See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228989316

Hammerstein model identification

Conference Paper · July 2002

CITATIONS READS

11 1,977

5 authors, including:

Fouad Giri Youssef Rochdi

Université de Caen Normandie National School of applied sciences, Ibn tofail university, Morocco
450 PUBLICATIONS 4,822 CITATIONS 54 PUBLICATIONS 512 CITATIONS

SEE PROFILE SEE PROFILE

Abdessamad Naitali
Mohammed V University of Rabat
20 PUBLICATIONS 155 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

reduction of interfacial tension View project

Intelligent transport systems ITS View project

All content following this page was uploaded by Fouad Giri on 20 May 2014.

The user has requested enhancement of the downloaded file.

Proceedings of the 10th Mediterranean Conference
on Control and Automation - MED2002
Lisbon, Portugal, July 9-12, 2002.

HAMMERSTEIN MODEL IDENTIFICATION

F. Giri, F.Z. Chaoui, M. Haloua, Y. Rochdi, A. Naitali
LAP, ISMRA, Caen, France
fax: +33231567340
e-mail: [email protected]

Keywords: System identification, Hammerstein recursive least squares type algorithms, (see e.g. [1,
models, Persistent excitation, Parameter 3, 12, 14]. The second direction, commonly
convergence. referred-to nonparametric, considers that the
nonlinear element of the model is not necessarily
polynomial. It may be any continuous function as
Abstract
in [15] or a measurable (in Lebesgue/Borel sense)
We are considering system identification based on function as [6, 7, 13]. However, even in this case
the Hammerstein model i.e. a non-linear static gain the identification process involves a truncated
in series with linear dynamics. The static gain series approximation either of the nonlinear
characteristic is any nonlinear function F. An element or of related functions. For instance,
identification scheme is designed to determine Fourier series approximation has been used in [6,
exactly the model of the plant dynamics and a set 9], polynomial series approximation (involving
of N different couples (x , F(x)), where N is Laguerre, Legendre or Hermite polynomials) has
arbitrarily chosen by the user. Given the previous been proposed in [8, 13], block-pulse functions
couples, one can build up an Nth degree polynomial have been used in [3]. Due to these finite series
or polygonal approximation of the static gain approximations, the identification problem
characteristic. The main feature of the amounts, just as in the parametric approaches, to
identification scheme is the design of a persistently estimating a finite number of parameters. The
exciting input sequence that makes it possible to nonparametric methods usually involves
identify exactly the unknown parameters, using a probabilistic tools in the estimation process of the
gradient algorithm and a singular value unknown parameters. The convergence of the
decomposition. parameter estimates has been analyzed, using
stochastic tools, both for parametric and
1 Introduction nonparametric methods. In [9] it is shown that
consistency can be achieved, with a parametric
Hammerstein models are composed of a static instrumental variable method, using as input a
nonlinear gain and a linear dynamics part. In some strictly persistently exciting sequence or a white
situations, they may be a good approximation for noise. Specific random inputs have been used in
nonlinear plants. The problem of identifying plants nonparametric methods to ensure consistency and
based on such a class of models has been given a other properties e.g. convergence of the nonlinear
great deal of interest over the last years, e.g. [1, 3, element estimates in the mean integrated square
6, 7, 9, 11, 12, 13, 14]. It has been approached error (MISE) or uniformly and MISE [11, 14]. In
following two major directions. The first one the light of the above discussion, it appears that
consists in supposing polynomial (or polygonal) the both parametric and nonparametric approaches
nonlinear element of the model. Then, the involve, explicitly or implicitly, orthogonal
identification problem turns out to be a parametric function series approximations of the nonlinear
one since it consists in estimating the parameters of element. This is a normal consequence of the fact
the model linear and nonlinear parts. Parameter that the plant nonlinear element is characterized by
estimation is generally performed, based on a continuous function F(v), defined on some real
adequate (known) structures of the model, using
