2011 American Control Conference
on O'Farrell Street, San Francisco, CA, USA
June 29 - July 01, 2011
Persistent Excitation by Deterministic Signals in Subspace MISO
Hammerstein System Identification
A. Naitali, Member, IEEE and F. Giri, Member, IEEE
Abstract— The persistent excitation issue is considered for
MISO Hammerstein system identification using the subspace
approach. A signal generator that provides deterministic
persistently exciting (PE) sequences is developed. These are
obtained by operating amplitude weighting and phase shifting
on a discrete Dirac Comb signal of finite length. The proposed
signal generation procedure presents quite interesting features
e.g. the signals data samples are deterministic, weakly
constrained, and their size do not need to be too large.
Theoretically, the signal generator design relies on a technical
lemma establishing the PE property for the considered
identified systems. The effectiveness of the proposed class of
exciting sequences is confirmed by simulation.
I. INTRODUCTION
B LOCK oriented nonlinear models has been proven to be
a quite interesting tool in capturing many biologic,
chemical and electrical nonlinear systems behavior. Among
these, the most popular models are those composed of a
linear dynamical subsystem and a static nonlinear element
connected in series namely Hammerstein and Wiener
models. Several identification schemes have been developed
for this class of systems following various approaches
including e.g. frequency, stochastic, blind, iterative
optimization and others (see e.g. Giri and Bai [1] and
references therein). Presently, the focus is made on the State
Space Subspace Identification approach which captured a
particular attention in control systems community. This class
of algorithms had been introduced and developed for MIMO
linear systems during the nineties by P. Van Overschee and
B. De Moor [2] and references therein, and extended a
decade after to both SISO and MIMO Hammerstein and
Wiener nonlinear systems by Lovera et al. [3], Gomez and
Baeyens [4] and Naitali et al. [5].
From a theoretical point of view, there is no denying that
subspace algorithms are accurate and robust. However, to
guarantee the consistency of the system parameter estimates,
the input sequence applied while the identification
experiment must be so that the virtual input sequence of the
equivalent multivariable linear system is Persistently
Exciting (PE) of appropriate order.
In the present work, parametric identification of multi-
input single-output (MISO) Hammerstein systems is
addressed based on the subspace approach. The focus is
specifically made on the achievement of the PE property that
Manuscript received September 22, 2010; accepted January 28, 2011
A. Naitali and F. Giri are with the GREYC Lab, University of Caen,
France. (a.naitali@ieee.org, fouad.giri@unicaen.fr).
978-1-4577-0079-8/11/$26.00 ©2011 AACC
guarantees the consistency of the subspace algorithm results,
by using deterministic input sequences. To this end, a
technical lemma is established showing that a class of PE
input sequences for MISO Hammerstein systems can be
generated by operating phase-shifting and amplitude-
weighting on a specific mother mono variable signal. The
latter has been previously constructed by the authors for
polynomial basis function (PBF) based subspace
identification of SISO Hammerstein system [5]. Compared
to white Gaussian noise (WGN) and multi-level pseudo
random sequence (MLPRS), the proposed vector input
sequence procedure presents quite interesting features: (i)
there is no (theoretical) need to large set of input and output
data samples; (ii) the design procedure provides the user
with some freedom, accordingly, the sequence wave form
can be shaped to meet some desirable requirements e.g.
uniform excitation of the input variation range, keeping the
system around an operation point.
The paper is organized as follows: the identification
problem is formulated in section II. In section III, the
estimation procedure is described. In section IV, the main
theorem, that describes key design elements for generating
PE sequences, is established. In section V The effectiveness
of the proposed sequence design procedure is illustrated by
numerical simulation.
II. IDENTIFICATION PROBLEM FORMULATION
A. Class of systems
We are considering MISO nonlinear systems that can be
well represented by the Hammerstein block-oriented multi
channel model of Fig. 1. The model consists of m channels
each one being composed of a SISO Hammerstein system.
η1
F1 G1(q-1) y1
u1 + + y
η2
G2(q-1) y2
u2 F2 +
: : ηm :
ym
um Fm Gm(q ) -1
+
Fig. 1. Multi-branch block-oriented nonlinear system
In discrete-time, the system model output y (t , ) is
related to the m scalar inputs, denoted u1 (t ),.., u m (t ) ,
through the nonlinear dynamical equation:
256
 m III. SUBSPACE IDENTIFICATION OF HAMMERSTEIN BLOCK
y t ,     Gr (q 1 ) Fr u r (t )  e(t ) (1) ORIENTED MULTI INPUT NONLINEAR SYSTEMS
r 1
In this section, the subspace identification approach is
Br (q 1 ) br ,1q 1  ..  br ,nr q  nr extended to Hammerstein multi-channel systems (Fig. 1).
Gr (q 1 )   (2)
Ar (q 1 ) 1  a r ,1q 1  ..  ar ,nr q  nr This is performed in four steps:
(i) The system model is reformulated as a series
Fr u r (t )   Fr ,d r  r ,d r u r (t )    r ,d r (t ) (3) combination of a known MIMO static nonlinear subsystem

