Automatica, Vol. 22, No. 5, pp. 555 560, 1986 Printed in Great Britain.

0005 1098/86 $3.00 + 0.00 Pergamon Journals Ltd. 1986 International Federation of Automatic Control

Estimation and Direct Adaptive Control of Delaydifferential Systems*

ROLF JOHANSSONt

Model matching of 2-D systems and parameter identification are combined for adaptive control of delay-differential systems.
Key Words--Adaptive control; delays; discrete systems;estimation.

Abstract--The problem of adaptive prediction and direct adaptive control for systems described by delay-differential systems is considered. A model reference approach to direct adaptive control is shown to be feasible for a class of 2-D systems. A reparametrization of the 2-D system is given and the identification problem is converted into a linear estimation problem. Some necessary conditions for existence of solutions to the adaptive prediction and direct adaptive control problem are also given. INTRODUCTION THE PROBLEM of direct adaptive control and adaptive filtering for systems described by delay-differential equations is considered. This kind of system description is naturally obtained for processes with flows, transport delays, and refluxes, (Morse, 1976; Emre et al., 1982). A quite different application described by similar system equations is image processing systems. This latter application will however not be considered here. A model reference approach to direct adaptive control is shown to be feasible for a class of 2-D systems. A reparametrization of the 2-D system description is given and the identification problem is converted into a linear estimation problem. Some necessary conditions for existence of solutions to the adaptive prediction and direct adaptive control problem are also given. PRELIMINARIES Let R[z] denote the ring of polynomials in z with real coefficients. Here systems defined over R[zl,z2], i.e. the ring of polynomials with real coefficients in two indelerminates - - zl and z2, will *Received 8 August 1985; revised 24 February 1986. The original version of this paper was presented at the 7th IFAC/ IFORS Symposium on Identification and System Parameter Estimation which was held in York, England during July 1985. The published proceeding of this IFAC meetingmay be ordered from Pergamon Books Ltd, Headington Hill Hall, Oxford OX3 0BW, England. This paper was recommended for publication in revised form by Associate Editor P. J. Gawthrop under the direction of Editor P. C. Parks. ~'Lund Institute of Technology, Department of Automatic Control, Box 725, S 22007, Lund 7, Sweden. 555

also be considered. Such systems are for example continuous-time systems with delays and 2-D image processing filters. The systems may as well be described by polynomials in R(zl)[z2] or R[zl] [z2], i.e. polynomials in z2 whose coefficients are rational functions and polynomials in zl, respectively. For the theory of polynomials of several variables, see K a m e n and Rouchaleau (1983) or Van der Waerden (1970). The algebraic properties of R[Zl,Zz] are different from those of R[z]. The ring of polynomials R[z] is e.g. a Euclidean domain, while R[zl,zz] is not, see Shapiro (1975). Traditional methods to solve model matching problems in linear systems theory, see Astr6m and Wittenmark (1984) or Kailath (1980) are based on the Diophantine equation

AR + BS = P,

(1)

which for given A, B, PeR[z] has solutions R, S~R[z] if A and B are coprime. In the 2-D case there are solutions R, S~R(zl)[Zz] for A, B, P ~ R[zx] [zz] when A and B are coprime, i.e. when the Bezout identity exists, see Morf et al., (1977). There are also solutions R, SER[z1][z2] for coprime A, B6R(Zl)[Z2] and PeR[Zl][Z2], see Khargonekar and Sontag (1982) and theorem 2.11 of Emre et al. (1982). For further investigations on the coprimeness problem of 2-D polynomials, see Morf et al. (1977). PROBLEM FORMULATION The problem of direct adaptive control of 2-D systems will now be considered. The treatment is based on a model reference approach to direct adaptive control. A closer account is given, e.g. in ~str6m and Wittenmark (1984) or Landau (1979). The model reference approach is closely connected with the problem of exact model matching. Treatments of the 2-D case of model matching are found in e.g. Eising and Emre (1979) and Sebek (1983).

556

R. JOHANSSON locations of zeros. Introduce for this reason the following factorizations
B = boBIB2;BI(O,O ) = 1,B2(I, 1)= 1.

