Anda di halaman 1dari 4

2012 IEEE Conference on Control, Systems and Industrial Informatics (ICCSII)

Bandung, Indonesia, September 23-26,2012

Online Subspace Identification Methods for MIMO


Model

1. w. lamaludin N. A. Wahab, M. F. Rahmat & S. Sahlan


Department of Mechatronics Control and Instrumentation Department,
Faculty of Electrical Engineering Faculty of Electrical Engineering
Universiti Teknikal Malaysia Melaka Universiti Teknologi Malaysia
Malaysia. Malaysia.
irma@utem.edu.my aliza@fke.utm.my

Abstract-Subspace model identification (SMI) has been a highly called MOESP methods [7] are considered in the first group.
interest of research topic in recent years. Subspace methods such The second group consists of method that aim at
as Multivariable Output Error State Space (MOESP) and
approximating the state sequence of the system and use the
Numerical algorithms for Subspace State Space System
approximate state in a second step to estimate the system
Identification (N4SID) algorithms are well known and often
matrices A, B, C and D. The methods that constitute the
implemented especially for multivariable identification due to the
use of robust numerical tools such as the QR decomposition and
second group are the N4SID methods [8] as well as CVA
singular value decomposition (SVD). SMI algorithms are methods [9].
attractive not only because of their numerical stability and Although those numerical tools are applicable for batch
simplicity, but also for the state space form that is very suitable processing, they are inappropriate for on-line identification
in estimating, predicting, filtering and control design. Several due to the computational complexity and storage costs. Thus
simulation studies for MOESP and N4SID algorithms performed to overcome this problem, an on-line subspace identification
in online version are presented. This paper focuses on which consist of recursive algorithm is needed. To make a
computation time, order selection and the stability from both
subspace state space system identification algorithm
methods when performed online.
recursively updated, the need for online in updating and
Keywords- subspace identification; singular value downdating the input output data is required. In this paper, the
decomposition; Hankel matrices; QR decomposition; cllOlesky output from online POMOESP and online N4SID methods
will be discussed.
I. INTRODUCTION

Subspace identification methods (SIMs) have enjoyed a TT. PROBLEM FORMULAnON


great development in the last two decades in both theory and
practice since they can identify system matrices of the state Consider the linear time invariant state space model:
space model directly from the input and output data. Subspace (1)
Xk+l = AXk + BUk + wk
methods are based on robust numerical tools such as QR
factorization and singular value decomposition (SVD) in Yk =Cxk +ek (2)
which make them attractive from the numerical point of view. where xkEffi", ukEffi'", YkEffi', wkEffi" and ekEffi' are
The most influential methods are Canonical Variate Analysis the system state, input, measured output, process noise and
(CVA), Multivariable Output Error State Space (MOESP) and measurement noise respectively. The goal is to estimate
Numerical Subspace State Space System Identification systems matrices A, B, C and D with appropriate dimensions.
(N4STD). The key problem is the consistent estimation of the column
Subspace methods are also have been developed for space of the extended observability matrix, r, which defined
certain classes of nonlinear system such as Hammerstein [1], as:
[2], Wiener [3], [4] and bilinear [5], [6]. In this paper, we
C
will consider only the subspace identification for linear
system. CA
Subspace identification methods for linear time invariant r= CAl (3)
systems basically can be classified into two groups. The first
consists of methods that aim at recovering the column space of
the extended observability matrix and use the shift invariant CA,-l
structure of this matrix to estimate A and C matrices. The so-

978-1-4673-1023-9/12/$31.00 ©2012 IEEE 22


One important equation in the derivation of subspace original MOESP does a QR decomposition on [Uf : Yf] and
state space system identification algorithms is the data
then a SVD on part of the R matrix. Part of the singular
equation relating (block) Hankel matrices constructed from the
matrix is taken as rf' based on which A and C matrices are
input-output data samples. The input (and output) block
Hankel matrices are defined as: estimated, while B and D are estimated based on least square
fitting.
Uj U2 U3 U.I
As mentioned before, subspace methods are based on
3
U2 U U4 Ui+!
robust numerical tools such as QR factorization and singular
value decomposition (SVD). However, they are inappropriate