  Ar , Br , Fr , d r r 1..m     n (4) and an unknown MISO linear subsystem whose inputs and
output sequences are absolutely known.
m
(ii) The state space matrices of the linear part are
n  m  (2nr  d r ) (5)
estimated using a subspace identification algorithm.
r 1
(iii) Estimates of the state space matrices of all channels
where Ar (q 1 ) , Br (q 1 ) designate the denominator and are recovered from those of the diagonal blocks of MISO
numerator of the transfer function Gr (q 1 ) that linear subsystem matrices given by their canonical
characterizes the dynamics of the linear subsystem of the realization in the modal form.
(iv) Finally, estimates of all channel nonlinearity vectors
r th channel; Fr () is a function describing its static nonlinear
and input matrices are obtained the linear MISO subsystem
element; d r is the dimension of the multivariable nonlinear input matrices estimates through singular values

function basis of the system inputs r , d r ()  r, k () k 1,.., d
decompositions (SVD).
r
A. Multivariable State Space Model
on which the line vector Fr ,d r is developed; e(t ) denotes
The input-output model (1) can be given a state space
the equation error resulting from the measurement noises  i representation of the form:
(i  1 m) , and  r , d r (t ) being a function standing for the
nonlinear element modeling error, and which tends to zero  xr (t  1)  Ar x r (t )  Br  Fr (u r (t ))   r (t )
 m
 y (t )  C x (t )   (t )  (8)
when d r tends to infinity. Obviously, none of Br (q 1 ) and
  r r r
Fr (.) is identically zero. Furthermore the system subject to  r 1

the following constraints: (i) the integers nr and d r are where xr (t ) denote the r th channel state vector,

bounded from above by known integers nr and d r ; and (ii) Ar   nr nr , Br   nr 1 and Cr  1nr stand for the state,
the considered linear subsystems are BIBO stable. input and output matrices of the linear dynamic subsystem
of the r th channel;  r (t )   n r and  r (t )   being real
B. Identification Problem Statement
noise sequences standing for the state and output errors.
The identification problem at hand is to determine
accurate parameter estimates of the linear and nonlinear
Consequently, considering the column vector r u r (t ) 
elements of all channels, based on a given set of consistent
input-output data of the system (1-5). Specifically, we seek as an intermediate input, the discrete state space equation of
 each channel can be expressed by:
the determination of the vector  that minimizes the output
estimation error criterion i.e.
xr (t  1)  Ar xr (t )  Br ,d r  r ,d r u r (t )    r (t ) (9)
 where Br ,d r is a one-rank input matrix defined by :
  arg min V 2 () (6)