The problem is to find a continuous-time regulator described by S T u = ~ y + ~uc (2)

(7)

to control a continuous-time object described by the factorized input-output model B y = 5u,

The polynomial B~ is supposed to be such that all solutions s = si to the equation

1 s~aa'

= 0

(8)

(3) have Re(sl)< 0 while B2 contains the remaining factors. The factor bo is a scalar, i.e, a pure constant gain. The polynomial B1 may be called a greatest causal and stable factor of B. The decomposition of/3 such that (7), (8) hold is always possible to do since one may always choose B1 = 1 and all remaining factors in B 2. The problem to find a B 2 of least possible complexity will not be considered here but is of some practical importance. The reason for making this factorization is to assure internal stability in the controlled system. The factor B2 will play the same role of system invariant as the relative degree, time delay etc. in 1-D model matching problems, see Pernebo (1981) and /kstr6m and Wittenmark (1984).

with input u, output y and command signal uc. The problem of existence of polynomial factorizations is treated in Khargonekar (1982). Here A, B, R, S, T are polynomials with unknown real coefficients in some operator(s). The operators may be, e.g. the differential operator p and some time delay operator. In this paper the following operators will be used zl a p+a

and

z2 = e-~P

(4)

for some given, positive constant a. The operator z~ is a low pass filter operator and z2 represents a time delay ~. These operators are chosen because they are causal, stable operators with finite gain. Pernebo (1980) has developed an algebraic systems theory partly based on such operators and Johansson (1983) has demonstrated their relevance for adaptive control. Notice that these operators commute in the context of this presentation, i.e. the input-output point of view of signal processing. Notice also that the transformation results in new parameters which are linear with respect to the
original parameters.

ASSUMPTIONS

The polynomials A and B are then

A number of restrictions on the class of control objects will now be given in order to assure solutions with good properties. Assume that (A1) boo = 0, i.e. there is no direct term from input to output. (A2) The polynomial degrees nia , n j a , niB , njb of A and/3 are all known. (A3) B2 is known. (A4) A and B are coprime i.e. 3 M , N ~ R ( z l , z2):
A M + B N = 1.

i=oj=o B(21,22) = Z bijZilZJ2 /=Oj=O (5)

for some finite orders niA , njA , nia , njB. The condition on aoo is imposed for the purpose of normation and is possible if the transfer function of (3) is both causal and proper. The transfer operator of (3) may also be factorized as

(A5) The time delay z is known. It should be said that the requirement to know B 2 is impractical except when B 2 may be reduced to pure powers of Zl and z2, i.e. low pass filters and time delays in series. The condition (A4) is formally restrictive since any two polynomials of R [ z l , z 2 ] have generically a common zero.

SOLUTION

A(z

z2)((t) =

u(t)
(6)

The closed loop system of (2) and (3) with timeinvariant A, B, R, S, and T gives the transfer operator as
BT Y - A R + B S uc,

y(t) = B(za, Zz)~(t)

for some internal state variable 4. The polynomial B may be further decomposed into a number of factors. The main concern is the

(9)

where R, S, T may be tuned to fulfil specifications

Delay-differential systems in terms of a reference model. This model should reflect the design trade-offs and is of the type y,. =
B2Bm

557

data vector @. The vector ~b thus contains filtered and delayed inputs u and outputs y. Then we have
u(t) = - 0T(t)

~,. ug

Am(0,0) = 1,

(10) with

(16)

where A,,, B,, are desired pole- and zero-polynomials and where B 2 is the part of the B-polynomial of the control object which should not be cancelled in the closed loop system (9).
P = A,.B~ T = toB,.; to -- 7-, Oo
1

0 = (r 1o... rij... Soo... Ski... to)T

(17)

and
=

(1 la-b)

(Bmuc)) +.

(18)

where bo is the constant gain introduced in (7). The model matching problem is then solved via the Diophantine equation
A R + B S = P.

The estimation problem of direct adaptive control is now to calculate 0 from input-output data.

