2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac.

2013, Kuala Lumpur, Malaysia

N4SID and MOESP Subspace Identification Methods

I. W. Jamaludin1, N. A. Wahab2, N. S. Khalid2, S Sahlan2, Z. Ibrahim3 and M F. Rahmat2
1
Mechatronic Department, Faculty of Electrical Engineering, Universiti Teknikal Malaysia Melaka, 76100 Durian Tunggal,
Melaka, Malaysia
2
Department of Control and Mechatronic Engineering, Faculty of Electrical Engineering, Universiti Teknologi Malaysia, UTM
Skudai, 81310 Johor, Malaysia.
3
Faculty of Electrical & Electronic Engineering, Universiti Malaysia Pahang, Kampus Pekan, 26600 Pekan, Pahang, Malaysia
*
Corresponding author. E-mail: [email protected] (Tel: +607-5535418, Fax: +607-5566272)

Abstract— Multivariable Output Error State Space (MOESP) second group consists of methods that aim at approximating
and Numerical algorithms for Subspace State Space System the state sequence of the system and use the approximated
Identification (N4SID) algorithms are two well known subspace state in a second step to estimate the system matrices. The
identification techniques discussed in this paper. Due to the use of methods that constitute the second group are the N4SID
robust numerical tools such as QR decomposition and singular
value decomposition (SVD), these identification techniques are
methods [3], [7] as well as CVA methods [1]. In both groups,
often implemented for multivariable systems. Subspace before identifying system matrices, the calculation of state
identification algorithms are attractive since the state space form vector sequences make use of measured input-output data
is highly suitable to estimate, predict, filters as well as for control only. In the identification procedure, the main step is to
design. In literature, there are several simulation studies for compute the singular value decomposition (SVD) of a block
MOESP and N4SID algorithms performed in offline and online Hankel matrix, H constructed with input-output data. The
mode. In this paper, order selection, validity and the stability for computation of SVD can be done offline or online. The offline
both algorithms for model identification of a glass tube can be easily converted to an adaptive online version for a
manufacturing process system is considered. The weighting slow time varying system such as process control system.
factor α, used in online identification is obtained from trial and
error and particle swarm optimization (PSO). Utilizing PSO, the
Subspace methods have been developed for certain classes
value of α is determined in the online identification and a more
accurate result with lower computation time is obtained. of nonlinear system such as Hammerstein [8–10], Wiener [11–
14] and bilinear [15–17]. In this paper, model identification
Keywords- subspace identification; singular value is implemented to show the effectiveness of SVD
decomposition; Hankel matrices; QR decomposition; MOESP; computation. The N4SID and MOESP subspace algorithms
N4SID are used to identify state space models of a glass tube
manufacturing process system.
I. INTRODUCTION
Subspace identification methods (SIMs) have enjoyed a
great development in both theory and practice since they can II. PROBLEM FORMULATION
identify system matrices of the state space model directly from
the input and output data. Subspace methods are based on Consider the general linear discrete time invariant state
robust numerical tools such as QR factorization and singular space model:
value decomposition (SVD) which makes them attractive from
the numerical point of view. The most influential methods are x k +1 = Ak x k + B k u k + wk (1)
Canonical Variate Analysis (CVA) proposed by [1], (2)
Multivariable Output Error State Space (MOESP) by [2] and y k = C k x k + D k u k +v k
Numerical Subspace State Space System Identification
(N4SID) by [3]. Past research has revealed that the three where x k ∈ ℜ n , y k ∈ ℜ and u k ∈ ℜ n are the system state,
ny u

algorithms use exactly the same subspace to determine the

output and input respectively. While wk ∈ ℜ and v k ∈ ℜ
n ny
order and the extended observability matrix, but that the
weighting matrix, used to calculate a basis for the column are additional unknown noise sequences. They will be
space of the observability matrix is different for those cases. recognized as the residuals of the set of equations for
determining the system matrices and will be omitted for a
Subspace identification methods for linear time invariant while. Thus, consider the following linear discrete time
systems can be classified into two groups. The first consists invariant system;
of methods that aim to recover the column space of the
extended observability matrix and use the shift invariant
structure of this matrix to estimate A and C matrices, xk +1 = Ak xk + Bk u k (3)
subsequently the B and D matrices. The so-called MOESP y k = Ck xk + Dk u k (4)
methods [2], [4–6] are considered in the first group. The

978-1-4673-5609-1/13/$31.00 ©2013 IEEE 140

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.
 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac. 2013, Kuala Lumpur, Malaysia