interval (say [vmin, vmax]), while the plant input The present paper is organised as follows: the plant
sequence (say {v(t)}) is a discrete-time sequence to be identified is described in section 2, the
(i.e. t=0, 1, 2, …). Actually, a continuous real identification scheme is presented in sections 3 to 5
function F(.), which is defined by an uncountable and it is analysed in section 6. Finally, a simulation
number of values F(v) (vmin≤v≤vmax), cannot be comes to illustrate the effectiveness of the proposed
captured, except for particular cases, by a identification method.
countable set of values, namely {v(t); t=0, 1, …}.
2 Identification problem statement
In this paper, a parametric identification scheme is
designed to deal with the case where the static gain 2.1 Class of Plants
is any (non-identically null) nonlinear function F.
The proposed scheme identifies perfectly the model We are considering plants that can be described by
of the plant dynamics and a set of N points (vi, the following Hammerstein model:
F(vi)) (i=1, …, N) of the static gain characteristic F,
where N and the vi’s are arbitrarily chosen by the B(q −1 )
y(t) = u(t) + z(t) (2.1a)
user. The key ideas of the identification scheme are A(q −1 )
briefly presented in the following. The points (vi,
F(vi)) uniquely determine an Nth-degree polynomial u(t) = F(v(t)) (2.1b)
P(v) that coincides with F(v) on the vi’s, then
with
identifying the points (vi, F(vi)) amounts to
identifying P(v). Furthermore, letting the plant A(q-1) = 1 + a1q-1 + ... + anaq-na (2.2a)
input be chosen in the set {vi, i=1, …, N}, allows
substitution of P(v) to F(v) in the plant model; so B(q-1) = b1q-1 + ... + bnbq-nb (2.2b)
doing, the initial identification problem is where q-1 denotes the backward shift operator ; v(t)
converted to one where the nonlinear element is and y(t) are the plant input and output,
polynomial. The mentioned substitution leads to a respectively ; the internal signal u(t) is non-
plant representation that is linear in the unknown measurable. z(t) accounts for external disturbances
parameters, the estimation of which can be and modeling errors; it is supposed to be bounded.
performed by a least-squares type algorithm. It is The function F, which represents the static gain, is
worth noting, that the involved parameters are non-identically null and satisfies the following
bilinear functions of the desired (unknown) conditions:
parameters i.e. those of the plant dynamics, on one
hand, and those of the static gain, on the other F(0) = 0 (2.3a)
hand. To recover the desired parameters (from the
estimates), a procedure is built up based on matrix |F(v)| < ∞ for any (2.3b)
algebraic tools such us singular values Equality (2.3a) simply means that the origin is an
decomposition and properties of Vandermonde equilibrium of the plant. The plant dynamics are
matrices. Finally, a persistently exciting input further submitted to the following assumptions:
sequence is designed and shown to guarantee the
A1. There is a known integer n such that:
convergence of the estimates to their true values.
The proposed exciting input is an impulse type n ≥ max {degree A(q-1), degree B(q-1)}.
sequence just as in [6]. However, the present paper
extends the results of [6] in two main directions: (i) A2. A(q-1) and B(q-1) are coprime.
the class of the considered nonlinear elements is
enlarged, (ii) the introduction of singular value A3. All zeroes of qnaA(q-1) are strictly inside the
decomposition to recover the desired parameters unit circle.
from the estimated bilinear parameters.
Remarks 2.1. P(0)=0 (3.3a)
a) Assumption A2 ensures the controllability of the P(Vj ) = F(Vj ) for j=1, ..., N (3.3b)
-1 -1
transfer function B(q )/A(q ). Such a polynomial can be written as follows:
b) Except for assumptions A1-A3, the plant is
N N v−V
arbitrary: the dynamic parameters (ai, bi) are
unknown, the zeroes of B(q−1) may be unstable
P ( v ) = v ∑ j jd P ( v ) with P j ( v ) = Π i
j=1 i =1 V j − V i
(non minimum phase dynamics), the leading i≠ j
coefficients (b1, b2, ...) may be null i.e. the true (3.4)
plant delay is unknown (but not greater than n). It can be easily seen that:
c) It is worth mentioning that in [6], the nonlinear
element F(.) has been supposed to be a 1 if i = j
P j (Vi ) = δ ij =  (3.5a)
piecewise-affine continuous function with a 0 if i ≠ j
nonzero derivative at the origin, i.e. F’(0)≠0. In
the present paper, F(.) is just supposed to be P(V j )
and d j = (3.5b)
non-identically null. Vj