(12)

It is necessary to obtain polynomials R and S from (12) in order to have solutions that are suitable for direct adaptive control. Such solutions do not always exist, see Example 2 below. When there are polynomial solutions R, S in two indeterminates, it is possible to express them as
R(za,z2) = ~
i=0/=0

PARAMETER ESTIMATION The parameters of the regulator may now be estimated from a linear model obtained by manipulations of the factorization (6) using (12) and (6), respectively.
y = B~ = B R A + SB B p ~---fi(Ru+Sy).

(19)

Assume that B 2 is known, cf. (A3), and define

iz 2~. rijz1 ~ roo

=1

(~ = B2u,)7 = B2yetc.

(20)

his

S ( z , , z 2 ) = ~, ~ sijz]z
i=0j=O

(13)

Introduce the data vector q~

(D : ((z1/~)... (z1z2/~)....~...

(zkzl2y)...
(21)

for some finite numbers n~R, njR, n~s, njs, which are determined in the process of solving (12). The model matching control law (2) is then expressed explicitly
as

-- (Amy))

niR
U ~ -i=lj=l niR ,=1 j=l

in accordance with standard notation of recursive estimation, see Ljung and S6derstr6m (1983). Then it is possible to rearrange (19) to the scalar product
~(t) = - OXp(t).

(22)

his

"~ so(z]z~2y) + to(Bmuc).

i=Oj=O

(14)

The data filters

z]zJ2 Vi,j >1 0

(15)

require for their implementation the causal operators of low pass filters and time delays and are therefore realizable. Let the control law (14) be formulated as the scalar product of the parameter vector 0 and the

The constant parameters 0 may now be estimated by any suitable method for identification of parameters of a linear model with known components of ~o. Recall that the estimates 0 of the parameters 0 in (22) are intended for use in the continuoustime regulator (14). The parameters may however be estimated by discrete-time methods. One natural choice is to sample corresponding t~ and ~o and to fit parameter estimates to the sampled data by recursive identification. For further details on recursive estimation, see Ljung and S6derstr6m (1983). Remark. The suggested methods in two operators for continuous-time systems may of course also be used for identification of "conventional" continuous-time systems by discrete-time methods. The

558

R. JOHANSSON A polynomial description of the control object in terms of the operators zlis obtained via
1 -Z~x
Z1

operator in which to formulate the parametric model for identification is then the low pass filter operator Zl. The choice of bandwidth of z~ and the sampling period for the estimation have certainly many interesting information-theoretical aspects but these will not be elaborated on here. This also opens possibilities of bridging the gap between the two traditions of continuous- and discrete-time identification. This is one method for on-line identification and recursive estimation of a continuous-time system by discrete-time methods while keeping the structure of the continuous-time system. The use of the low pass filters differs from that of traditional continuous-time methods, see Eykhoff (1974), where the filter assembly is a part of the hardware for estimation. All parameter estimation is however made by discrete-time recursive methods in the approach of this paper. ADAPTIVE FILTERING An adaptive filter or "predictor" is obtained as a spin-off from the solution of the direct adaptive control problem. An observer for the quantity PC is obtained as
Ru + Sy.

1 p+l

and

z2=e

~v

(28)

= - a ~ x + a222x nt- bu

(29)

y=x. This yields the transfer function y(t) =

bz 1

I -1- (a 1 -- l)z 1 -- azz1Z 2

u(t)

_ B(zl ,z 2) u(t). A(z~, z2)

(30)

It is here easy to see that A and B are coprime. The zero of B appears at z 1 =0, (31)

(23)

This is seen by the following argument.

where the polynomial A certainly is non-zero. The zero of (31) is---in terms of the differential o p e r a t o r - - a n infinite zero of degree 1 of the transfer function. The reference model

Ru + Sy = (RS) B ~ = P ~

(A)

(24)

9.5zl l = 2p + I 1 -- 0.5Z

B_ _A Z:~zl
1 --

(32)

A m

It is possible to carry through all of the arguments above with the purpose of deriving adaptive filters for PC with any P such that there is a solution to (12). There is in this case no need to factorize B into B 1 and B2 and assumptions on common zeros and prior knowledge are less restrictive.
Example 1. Consider the following delay-differential control object

has the same type of zero and it is therefore possible to find a solution to the pole placement problem with a regulator (26) which is causal and without derivatives. A solution with A,, = l-0.5z~ and B,, = 0.5 is obtained for R=I =So0 +S01Z2

