| B1 + C1 , resulting in (10) for t ts . For the rst phase we have |u(t)| |q | + Kp,1 + Kd,1 , therefore we have |u(t)| for all t 0 with + Kp,2D0 + Kd,2 D1 } = max{|q | + Kp,1 + Kd,1 , B (15)

It may be observed that the upperbound (15) is not very accurate and thus possibly a better bound exists. 2 6

149

Simulations

To support our results, we simulated with MATLABTM the system (1) with the controller (4). The Dung parameters were selected as p = 0.4, p1 = 1.1, q = 2.1 and = 1.8, in which case the Dung equation displays chaotic behaviour [2]. In our rst simulation we initialize our system in (x(0), x (0)) = (0, 0) and the desired trajectory we want to track is an orbit of the uncontrolled system = 0), initiated in (xd (0), x d (0)) = (0, 1). To see clearly the dierence (i.e. B between the chaotic uncontrolled motion and the regulated system, no control is applied during the rst 30 seconds. At t = 30s we initiate the composite controller (14) with Kp,1 = 5, Kd,1 = 3, Kp,2 = 5 and Kd,2 = 7. The switching moment ts is taken as the rst moment t at which both |x(t)| C0 = 0.1 and |x (t)| C1 = 0.1. The resulting performance is depicted in gure 1 We can see in gure 1a that within ve seconds the system perfectly follows the desired trajectory. In gure 1b we can see the control-eort is initiated at t = 30. At t = 32.6 the state is small enough to switch to the second phase controller in order to establish tracking of the desired trajectory. To show that we are able to track any desired trajectory, gure 2 shows an anologous simulation, only our desired trajectory xd (t) has been taken as xd (t) = sin(t). We again see that within ve seconds, our controller succeeds in tracking the desired trajectory, but, of course, a nonzero control-action in order to follow the desired trajectory is needed. With the above parameter setting, we may compute, using Corollary 3 an upperbound for umax . For the rst phase we have |u(t)| 5 + 3 + 2.1 = 10.1. We switch to the second phase at a moment ts at which both |x(ts )| 0.1 and |x (ts )| 0.1 such that in the rst simulation we guarantee that |e(ts )| 2.2 and |e (ts )| 3.0 (since B0 = 2.1 and B1 = 2.9). Therefore V (ts ) 110.5 | 19.3 implying that and so for all t ts we have that |e(t)| 3.1 and |e |u(t)| 151; a very poor estimate, which in the simulations turns out to be about 15 times smaller. To illustrate that the transient behaviour of the controlled system can be inuenced by a suitable choice of the gains Kp and Kd , we did the same simulation as the rst one (with the same initial conditions), only now using Kp,1 = Kp,2 = 2 and Kd,1 = Kd,2 = 8. The resulting performance is depicted in gure 3. The main reason for the poor estimate on the upperbound is caused by large |e(ts )|. Therefore, a way to reduce the value of 151 is to switch at a moment that |e(ts )| is small, for instance as |e(ts )| 0.1. Notice we are only able to 7

150

do this if the desired trajectory comes innetely many times innitely close to the origin. The anology to the OGY-method seems obvious, allthough this propery not only holds for chaotic reference trajectories, but also for other signals. To illustrate this idea, gure 4 shows the resulting performance when we redo our second simulation, tracking of xd (t) = sin(t), but this time we switch as soon as both |e(t)| 0.1 and |e (t)| 3.0. The resulting estimate for the upperbound on the control-eort can be calculated to satisfy |u(t)| 35.7.