2.2 Identification Objective Identifying the N couples (Vj, F(Vj)) amounts to

identifying the polynomial P(v) i.e. the coefficients
Our purpose is to design an on-line identification dj (j=1, …, N). Furthermore, if the plant input v(t)
scheme that provide asymptotically the true is forced to take its values in the set {0, V1, …, VN}
dynamics model B(q-1)/A(q-1) and N points (xi, then, due to (3.3b), the plant (2.1a-b) can also be
F(xi)) (i=1, …, N and xi≠0) of the static gain represented by the following model:
characteristic. The main difficulty comes from the
fact that the internal sequence u(t) is not A(q-1)y(t) = B(q-1)u(t)+ A(q-1)z(t) (3.6a)
measurable, i.e. only the input sequence {v(t)} and
the output sequence {y(t)} should be used in the u(t) = P(v(t))= F(v(t)) (3.6b)
identification scheme (fig. 1). Notice that the substitution of P(.) to F(.) in (3.6a-
The identification scheme is designed in three b) generates no new error. As long as v(t) belongs
steps: i) first, an adequate plant model to {0, V1, …, VN}, the models (2.1a-b) and (3.6a-b)
parameterization is determined; ii) second, the are equivalent and, consequently, the initial
unknown model parameters are estimated based on identification problem (that consisted in identifying
the proposed parameterization; iii) finally, an the N couples (Vj, F(Vj)) and the parameters (ai, bi)
exciting input sequence that satisfies the PE amounts to identifying the parameters ai, bi and dj.
property is generated . Indeed, the N couples (Vj, F(Vj)) can be computed
from the dj’s using (3.5), i.e.:
3 Plant model reparameterization F(Vj) = P(Vj) = dj .Vj ; for j=1, ..., N (3.6c)
Our objective is to determine exactly the dynamics
model B(q-1)/A(q-1) and N couples (Vj, F(Vj)) (j=1, Now, a regression form will readily be derived for
..., N), where the Vj’s are arbitrarily chosen. (3.6a-b), Indeed, one has successively:
Without loss of generality we suppose that: A(q−1)y(t) = B(q−1)P(v(t)) + A(q-1)z(t)
V1 < V2 < …< VN (3.1) n N
= ∑ b i q −i ∑ d j Pj (v(t)) v(t) + η(t)
Vj ≠0 (j =1, ..., N) (3.2) i =1 j=1
n N
Using a polynomial interpolation, we associate to
the couples (Vj, F(Vj)) a unique Nth-order = ∑ ∑ ij v(t − i)Pj (v(t − i) ) + η(t)
i =1 j=1
polynomial P such that:
(3.7a)
where η(t) = A(q-1)z(t) and: Proposition 3.1. Let M∈RnxN be any rank-1 real
matrix. Then its SVD decomposition has the
µij = bidj (i=1, ..., n; j=1, ..., N) (3.7b)
following form:

Equation (3.7a) can be given the following σ1 0  0

0 0 0
regression form:
M =Γ Σ (3.11)
y(t) = - W , + η(t)
T *
(3.8a)  
 
0 0  0
Φ(t)T = [−y(t−1) … −y(t−n)
where Γ∈Rnxn, Σ∈RNxN and σ1 is the unique
v(t-1)P1(v(t-1)) ... v(t-n)P1(v(t-n)) … nonzero singular value of M. Furthermore, M can be
... v(t-1)PN(v(t-1)) ... v(t-n)) PN(v(t-n))] uniquely decomposed as follows:
(3.8b)
 b1 
Θ* = [a1 … an µ11 ... µn1 ... ... µ1N ... µnN]T
(3.8c) M =   [d 1  d N ] (3.12a)
As Θ* comes in linearly, equation (3.8a) turns out b n 
to be an adequate parameterisation to get estimates
with:
of the parameters ai and µij, using a least squares
1 1 
type algorithm. But, how can we obtain estimates 0
of (bi, dj) from those of estimates of µij ?. Before + 
answering this question let us observe that (3.7b) 
can be rewritten as follows :  
[b1  b n ] =   sign(γ);
T 0
 11   1N   b1  1 1 

