Professional Documents
Culture Documents
Persistent Excitation by Deterministic Signals in Subspace MISO Hammerstein System Identification
Persistent Excitation by Deterministic Signals in Subspace MISO Hammerstein System Identification
Abstract— The persistent excitation issue is considered for guarantees the consistency of the subspace algorithm results,
MISO Hammerstein system identification using the subspace by using deterministic input sequences. To this end, a
approach. A signal generator that provides deterministic technical lemma is established showing that a class of PE
persistently exciting (PE) sequences is developed. These are input sequences for MISO Hammerstein systems can be
obtained by operating amplitude weighting and phase shifting
generated by operating phase-shifting and amplitude-
on a discrete Dirac Comb signal of finite length. The proposed
signal generation procedure presents quite interesting features weighting on a specific mother mono variable signal. The
e.g. the signals data samples are deterministic, weakly latter has been previously constructed by the authors for
constrained, and their size do not need to be too large. polynomial basis function (PBF) based subspace
Theoretically, the signal generator design relies on a technical identification of SISO Hammerstein system [5]. Compared
lemma establishing the PE property for the considered to white Gaussian noise (WGN) and multi-level pseudo
identified systems. The effectiveness of the proposed class of
random sequence (MLPRS), the proposed vector input
exciting sequences is confirmed by simulation.
sequence procedure presents quite interesting features: (i)
I. INTRODUCTION there is no (theoretical) need to large set of input and output
data samples; (ii) the design procedure provides the user
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
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
these, the most popular models are those composed of a system around an operation point.
linear dynamical subsystem and a static nonlinear element The paper is organized as follows: the identification
connected in series namely Hammerstein and Wiener problem is formulated in section II. In section III, the
models. Several identification schemes have been developed estimation procedure is described. In section IV, the main
for this class of systems following various approaches theorem, that describes key design elements for generating
including e.g. frequency, stochastic, blind, iterative PE sequences, is established. In section V The effectiveness
optimization and others (see e.g. Giri and Bai [1] and of the proposed sequence design procedure is illustrated by
references therein). Presently, the focus is made on the State numerical simulation.
Space Subspace Identification approach which captured a
particular attention in control systems community. This class II. IDENTIFICATION PROBLEM FORMULATION
of algorithms had been introduced and developed for MIMO
linear systems during the nineties by P. Van Overschee and A. Class of systems
B. De Moor [2] and references therein, and extended a We are considering MISO nonlinear systems that can be
decade after to both SISO and MIMO Hammerstein and well represented by the Hammerstein block-oriented multi
Wiener nonlinear systems by Lovera et al. [3], Gomez and channel model of Fig. 1. The model consists of m channels
Baeyens [4] and Naitali et al. [5]. each one being composed of a SISO Hammerstein system.
From a theoretical point of view, there is no denying that
subspace algorithms are accurate and robust. However, to
η1
guarantee the consistency of the system parameter estimates, F1 G1(q-1) y1
u1 + + y
the input sequence applied while the identification
experiment must be so that the virtual input sequence of the η2
G2(q-1) y2
equivalent multivariable linear system is Persistently u2 F2 +
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 1nr 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.
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
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,..,0T , Fr,dr r 1 0 .. 0VrT , 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 NN
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 i1, j ,U0 2 i1, 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 2i1, j 1V0T2i1, j 1
V0 2i1, j 2 V0T2i1, j ... V0 2i1, j mV0T2i1, j
1
0 otherwise
2V 1 T 2
V0 2i1, j 2 V0T2i1, j ... 2 V0 2i1, j mV0T2i1, j (46)
Consider a md variable signal U (t ) defined on the V
VVT 0 2i1, j 0 2i1, j
interval t 0,2i j 2 , for some integers i and j , by: ... ... ... ...
mV0 2i1, j 1V0T2i1, j m
V0 2i1, j 2V0T2i1, j ... mV0 2i1, j mV0T2i1, 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 di2 di , r c, (47)
(i) rank R2i r ,d (39)
r
It is checked that any non diagonal block V0 2i1, 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 2i1, j c V0T2i1, j r,d ur ( p k)c,d uc (q k)T
p0..2i1,
2id2id (48)
k0 q0..2i1
V (p,q)
U (0) U (1) .. U ( j 1) j1
U (1) U ( 2) .. U ( j) 2 i d j (42)
r c T
0 2i1, j V0 2i1, j
r,m ur ( p k) c,d uc (q k)T dd (49)
U 0 2i 1, j
.. .. .. .. k0
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,1ur ( p k) c,1uc (q k) ... r,1ur ( 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,1uc (q k) ... r,2 ur ( p k) c,d uc (q k) (50)
..