Ns 1 Br , d r  Br  Fr ,d r   nr d r (10)
V 2 ( )    y(t )  y(t, )2
t 0
(7)
with  r (t )   nr 1
is a zero-mean and bounded noise.

where    n is the related search space and where N s

 r (t )   r (t )   r (u r (t )) (11)
designates the number of measured data samples. The main
difficulties in such minimization problem are three folds: (i)
Now, let U (t ) be the vector input obtained by the
the multivariable feature and its high dimensionality; (ii) the
strong nonlinearity of the output error with respect to the concatenation of all function basis of all channel inputs i.e.
unknown parameters vector  ; (iii) the design of an input
sequence ensuring the necessary PE requirement.  
T
U (t )  (u(t ))  1, d r u1 (t )T ,.., m, d r um (t )  M 1 (12)
with:

257
 m ˆ  p  1: p  n , q  1: q  d 
Bˆ r ,d r   (25)
u (t )  u1 (t ), u 2 (t ),.., u m (t )T   m and M   d r (13-14) c r r r r r r

r 1 r 1 r 1

and let X (t ) be the global state vector by: where pr   k 1

nk and qr  d
k 1
k (26-27)

 
m
 nr (15-16)
T
X (t )  x1T (t ), x2T (t ),.., xmT (t )   N 1 with N   The output matrix is given by the r th row block of ̂c :
r 1
Then, it follows from (9), (12) and (13) that the initial Cˆ r  ˆ c  pr  1 : pr  nr  (28)
system input-output model (1-5) can be compacted in the
 The polynomial vector Fr , d r and the input matrix Br are
following state space form:
recovered up to an arbitrary multiplicative constant r
 X (t  1)  X  U (t )  X (t ) from the SVD of the input matrix estimate Bˆ (29-32)
 (17) r ,d r
 y (t )  X  i (t ) where  r is its unique nonzero singular value:
where:
   N N ,    N M and   1 N designate the global Bˆ r ,d r  U r S rVrT (29)
state, input and output block-diagonal matrices of the linear  1
Br  Ur r ,0,..,0T , Fr,dr  r 1 0 .. 0VrT , r  * (30-32)

part in (16), they are given by: r
Theoretically, these formulas show that all channel
A  diag  A1 ,.., Am  ; (18)
parameters can simultaneously be identified from input

  diag B1, d1 ,..Bm, d m ; (19) output data. However, to be really effective the
  C1 ,.., C m  (20) multivariable input u (t ) must provide the required PE
X  N 1
and i   are bounded and zero-mean properties to the virtual input U (t ) . This is the subject of
equation errors. The new system model (17) shows that the the next section where design elements of a class of PE
initial multi-channel Hammerstein system (Fig. 1) includes deterministic pulse sequences are capitalized as a technical
two subsystems connected in series: (i) a fully known. static lemma.
nonlinear operator  () and (ii) an unknown linear
IV. VECTOR INPUT SEQUENCE DESIGN FOR PE PURPOSE
subsystem represented by the triplet ( , , ) , with input
U (t ) and output y ; and which can be estimated by using A. Persistent Excitation
the state space subspace identification Algorithm To guarantee the consistency of the parameter estimator
described in subsections III.B, the virtual multivariable input
B. Computing the Channels Matrices Estimates
sequence U (t ) must meet the PE requirement stated in the
Let ̂   N  N , ̂   N  M , and ˆ  1 N be the system following definition [2].
matrices estimates provided by the robust version of the
combined deterministic-stochastic subspace identification Definition 1. Let U (t ) be any real multivariable sequence
algorithm namely subid [2] of the linear multivariable
of finite dimension M and U 0 2i 1, j the block Hankel matrix
system whose multivariable input is U and whose scalar
output is y defined by (12) and (17) respectively, and let of U (t ) having 2iM rows and j columns. Then, the input
c   NN
be a similarity matrix which transforms the sequence U (t ) is persistently exciting of order 2i if the
system matrices estimates to the canonical state space input covariance matrix R2Ui ,U is of full rank, where
realization given in the modal form (21-23).
ˆ   1 
 c c
ˆ  ; ˆ   1
c c c
ˆ ; ˆ  ˆ 
c c (21-23)