[�� l
U, Ui+1 Uj+2 Ui+/_1
for online identification because of the computational
u= = (4) complexity and storage costs. Thus, an online subspace
Ui+1 Uj+2 Uj+J Ui+J
identification which consist recursive algorithm is developed.
Ui+2 Ui+J Ui+4 Ui+J+1
The most important thing is to online update the estimation for
the extended observability matrix when the new data arriving.
U2i U2i+1 U2i+2 U2i+J-1
' '
The subscript p and 'f stand for 'past' and 'future'. A. Online N4SJD method [12]
Step 1: Construct data block Hankel matrices, i.e. UI; Up, Yt; Yp
Then, the block Toeplitz matrix is defined as: from the given input-output data.
Step 2: Compute EWr {Yj I Wp} and the QR decomposition from
D 0 o
the Hankel matrices.
CB D o
Step 3: Calculate the SVD of the matrix EWr {Yj I wJ.
H= CAB CB o (5)
Step 4: Compute the matrices A, B, C and D by solving the
least square problem.
D Step 5: Update the Hankel matrices by adding the input-output
data and removing the older ones. Go to Step 2 and repeat.

The success of SIMs is mainly due to the fact that it relies B. Online POMOESP method
on a simple matrix decomposition for which reliable Step 1: Construct the block Hankel matrices from the given
numerical algorithms are available. However, the major input-output data.
drawback is large data matrices are involved and may lead to Step 2: Compute the LQ decomposition.
high computing and storage costs. For instance, the subspace Step 3: Compute the SVD.
identification algorithm [10] involves SVD and eigen Step 4: Obtain matrices C and A.
decomposition which require O(MN2 + M2N + M3) flops and Step 5: Estimate B and D by the least square method.
O(MN + M2 + N2) storage where Nis the data length and Mis Step 6: Update the Hankel matrices by adding the input-output
a so called sliding window size. When N 1000 and M 100 = =
data and removing the older ones. Go to Step 2 and repeat.
as could often be the case, the numbers are approximately 109
flops and 10 Mbytes. Later, [11] proposed a faster
identification method by taking advantage of the Hankel TTT. SIMULATION
structure of the data matrix and the low rank shift structure of Fig. 1 shows the simulation block for solving the matrices
the covariance matrix. The method only require O(MN) flops A, B, C and D. It consists of two inputs and two outputs data
and O(N+ M2) storage space. where both in size 400. The block Simulink was simulating
with online POMOESP and N4STD methods. For N4STD
A typical state space subspace system identification method, the computation time is being recorded for different
algorithm involves two steps: ways of constructing the Hankel matrices which the first will
1) Identification of the extended observability matrix and a use QR decomposition and the second will use cholesky
block triangular Toeplitz matrix. method. The order of the system will be determined by the
2) Estimation of the system matrices A, B, C and D from the number of singular values where the change of gradient is
identified observability matrix and Toeplitz matrix. highest [13].

N4STD projects Yr onto [�;Up;Uj] and does a singular


value decomposition (SVD) on the part corresponding to the
past data, the right singular vectors are estimated as state
variables and fit to the state space model. Meanwhile, the

23
45

y1 40

35 /
/

� 30
...
6 25
/ /
...
!! 20 /' -c h o l esky
/V
"

E 15 R
...

o -O
u
10 //
/
/
o
401 410 450 500 1000 1500 2000 2500

Figure 2. Differences in computation time for QR decomposition and


Tapped Delay3
cholesky method.

y4

Table 1. Order selection from POMOESP and N4SID algorithms


Figure I. Simulink model with 2 inputs and 2 outputs

Order

Simulation time, n POMOESP N4SlD


TV. RESULTS AND DISCUSSIONS
410 4 4
The conventional algorithm for N4STD, QR decomposition
450 4 3
of the Hankel matrices is used and it consumed a lot of time.
By applying the cholesky method via sliding window 500 4 3
approach, the computation time can be reduced. It can be seen 750 4 2
from Fig. 2 that applying the cholesky to the matrices gave the
faster computation time instead of QR decomposition. The 1000 I 3

results obtained from the offline identification. 1250 2 2

For online identification, it is very important to update the


order of the system. As mentioned before, the order of the
system can be determined from the number of singular values
below the point where the rate of change is highest. Both 0.8

algorithms gave mostly different order of the system since the 0.6 .

input-output data for constructing the Hankel matrices are not 0.4 _

in the same way. From Table I, the average order of the system 0.2

is 4. However, the order of the system is calculated as 3. o .

