System Identification using Control Theory.

T.J.Moir
School of Engineering
AUT University
Auckland, New-Zealand
[email protected]

Abstract—This paper considers preliminary results for a novel where a signal (normally a step) is driven into an unknown
approach to the identification of finite-impulse response (FIR) or system and the same signal is driven into a convolution which
autoregressive (AR) models. Whereas traditional methods have occurs in the feedback path of deconvolution loop.
employed a cost function such as least-squares or steepest
descent, the new method uses deconvolution to split the unknown uk
b( z 1 )
sk + ek
K
Ŵk
parameters from the regressors. This is achieved by using a ( z 1 ) - 1  z 1
I

convolution in the feedback path of a high-gain control-system. yk

Keywords—system identification; feedback; autoregressive

modelling
* z1I
Ŵk1
Introduction
The method of least-squares has been around since the time Figure 1. System identification using feedback.
of Guass and in its various derivative forms has been the
cornerstone for system identification during the 20th century. The unknown system is considered to be additive noise-free.
For example, for identification of finite-impulse response (FIR) A general scalar system output is given in scalar terms as a
systems for channel equalization, the least-mean squares pole-zero model
(LMS) approach which uses steepest descent has proved one of b(z-1 )
the most popular in signal processing applications due to its sk  G(z-1 )u k  uk (1)
computational ease of implementation[1]. For autoregressive a(z-1 )
(AR or all-pole) or autoregressive moving-average (ARMA or -1
pole-zero) identification the recursive-least squares (RLS) where z is the z-transform backwards shift operator defined
method[2] has proved popular and has a much faster rate of for a scalar discrete signal at sample instant k as
convergence than that of LMS at the expense of much y k-1 = z-1y k . The polynomials are given by
2
increased computation (approximately O(n ) versus 1 2  na
a(z ) = 1+ a1z  a 2 z  ...  a na z
-1
which has all of
O(2n) multiplications for n parameters). Furthermore the its roots within the unit circle of the z-plane and
RLS algorithm can also lead to instability over prolonged
periods and has poor tracking ability as compared with LMS. b(z -1 ) = b 0 + b1z 1  b 2 z 2  ...  b nb z  nb which must be
Certain ad-hoc methods using forgetting factors are minimum phase. Both the input and output of (1) are assumed
commonplace in the literature[3] when RLS is used in an to be measureable.
attempt to estimate time-varying system parameters. Such This method cannot identify a pole-zero model, but can
estimation algorithms and their variants have application in a identify either an all-zero or all-pole model of the system.
wide range of industries including wireless communication[4], Hence we consider two possible cases when i) na=0,
active noise-cancellation[5],hearing-aids[6] and antenna
arrays[7]. A condition for convergence of these types of a(z -1 ) = 1 and sk  b(z-1 )u k and ii) nb=0 so that
algorithms is that the input be persistently exciting. The very b
nature of the problem normally ensures this is the case due to sk  0-1 u k
the random signals involved. The following algorithm has a(z )
similar complexity to LMS yet can be used to estimate AR
models as well and has a similarly fast convergence rate to that Case 1: FIR Model sk  b(z-1 )u k
of RLS. It does not require a persistently exciting input A vector output is obtained from the block convolution “*” of
however. the system input and the impulse response vector
sk  W*uk (2)
I. FEEDBACK DECONVOLUTION
By using convolution in the feedback path of a system with
high gain we achieve the inverse property of convolution,
which is known as deconvolution. Fig 1 illustrates the method
where the weight and input vectors are defined respectively as inferior. In fact this is a classical approach to deconvolution
W  [b0, b1...b nb ] u k  [u k, u k 1...u k nb ] and
T T using feedback in a similar fashion that optimal control is to
classical controllers.
s k  [s k,s k 1...s k  nb ]T giving b0
Case 2: ALL-Pole Model sk  uk
nb
a(z-1 )
W * u k   w ju k  j
j 0
(3) Without losing generality we have a(0)  1 , since if this
The control-loop forward-path consists of a matrix of parameter is non-zero then it can be absorbed into the
integrators with gain K. Its error vector is found from Figure 1 numerator parameter b0 . The system output can be written as
to be
ˆ sk  a1sk 1  a 2sk 2 ...  a na sk na + b0 u k
ek =s k -y k = s k -u k *W k (4)
which for na samples has a matrix form

Where * in (4) denotes block convolution of two vectors

s k  T2 Wk (7)
uk and Ŵk and the latter vector represents the estimate of
the true weight vector Wk . The convolution within (4) is of  s0   u 0 0 0 . 0   b0 
0
two vectors giving a new vector s  u s0 0 0 0 0   a1 
y k  [yk , yk 1...yk nb ]T whose elements consist of  1  1
 s2   u 2 s1 s0 0 0 0   a 2 
nb
   