R2Ui ,U   U0 2 i1, j ,U0 2 i1, j   E j U 0 2i 1, j  UT0 2i 1, j  (33)

Then, the state space matrices estimates of the r th channel with E j  the expectation operator:
can be retrieved from the multivariable state space matrices  1
estimates as follows: E j   lim  (34)
j   j
 The state and input matrices estimates  and Bˆ rare r ,d r The input design issue is to determine a real vector input
th
forwardly obtained by the r diagonal blocks of size sequences u (t ) that provide the virtual input U (t ) with the
nr  nr and nr  d r of the canonical state and input necessary PE property.
matrices ̂ c and ̂ c respectively (24-25). B. Identification Experiment Design
ˆ  p  1: p  n , p  1: p  n  Lemma. Let s1,k 
, s2, k ,.., s m, k  be any m real
Aˆ r   c r r r r r r (24)

258
sequences of length d and let  r  be any strictly increasing from each other by rows permutations, the dimensions of the
integer sequence of length m . subspaces spanned by these matrices are equal, therefore if
Define m scalar signals u1 (t ),.., u m (t ) derived, for some the covariance matrix of one of them is of full rank so will

 
be the case for the other. Hence, the full rank property of the
integer number i * , on the interval t  0,2i d as follows: covariance matrix R2Ui ,U can be guaranteed by ensuring the
d 1
u r (t )  s r , k  (t  r  2ki* ), r  1,.., m (35) full rank property of the following covariance matrix:
1
k 0 R2VV
i  lim VV T (45)
j   j
1 if t  0
where  (t )   (36)
 1V0 2i1, j 1V0T2i1, j 1
V0 2i1, j 2 V0T2i1, j ... V0 2i1, j mV0T2i1, j 
1
0 otherwise 
 2V 1 T 2
V0 2i1, j 2 V0T2i1, j ... 2 V0 2i1, j mV0T2i1, j  (46)
Consider a md  variable signal U (t ) defined on the V
VVT   0 2i1, j 0 2i1, j 
interval t  0,2i  j  2 , for some integers i and j , by:  ... ... ... ... 
 mV0 2i1, j 1V0T2i1, j m
V0 2i1, j 2V0T2i1, j ... mV0 2i1, j mV0T2i1, j 
U (t )    T 
rT, d r 1,..m   r ,k u r (t )  T
k 1,.., d ; r 1,.., m
 md
(37)
 

where  r , k k 1,.., d is any zero-kernel functions basis i.e: Therefore, if for any r the covariance matrix of
 r ,k 0   0 , r , k   1,.., m 1,.., d  (38) r , d (u r (t )) is of rank 2di  (43), it is so for any i  i  .

and let U 0 2i 1,0 j 1 be the Hankel matrix of U (t ) starting Consequently, all diagonal blocks of V V T are of full

from t=0 and having 2i block-rows and j columns. rank. Hence, to guarantee the full rank property of
V V T and thereby of R2Ui ,U , it suffices to find a weak
Then, there exist two integers i0 , j0 such that, for any
condition on the sequences u r (t ) that guarantees the
i   i0 , and any j  j0 , the covariance matrix R2Ui ,U is of
nullity of all off-diagonal blocks in (46), i.e. one seeks the
full rank, provided that:
property:

