Seminar LSD 2005 BW

Research Seminar
Subspace Identification of
Hammerstein and Wiener Models
Speaker
Juan C. Gómez
Laboratory for System Dynamics and Signal Processing
FCEIA, Universidad Nacional de Rosario
ARGENTINA
jcgomez@fceia.unr.edu.ar
LSD - September 2005 J.C. Gómez 1
Outline
Outline
Introduction: Motivation, New results

A (very)
very brief review on Subspace State-Space System IDentifi-
cation Methods
Block-oriented Nonlinear Models
Subspace Identification of Hammerstein Models
Subspace Identification of Wiener Models
Simulation Examples
Conclusions

Introduction
Introduction
Motivation for Nonlinear ((Subspace)
Subspace) Identification
• Most physical processes have a nonlinear behaviour, except in a
limited range where they can be considered linear.
• The performance of controllers designed from a linear approxi-
mation is strongly influenced by a change in the operating point
of the system.
• Nonlinear models are able to describe more accurately the global
behaviour of the system, independently of the operating point.
• Many dynamical systems can be represented by the inter-
connection of static nonlinearities and LTI systems. These models
are called block-oriented nonlinear models.
• Subspace Methods have been very successful for the identification of

LTI models in many practical applications.
• Although there is a well developed theory for Subspace Identification
methods for LTI systems, this is not the case for nonlinear systems.
Some recent contributions in this area are: (Verhaegen & Westwick,
1996) in Subspace Identification of Hammersterin and Wiener models,
and (Chen & Maciejowski, 2000) and (Favoreel et al., 1999) in
Subspace Identification of bilinear systems.

The new results (Gomez & Baeyens, 2005)
• New subspace algorithms for the simultaneous identification of the
linear and nonlinear parts of multivariable Hammerstein and Wiener
models are presented.
• The proposed algorithms consist basically of two steps:
Step 1: a standard (linear) subspace algorithm applied to an
equivalent linear system whose inputs (outputs) are filtered (by
the basis functions describing the static nonlinearities)
versions of the original inputs (outputs).
Step 2: a 2-norm minimization problem which is solved via
an SVD.
• Provided the conditions for the consistency of the linear subspace
algorithm used in Step 1 are satisfied, consistency of the estimates can
be guaranteed.
References
1. Gómez, J.C. and Baeyens, E.. Subspace Identification of

Multivariable Hammerstein and Wiener Models, European
Journal of Control, Vol. 11, No. 2, 2005.
2. Gómez, J.C., Jutan, A. and Baeyens, E.. Wiener Model
Identification and Predictive Control of a pH Neutralization
Process. IEE Proceedings on Control Theory and Applications,
Vol. 151, No. 3, pp. 329-338, May 2004.

SSubspace
ubspace SState-Space
tate-Space SSystem
ystem ID entification
IDentification
4SID Methods
Properties
They combine tools of System Theory,
Theory Numerical Linear Algebra
and Geometry (projections).
They have their origin in Realization Theory as developed in the
60/70s (Ho & Kalman, 1966).
They provide reliable state-space models of multivariable LTI
systems directly from input-output data.
They don’t require iterative optimization procedures Æ no problems
with local minima, convergence and initialization.
They don’t require a particular (canonical) state-space realization Æ

numerical conditioning improves.
They require a modest computational load in comparison to tradi-
tional identification methods like PEM.
The algorithms can be (they have been) efficiently implemented in
software like Matlab.
Matlab
Main computational tools are QR and SVD.
All subspace methods compute at some stage the subspace spanned
by the columns of the extended observability matrix.
The various algorithms (e.g., N4SID, MOESP, CVA) differ in the
way the extended observability matrix is estimated and also in the way
it is used to compute the system matrices.

The system model
xk +1 = Axk + Buk + Kek State-space model in

y k = Cxk + Duk + ek innovation form
The identification problem

To estimate the system matrices (A, B, C, D) and K , and the model
order n, from an (N+α-1)-point data set of input and output
measurements
{uk , yk }kN=+1α −1
Realization-based 4SID Methods

For a LTI system, a minimal state-space realization (A, B, C, D)
completely defines the input-output properties of the system through
∞
yk = ∑ hluk −l convolution sum
l =0
where the impulse response coefficients hl are related to the system

matrices by
⎧ D , l=0
hl = ⎨ l−1
⎩CA B , l > 0

⎡ h1 h2 L hj ⎤
⎢h h3 h j +1 ⎥⎥
H ij = ⎢
2 L H ij = Γi C j
⎢M M O M ⎥
⎢ ⎥
⎢⎣ hi hi +1 L hi + j −1 ⎥⎦
Impulse Response Extended Extended
Hankel Matrix Observability Controlability
Matrix Matrix
(i>n) (j>n)
An estimate of the extended observability matrix can be computed by a

full rank factorization of the impulse response Hankel matrix. This
factorization is provided by the SVD of matrix Hij.
⎡Σ1 0 ⎤ ⎡V1T ⎤
H ij = [U1 U 2 ]⎢ ⎛ 2 ⎞⎛ 2 T ⎞
1 1
⎢
⎥ T ⎥ ≈ U Σ V T
= ⎜ U Σ ⎟⎜ Σ1 V1 ⎟
⎣ 0 Σ 2 ⎦ ⎣V2 ⎦
1 1 1
⎝
1
1 1
424 31 ⎠⎝ 424 3⎠
Γˆ i Ĉ j
rank reduction
In the absence of noise, Hij will be a rank n matrix, and Σ1 will contain
the n non-zero singular values → model order is computed. In the
presence of noise, Hij will have full rank and a rank reduction stage will
be required for the model order determination.
Problems: it is necessary to measure or to estimate (for example, via
correlation analysis) the impulse response of the system → not good