yk   w
ˆ ju k  j , k  0,1...nb  .   . s2 s1 . 0 0 . 
j 0
 .   . . . . . 0 . 
Although the unknown system is itself scalar in nature, the     
feedback estimator is multivariable. Assume that the closed- s na   u na s na-1 s na-2 . . s 0   a na 
loop multivariable system is stable. Then the error for large
(8a)
gain K becomes (via the usual action of negative feedback)
In other stochastic approaches (eg the Yule-Walker equations)
ek  0 so that u k *W ˆ  s and hence W
k k
ˆ W
k we would instead have a correlation matrix. The eigen-values
as k   . If we assume the closed-loop system is always of such a matrix are given by the diagonal elements.
Herein we see why this method is successful. Due to the
stable then for a time-varying system, the control-loop output
nature of convolution, the weights and inputs are swapped and
must track any such changes. We can for future reference also
provided the closed-loop system can be made stable then the
write the system model (2) as
integrators must converge to the true values of the parameters.
The method is aided by the form of the matrix T in (8) (or (6)
s k  T1 Wk (5) ). With any other form of matrix this would not be achievable
where and this is also why we cannot apply the current method to
pole-zero (ARMA) models. Moreover, if there are any pure
 s0   u 0 0 0 . 0   b0 
0
s  u delays in the transfer-function, e.g. b0  0 the method will
 1  1 u0 0 0 0 0   b1  fail. To overcome this problem we need to delete the first
 s2   u 2 u1 u0 0 0 0   b2  data-point s0 and treat the second data point as the first, as in
   
 .   . u2 u1 . 0 0  . 
 .   . . . . . 0  .   s1   u1 0 0 . 0   b1 
0
     s  u s1 0 0 0 0   a1 
s nb   u nb u nb-1 u nb-2 . . u 0   b nb   2  2
(6)  s3   u 3 s2 s1 0 0 0   a 2 
The lower triangular matrix T1 of degree na+1 is known as a
   
 .   . s3 s2 . 0 0  . 
convolution matrix.  .   . . . . . 0  . 
Furthermore, if s k and uk are reversed as inputs to the     
s na   u na s na-1 s na-2 . . s1   a na 
control-system, then the inverse of the system impulse
response will be estimated instead, provided of course enough (8b)
weights are allocated. The above algorithm is not optimal in where b1 is now estimated instead of b0 . This idea can be
any least-squares sense since no cost function is minimised.
extended for any number of steps delay.
This does not however imply the method is in any way
 II. STABILITY AND CONVERGENCE Substitute (13) into (12) and re-arranging, gives the
We start by replacing the convolution operation in the multivariable transfer function matrix relating the estimated
feedback path of Figure 1 with an alternative yet equivalent vector with the true vector as input.
operation. ˆ (z -1 )  [I  (I  KTz-1 ]-1 KTW(z-1 )
W
(14)
y k  Ti W
ˆ
k (9) The roots for z of the determinant of the return-difference
Where the lower diagonal matrix T depends on the particular matrix satisfy
case c (z)
det[zI-(I-TK)] = 0
 u0 0 0 0
. 0 o (z) (15)
u 0 0 0 
 1 u0 0 where c (z) and o (z) are respectively the closed-loop and
u u1 u0 0 0 0 open-loop characteristic polynomials of the closed-loop
T1   2  estimator. That is, the closed-loop characteristic polynomial of
 . u2 u1 . 0 0
the matrix I-TK must have roots (eigen-values) which lie
 . . . . . 0 within the unit-circle on the z-plane. Now since T is a lower-
 