{~(t) = --alx(t ) + a2x(t -- z) + bu(t) y(t) = x(t),

(25)

1 T = ~'0.5 = toB m.

(33)

with input u and output y. The time delay r is supposed to be known. The problem is to design an adaptive controller
S(zl,z2) . . T(z~,z2) . . ~-(~;~2)Y(t) + ~ u A t )

Input-matching estimation based on the linear model

(zlu) = -(Soo + SoGz)(zxy ) + to(Amy )

(34)

u(t) -

(26)

or shorter a(t) = with

~o = (SooSoltoF
-

such that the closed loop system matches the reference model
y,,(t) - 2p + 1 Uc(t).

0T~p"j

(27)

(35)

Delay-differential systems and ~o

=

559

The A- and B-polynomials now both vanish at ((z,y)(zzzly) -- (Amy))T.

Here it is possible to use identification methods for estimation of parameters of a linear model if the filtered input-output data (zlu), (zly), (zlz2y) are available. The following simulations show the results when (34) has been sampled with sampling period h = 1. See appendix for details. Recursive least-squares identification has been applied to estimate the parameters. Example 2. Consider the following delay-differential equation for a control object :t(t) = - a l x ( t ) + a2x(t - ~) + bu(t - r) (36) y(t} x(t). This equation differs from that of the previous example in the time delay of u. The polynomial representation now becomes bzzz2 y(t) = 1 + (al - 1)zl u(t).
-a2z1z2

For this reason it is not possible to solve the polynomial Diophantine equation AR + BS = P (39)

for an arbitrary P. It is necessary to include in P a factor with a zero at (38). The factor
(I + (a I - - l ) z l ) (40)

(37)

O.ao.6 i () o,t,
().2 0

would be suitable when it is stable and stably invertible, see Sebek (1983). This presents no particular problem in a pure model matching problem with a known process to control. It would then be possible to include (40) as a stable common factor of both the numerator and the denominator in the reference model. This trick cannot be used in the adaptive case since the zero (38) is not known to the designer. This lack of coprimeness therefore puts a limit to the applicability of model reference adaptive control to delay-differential systems. Simulations do however show that the performance is often also very good when there is no solution to the Diophantine equation. The "least squares solutions" show slower convergence but good tracking capabilities are still exhibited.
CONCLUSIONS

50

100

150

200
It has been possible to formally carry over many of the algebraic results from 1-D direct adaptive control. There is however a strong practical difference between the required prior information in the two cases. The question of coprimeness is also more complicated in the 2-D case. Any solution (zl,z2) = (a,b) to the equation A(q,z2) = 0 B(z1, z2) -- 0

o
2 (C)

,oo

~6o

3oo

4OO

(41)

-2
0 100 200 300 400

puts the constraint on P that P(a, b) must be zero in order to assure the existence of a solution to (12). In contrast to the 1-D case it is not in general possible to get rid of the common zero by polezero cancellation between A and B, see example 2 and Morf et al. (1977).
Acknowledgement--This work was supported by the Swedish
Board of Technical Development under contract 82-8430. REFERENCES Astr6m, K. J. and B. Wittenmark (1984). Computer-controlled Systems-- Theory and Design. Prentice-Hall, Englewood Cliffs, New Jersey.

loo

200

300

z.oo

FIG. 1. Simulations. (a) Impulse response of open-loop system with a 1 = 1, a: = 1, b = 1 and z = 10; (b)parameters estimates 01, 02, 0 s from recursive LS identification with sampling period h = 1 s and ~ = 10; (c)input u; (d)output y and reference signal
Ym"

560

R. JOHANSSON
a (partial) updated survey. Proc. I F A C 1981, Kyoto, pp. 325 330. Van Der Waerden, B. L. (1970). Algebra. Frederick Ungar. New York. APPENDIX
Details of the simulations