   2di
r
 (u r ), r ,d (u r ) * V0 2i 1, j c V0T2i 1, j  0 2 di2 di , r  c, (47)
(i) rank R2i r ,d (39)

 
r
It is checked that any non diagonal block V0 2i1, j c V0T2i 1, j
(ii)  m  1  2 i   i (40)
(iii) min r  c   2i (41) of the covariance matrix V0 2i 1, j V0T2i 1, j (48), contains the
r  2,..m ,c 1..r 1
element (49), at the intersection of the p th -row and the q th -
Proof: Let U 0 2i 1, j and r
V0 2i 1, j be the block Hankel column of the block located at the r th -row block and the
c th -column block,
matrices, of U (t ) and of r ,d (u r (t )) respectively, starting
 j 1 
from t 0  0 with 2i row blocks and j columns: r

V0 2i1, j c V0T2i1, j   r,d ur ( p  k)c,d uc (q  k)T 
p0..2i1,
2id2id (48)
 k0 q0..2i1

V (p,q) 
 U (0) U (1) .. U ( j  1) j1
 
 U (1) U ( 2) .. U ( j)  2 i  d  j (42)
r c T
0 2i1, j V0 2i1, j 
r,m ur ( p  k) c,d  uc (q k)T dd (49)
U 0 2i 1, j  
.. .. .. .. k0
 
 U ( 2i  1) U ( 2i) .. U ( j  2i  2)  This can be further developed as follows:
 

r
V0 2i 1, j c V0T2i 1, j ( p, q) 
 r,d (ur (0)) r,d (ur (1)) .. r,d (ur ( j 1))   r,1ur ( p  k) c,1uc (q  k) ...  r,1ur ( p  k) c,d uc (q  k) 
 
 r,d (ur (1)) r,d (ur (2)) r,d (ur ( j))  j 1  
 r,2 ur ( p  k) c,1uc (q  k) ...  r,2 ur ( p  k) c,d uc (q  k) (50)
..
r (43)
V02i1, j  

... ... .. ... 

  ... ... ... 
k 0
r,d (ur (2i 1)) r,d (ur (2i)) ... r,d (ur (2i  j 2)))  
  u
 r, d r ( p  k ) c,1uc (q  k) ...  r, d ur ( p  k) c, d uc (q  k)

The component at the  th -row and  th -column is explicitly

Consider the matrix V obtained by concatenating all
given by:
the r V0 2i 1, j ’s block Hankel matrices, i.e.:
V (p,q)(,)  
j 1
r c T

r, ur ( p  k) r, uc(q  k) (51)

V  1 V0T2i 1, j , 2 V0T2i1, j ,.., m V0T2i 1, j  
T 2imd  j
(44)
0 2i 1, j V0 2i 1, j
k 0
and which can be forced to zero by making u c q  k   0
Then, since the matrices V and U 0 2i 1, j can be obtained when u r  p  k   0 and vice versa, since that

259
 r , k 0  0 (38). Having (35), This amounts to let the
integer i and the integer sequence r  be such that the two

equalities (52) do not simultaneously hold where the
notation (mod 2i  ) stands for the congruence modulo 2i  .
 p  k   r  0 (mod 2i  )
 
 q  k  c  0 (mod 2i )
Bearing this in mind and knowing that both p and
strictly between 0 and 2i-1, if (40) and (41) hold, it
that the following double inequality (53) is satisfied,
0  r  c  ( p  q)  2i  (53)
Consequently (52) do not holds, therefore all the off
diagonal blocks of V V T are zeros. Furthermore, to make
possible the full rank property of R2Ui ,U the number columns
of the Hankel matrix U 0 2i 1, j must be enough large,
i.e. j  2mdi . Consequently there exist
i0*  i  ( m  1 ) / 2 such that , for any i  i0 , and any

j  j0 , the covariance matrix R2Ui ,U is of full rank and
which is equal to 2mdi , which establishes the lemma ■
(iv) j  j0  max2i (d  1)  (m  1)0 ,2i (md  1)  (63)
Proof: It was proved in [5] that, if sr , k  sr , k  for all k  k ,
and any r  1, ..., m , then the covariance matrix of the vector
sequence r ,d (u r (t )) defined by (55-56) is of full rank,
(52) provided that j is enough large so that any signal u r (t )
contains at least d pulses, i.e.:
q vary
comes
 