M= 
  
  =   [d1  d N ] (3.9) 0
 + 
 n1   nN  b n  
 
Notice that M is a rank-1 matrix and that the bi’s 0
and dj’s cannot be determined uniquely from the
µij’s, unless extra conditions are imposed on the 1 1
d  0
bi’s or dj’s. Uniqueness of the solution of (3.9) can 1
 
 = +   . [1 0  0] .sign(γ) (3.12b)
be achieved imposing, for instance, the couple of   
conditions: d   
 N
n 0
∑ b i2 = 1 and ρ([b1 ... b n ]) > 0 (3.10) where γ = ρ Γ[σ 0  0]T .
( )
i =1 1

where ρ([b1 ... b n ]) denotes the first component of The vector [b1  b n ] thus obtained is the only
T

the vector [b1 ... b n ] that satisfies: solution of (3.12a) that satisfies:

ρ([b1 ... b n ]) = sup b j .

n

1≤ j≤ n
∑ b i2 = 1 ; ρ([b1 ... b n ]) > 0 (3.13)
i =1
Based on the above observations a procedure to Proof. See [4] .
solve (3.9) is designed using singular value
decomposition (SVD). This is described in the
following proposition.
4 Model parameter estimation { ,̂(t) − ,̂(t − 1) }, { e (t) }∈S( K 
p 1

2) In the ideal case ({z(t)}={0}) one

4.1 Plant Parameter Estimation { } {
has: ,̂(t) − ,̂(t − 1) ∈ l2 , ep (t) ∈ l2 . }
Estimation Algorithm. Estimate ,̂ (t) of ,* can be
Proof: this can be found any many places, see for
recursively obtained using, for instance, the example [2], [7].
following algorithm :
-WHW The proposition shows that the quality of the
, ' (t) = ,̂(t − 1) + (asymptotic) input-output behavior of Θ̂( t )
1+ - W T - W
(4.1a) depends on the (mean) size µ of the modeling error
* ' , ' (t) η(t). The smaller µ the better the (asymptotic)
,̂(t) = min(! , , (t) ).
, ' (t) model. Interestingly, the above properties hold
whatever the input sequence {v(t)}. However, Θ̂( t )
e(t) = y(t) - Φ(t)T,̂ (t − 1) (4.1b) will not converge to Θ* unless the input sequence
{v(t)} is sufficiently rich. An example of such
where the initial vector ,̂ (0) is arbitrary and the
sequence is proposed in the subsection 5.
*
real constant ! is any upper bound on Θ . In
*

4.2 Reconstruction of A(q-1), B(q-1) and F

fact, equations (4.1a-b) define a gradient algorithm
with parameter projection on the sphere centered Given the parameter vector estimate:
on the origin with radius ! * . Such a projection Θ̂( t ) = [â1(t) ... â n −1 (t) ˆ 11(t) ... ˆ n1 (t) ˆ 12 (t)
prevents ,̂ (t) from diverging, despite the
modeling error {η(t)}. The quality of the estimates ˆ 12 (t) ... ˆ n2 (t) ... ... ˆ 1N (t) ... ˆ nN (t)] T (4.2a)
Θ̂( t ) depends, at least partly, on the mean size of one can immediately construct an estimate of
{η(t)}. A(q-1)