Eising, R. and E. Emre (1979). Exact model matching of 2-D systems. I E E E Trans. Aut. Control, AC-24, 134-135. Emre, E. (1983). On adaptive control/observers/parameter estimation of multi-variable systems. Proc. 22nd 1EEE CDC, 1380-1381. Emre, E. and P. P. Khargonekar (1982). Regulation of split linear systems over rings: coefficient assignments and observers. I E E E Trans. Aut. Control, AC-27, 104 113. Emre, E., A. B. Ozguler and P. P. Khargonekar (1982). Systems over rings: output regulation and tracking. Proc. 21st I E E E CDC, 408--413. Eykhoff, P. (1974). Systems ldent!lication. John Wiley, London. Johansson, R. (1983). Multivariable Adaptive Control. Lurid Inst. Tech., Lurid, Sweden. Jury, E. I., V. R. Kolavennu and B. D. O. Anderson (1977). Stabilization of certain two-dimensional recursive digital filters. Proe. IEEE, 65, 887-892. Kailath, T. (1980). Linear Systems. Prentice-Hall, Englewood Cliffs, New Jersey. Kamen, E. W. and Y. Rouchaleau (1983). Theory and Applications qf Linear Systems Over Rings and Algebras. Macmillan, New York. P. P. Khargonekar, (1982). On matrix fraction representations for linear systems over commutative rings. S I A M J. Control Opt. 20, 172 197. Khargonekar, P. P. and E. D. Sontag (1982). On the relation between stable matrix fraction factorizations and regulable realizations of linear systems over rings. I E E E Trans. Aut. Control, AC-27, 627-628. Landau, Y. D. (1979). Adaptive Control The ModeI Reference Approach. Marcel Dekker, New York. Lee, E. B. and S. H. Zak (1983). Smith forms over R[zl,z2]. I E E E Trans. Aut. Control, AC-28, 115 118. Ljung, L. and T. S6derstr6m (1983). Theory and Practice of Recursive Identification. MIT Press, Cambridge, Massachusetts. Morf, M., B. C. L6vy and S. Y. Kung (1977). New results in 2-D systems theory, Part 1: 2-D polynomial matrices, factorization, and coprimeness. Proc. IEEE, 65, 861-872. Morse, A. S. (1976). Ring models for delay-differential systems. Automatica, 12, 529-531. Pernebo, L. (1981). Algebraic control theory for linear multivariable systems. Parts I-II. I E E E Trans. Aut. Control, AC26, 171 193. Sebek, M. (1983). 2-D exact model matching. I E E E Trans. Aut. Control, AC-28, 215 217. Sebek, M. (1985). On 2-D pole placement. 1EEE Trans. Aut. Control, AC-30, 819-822. Shapiro, L. (1975). Introduction to Abstract Algebra. McGraw Hill, New York. Sontag, E. D. (1981). Linear systems over commutative rings

Example 1 involves both continuous-time and discrete-time simulation. The sampling interval is h = I and the time variables are called t and k, respectively. The relation is k < t < k + h between the samples. The example requires simulation of the following equations.
Object

continuous time
dx(t)/dt = - a l x l t t + a2x(t

r) + bu(t)

y(t) - x(t) x ( 0 ) - 1.
Controller

continuous time

u(t) = -~oo(k)y(t) - gox(k)y(t - r) + 0.5fo(k)u~(t ).

Delay z = 10 assumed to be known

uc(t) = sign(sin(0.1 t)). ldent(fication model (~(t) = sooy(t)

continuous time

s01y(t - z) + to(ylt ) - 0.59(t))

~(t) + u(t)

with
d(a(t))/dt = d(~)(t))/dt = - ,)(t) + y(t). Identification model

sampled at times t = 0, 1, 2..... k....

u(k) = Soof~(k) - Solf~(k - r) + to(y(k ) - 0.53~(k))

fi(k) = -good(k) - ~0D~(k -~:) + to(y(k ) - 0.5~(k))

e(k) = ~(k) LSestimation (~(k).

discretetime

() ~-~ (?'OOSO 1 to) v; (~ = ('~ooSo 1 f0) T ~p(k) = (~(k)f~(k - r) - (y(k) - 0.537(k)))x O(k) - O(k - 1) + P(k)~o(k)e,(k) P(k) l ( p( k _ . , P l k - 1,~o(k)q)'(k)P(k ~ 1!~ l~ -2 ~ o T ( k ) p ( k _ l)~o(k) ]"