rank R2i r ,d
(u r ), r ,d (u r )
  2di *
(64)
2i  j  1  2i  ( d  1)  ( m  1) (65)
Therefore since that r , d 0  0 d 1 , and that the arithmetic
progression sequences  r (58) is strictly increasing, all the
conditions (38-41) of the lemma are satisfied provided that
  2i . Consequently, and knowing that in the subspace
identification it is assumed on one hand that the half number
( i ) of block rows of the input Hankel matrix must be strictly
j0  2mdi and larger than the system order ( N ), and on the other hand that
its number of columns ( j ) must be larger than the number
of rows ( 2i (md  1) ) of the matrix concatenating the input
and output Hankel matrices [2], there exist i0  N  1 ,

C. Pulse Based Input PE Signals for Subspace
Identification of MISO Hammerstein Systems
th
Theorem. Consider the N  order linear input-output state
space model of the MISO Hammerstein system (17) subject
to the multivariable input signal U (t ) defined on the
interval t  0,2i  j  2 , for some integers i and j , by:
 0  2i , i0*  i  (m  1) / 2 and
j0  max(2i(md 1),2i (d 1)  (m 1)  2i 1) , such that for any
i  i0 ,    0 , i   i0 and j  j0 , the covariance matrix
R2Ui ,U is of full rank ( 2mdi ) and thereby the input sequence
U (t ) is PE of order 2i , which establishes the theorem ■


U (t )  r , d (u r (t ))T 
T
r 1,.., m
  md 1 (54) V. NUMERICAL EXAMPLE
 To confirm the usefulness of the above theorem we
where r,d (ur (t)) 1ur (t)  2 ur (t) ..  d ur (t)T (55)
consider the following numerical example:
and where: 1 0 .1251 q  1  0 .1166 q  2
F1 ( v )  sin( 2 v ); G1 ( q 1 )  (66)
x  , k x   x k , k  1,.., d  (56) 2 1  1 . 569 q 1  0 . 811 q  2
the signal u r (t ) being defined, for some integer numbers   8v  0.07417q 1  0.07224q 2
F2 (v)  arctan ; G2 (q 1 )  (67)
i * and  , by: 8   1  1.778q 1  0.9245q  2
d 1 0.0723 q 1  0.06859 q 2
F3 (v )  v exp( 2v 2 )G3 ( q 1 )  (68)
u r (t )  s
k 0
r , k  (t  r  2ki ), r  1,.., m
*
(57) 1  1.714 q 1  0.8546 q  2

where  r  (r  1) (58) A. Input sequence design

and where the real sequences s1, k ,.., sm, k  of length d are The system (66-68) is excited by a 24-order PE sequence
generated according to the proposed theorem by letting
such that:
i  12; i *  36;   24 for m  3 , n*  2 and d *  7 , yielding
sr , k  sr , k  k  k  (59)
M  21 , j  528 and a total number of samples N s  599 .
Then, the multivariable input signal U (t ) (54) is PE for the
As mentioned in the theorem, any real sequence composed
system (17) provided that:
of different samples can be used to weight the samples the
(i) i  i0  N  1 (60)
input sequences. Presently, the following simple alternated
(ii)    0  2i (61) saw-tooth weighting sequence is used:
  sr , k   r (1) k ((m  1) 0  2ki  ) (69)
(iii) i   i0  i  m  1  (62)
 2 where  r is a real number which chosen in accordance with

260
the input range of the r th input channel of the system. The 0.5

input sequence thus defined as well as the resulting output 0.4

sequence, are plotted in figure 3. 0.3

0.2

nonlinearities output level

1st channel
0.2 0.1
output

0
-0.2 0
0 50 100 150 200 250 300 350 400 450
time -0.1
1
2nd channel
1st input

-0.2
0

-1 -0.3
50 100 150 200 250 300 350 400 450 3rd channel
1 -0.4
2nd input

0 -0.5
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1 nonlinearities input level
50 100 150 200 250 300 350 400 450
1
Fig. 4. Graphical representation of the system nonlinearities. Solid: true
3rd input

0 nonlinearities. Dotted: estimated nonlinearities.