Concluding remarks

We have presented a bounded feedback controller for trajectory tracking in the forced Dung equation. The methodology under which we have set up the controller (14) is subject to various modications (i) In the rst stage of the composite controller we steer the system towards the origin; if the desired trajectory is far away from the origin it may become more protable to steer the system with a slightly modied controller towards a point closer to the desired trajectory. (ii) The switching time ts in the controller (14) is in some sense very important. Switching immedeately once the controlled system has reached (ts )| C1 may lead to relatively large the box where |x(ts )| C0 and x (t0 )). Thus it may control actions due to relatively large errors (e(t0 ), e be more appropriate to switch the controller at a later time for which ||(e(t0 ), e (t0 ))|| is smaller. Clearly this issue is strongly related with (i) (t0 ))|| bewhere we assure via a dierent initial controller that ||(e(t0 ), e comes small. (iii) The upperbound (15) is not tight and can certainly be improved. However in any case the main designer freedom in the composite controller (14) lies in the selection of the gains Kp,i, Kd,i , i = 1, 2 (and as mentioned the selection of ts ). The larger the gains, the faster the controller becomes. (iv) Our controller is globally dened and the trajectory tracking was established for any initial state (x(0), x (0)) and any desired trajectory. This was achieved via a careful Lyapunov analysis. Clearly, other Lyapunov functions may lead to other controllers and dierent convergence rates, but this is a usual problem of how to select a candidate Lyapunov function. (v) It is worthwhile to note that for the second part of the controller (14) the V for some > 0 with V given by closedloop system satises V (6). This implies that in this phase the controller guarentees exponential convergence towards the desired trajectory.

151

References

[1] G. Chen. Control and synchronization of chaotic systems (a bibliography), 1996. available from ftp://uhoop.egr.uh.edu/pub/TeX/chaos.tex (login name anonymous password: your email address). [2] G. Chen and X. Dong. On feedback control of chaotic continuous-time systems. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 40(9):591601, 1993. [3] Y. Chitour, W. Liu, and E. Sontag. On the continuity and incremental-gain properties of certain saturated linear feedback loops. International Journal of Robust and Nonlinear Control, 5:413440, 1995. [4] W. Hahn. Stability of Motion. Springer-Verlag, Berlin, 1967. [5] H.K. Khalil. Nonlinear Systems. Macmillan, New York, 1992. [6] H. Nijmeijer and H. Berghuis. On Lyapunov control of the Dung equation. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 42(8):473477, 1995. [7] E. Ott, C. Grebogi, and J. Yorke. Controlling chaos. Physical Review Letters, 64(11):11961199, 1990. [8] E. Ott, T. Sauer, and J. Yorke, editors. Coping with chaos. Wiley, 1994. [9] M. Vidyasagar. Nonlinear Systems Analysis. Prentice Hall, Englewood Clis, second edition, 1993.

152

a 15 2 1 x and xd 0 1 2 0 20 t c 3 2 2 1 0 1 2 2 3 0 20 t 40 60 4 0 20 e_dot 0 e 4 40 60 5 0 20 u 10 5 0

40 t d

60

40 t

60

Fig. 1. xd solution of Dung, Kp,1 = Kp,2 = 5, Kd,1 = 3, Kd,2 = 7.

10

153

a 15 2 1 x and xd 0 1 2 0 20 t c 3 2 2 1 0 1 2 2 3 0 20 t 40 60 4 0 20 e_dot 0 e 4 40 60 5 0 20 u 10 5 0

40 t d

60

40 t

60

Fig. 2. xd (t) = sin(t), Kp,1 = Kp,2 = 5, Kd,1 = 3, Kd,2 = 7.

11

154

a 15 2 1 x and xd 0 1 2 0 20 t c 3 2 2 1 0 1 2 2 3 0 20 t 40 60 4 0 20 e_dot 0 e 4 40 60 5 0 20 u 10 5 0

40 t d

60

40 t

60

Fig. 3. xd solution of Dung, Kp,1 = Kp,2 = 2, Kd,1 = Kd,2 = 8.

12

155

a 15 2 1 x and xd 0 1 2 0 20 t c 3 2 2 1 0 1 2 2 3 0 20 t 40 60 4 0 20 e_dot 0 e 4 40 60 5 0 20 u 10 5 0

40 t d

60

40 t

60

Fig. 4. xd (t) = sin(t), Kp,1 = Kp,2 = 5, Kd,1 = 3, Kd,2 = 7, OGY-like switching strategy.

13

156