-0.2
Also, as can be seen from Fig.3, the eigenvalues for A
-0.4
matrix is within a unit circle. Thus, both methods are
considered stable. -0.6 _

-0.8 _

-1

-1 -0.5 0.5
Real Part

(a)

24
[6] W. Favoreel,B. De Moor, and P. Van Overschee, "Subspace
1 _
identification of bilinear systems subject to white inputs," IEEE
Transactions on Automatic Control, voL 44, no. 6,pp. 1157-1165,
0.8
1999.
0.6 _

0.4 [7] M. Verhaegen and P. Dewilde, "Subspace Model Identification .l.


;; 0.2 The Output-Error State-Space Model Identification Class of
"-
Algorithms," Int Journal of Control, voL 56,no. 5,pp. 1187-1210,
0
� 1992.
.� -0.2 _

-0.4 [8 ] M. Moonen,B. De Moor, L. Vandenberghe, and 1. Vandewalle,
-0.6 _
"On-and off-line identification of linear state-space models,"
International Journal of Control, voL 49, no. 1,pp. 219-232,1989.
-0.8

-1 _
[9 ] W. Larimore, "Canonical variate analysis in identification,filtering,
-1 -0.5 0.5 and adaptive control," in Decision and Control, 1990 .. Proceedings
Real Part
oj, 1990.

(b) [10] B. De Moor, M. Moonen,L. Vandenberghe,and.T. Vandewalle, "A


Figure 3.(a) Stability of the N4STD system (b) Stability of POMOESP geometrical approach for the identification of state space models
method with singular value decomposition," in Acoustics, Speech, and
Signal Processing, 1988. ICASSP-88., 1988 International
Coriference on, 1988,pp. 2244-2247.

[11] Y. M. Cho,G. Xu,and T. Kailath, "Fast identitlcation of state-space


V. CONCLUSIONS models via exploitation of displacement structure," Automatic
Control, IEEE Transactions on, vol. 39, no. 10, pp. 2004-2017,
An online N4SID and POMOESP algorithms are proposed
1994.
in this paper. The cholesky method gives faster result
compared to QR decomposition of the Hankel matrices. Order [12] N. Wahab, M. Katebi,M. F. Rahmat,and S. Bunyamin, "Data­
of the system is determined by the number of singular values Driven Indirect Adaptive Model Predictive Control," Jurnal
where the gradient is the highest. In term of stability, both Teknologi, vol. 54, pp. 141-163,2011.
results are within a unit circle.
[13] B. Schotleld, "Subspace Based Identification for Adaptive Control,"
2003.
ACKNOWLEDGMENT

The authors wish to thank the Fundamental Research Grant


Scheme (FRGS) by Ministry of Higher Education (MOHE)
vote 4F052 and Research University Grant (GUP) vote
Q.J130000.2533.02H70, Universiti Teknologi Malaysia and
Universiti Teknikal Malaysia Melaka (UTeM) for their
financial support. Also thanks to Prof. Dr. Ir. Michel
Verhaegen and Gijs van der Veen for LTT SI toolbox for
Matlab.

REFERENCES

[I ] C. Xi and F. Haitao, "Recursive Subspace Identification Method for


Hammerstein Systems," in Proceedings of the 29th Chinese Control
Coriference, 20I0,pp. 6197-6202.

[2] L. Balm,G. Mercere,S. Lecoeuche, and M. Lovera, "Recursive


subspace identitlcation of Hammerstein models based on least
squares support vector machines," let Control Theory And
Applications, vol. 3,no. 9,pp. 1209-1216,2009.

[ 3] C. Xi, F. Hai-tao, and W. Xin, "Subspace Identification for Wiener


Systems with General Nonlinearity," System, pp. 1696-1701,2011.

[4 ] M. Lovera,T. Gustafsson,and M. Verhaegen, "Recursive subspace


identification of linear and non-linear Wiener state-space models,"
Automatica, vol. 36,no. April,pp. 1639-1650,2000.

[5 ] H. Chen and.T. M. Maciejowski, "Subspace identitlcation of


deterministic bilinear systems," in IEEE Coriference on Decision
and Control, 2000,vol. 3,pp. 1797-180I.

25