-1
50 100 150 200 250 300 350 400 450
time
Finally, Fig.5 shows the true system output and the
Figure 3 wave forms of the recorded input and output sequences estimated model output, when both are excited by the input
sequence generated in Subsection A. It is readily seen the
B. Identification results estimated model matches well the true system. The above
The robust version of the Subspace Algorithm (subid) is observations confirm the high accuracy of the subspace
applied using the previously generated input and output data identification when the system is excited by sequences
by letting i  12 and j  528 . The resulting estimates for the generated according to the input design theorem of Section
linear subsystems are the following: IV.C.
0.2

 0.785 0.442   1.063 

Aˆ1    , C1   0.078 -0.0066 (70)
0.15
, B1  
-0.442 0.785  6.245
the system true and predicted outputs

0.1

0.889 0.366    3.01 

Aˆ2    , C2   0.075 0.057 (71)
0.05
, B1  
-0.366 0.889  2.67  0

 0.857 0.347   2.631 

Aˆ3    , B1  2.527 , C3  0.0858  0.0607 (72)
-0.05

-0.347 0.857   -0.1

-0.15

The corresponding transfer functions are: -0.2

0 .1255 q  1  0 .1163 q  2
0 100 200 300 400 500 600

Gˆ1 ( q 1 )  (73) time

1  1 . 569 q 1  0 . 8111 q  2
1
0 .07425 q  0 .07217 q
Gˆ 2 ( q 1 ) 
1  1 . 778 q 1  0 . 9245 q  2
0 .07218 q  1  0 .0687 q  2
Gˆ 3 ( q 1 ) 
1  1 . 714 q 1  0 . 8547 q  2
It is readily seen, comparing (73-75) and (66-68), that the
transfer function coefficients are accurately estimated. The
coefficient estimates of the 7th degree-polynomial
nonlinearities are:
Fˆ1,7  1 0.041  6.255  0.316 10.213 0.431 5.122  (76)
Fˆ2,7  1 0.03  1.596  0.188 2.207 0.221 1.197
Fˆ3,7  1  0.002  1.99 0.02 1.83  0.036 0.77 (78)
Fig. 4 shows the graphical characteristics of the system
polynomial static nonlinearities and their estimates. It is seen
that the estimated curves are quite close to the true curves.
Fig. 5. True (-) and predicted (..) output sequences of the considered 3-
2 channels Hammerstein system
(74)
REFERENCES
(75)
[1] F. Giri, E.W. Bai (Eds), Block-oriented Nonlinear system
Identification, Springer, U.K., 2010.
[2] P. Van Overschee and B. De Moor, Subspace identification for linear
systems: Theory-Implementation-Applications, Kluwer Academic
Publishers, pp 95-134, 1996.
[3] M. Lovera, T. Gustafsson and M. Verhaegen, “Recursive subspace
identification of linear and non-linear Wiener state-space models”,
Automatica, 36, pp 1639-1650, 2000.
[4] J.C. Gómez, and E. Baeyens, “Hammerstein and Wiener Model
(77)
Identification using Rational Orthonormal Bases”. Latin American
Applied Research, Vol. 33, No. 4, pp. 449-456, 2003.
[5] A. Naitali, F. Giri, FZ Chaoui, M. Haloua, Y. Rochdi “Parameter
Identification Based on Hammerstein Models -A subspace Approach”,
IFAC Symposium on System Identification (SYSID), Rotterdam, The
Netherlands, pp 1327-1332, 2003.

261