⎡ uk u k +1 u k + j −1 ⎤
The goal is to estimate systems matrices A, B, C and D with ⎢ ⎥
appropriate dimensions. One important equation in the ⎢ yk y k +1 y k + j −1 ⎥
derivation of subspace state space system identification ⎢ u k +1 u k +2 uk + j ⎥
algorithms is the data equation relating (block) Hankel ⎢ ⎥
H p = ⎢ y k +1 yk + 2 y k + j ⎥. (11)
matrices constructed from the input-output data samples. The ⎢ ⎥
output block Hankel matrices are defined as: ⎢ ⎥
⎢u k +i −1 u k +i u k + j +i − 2 ⎥
⎡ y1 y 2 y3 yj ⎤ ⎢y y k +i y k + j +i −2 ⎥⎦
⎣ k +i −1
⎢ y ⎥
⎢ 2 y3 y4 y j +1 ⎥
⎢ ⎥ ⎡ u k +i uk +i+1 uk + j −1 ⎤
⎢ ⎥ ⎢ ⎥
y yi +1 yi + 2 yi + j −1 ⎥ ⎡ Yp ⎤
Y =⎢ i = . (5) ⎢ y k +i yk +i +1 y k + j −1 ⎥
⎢ yi +1 yi +2 y i +3 yi + j ⎥ ⎢⎢ Y f ⎥⎥ ⎢ u k +i+1 uk + j ⎥
⎢ ⎥ ⎣ ⎦ u k +i + 2
⎢ yi + 2 yi +3 yi + 4 yi + j +1 ⎥ ⎢ ⎥
H f = ⎢ yk +i+1 y k +i + 2 yk + j ⎥. (12)
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ y2 i y 2i +1 y 2i +2 y 2 i + j −1 ⎥⎦ ⎢ ⎥
⎢uk +2 i−1 uk +2 i u k + j +i −2 ⎥
⎢y yk +2i yk + j +i−2 ⎥⎦
where the subscript ‘p’ and ‘f’ stand for ‘past’ and ‘future’ and ⎣ k + 2i−1
i block rows and j columns. The input block Hankel matrices
Up and Uf can also be defined in the same way. The matrix
input-output plays an important role in the problem treated in A typical state space subspace system identification
linear subspace identification and it can be obtained by algorithm involves two steps:
recursive substitution of (6) and (7).
1) Identification of the extended observability matrix and a
Y p = Γi X + H tU p (6) block triangular Toeplitz matrix.
(7) 2) Estimation of the system matrices A, B, C and D from the
Y f = Γi X + H tU f
identified observability matrix and Toeplitz matrix.

The extended observability matrix is defined as N4SID projects Y f onto [Yp ;U p ;U f ] and does a singular
value decomposition (SVD) on the part corresponding to the
⎡ C ⎤ past data, the right singular vectors are estimated as state
⎢ CA ⎥ variables and fit to the state space model. Meanwhile, the
⎢ ⎥
Γ = ⎢ CA 2 ⎥. (8) original MOESP does a QR decomposition on [U f : Y f ] and
⎢ ⎥ then a SVD on part of the R matrix. Part of the singular
⎢ ⎥ matrix is taken as Γf , based on which A and C matrices are
⎢⎣CA i −1 ⎥⎦
estimated, while B and D are estimated based on least square
fitting. As mentioned before, subspace methods are based on
The sequences of state vector, X as follows: robust numerical tools such as QR factorization and singular
value decomposition (SVD). However, they are inappropriate
X = (xk xk +1 x k + 2 … x k + j −1 ). (9) for online identification because of the computational
complexity and storage costs. Thus, an online subspace
The lower block Toeplitz matrix, H, is defined as: identification which consist of recursive algorithm is
developed. A significant contribution is the ability of the
0 0⎤ algorithm to update the estimation for the extended
⎡ D
observability matrix online when new data arrives.
⎢ CB
⎢ D 0 ⎥⎥
H = ⎢ CAB CB 0 ⎥. (10)
⎢ ⎥
⎢ ⎥ A. Offline subspace identification method [18]
⎢⎣CAi − 2 B CAi −3 B D ⎥⎦
The block Toeplitz matrix, H is the concatenation of Hp and Step 1: Calculate U and S in the SVD of H as shown in
Hf, defined by 11 and 12. equations (8) and (9). Singular value decomposition produces

141
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.
 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac. 2013, Kuala Lumpur, Malaysia

a diagonal matrix S, of the same dimension as H and with III. SIMULATION RESULTS
nonnegative diagonal elements in decreasing order.
A. Offline subspace identification
⎡U U 12 ⎤ ⎡ S11 0⎤ T
H = U ⋅ S ⋅ V = ⎢ 11
T
⎥⋅⎢ ⋅V
⎣U 21 U 22 ⎦ ⎣ 0 0⎥⎦ There are several ways to estimate the order of the system.
In this project, singular value decomposition (SVD) is applied.
Figure 1(a) and (b) show the SVD based on N4SID and
Step 2: Calculate the SVD of U 12T ⋅ U 11 ⋅ S11 MOESP method respectively. The best order for both N4SID
and MOESP are six.
⎡S 0⎤ ⎡ VqT ⎤
[
U 12T ⋅ U 11 ⋅ S11 = U q ]
U q⊥ ⎢ q ⎢ T ⎥
0⎥⎦ ⎣(Vq⊥ ) ⎦
⎣0