The mean size of real sequences can be evaluated A t (q −1 ) = 1 + a 1 (t )q −1 + ... + a n ( t)q − n (4.2b)
using the smallness-in-the-mean concept [5].
Accordingly, a real sequence {s(t)} is said to be α- The estimates (b̂ i , d̂ j ) of (bi, dj) are obtained from
small in the mean, and we note {s(t)}∈S(α), if: µ̂ ij through a singular value decomposition of the
1 h+k following matrix:
l im sup ∑ s( t ) ≤ α
k →∞ k t = h +1  ˆ 11 (t)  ˆ 1N (t) 
Let µ be the smallest real such that: M̂(t) =    
{η(t)}∈S(µ) (4.1c) ˆ n1 (t)  ˆ nN (t)
(4.3)
Note that µ exists because {η( t ) } is bounded. The That is:
properties of the estimation algorithm (4.1a-b) can 1 (t)

 0  0 

1
 
now be stated (see Proposition 4.1) in terms of µ 
0 1 (t)   
and the posterior prediction error defined by: M̂(t) = + W 

2 O.Σ(t)


   0 

ep(t) = y(t) - Φ(t)T Θ
ˆ (t ) (4.1d) 
0
 

0 1n (t) 


Proposition 4.1. (General properties of Θ̂( t ) ) (4.4)

1) There exists a real constant K1, independent of
Then, Proposition 3.1 suggests the following
µ , such that:
estimates for the parameters (bi, dj):
 1 1 (t)
 0 
( )
and consequently the b̂ i (t), d̂ j (t) will converge to
there true values (bi, dj).
+ W  
   Proof: see [4]
 
[b̂ (t)
1  b̂ n (t) ]
T
=  0 
1 1 (t)
sign(γ(t)) ( )
Given the estimates b̂ i (t), d̂ j (t) , equations (2.2b),
 0  (3.4) and (3.6b) suggest the following estimates for
+ W   B(q-1) and F(.):
  
  B̂(q −1 ) = b̂1 (t)q −1 + ... + b̂ n (t)q − n (4.6a)
 0  N
(4.5a) F̂ (v) = v∑ d̂ j (t)Pj (v) (4.6b)
T j=1
 d̂1 (t) 
 
   =
Remark 4.1
d̂ (t) In comparison with [6], the above singular value
 N  decomposition is a new feature of the present
σ 1 (t) paper.
 0 
+ (t)   . [1 0  0] (t) .sign(γ(t)) 5 Exciting input sequence
  
 
 0  The input sequence {v(t)} should be generated so
(4.5b) that its values belong to the set {0, V1, …, VN} and
(
γ ( t ) = ρ Γ[σ1 ( t ) 0  0]T ) the resulting regression vectors Φ(t) is persistently
(4.5c) exciting (PE). Bearing this in mind, {v(t)} is
chosen to be a periodic sequence, with period
The estimates thus obtained are the only ones that
T=n(N+3), defined as follows:
satisfy the condition
for all integer k and all t in the interval tk ≤ t <tk+1
n 2
∑ b̂i = 1 with tk=kT:
i =1 V1 for t = t k + 2n
and ([ ])
ρ b̂ (t) ... b̂ (t) > 0 .
1 n


v(t) = 
Vj for t = t k + (2 + j)n; j = 2, ..., N
Notice that the singular values σ2(t), …, σn(t)  0 otherwise
have not been accounted for in the rules (4.5a-b).
This has no effect when the µ̂ ’s converge to there (5.1)
ij
true values, since then the matrix M̂ ( t ) has
6 Parameter convergence analysis
asymptotically a rank equal to 1. This is made
precise in the following proposition:
The sequence {v(t)} described above is shown in
{ }
Proposition 4.2. Let M̂(t) be the real matrix this section to provide the regression vector {Φ(t)}
with the PE property. First let Φ(t) be decomposed
[ ]
sequence defined by (4.3). Let b̂1 ( t )  b̂ n ( t ) as follows:
[ ]
and d̂1 (t)  d̂ N (t) be the vectors obtained from Φ(t) =[ φ (t)T φ (t)T … φ (t)T]T
1 2 (6.1a)
N
M̂(t) according to the rules (4.5a-c). In the ideal where:
case, i.e. when {z(t)}={0}, the µ̂ ij ’s converge to
φ1(t)T = [y(t−1) ... y(t−n)
there true values µij, due to Proposition 6.4 (Part 2),
 v(t-1 )P1(v (t−1)) ... v(t-n )P1(v(t−n))] (6.1b) The PE property for the global regression vector
Φ’(t) will now be established using the fact that
φj(t)T = [v(t-1 )Pj(v(t−1)) ... v(t-n )Pj(v(t−n))] Φ’(t) is the concatenation of N PE subvectors,
namely the φj(t)’s. This will prove to be possible
for j=2, …, N (6.1c) since not all the subvectors act simultaneously.
Furthermore, we define Φ’(t) to be the unperturbed Proposition 6.3 (Persistent excitation of Φ’(t) ).
version of Φ(t), i.e.: There exists a real ε>0 such that one has, for any
Φ’(t) =[φ’1(t)T φ2(t)T ….φN(t)T]T (6.1d) integer k and any unit vector W:
where:
max -' (t k + 2 ) T W > 0 (6.5)
1≤2 ≤ n(N + 3)