r (43)
V02i1, 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,1uc (q k) ... r, d ur ( p k) c, d uc (q k)
259
r , k 0 0 (38). Having (35), This amounts to let the (iv) j j0 max2i (d 1) (m 1)0 ,2i (md 1) (63)
integer i and the integer sequence r be such that the two
Proof: It was proved in [5] that, if sr , k sr , k for all k k ,
equalities (52) do not simultaneously hold where the
and any r 1, ..., m , then the covariance matrix of the vector
notation (mod 2i ) stands for the congruence modulo 2i .
sequence r ,d (u r (t )) defined by (55-56) is of full rank,
p k r 0 (mod 2i )
(52) provided that j is enough large so that any signal u r (t )
q k c 0 (mod 2i )
contains at least d pulses, i.e.:
Bearing this in mind and knowing that both p and
strictly between 0 and 2i-1, if (40) and (41) hold, it
q vary
comes
rank R2i r ,d
(u r ), r ,d (u r )
2di *
(64)
that the following double inequality (53) is satisfied, 2i j 1 2i ( d 1) ( m 1) (65)
0 r c ( p q) 2i (53) Therefore since that r , d 0 0 d 1 , and that the arithmetic
Consequently (52) do not holds, therefore all the off
progression sequences r (58) is strictly increasing, all the
diagonal blocks of V V T are zeros. Furthermore, to make
conditions (38-41) of the lemma are satisfied provided that
possible the full rank property of R2Ui ,U the number columns 2i . Consequently, and knowing that in the subspace
of the Hankel matrix U 0 2i 1, j must be enough large, identification it is assumed on one hand that the half number
( i ) of block rows of the input Hankel matrix must be strictly
i.e. j 2mdi . Consequently there exist j0 2mdi and larger than the system order ( N ), and on the other hand that
i0* i ( m 1 ) / 2 such that , for any i i0 , and any
its number of columns ( j ) must be larger than the number
j j0 , the covariance matrix R2Ui ,U is of full rank and of rows ( 2i (md 1) ) of the matrix concatenating the input
which is equal to 2mdi , which establishes the lemma ■ and output Hankel matrices [2], there exist i0 N 1 ,
C. Pulse Based Input PE Signals for Subspace 0 2i , i0* i (m 1) / 2 and
Identification of MISO Hammerstein Systems j0 max(2i(md 1),2i (d 1) (m 1) 2i 1) , such that for any
th
Theorem. Consider the N order linear input-output state i i0 , 0 , i i0 and j j0 , the covariance matrix
space model of the MISO Hammerstein system (17) subject
to the multivariable input signal U (t ) defined on the R2Ui ,U is of full rank ( 2mdi ) and thereby the input sequence
interval t 0,2i j 2 , for some integers i and j , by: 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)) 1ur (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
260
the input range of the r th input channel of the system. The 0.5
0.2
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
0.1
-0.15
0 .1255 q 1 0 .1163 q 2
0 100 200 300 400 500 600
1 1 . 569 q 1 0 . 8111 q 2 Fig. 5. True (-) and predicted (..) output sequences of the considered 3-
1 2 channels Hammerstein system
0 .07425 q 0 .07217 q
Gˆ 2 ( q 1 ) (74)
1 1 . 778 q 1 0 . 9245 q 2
0 .07218 q 1 0 .0687 q 2 REFERENCES
Gˆ 3 ( q 1 ) (75)
1 1 . 714 q 1 0 . 8547 q 2 [1] F. Giri, E.W. Bai (Eds), Block-oriented Nonlinear system
Identification, Springer, U.K., 2010.
It is readily seen, comparing (73-75) and (66-68), that the [2] P. Van Overschee and B. De Moor, Subspace identification for linear
transfer function coefficients are accurately estimated. The systems: Theory-Implementation-Applications, Kluwer Academic
coefficient estimates of the 7th degree-polynomial Publishers, pp 95-134, 1996.
nonlinearities are: [3] M. Lovera, T. Gustafsson and M. Verhaegen, “Recursive subspace
identification of linear and non-linear Wiener state-space models”,
Fˆ1,7 1 0.041 6.255 0.316 10.213 0.431 5.122 (76) Automatica, 36, pp 1639-1650, 2000.
Fˆ2,7 1 0.03 1.596 0.188 2.207 0.221 1.197
[4] J.C. Gómez, and E. Baeyens, “Hammerstein and Wiener Model
(77)
Identification using Rational Orthonormal Bases”. Latin American
Fˆ3,7 1 0.002 1.99 0.02 1.83 0.036 0.77 (78) 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”,
Fig. 4 shows the graphical characteristics of the system IFAC Symposium on System Identification (SYSID), Rotterdam, The
Netherlands, pp 1327-1332, 2003.
polynomial static nonlinearities and their estimates. It is seen
that the estimated curves are quite close to the true curves.
261