Step 3: Solve the following set of linear equation to get A, B,

C and D matrices

⎡U qt ⋅ U 12t ⋅ U (m + l + 1 : (i + 1)(m + l ),:) ⋅ S ⎤

⎢ ⎥
⎣ U (mi + li + m + 1 : (m + l )(i + 1),:) ⋅ S ⎦ (a) (b)
⎡ A B ⎤ ⎡ U qt ⋅ U 12t ⋅ U (1 : mi + li,:) ⋅ S ⎤
=⎢ ⎥⎢ ⎥
⎣C D⎦ ⎣U (mi + li + 1 : mi + li + m,:) ⋅ S ⎦
Figure 1. (a) Singular Value Decomposition (SVD) based on N4SID (b)
Singular Value Decomposition (SVD) based on MOESP

B. Online Subspace Identification

Figure 2 (a) and (b) show the validation based on N4SID
and MOESP. In this paper, 3000 data are used for identification
Step 1: Construct new column column to be added to H, using and 500 data are used for validation.
the 2i latest I/O-measurements.

Step 2: Calculate SVD

U k ⋅ S k ⋅ VkT = [α ⋅ U k −1 ⋅ S k −1 column]

where α is a weighting factor and partition

⎡U U 12 ⎤ ⎡ S11 0⎤
U k ⋅ S k = ⎢ 11 ⎥⋅⎢
⎣U 21 U 22 ⎦ ⎣ 0 0⎥⎦

Step 3 and Step 4: Refer to an off-line algorithm step 2 and 3.

C. Particle Swarm Optimization (a)

Particle swarm optimization (PSO) is a population-based

heuristic optimization method inspired by animal behaviour
such as bird flocking. The population is made up of potential
solutions to the optimization problem referred as particles and
the collection of particles is called swarm. The particles are
initialized randomly which searches for best position in N-
dimensional problem space. Each particle updates its velocity
and position during searching based on its best experience and
that of the whole population. This updating principle makes
the swarm move closer to the location of best fitness value and
eventually all the particles will converge to that point. In this (b)
paper, PSO is used in order to obtain the value of α.
Figure 2. (a) Validation based on N4SID (b) Validation based on MOESP

142
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.
 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac. 2013, Kuala Lumpur, Malaysia

The best fitness for output 1 and 2 are 96.71% and 84.79%
respectively based on N4SID. Meanwhile, for MOESP
method, the best fitness for output 1 and output 2 are 93.07%
and 80.23% respectively. Figure 3 (a) and (b) show poles of the
system based on N4SID and MOESP respectively. The
stability is determined by checking the poles of the system or
eigenvalues of the A matrices. The eigenvalues must be inside
a unit circle. All the eigenvalues are in the unit circle, hence the
systems for both methods are stable.

(a)

(a)

(b)
Figure 5. (a) Validation based on N4SID (b) Validation based on MOESP

From Figure 5(a), the best fitness for output 1 and

output 2 are 82.79% and 72.66% respectively. Meanwhile,
Figure 5 (b) shows that the best fitness for output 1 is 79.02%
and for output 2 is 67.04%. The stability of the system is
(b)
tested by determine the poles or eigenvalues. Since all the
Figure 3. (a) Poles of the estimated system based on N4SID (b) Poles of the eigenvalues are inside the unit circle, the systems are stable for
estimated system based on MOESP both methods.

B. Online Subspace Identification Using Weighting Factor,

α=1.
Figure 4 (a) and (b) show the SVD based on N4SID
and MOESP respectively. The result is taken when data
sequence (k) is equal to 60. The best order for both methods is
five.
(a)

(a) (b)

Figure 4. (a) Singular Value Decomposition (SVD) based on N4SID (b) SVD
based on MOESP
(b)

Figure 6. (a) Poles of the estimated system based on N4SID (b) Poles of the
estimated system based on MOESP

143
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.
 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac. 2013, Kuala Lumpur, Malaysia

C. Online Subspace Identification Using PSO ACKNOWLEDGMENT

The authors wish to thank Ministry of Higher Education
The value of weighting factor, α is chosen based on (MOHE), Universiti Teknologi Malaysia and Research
the value of global best or also known as gbest. The final University Grant (GUP) vote Q.J130000.2533.02H70 for their
value of α is chosen with the highest value of gbest PSO. 500 financial support. This support is gratefully acknowledged.
particles and 150 iterations are used for PSO part. The highest
gbest value for N4SID method is 134.8900 which gives
α= 0.4859 and for MOESP method the highest gbest is
REFERENCES
141.5678 which gives α=0.3123.
[1] W. Larimore, “Canonical variate analysis in identification, filtering,
and adaptive control,” in Decision and Control, 1990., Proceedings
of, 1990.

