TA08 11:40 Proceedings of the 37th IEEE

Conference on Decision & Control

Tampa, Florida USA December 1998

A Comparison of Some Subspace Identification Methods

'I'ohru Katayamat, Shogo Omorit and Giorgio Piccis
t Department of Applied Mathematics and Physics, Kyoto University
Kyoto 606-0 1, Japan; e-mail: kat ayama@kuamp .kyot 0-u .ac.j p
$ Department of Electronics and Informatics, University of Padova
35131 Padova, Italy; e-mail: [email protected]

Abstract where II, @ are given by

-1
Recently, we have derived stochastic realization
methods for a system with exogenous inputs [4, 11 and [rI 'PI = [E,, C,,] [ Cup C,, ] (2)
the relevance of stochastic realization to subspace iden-
tification of state-space systems has shown in [4]. In this It can be also shown that the operators ll and 0 satisfy
paper, we briefly review the basis of stochastic subspace the discrete Wiener-Hopf type equations
identification algorithm 11, 21 and present some simula-
tions results to compare the performance and compu- HCppIyu = Cfplul @'cuulp = Cfulp (3)
tational loads of the realization based algorithms, the where Cpplu,Cuulpare the conditional covariance opera-
CLS algorithm [3], and the basic 4SID [7]. tors of the past vector p ( t ) given u+(t)and of the future
input u + ( t ) given the past p ( t ) , and are defined by
Stochastic Realization with
Exogenous Inputs C,alc := E{(aICL)(blCL)T} = c a b - C,,C,-,'Cca
In this section, we summarize the stochastic realiza- where aJc* := a - (ulc). 0
tion method based on [I, 21. Consider a discrete-time Let {y(t), u ( t ) } be the jointly stationary regular full
stochastic linear system with the rn x 1 input vector rank process. Suppose that there is no feedback from
u ( t ) and the p x 1 output vector y(t). It is assumed that y t o U . Then it can be shown that is block lower-
{ u ( t ) , y(t), t = O , f l , .. are jointly wide sense sta-
9) triangular, so that it is a causal operator.
tionary processes with zero mean and finite covariance Let rankCjplu = n. Consider the Cholesky factor-
matrices. izations CP,lu = LpLT and C f f l , , = L j L F . Define
Let t be the present time and k a positive integer. E + @ ) := L;1(fIU+L)(t), a - ( t ) := L,l(plu+*)(t). It
We then define the stacked vectors of past and future then follows that E{E+(t)E?(t)}= LflCj,luL;T. Sup-

[ :y] [ ]
inputs as pose that the SVD of the normalized block Hankel ma-
trix LT'Cjpl,L;T be given by
u-(t) := , u + ( t ) := fl)
(4)

u(t + k - 1) 2
where UTU = I,,, VTV = I,, and = diag(81,...,8,,)
is a diagonal matrix with nonzero singular values (1 1
and y-(t) and yt(t), the stacked vectors of past and 51 1 ... 2 5,, > 0). We see that 6:s are the canonical
future of outputs, are defined similarly. For notational correlation coefficients between the conditional random
simplicity, we also define the past and fiture its vectors (flu+')(t> and ( p l u + l ) ( t ) .
For the SVD of (4), we define the extended observ-
d t ):= [ y-(t) ] , f(t) := y+(t) ability and controllability matrices as

0 := L f U W 2 , c := j y 2 V T L TP (5)
Theorem 1 Suppose that p ( t ) n ut(t) = 0. Then the
optimal LS predictor f^(t), of the future output vector where rankU = rankC = n. Then the block Hankel
f(t), based on the past input-output data p ( t ) and fu- matrix Cfplu has a decomposition Ejplu = OC. Since
ture inputs u+(t), is given by the orthogonal projection II = C,pluC&, the oblique projection is expressed as
i(t) = f(t)lP(t) v u+(t)= nP(t)+ @ut(t) (1) rrp(t) = O z ( t ) (6)
$1 0.000 1998 IEEE
0-7803-4394-8198 1850
where the state vector is now defined to be the n x 1
vector Table 1: The number of flops for 50 simulation runs

z ( t ) = cc-'
P P b At) = 5'/'VTL,'p(t) (7)
COV-a CLS-a Basic 4SID LQ
Theorem 2 Suppose that there is no feedback from Flops 1.58 x 10' 2.23 x log 5.26 x lo7 6.42 x lo8
the output y(t) to the input. u ( t ) . We assume that
rankCjplu = n. Then in terms of a state vector z ( t )
of (7), we have a stochastic realization of the form-
In the present simulation studies, the input is cho-
z(t + 1) = A z ( t ) + B u ( t )+ K e ( t ) (8) 10
sen as u ( t ) = U O C ~&(wit),
= ~ where the frequencies
y(t) +
= Cz(t) Du(t) e ( t )+ (9)
wi's are uniformly spaced in the interval (O.l,3)(rad)
and where U0 is adjusted to yield U," = 1. The noise
variances are chosen as U: = cr," = (0.05)'. It follows
from the PE condition for U012k-1 that IC 5 10. The
Simulation Results performance is evaluated by the mean square error

