Professional Documents
Culture Documents
Abstract
The choice of input signal in an identification experiment has a significant bearing on the final result. At
one extreme, an input can result in violent system disturbances and the subsequent loss of product or damage
to the system. At the other extreme, an input can be so conservative as to yield little or no information
about the system's dynamic behaviour. The optimal input signal achieves a compromise between these
extremes. This paper describes a design procedure for synthetising optimal input signals. A state-space
description of the system is used, and the optimal input is chosen to minimise the variance of the parameter
and state estimates. The method applies to both linear and nonlinear systems and the case of both white and
coloured output noise has been considered. The optimal input signals are shown to result in a significant
improvement in the identification accuracy.
r=
(X2
Introduction
x(k
922
= f{x(k), u(k),p, k)
k=
...,N-l
*(D=a
(1)
(2)
k=l,2,...,N
(3)
k2)
k2=
3
Paper 6445 C, received 11th January 1971
Dr. Goodwin is with the Department of Computing & Control,
Imperial College, Exhibition Road, London SW7, England
Model equations
1,2,..., N
(4)
Minimum-variance estimation
x(k) ~ x(k) +
- P] k = 1, 2 , . . . , N
k=l,2,...,N
(7)
(8)
P=P
(9)
k=\
{ 2 Xj(.kx)H^{.kx)W{kx, k)\
(10)
2
where / is an r x r identity matrix and is a Kronecker delta
function.
It is well known 9 that the covariance of the estimate p is
given by
covP == M~l
(12)
. . . .
(14)
A/ = 2 gr(k)&u(k)
-s - 1
V{x{k), u{k), k)
J=
We) = { S
Performance index
(6)
+ 1) =
(15)
(13)
Examples
(16)
(18)
(19)
923
.Pi.
' 0"
1
(20)
1_
Coloured noise
To illustrate the dependence of the optimal input on
the nature of the noise, a final example is presented for which
the noise is assumed to have the autocorrelation shown in
Fig. 3. The input constraint \u\ < 1 was imposed and Fig. 4
White noise
u 0
Fig. 2
Optimal input and model response for lst-order system with white
noise and both state and input constraints
a Optimal input
b Model response
x 0
-1 Fig. 1
Optimal input and model response for lst-order system with white
noise and input constraint
a Optimal input
b Model response
Fig. 3
Coloured-noise autocorrelation function
State constraint
This example is similar to that in Section 6.1, save that
the additional state constraint |x| < 0-5 is imposed. This
constraint may be added to the performance index as follows:
k^
64
O0
10
-1
Fig. 4
Optimal input when the output noise of the system has the autocorrelation shown in Fig. 3
+ p4u3(k)+ p5x3(k)}
x2(k + 1) = x2(k) + {p6
1-2
x3(k
0-8 -
= x4(k)
{/16 +
pxlxx(k)
+ pl9u2(k)+
where xv =
x2 =
x3 =
x4 =
, =
u2 =
w3 =
0-4
0-42 h
0-34 h
R{kuk2) =
49
0
0
0
0
0
1 0
0 36
0
0
0'
0
hkxk2
0
256
0-26 r-
0-18
(23)
0-8
The estimated parameter values were found to have magnitudes differing by factors of 104. Thus, to make a comparison
between their variances meaningful, the parameters were
rescaled to unity. The sum of the variances of the scaled parameters was calculated from the data-logging record, and was
found to be 7-9.
In anticipation of the design of an optimal input signal, a
study was made of the restrictions on the system operation.
The important constraints were found to be
(i)
(ii)
(iii)
(iv)
(v)
_l_
a limit on fuel
flow
0-4 < ! < 1 -2 *J input
a limit on steam flow 0-2 < u2 < 0-42 > cona limit on feedwater flow 0 < u3 < 1 -4 J straints
a safety limit on steam pressure 110 < x, < 210\
a safety limit on drum water level 3 < x2 < 3 /
state constraints
0-4
200
400
600
t,s
800
1000
Fig. 5
Optimal input for model of steam generator
a Fuel mass-flow rate
b Steam mass-flow rate
c Feedwater mass-flow rate
Suboptimal inputs, based on the true optimal inputs discussed in this paper, may lead to even simpler methods of
implementation, but this possibility has not yet been fully
explored.
9
>
where ^[M"" 1 ] is the trace of the 21 x 21 parameter covariance matrix, and the second term has been included to
account for the state constraints.
The optimal input is illustrated in Fig. 5, and the optimal
trace of the parameter covariance matrix was found to be
1-3. The optimal value of 1-3 is one-sixth of the sum of
variances obtained with the data-logging input. The improvement is especially significant, as the data-logging input had
been carefully selected (manually) to excite fully the plant
dynamics.
Methods of implementation
Conclusions
A design procedure has been described for the synthesis of the input signal which leads to minimum variance
estimates of the parameters and states in models of nonlinear
dynamic systems. The allowable set of input signals has been
restricted by amplitude constraints on the inputs and states of
the system. The case of both white and coloured noise has
been considered, and several examples have been used to
illustrate the resultant optimal input signals. The examples
have shown that an order-of-magnitude improvement in the
parameter and state variances can result from the use of the
optimal input signal. The design procedure has been shown
to be tractable for large-scale systems by an application to the
model of an industrial-scale steam generator.
10
Acknowledgments
11
References
2
k=\
2 vj
k=l
2
A:=l
. . . .
(27)
We can now use 'summation by parts' to remove the dependence of eqn. 27 on Ax(k + 1) and AX$(k + 1). The appropriate summation formulas are
k=i
optimal control problems with a state variable inequality constraint', IEEE Trans., 1969, AC-14, pp. 457-464
12 PAGUREK, B., and WOODSIDE, c. M. : 'The conjugate gradient method
for optimal control problems with bounded control variables',
Automatica, 1968, 4, pp. 337-349
13
1 = 2
+ <r(A0, AXpiN + l)> - <r(0), AAp(l)> (29)
Note in eqns. 28 and 29 that the initial conditions Ax(l) and
AXg(l) are both zero, since the boundary conditions given in
eqns. 2 and 8 are fixed. Using this result and substituting
eqns. 28 and 29 into eqn. 27, leads to
- 2 2 { S <HHkW(k,k2)H(k2)X&(k2)M-2,AX&(k)y\
Appendix
12
+ 2{
2 {V
k=l
2^
+ V{x(k), u{k), k)
(25)
) + Fu(k)Au(k)}
k=l
k=\
), Fx(k)X&(k) +
- 2
2 <Fj(k)T(k),
k=l
- 1),
k=l
(30)
, Fx(k) X&(k)
. . . .
(26)
+ F&(k)>Au(k)
k=l
- <rT(A0,
k=\
l)Ax(k)
+ V{x(k), u(k), k}
Jc =
- 1) = Fj(k)Xk) + VxV(k)
\ Fx(k)X0)
k = N, N
(31)
(32)
2 =i
k = N,N- 1, . . . , 2 (33)
T(N) = 0
. (34)
-22 { 2
k=\ ^2=1
2{
926
k=l
(35)
k=\