INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING, VOL.

11, 641—648 (1997)

ADAPTIVE PARAMETER ESTIMATORS FOR A CLASS OF

NON-LINEAR SYSTEMS

Y. D. SONG
Department of Electrical Engineering, North Carolina A&T State University, Greensboro,
NC 27411, U.S.A.

SUMMARY
Adaptive algorithms for parameter estimation of a class of non-linear systems are developed in this paper.
The salient feature of these algorithms is that there is no need to solve any Lyapunov equation, which may
prove useful in dealing with high-dimensional systems. Since the algorithms are constructed by using the
accessible system signals, the derivatives of these signals are not required. It is shown that the presented
algorithms include the existing method as a special case. A numerical example is included to demonstrate the
method. ( 1997 by John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process., 11, 641—648 (1997)

No. of Figures: 3 No. of Tables: 0 No. of References: 9

Key words: adaptive estimation; non-linear system; Lyapunov equation

1. PROBLEM STATEMENT
Practical systems may contain parameters that are not known a priori. To analyse, monitor and
control such systems, it is necessary to develop a strategy to estimate the unknown parameters.
This represents an interesting area in system identification and has been the subject of research in
system engineering. There has been much progress in linear system estimation (see References 1—3
for a brief review). In this work the problem of parameter estimation of a class of non-linear
systems is investigated and an adaptive method for parameter estimation is presented. This
method has its origin in Reference 1, where new algorithms for estimation of linear time-invariant
systems are described. These algorithms are further simplified in Reference 4. The key idea behind
the work reported therein is that instead of using the Lyapunov function candidate of the form

»"eTPe#tr ('T'#(T()

a modified Lyapunov function candidate

»"eTPe#tr('TP'#(TP()

¹his paper was recommended for publication by editor M. J. Grimble

Correspondence to: Y. D. Song, Department of Electrical Engineering, North Carolina A&T State University,
Greensboro, NC 27411, U.S.A.

CCC 0890—6327/97/070641—08$17.50 Received 9 April 1993

( 1997 by John Wiley & Sons, Ltd. Accepted 30 December 1995
642 Y. D. SONG

is utilized in addressing the stability, where P is the solution of the Lyapunov equation. Such
a consideration removes P from the estimation algorithms, which simplifies the calculation
because one does not need to solve the Lyapunov equation. This represents an attractive feature,
especially when the system is of high dimension.
It is the purpose of this paper to apply the above idea in the construction of adaptive
algorithms for parameter estimation of a class of non-linear systems. The system considered is
described by

xR "f (x, u)"At (x)#B/(u)#Cm(x, u), x(t )"x (1)

0 0
where x3Rn is the system state vector, x 3Rn is the initial condition, u3Rm is the control input
0
vector, A3R n]n, B3R n]m and C3R n]r are system parameter matrices and t(x)3Rn, /(u)3Rm
and m(x, u)3Rr are known non-linear and bounded vector functions of x and/or u. It is assumed
that the states of the system are physically measurable and the system is BIBS-stable.5
The problem to be addressed here is to find appropriate algorithms that adaptively estimate the
system parameter matrices A, B and C without explicit use of the derivatives of the accessible
signals. The results reported in this paper are motivated in part by the early work of Song and
Gao6 on robotic systems. They are also closely related to the work of Rimon and Narendra4 on
adaptive estimation of linear systems. In fact, the results given in this paper can be viewed as
a natural extension of the work in References 4 and 6.

2. ADAPTIVE ALGORITHMS FOR PARAMETER ESTIMATION

We construct the following dynamic model for the given non-linear system (1) by making use of
the estimated matrices AK , BK and CK and the accessible signals x and u:

xQL "A xL #U(AK , BK , CK , x, u) (2)

.
]
where A. 3Rn n is a Hurwitz matrix chosen arbitrarily and U(AK , BK , CK , x, u) is a generalized input
defined by
U(AK , BK , CK , x)"AK t(x)#BK /(u)#CK m(x, u)!A x (3)
.
The corresponding block diagram is shown in Figure 1. To make the model (2) implementable, we
need to develop feasible estimation algorithms for AK , BK and CK . To this end we define

e"xL !x, '"AK !A, ("BK !B, !"CK !C.

Consequently, the following error model is obtained:

eR "xLQ !xR "A e#'t(x)#(/(u)#!m(x, u) (4)

.
Based on this error model, the following algorithms are constructed.