-
Fig. 1 The plant model
where N = 200,400,1000,2000, and B j denotes the true
parameter and e,(/, N ) is the estimate of B, at I-th run
with the number of data N, and where M denotes the
number of simulation runs.
Fig. 2 depicts the performance of five algorithms,
Some results of computer simulations are presented
where IC = 8, M = 100. In this case, COV-a and COV-b
to show the performance of five subspace identification
show similar performance, but the performance of CLS-
algorithms.
a and CLS-b is rather different. In order to analyze
Basic 4SID is due to Verhaegen[7], where the this fact, we have simulated CLS-a and CLS-b methods
Cholesky factorization is used to get L factor. for several different IC's, where the input is a sum of
15 sinusoids and it4 = 50, N = 1000. We see from
COV-a is the algorithm based on the stochastic re- Fig. 3 that both methods give similar performance for
alization, where the system matrices are estimated IC greater than 10, but for the smaller IC, CLS-a shows
by using the estimate of state vector (see (7)). better performance. In Figs. 4 and 5, the pole estimates
by COV-a and CLS-a methods are depicted for IC = 8,
COV-b is the algorithm based on the stochastic re- N = 1000. We see that COV-a gives a rather scattered
alization, where the system matrices are estimated
pole estimates, but CLS-a yields better pole estimates
by using 0 and 0 (see (3) and (5)).
with a smaller variability.
CLS-a is the algorithm based on the constraint Table 1 shows the number of flops of four algorithms,
least-squares algorithm due to Peternell et a1.[3] where LQ denotes the algorithm based on the LQ fac-
and using the estimation of state vector (see (7)). torization of the Hankel matrix [5, 6, 71. The number of
flops includes all the computations for the whole simu-
CLS-b is the constraint least-squares algorithm and lations by each algorithm for IC = 8, M = 50, N = 1000.
using 0 and CP (see (3) and (5)). It therefore follows that by using the Cholesky factor-
ization [2], we get a great computational saving over the
We consider a 5th-order SISO system shown in Fig.
method based on LQ factorization. Also, it is rather sur-
1 [9], where u ( t ) is the input, and w(t) and v(t) are
prising to find that CLS-a is ten times more expensive
white noises with mean zeros. The transfer function is
than COV-a.
given by G ( z )= B ( z ) / A ( z ) ,where

B ( z ) = 0 . 0 2 7 5 ~ -+~ 0 . 5 5 1 ~ - ~
A (. z.) = 1 - 2.3443~-1+ 3.081z-' - 2 . 5 2 7 4 ~ - ~
References
+ 1 . 2 4 1 5 ~ --~ 0 . 3 6 8 6 ~ - ~ [l] T. Katayama and G . Picci, "An Approach to Real-
ization of Stochastic Systems with Exogenous Input",
The G ( z ) has a zero at z = -2 and poles at z = 0.9, Preprints of 11th IFAC Symposium on System Identifi
0 . 8 e f J ,0.8ef1.'j. cation, Kitakyushu, Japan, July 1997, pp. 1107-1112.

1851
T. Katayama and G. Picci, “Realization of Stochastic P. Van Overschee and B. De Moor, Subspace Identifi-
Systems with Exogenous Inputs and Subspace Identifi- cation for Linear Systems, Kluwer Academic Publica-
cation Methods,” 1998 (submitted). tions, 1996.
K. Peternell, W. Scherrer and M. Deistler, “Statisti- M. Verhaegen and P. Dewilde, “Subspace Model Identi-
cal Analysis of Novel Subspace Identification Methods,” fication (Parts 1 and 2),” Int. J . Control, vol. 56, 1992,
Signal Processing, kol. 52, no. 2, July 1996, pp. 161-177. pp. 1187-1210 & pp. 1211-1241.
G. Picci and T. Katayama, “StochasticRealization with M. Verhaegen, “Identification of the Deterministic Part
Exogenous Inputs and “Subspace Methods” Identifica- of MIMO State Space Models given in Innovations Form
tion,” Signal Processing, vol. 52, no. 2, July 1996, pp. from Input-Output Data,” Autornatica, vol. 30, no. 1,
145-160. 1994, pp. 61-74.
P. Van Overschee, P. and B. De Moor, “N4SID - Sub- M. Viberg, “Subspace-based Methods for the Identifi-
space Algorithms for the Identification of Combined De- cation of Linear Time-invariant Systems,” Autornatica,
terministic - Stochastic Systems,” Automatica, vol. 30, vol. 31, no. 12, 1995, pp. 1835-1851.
no, 1, 1994, pp. 75-93.

[-Bulc :COV 4 L S o(da10) x(PM) &hueoldwlvd.w25l

0.3 -

0.2 -
0.1 -
0
0.2 0.4 0.6 0.1
I I .2
Numbar ol DU. N Red Rn

Fig. 2 The performance of five algorithms Fig. 4 The pole estimates by COV-a

I
10-1
7 8 0 10 11 12
(run~t!-mMIXk

Fig. 3 The performance vs. number of rows Fig. 5 The pole estimates by CLS-a

1852