triangular Toeplitz matrix, it follows I-TK is also a lower-
 u nb u nb-1 u nb-2 . . u 0 
triangular Toeplitz matrix. Furthermore, it is well established
(10) that the eigen-values of such a matrix are just the its diagonal
for all-zero models and elements which in this case are the repeated roots depending
 u0 0 0 . 0 0 on what model we use:
u s0 0 0 0 0 
 1 Case i: All-Zero Model
u s1 s0 0 0 0 For this case T=T1 in (10)
T2   2  c (z)  (z  1  u 0 K) nb 1  0 giving
 . s2 s1 . 0 0
 . . . . . 0 zi  1  u 0 K,i  1, 2...nb  1 (16)
  For stability all roots must lie within the unit circle in the z-
 u na s na-1 s na-2 . . s 0 
plane. This gives us
(11)
for all-pole models.
1  u 0K  1
We now obtain a new equivalent system as shown in Fig 2. (17)
Excluding the case when u 0  0 , the gain K must satisfy:
Wk + Ŵk
T K 0  u0K  2 (18)
- k 1  z 1
I
Equation (19) clearly poses no problem provided u0 and K are
Ŵk1
always positive. Then
2
0K (19)
u0
z1I However, if u0 is negative then K (from 18) must also be
negative for stability. This will require a time-varying K
Figure 2. Equivalent Multivariable System. whose sign tracks the same sign as u0 . Whilst applying a
positive step to an unknown system this does not present a
From Fig 2, the weight vector estimate is found from the problem.
multivariable system in the z-domain according to, Case ii: All-Pole Model
ˆ (z -1 )  K A similar argument holds for when T=T2 in (11) and we arrive
W Tε(z -1 )
1  z 1 (12) at the condition for stability as
1  u 0K  1
and the weight error vector is found from (20)
1  s0 K  1
(21)
-1 ˆ (z-1 )
ε(z )=W(z )-z W -1 -1
(13) provided u0 ,so  0
The condition for stability is thus
 2 III. EXAMPLES
0K (22) Consider the convergence properties of the parameters of a
max(u o ,so )
system under ideal conditions. A unit step is applied to second
Provided both u0 and s0 are both positive, (22) holds and the order system with transfer function
estimator will converge to the true values. The converse is true 0.9
G(z -1 )  . The estimated weights are shown
if both are negative (and the gain K changes sign). However, if 1- z  0.9z 2
-1
any one of them is negative and the other positive the method
in Fig 3 below for a step size of K=0.5.
must fail. Such a system will arise if the original system is
ARMA and non-minimum phase. Such systems can have an
initial negative going output for a positive input which then
reverses back to being positive again. The period when the
output is negative (an undershoot) would result in instability.
This is equivalent to trying to approximate an AR model to a
non-minimum phase ARMA model. This is not possible since
the pole polynomial would be non-convergent. Also note that
for systems with pure time-delay (22) is modified (since
so  0 ). e.g. when b1 is now estimated instead of b0 (as in
2
(8a)), then (22) becomes 0  K  .
max(u o ,s1 )
It is interesting to see for Case i that the stability is only Figure 3. Convergence of weights for second order example.
dependant on u0 and is also independent of the system order.
The convergence is fast, in fact as fast as recursive-least
This is different from the LMS case where the step-size
squares (RLS)[2] which converges in as many steps as are
depends on the variance of the input signal and the number of
parameters to estimate. Note that the pole-polynomial
weights. Of course u0 is renewed with every new buffer of parameters are shown as negative in Fig 3 as per equation
data and so it is not just one value which the stability is (8a). An equivalent RLS identification algorithm was run on
dependent on. the same system output and did not converge to the right
Now we assume that the true weight vector has a constant but values. This was because the RLS algorithm requires a
persistently exciting input, rich in harmonics. This can be
unknown set of weights, say W0 . This is modeled by a step achieved with either white-noise or a square-wave as the
1 driving signal. White noise was chosen and the parameters
z-transform vector W0 .
1  z 1 converged for RLS as per Fig 4. The forgetting factor was
chosen as unity and the initial error-covariance matrix was
We now examine the convergence by applying the step input
chosen as a diagonal matrix of value 10 for the diagonal
1 1 elements. The new method convergence characteristic did not
vector in the z-domain to (14) thus: W (z ) =W0
1  z 1 change for a white-noise input, but the gain had to be switched
1 in sign when the input and output changed sign to maintain
ˆ (z -1 )  [I  (I  KTz -1 ]-1 KTW
W stability. This is not a definitive comparison however as there
1  z 1
0
(23) are a great many variables which can affect the outcome in the
By applying the discrete-time final-value theorem to (23) we case of RLS.
get
Lim ˆ =
W Lim ˆ (z -1 )(1- z-1 )
W (24)
k  k z 1

Ŵk  Wo (25)
So that the weights converge to the true values provided the
closed-loop system is stable.
We can also show that the error vector satisfies
ε(z-1 )  [I  (I  KTz-1 ]-1 (1  z1 )W(z-1 ) (26)

Figure 4. Convergence of the parameters using the RLS

algorithm
By defining the error norm vector given by Signal Processing, Communications and Control):
Wiley Interscience, 2003.
εk  Wk -W
ˆ
k-1 (27) [2] S. Haykin, Adaptive filter theory. Englewood Cliffs
we can directly compare the convergence of both algorithms. New Jersey: Prentice Hall, 1986.
[3] W. Junfeng, "A variable forgetting factor RLS
adaptive filtering algorithm," in Microwave, Antenna,
Propagation and EMC Technologies for Wireless
Communications, 2009 3rd IEEE International
Symposium on, 2009, pp. 1127-1130.
[4] G. G. Yin and V. Krishnamurthy, "Least mean square
algorithms with Markov regime-switching limit,"
Automatic Control, IEEE Transactions on, vol. 50,
pp. 577-593, 2005.
[5] S. Elliott, I. Stothers, and P. Nelson, "A multiple
error LMS algorithm and its application to the active
control of sound and vibration," Acoustics, Speech
and Signal Processing, IEEE Transactions on, vol.
35, pp. 1423-1434, 1987.
[6] J. E. Greenberg and P. M. Zurek, "Evaluation of an
Figure 5. Error norm comparison of RLS with feedback Adaptive Beamforming Method for Hearing-Aids,"
method. Journal of the Acoustical Society of America, vol. 91,
pp. 1662-1676, Mar 1992.
Fig 5 shows that the rate of convergence is very similar with [7] K. Yamamoto, K. Ohno, and M. Itami, "Improving
the feedback method producing a much smoother graph. performance of Inter-Vehicle Communication using
LMS adaptive circular array antenna," in Intelligent
References Vehicles Symposium (IV), 2010 IEEE, pp. 1128-1133.
[1] S. Haykin and B. Widrow, Least-Mean-Square
Adaptive Filters (Adaptive and Learning Systems for

View publication stats