Algorithm I
Consider the estimate model (2), (3). If the algorithm

AK Q "!etT(x), BK Q "!e/T(u), CK Q "!emT(x, u)

Int. J. Adapt. Control Signal Process., 11, 641—648 (1997) ( 1997 by John Wiley & Sons, Ltd.
 ADAPTIVE PARAMETER ESTIMATORS 643

Figure 1. Adaptive parameter estimation block diagram

is used to update AK , BK and CK , then global uniform stability is ensured and

lim e(t)"0
tPR
Furthermore,
lim AK "A, lim BK "B, lim CK "C
tPR tPR tPR

provided that [tT, /T, mT]T is persistently exciting.

Proof. Choose the Lyapunov function candidate

»"eTPe#tr ('TP')#tr ((TP()#tr(!TP!)
where P is the solution of the Lyapunov equation
AT. P#PA "!Q(0 (5)
.
Note that since A is a Hurwitz matrix, it is guaranteed that for any given Q"QT'0 there exists
.
a unique symmetric and positive definite matrix P in (5).7 Taking the derivative of » w.r.t. time
along the vector field defined by (4) gives
»Q "!eTQe#2eTP['t(x)#(/(u)#!m(x, u)]#2tr('TP'0 )#2tr ((TP(0 )#2tr (!TP!0 )
(6)
By using the matrix trace properties
tr(AB)"tr (BA), tr(AT)"tr(A), a"tr(a), a is a scalar

( 1997 by John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process., 11, 641—648 (1997)
644 Y. D. SONG

we may rewrite (6) as6

»Q "!eTQe#2tr M'TP['0#etT(x)]N#2tr M(TP[(0#e/T(u)]N#2tr M!TP[!0#emT(x, u)]N
(7)
In view of Algorithm I, (7) reduces to
»Q "!eTQe)0
which implies that

P P
t t
»(t)"»(0)! eTQ e dq)»(0)!j (Q) EeE2 dq ∀t*0
.*/
0 0
Therefore

P
»(0) t
EeE2 dq) (R
j.*/ (Q)
0
where j.*/ (Q) denotes the smallest eigenvalue of Q, which is obviously positive because Q is
symmetric and positive definite. The above relations infer that e3¸ W¸ , '3¸ , (3¸ and
2 = = =
!3¸ . By (4), eR is bounded, which ensures that e is uniformly continuous, hence it follows that
=
e(t)P0 as time approaches infinity. With the condition of persistent excitation the convergence
of the estimated parameter to the true value can be addressed as in Reference 8. K

Remarks
1. The significance of the proposed algorithms is that there is no need to solve the Lyapunov
equation since P is not explicitly used in the algorithms. All that is required is the fact that
such a P exists.
2. The result presented herein is an extension of that reported in Reference 4. In fact, when
applied to a linear time-invariant system
xR "Ax#Bu
where A and B are unknown matrices, we have /(x)"x, t(u)"u and m(x, u)"0,
Algorithm I reduces to
AK Q "!exT, BK Q "!euT
and the estimate model (3) takes the form
xLQ "A xL #(AK !A ) x#BK u
. .
which corresponds to the result described in Reference 4.

Algorithm II
Assume the same conditions as in Algorithm I. Use the following algorithm to estimate A,
B and C:
d
AK Q "!etT(x)! etT(x)
dt

Int. J. Adapt. Control Signal Process., 11, 641—648 (1997) ( 1997 by John Wiley & Sons, Ltd.
 ADAPTIVE PARAMETER ESTIMATORS 645

d
BK Q "!e/T(u)! e/T(u)
dt
d
CK Q "!emT(x, u)! emT(x, u)
dt
Then asymptotically stable estimation is ensured.

Proof. To prove this result, define

A"AK !A#etT"'#etT
B"BK !B#e/T"(#e/T
C"CK !C#emT"!#emT
Consider the Lyapunov function candidate
»"eTPe#tr(ATPA)#tr (BTPB)#tr (CTPC) (8)
It can be shown that

A B
d
»Q "!eTQe#2eTP['t#(/#!m)#2tr('#etT) P AK Q # etT
dt

A B A B
d d
#2tr((#e/T) P BK Q # e/T #2tr(!#emT) P CK Q # emT
dt dt

C A BD C A BD
d d
"!eTQe#2tr 'TP AK Q # etT#etT #2tr etTP AK Q # etT
dt dt

C A BD C A BD
d d
#2tr (TP BK Q # e/T#e/T #2tr e/TP BK Q # e/T
dt dt

C A BD C A BD
d d
#2tr !TP CK Q # emT#emT #2tr emTP CK Q # emT (9)
dt dt
Inserting Algorithm II into (9) leads to
»Q "!eTQe!2g !2g !2g
1 2 3
where
g "tr [(etT) P(etT)]*0
1
g "tr [(e/T) P(e/T)]*0
2
g "tr [(emT) P(emT)]*0
3
Using the same argument as before, the result is established. K

Remarks
1. It is seen that in addition to the integral term, Algorithm II contains a proportional term. As
discussed in Reference 9, this PI estimate algorithm gives a better convergence property than
Algorithm I. This point is also confirmed by simulation. Nevertheless, Algorithm I is slightly
simpler than Algorithm II.