φ’1(t)= [−x(t-1) … −x(t−n) v(t-1 ) Proof: see [4]

P1 (v(t−1)) ... v(t-n ) P1 (v(t−n))]T
and Proposition 6.4. (Convergence of Θ ˆ (t ) ).
1) General case, i.e. {z(t)} arbitrary but bounded:
B(q −1 )
x(t) = u(t) = y(t)−z(t) (6.1e) there exists a real constant K2, independent of µ,
A(q −1 ) such that:
From (6.1a-e) one gets:
~
,̂(t) − , * ∈ S(K 2  ) { }
φ1(t)= φ’1(t)+ φ ( t ) (6.2a)
2) The ideal case, i.e. {z(t)}={0}:
~
with φ ( t ) = [−z(t−1) … −z(t−1) 0 … 0] , and:
{ }
T
,̂ (t) − Θ * exponentially converge to
Φ(t) = Φ’(t) + Φ(t) (6.2b) zero.
with: Proof: see [4]
Φ(t) =[φ(t) T 0T… 0T] T.
The following propositions stipulate that the 7 Simulation
vectors φ’1(t) and φj(t) (j=2, …, N) have the PE
property. To illustrate the effectiveness of the above
identification scheme, let the true plant model be
Proposition 6.1. (Persistent Excitation of φ’1(t)). described by (2.1a-b) where the dynamic part is
There exists a positive real ε0, such for all k: characterized by:
4n
A(q-1) = 1−1.5 q-1 + 0.56 q-2; B(q-1) = q-1−2 q-2;
∑ φ ′(t
2 =1
1 k + 2 )φ1′ (t k + 2 ) T > 0 0 I2n (6.3)
and the static element F(.), plotted in Fig. 2, is
where I2n is the identity matrix 2nx2n. defined by the polynomial function:
Proof: see [4] . N
F(v) = ∑ c j v j with N=6, c1=1, c2=0.0,
Proposition 6.2. (PE of φj(t), j=2,..,N). one has for j=1
all k: c3=0.2, c4=0.0, c5=0.009, c6=0.0
n

