- Home
- Documents
*Optimum Realizations of Sampled-data Controllers .Optimum Realizations of Sampled-data Controllers*

prev

next

out of 13

View

213Download

0

Embed Size (px)

Pergamon 0005-1098(94)00112-X Aufomarica, Vol. 31, No. 3, pp. 367-379, 1995

Ekvier Science Ltd Printed in Great Britain

mo5-1098/95 $9.50 + 0.00

Optimum Realizations of Sampled-data Controllers for

FWL Sensitivity Minimization*

ANTON G. MADIEVSKI,? BRIAN D. 0. ANDERSON? and MICHEL GEVERSS

A procedure for choosing a realization of a digital compensator of known transfer function is described, which ensures that the errors introduced into a sampled-data closed loop by using jinite-word-length arithmetic in the

compensator operation are minimized.

Key Words-Closed-loop systems; control systems; convergence of numerical methods; digital control; linear systems; numerical methods; sampled data systems; time-varying systems.

Abstract-The problem of an optimal finite-word-length state-space realization of a digital controller is investigated. The closed loop to be considered consists of a continuous-time plant, a discrete-time controller, a sampler, a zero-order hold and an antialiasing filter. An effective algorithm is proposed to find the optimal sampled-data controller realization minimizing the sensitivity of the closed-loop performance with respect to coefficient errors in the state variable matrices of the controller realization. In order to get a tractable problem, a two-step procedure is to be used: very fast sampling at a multiple of the sampling frequency followed by blocking or lifting to obtain a single-rate system. The procedure allows consideration of the systems intersample behaviour.

1. INTRODUCTION

In order to achieve the desired characteristics of a closed-loop system, a controller is to be used. It is well-known that a desired controllers transfer function can be implemented by any one of an infinite set of realizations of the controller. Though all these realizations are in principle equivalent, since they yield the same transfer function, they have different numerical pro-

* Received 7 March 1994; revised 9 June 1994; received in final form 19 July 1994. The original version of this paper was presented at the 12th IFAC World Congress, which was held in Sydney, Australia during 19-23 July 1993. The Published Proceedings of this IFAC Meeting may be ordered from: Elsevier Science Limited, The Boulevard, Langford Lane, Kidlington, Oxford OX5 lGB, U.K. This paper was recommended for publication in revised form by Associate Editor K. Uchida under the direction of Editor T. Bagar. Corresponding author Dr Anton Madievski. Tel (61) 249 4581; Fax (61) 6 249 2698; E-mail antlOl@syseng.anu.edu.au.

t Department of Systems Engineering, Australian National University, Canberra, ACT 0200, Australia.

$ Centre dhrgeneirie des Systemes, dAutomatique et de Mecanique Appliquee, Universite Catholique de -Louvain, BLtiment Euler, 4-6, avenue Georges Lemaitre, B-1348 Louvain-la Neuve, Belgique.

perties due to finite-word-length effects when they are implemented by a digital device. Such factors as sensitivity and error propagation strongly affect closed-loop performance, and are responsible for differences between desired ideal closed-loop characteristics and those actually obtained. A problem of great importance is to find the realization of the controller that achieves the best performance of the closed-loop system, i.e. that gives the best approximation of the ideal closed-loop behaviour.

Results on optimal realizations of filters (or open-loop systems) minimizing some measure of performance degradation due to FWL errors date back to the late 1970s. The first results were on realizations that minimize roundoff error propagation (Hwang, 1977; Mullis and Roberts, 1976). Realizations minimizing some measure of the transfer function sensitivity to coefficient errors took much longer to emerge (Thiele, 1986).

It was not until the late 1980s that the problem of optimal controller realization minimizing closed-loop performance degradation due to numerical errors was addressed. Solutions were proposed first for specific control schemes (LQG, pole placement), and more recently for general two-degree-of-freedom controllers (Li and Gevers, 199Oa, b, 1991; Liu and Skelton, 1990; Liu et al., 1992; Williamson and Kadiman, 1989). The last three references provide an optimal FWL-LQG design (which includes an optimal realization in the design process). Liu and Skelton (1990) and Liu et al. (1992) provide an optimal approach, while Williamson and Kadiman (1989) provide a suboptimal approach.