[2] M. Verhaegen and P. Dewilde, “Subspace Model Identification .1.

The Output-Error State-Space Model Identification Class of
Algorithms,” Int Journal of Control, vol. 56, no. 5, pp. 1187–1210,
1992.

[3] P. Van Overschee and B. De Moor, “N4SID: Subspace Algorithms

for the Identification of Combined Deterministic-Stochastic
Systems,” Automatica, vol. 30, no. 1, pp. 75–93, 1994.

[4] M. Verhaegen, “Identification of the deterministic part of MIMO

state space models given in innovations form from input-output
(a) data,” Automatica, vol. 30, no. 1, pp. 61–74, 1994.

[5] D. Bauer and M. Jansson, “Analysis of the asymptotic properties of

the MOESP type of subspace algorithms,” Automatica, vol. 36, no.
4, pp. 497–509, 2000.

[6] X. Wei, M. Verhaegen, and I. Houtzager, “Fault Diagnosis for LTI

Systems Based on Subspace Identification Technique,” in Fault
Detection, Supervision and Safety of Technical Processes, 2009, no.
1, pp. 324–329.

[7] M. Moonen, B. De Moor, L. Vandenberghe, and J. Vandewalle,

“On-and off-line identification of linear state-space models,”
International Journal of Control, vol. 49, no. 1, pp. 219–232, 1989.
(b)
[8] M. Verhaegen and D. Westwick, “Identifying MIMO Hammerstein
Figure 7. (a) Validation based on N4SID (b) Validation based on MOESP
systems in the context of subspace model identification methods,”
International Journal of Control, vol. 63, no. 2, pp. 331–349, 1996.
Figure 7 (a) and (b) show the validation based on N4SID and
MOESP. The best fitness for output 1 and output 2 with [9] C. Xi and F. Haitao, “Recursive Subspace Identification Method for
N4SID method are 74.42% and 89.29% respectively. Hammerstein Systems,” in Proceedings of the 29th Chinese Control
Meanwhile, by MOESP method, the best fitness for output 1 Conference, 2010, pp. 6197–6202.
and 2 are 79.26% and 87.63%.
[10] L. Bako, G. Mercère, S. Lecoeuche, and M. Lovera, “Recursive
subspace identification of Hammerstein models based on least
IV. CONCLUSIONS squares support vector machines,” Iet Control Theory And
Applications, vol. 3, no. 9, pp. 1209–1216, 2009.

In this paper, an offline and online subspace identification [11] D. Westwick and M. Verhaegen, “Identifying MIMO Wiener
models of a glass tube manufacturing process system are systems using subspace model identification methods,” Signal
obtained based on two subspace algorithms which are N4SID Processing, vol. 52, no. 2, pp. 235–258, 1996.
and MOESP. For offline identification, the order of the system
is fixed and for the online identification, it can track every [12] C. T. Chou, “Identification of Wiener models with process noise,”
change of order and system matrices in the system. The in Electrical Engineering, 1999, no. December, pp. 598–603.
weighting factor, α, in online identification can be determined
by using particle swarm optimization. In term of stability, since [13] C. Xi, F. Hai-tao, and W. Xin, “Subspace Identification for Wiener
Systems with General Nonlinearity,” System, pp. 1696–1701, 2011.
both results are within a unit circle, yields a stable system.

144
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.
 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, 8 - 10 Mac. 2013, Kuala Lumpur, Malaysia

[14] 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.

[15] W. Favoreel, B. De Moor, and P. Van Overschee, “Subspace

identification of bilinear systems subject to white inputs,” IEEE
Transactions on Automatic Control, vol. 44, no. 6, pp. 1157–1165,
1999.

[16] H. Chen and J. M. Maciejowski, “Subspace identification of

deterministic bilinear systems,” in IEEE Conference on Decision
and Control, 2000, vol. 3, pp. 1797–1801.

[17] V. Verdult, M. H. Verhaegan, C. T. Chou, and M. Lovera,

“Efficient and systematic identification of MIMO bilinear state
space models,” Proceedings of the 37th IEEE Conference on
Decision and Control Cat No98CH36171, no. December, pp. 1260–
1265, 1998.

[18] N. Wahab, M. Katebi, M. F. Rahmat, and S. Bunyamin, “Data–

Driven Indirect Adaptive Model Predictive Control,” Jurnal
Teknologi, vol. 54, pp. 141–163, 2011.

145
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on April 20,2025 at 15:37:17 UTC from IEEE Xplore. Restrictions apply.