∑ φ (t + n(j+ 2) + 2 )φ j (t k + n(j+ 2) + 2 ) ≥ 0 1 .In

T
j k So, the true values of the parameters to be
2 =1
identified are the following:
for j = 2, ..., N (6.4)
where 0 1 = Inf (Vj ) and In is the nxn identity
2 a1 = 0.56, a2 = -1.5, b1 = 1, b2 = -2
j= 2.. N
Also, and the N couples (Vj, F(Vj)) to be estimated
matrix.
have the following true values
Proof: see [4] .
V1=−5, V2=−3, V3=−1,
V4=1, V5=3, V6=5
 F(V1)= −58.12, F(V2)= −10.58, F(V3)= −1.20, [5] F. Giri, M. M'Saad, J.M. Dion and L.
F(V4)=1.20, F(V5)=10.58, F(V6)=58.12 Dugard, "On the robustness of discrete-time
indirect linear adaptive controllers",
To check the actual convergence of the estimates to
Automatica, vol. 27, n°1, pp. 153-160, (1990).
their true values, we have let z(t)=0. The input
sequence, has been generated according to the law [6] F. Giri , F. Z. Chaoui and Y. Rochdi
(5.1). As one can see it in Figures 3 to 5, the ,’Parameter identification of a class of
parameter estimates provided by the identification Hammerstein plants’, Automatica, vol. 37,
algorithm (3.11a-b) do converge to their true 749-756, (2001).
values. Fig 4 represents the relative error estimation [7] G. C. Goodwin and K. S. Sin, ‘Adaptive
of F(.), as we can see it is always less than 0.1 %. filtering, prediction and control’, Prentice-Hall
(1983).
Conclusion
[8] M. Pawlak, “On the series expansion approach
We have considered nonlinear system identification to the of identification of Hammerstein
based on the Hammerstein model. The only systems’, IEEE-TAC, vol. 36, pp. 736-767,
assumptions are: (i) F is not identically null and, (1991).
(ii) F(v) takes a finite value for any v and F(0)=0.
We have designed an identification scheme that [9] P. Stoica and T. Söderström ‘Instrumental
determines precisely the parameters ai and bi (i=1, variable methods for identification of
-1
…, n) of the dynamic part (B(q )/A(q )) as well as -1 Hammerstein systems’, Int. J. control, vol. 35,
N points (Vj, F(Vj)), where N and the Vj’s are pp. 459-476 , (1982).
arbitrarily chosen. Parameter estimation is [10] S.A. Billings and S.Y. Fakhouri, ‘Nonlinear
performed using a gradient type algorithm (4.1a-b), system identification using the Hammerstein
based on the parameterization (3.8a-c), which is model’, Int. J. Syst. Sci., 10, pp. 567-578,
linear in the composed parameters µij. Then, the (1979).
( )
estimates of interest â i ( t ), b̂ i ( t ), d̂ i ( t ) are got by [11] W. Greblicki, ‘Non parametric orthogonal
performing a singular value decomposition of the series identification of Hammerstein systems’,

matrix M ( t ) = [µˆ ij ( t )] (see (4.4)-(4.5a-c)). The Int. J. Systems Sciences, vol. 20, pp. 2355-
convergence of the parameter estimates to their true 2367, (1989).
values has been ensured by resorting to the input [12].W. Greblicki, and M. Pawlak, ‘Nonparametric
sequence (5.1), that ensures the persistent identification of a cascade nonlinear time
excitation property (6.5). series system’, Signal Processing, vol. 22, pp.
61-75 (1991).
[13].W. Greblicki, and M. Pawlak, ‘Nonparametric
References recovering nonlinearities in block oriented
[1] A. Krzyzak, ‘Identification of discrete systems with the help of Laguerre
Hammerstein systems by the Fourier series polynomials’, Control-Theory and Advanced
regression estimate’, Int. J. Systems Sciences, Technology, vol. 10, part 1, pp. 771-791,
vol. 20, pp. 1729-1744 , (1989). (1994).
[2] E. Egardt, Springer Verlag, ‘Stability of [14] Z.Q. Lang, ‘Controller design oriented model
adaptive controllers’, (1979). identification method for Hammerstein
systems’, Automatica, vol. 29, pp. 767-771,
[3] F.C. Kung and D.H. Shinh, ‘Analysis and (1993).
identification of Hammerstein model nonlinear
systems using block-pulse function expansion’, [15] Z.Q. Lang ‘A nonparametric polynomial
Int. J. Control, vol. 43, pp. 139-147, (1986). identification algorithm for the Hammerstein
system’, IEEE-TAC, vol.42, pp.1435-1441,
[4] F. Giri et al: the full version of this paper. (1997).
 Dynamic gain
Static Gain
y(t)
v(t) u(t) B(q −1 )
F(.)
A(q −1 )

e(t) +
Identification
Algorithm _

û(t) B̂(q −1 ) Fig 3: Estimates of linear Parameters

F̂(.)
Â(q −1 ) ŷ (t)
Fig 1: Identification scheme

Fig 4: Estimation error of F

Fig 2: Non linear gain F

Fig 5: Estimates of Non linear Parameters

View publication stats