A survey of these results can be found in

367

368 A. G. Madievski et al.

Gevers and Li (1993). The methods essentially differ in the choice of performance measure (either roundoff error propagation or transfer function sensitivity) and in the norms used to evaluate this performance degradation. In Gevers and Li (1993) a synthetic measure of performance degradation of a closed-loop system, incorporating both roundoff errors and coefficient errors, was minimized with respect to all compensator realizations. The results on closed-loop sensitivity minimization in the above references all pertain to sensitivity measures of the closed-loop transfer function with respect to controller parameter errors. In Li and Gevers (1993) a weighted sensitivity measure of the closed-loop poles with respect to controller parameter errors is minimized.

The common feature of all these optimal controller realization results is that the system to be controller is assumed to be described by a discrete-time transfer function H(z). In most practical applications a digital controller is used to control a continuous-time plant, using both a sampler and a hold device.

Any optimization using solely a discrete-time transfer function of the closed loop neglects the intersample system behaviour and particularly intersample ripple. The novel contribution of this paper is to pose and solve a discrete-time compensator realization problem for a continuous-discrete closed-loop system, in which the digital controller acts on the continuous-time plant via a zero-order hold device, and in which the tracking error of the continuous system is passed through an antialiasing filter and then sampled. With this continuous-discrete set-up, the performance measure involves, of necessity, a hybrid operator: it is a measure of the sensitivity of the closed-loop input-output operator to the parameters of the compensator realization.

The outline of this paper is as follows: in Section 2 we establish the definitions of sensitivity functions (operators) and the 2. sensitivity measure of a closed-loop system. In Section 3 we study the finite-word-length optimal realization minimizing a measure of the sen- sitivity of the closed-loop operation with respect to controller coefficient errors. (No claim is made about FWL roundoff noise effects.) The existence and uniqueness of an optimal solution are established. A recursive algorithm for obtaining the optimal solution is given. A two-step procedure (fast sampling followed by blocking) that allows consideration of inter- sample behaviour of a closed-loop system is described and studied in Section 4. Two numerical examples to confirm theoretical results

are given in Section 5, followed by some concluding remarks in Section 6.

2. SENSITIVITY MEASURE OF A REALIZATION

First consider a discrete linear time-invariant multi-input, multi-ouput controller having a transfer function K(z) that can be expressed in terms of matrices A, B, C and D of a minimal state-space realization as follows:

K(z) = C(& - A)-B + D, (1)

where A E [WRxR, B E [WRxL, C E IWMxR, D E IWMx. and K E CMxL. Clearly, if the matrices A, B, C and D satisfy (1) then, for any similarity transformation T, the matrices T-AT, T-B, CT and D also satisfy (1). This means that there exist an infinite number of representations of the system. All these representations are equivalent insofar as they yield the same transfer function. However, different realizations have different numerical properties such as sensitivity to coefficient errors and propagation of signal roundoff errors. This means that in the finite-precision case all these realizations are no longer equivalent. In practice it is impossible to realize the matrices A, B, C and D exactly owing to finite-word-length (FWL) constraints. As a result, the transfer function given by (1) and the transfer function with the matrices A, B, C and D replaced by their FWL versions are different. Since different FWL realizations have different sensitivities, our task is to search for those realizations that minimize the sensitivity in some appropriate measure reflecting the overall control objective.

In order to define such a measure, we shall use the derivatives of elements of the controller transfer function matrix at an arbitrary but fixed value of z with respect to the elements of the matrices A, B and C of the realization:

where a, b, c, k, g and f are elements of the matrices A, B, C, K, G and F, respectively, with

C = C(& -A)- E cMxR, (34

F = (& - A)-B E cRxL, (3b)

1, i = 1,2, . . , L, m,j=l,2 ,..., M, r, 9 = 1, 2,. , R and 6 is the Kronecker delta. Note that the matrix D is coordinate-independent and

Sampled-data controllers and FWL sensitivity 369

has nothing to do with the optimal realization time-varying operator 2 with associated causal problem. impulse response 2?(t, s), such that

Our major goal is to find the optimal implementation of the controller for achieving the best performance of the closed-loop system where the controller is implemented with FWL. Best performance can mean many things. As made more precise below, we shall consider the accuracy of implementing the input-output operator for the closed loop.

I I

Y(t) = gtt, s&(s) b,