( 1997 by John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process., 11, 641—648 (1997)
646 Y. D. SONG

2. It can be readily verified that stable estimation is still ensured if AK , BK and CK are computed via
d
AK Q "!c etT (x)!b etT (x)
1 1 dt

d
BK Q "!c e/T (u)!b e/T (u)
2 2 dt

d
CK Q "!c emT (x, u)!b emT (x, u)
3 3 dt
where c '0 and b *0 are scalars chosen by the designer. Different choices of these
i i
parameters lead to different properties of convergence.

3. A NUMERICAL EXAMPLE
To verify the validity of the estimation algorithms described in Section 2, a simulation study was
performed for the simple non-linear system
xR "!x#at(x), x "0
0
where
t(x)"3#sin (x3)#xe !x2

Figure 2. Adaptive parameter estimation (Algorithm I)

Int. J. Adapt. Control Signal Process., 11, 641—648 (1997) ( 1997 by John Wiley & Sons, Ltd.
 ADAPTIVE PARAMETER ESTIMATORS 647

Figure 3. Adaptive parameter estimation (Algorithm II)

and a is an unknown parameter to be estimated. It is seen that the system is BIBS-stable.

Therefore the parameter estimation algorithms proposed in Section 2 are applicable to the
system. In this example both Algorithms I and II were used in the simulation and the control
input u is set zero (open-loop estimation).

Algorithm I
xQL "a (xL !x)!x#aL t(x), xL (0)"0
.

P
t
aL "aL (0)!c (xL !x) t(x) dr
1
0
The design parameters are aL (0)"0, c "3·5, a "!3.
1 .
Algorithm II
xQL "a (xL !x)!x#aL t(x), xL (0)"0
.

P
t
aL "aL (0)!b (xL !x)t(x)!c (xL !x) t (x) dq
1 1
0
The design parameters are aL (0)"0, c "3·5, b "1, a "!3.
1 1 .
( 1997 by John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process., 11, 641—648 (1997)
648 Y. D. SONG

Both Algorithms I and II were tested. The results by using these algorithms are satisfactory. In
the simulation the true value of the system parameter is a"1. As predicted by the theory,
although both algorithms give asymptotic estimation of the unknown parameter, Algorithm II
exhibits a better convergence property in that aL converges to a in less than 3 s with Algorithm II
(Figure 3) instead of 5 s with Algorithm I (Figure 2).

4. CONCLUSIONS
Adaptive algorithms for parameter estimation of a class of non-linear systems are introduced in
this paper. The basic procedure to implement these algorithms is to choose a Hurwitz matrix A
.
and set up the estimation model (2), (3). There is no need to solve any Lyapunov equation. These
algorithms are evaluated by a numerical example.

ACKNOWLEDGEMENTS

Support from the NREL HBCU-PPWER project is gratefully acknowledged.

REFERENCES
1. Narendra, K. S. and A. M. Annaswamy, Stable Adaptive Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989.
2. Landau, I. D., ‘Estimation of the real positivity condition in the design of parallel MRAS’, IEEE ¹rans. Automatic
Control, AC-23, (1978).
3. Ljung, L., System Identification: ¹heory for ºser, Prentice-Hall, Englewood Cliffs, NJ, 1987.
4. Rimon, E. and K. S. Narendra, ‘A new adaptive estimator for linear systems’, IEEE ¹rans. Automatic Control, AC-37,
(1992).
5. Byrnes, C. I. and A. Isidori, ‘Local stabilization of minimum-phase nonlinear systems’, Syst. Control ¸ett. 11, (1988).
6. Song, Y. D. and W. B. Gao, ‘Comment on adaptive control of robot manipulators’, Int. J. Control, 49, (1989).
7. Chen, C. T., Introduction to ¸inear System ¹heory, Holt, Rinehart and Winston, New York, 1970.
8. Narendra, K. S. and P. Kudva, ‘Stable adaptive schemes for system identification and control—Part I’, IEEE ¹rans.
Syst. Man, Cyber. SMC-4, (1974).
9. Landau, I. D., Adaptive Control: ¹he Model Reference Approach, Marcel Dekker, New York, 1979.

Int. J. Adapt. Control Signal Process., 11, 641—648 (1997) ( 1997 by John Wiley & Sons, Ltd.