Direct 4SID Methods
Yα = Γα X + H α Uα + Nα fundamental equation (1)
⎡ y1 y2 L yN ⎤ Output block Hankel matrix

⎢ y y3 L y N +1 ⎥⎥
Yα = ⎢ 2 (In a similar way are defined the
⎢ M M M M ⎥ Input block Hankel matrix Uα and
⎢ ⎥
⎣ yα −1 yα L y N +α −1 ⎦ the Noise block Hankel matrix Nα.)
X = [x1 , x2 , L , x N ]
⎡ C ⎤
⎢ CA ⎥
Γα = ⎢ ⎥ Extended (α > n)
⎢ M ⎥ Observability Matrix State Sequence Matrix
⎢ α −1 ⎥
⎣CA ⎦
⎡ D 0 0 L 0⎤
⎢ CB D 0 L 0 ⎥⎥ Lower triangular block
⎢
H α = ⎢ CAB CB D L 0⎥ Toeplitz matrix of impulse
⎢ ⎥ responses (unknown).
⎢ M M M O M⎥
⎢⎣CA B CA B CA − 4 B
α − 2 α − 3 α
L D ⎥⎦
The main idea of Direct 4SID methods

In the absence of noise (Nα = 0), eq. (1) becomes
Yα = Γα X + H α Uα (2)
and the part of the output which does not emanate from the state can
be removed by multiplying (from the right) both sides of eq. (2) by the
orthogonal projection onto the null space of Uα, i.e. by

( )
∆ ∆
⊥ T −1
Π UαT
= I − Uα Uα Uα
T
Uα = Uα⊥ orthogonal projection
such that Uα Uα⊥ = I

This yields
Yα Uα⊥ = Γα XUα⊥ (3)
Note that the matrix on the left depends exclusively on the input-output
data. Then, a full rank factorization of this matrix will provide an
estimate Γ̂α of the extended observability matrix. Estimates of the
corresponding system matrices can be obtained by resorting to the shift
invariance property of the extended observability matrix, and by solving
a system of linear equations in the least squares sense.
The factorization is provided by the SVD of the matrix on the left side
⎡Σ1 0 ⎤ ⎡V1T ⎤
Yα Uα = [U1 U 2 ]⎢ ⎛ 2 ⎞⎛ 2 T ⎞
1 1
⊥
⎢ T ⎥ ≈ U1Σ1V1 = ⎜⎝U1Σ1 ⎟⎠⎜⎝ Σ1 V1 ⎟⎠ (4)
T
⎥
⎣ 0 Σ 2 ⎦ ⎣V2 ⎦ 1 4243
Γˆα
rank reduction
(model order estimation)
(In the absence of noise Σ2 = 0)
Weighting Matrices
Row and column weighting matrices can be introduced in (4) before
performing the SVD of the matrix in the left hand side. Any choice of
positive-definite weighting matrices Wr and Wc will result in consistent
estimates of the extended observability matrix.
⎡Σ1 0 ⎤ ⎡V1T ⎤
Wr Yα UαWc = [U1
⊥
U 2 ]⎢ ⎥⎢ T ⎥ ≈ U Σ V T
= ⎛⎜U Σ 12 ⎞⎟⎛⎜ Σ 12V T ⎞⎟
⎣ 0 Σ V
2 ⎦⎣ 2 ⎦
1 1 1
⎝
1
1 1
424 3 ⎠ ⎝ 1 1
⎠
Γˆα
change of coordinates in state-space

Existing algorithms employ the following choices for matrices Wr and Wc ,
−1
• MOESP (Verhaegen, 1994): Wr = I , Wc = ⎛⎜ ΦΠ U⊥ T Φ T ⎞⎟ ΦΠ U⊥ T

1
⎝N α
⎠ α
−1 −1
⎛1 ⊥ T ⎞
2
⎛1 ⊥ T ⎞
2
• CVA (Larimore, 1990): Wr = ⎜ Yα ΠU T Yα ⎟ , Wc = ⎜ ΦΠ U T Φ ⎟
⎝N α
⎠ ⎝N α
⎠
• N4SID (Van Overschee and de Moor, 1994):

−1
⎛1 ⎞
Wr = I , Wc = ⎜ ΦΠ U⊥ T Φ T ⎟ Φ
⎝N α
⎠
Computation of the system matrices
Given an estimate Γ̂α of the extended observability matrix, estimates

of the system matrices can be computed as:
• Ĉ : first row block of Γ̂α

• Â : solving in the least squares sense
Γα = Γα Â shift-invariance property
• Bˆ and D
ˆ : solving a system of linear equations

Presence of noise
In the presence of noise
Yα = Γα X + H α Uα + Nα
and
Yα Uα⊥ = Γα X + Nα Uα⊥
noise term needs to be

removed
The noise term can be removed by correlating it away with a suitable
matrix. This can be interpreted as an oblique projection.
Block-oriented Nonlinear
Block-oriented Nonlinear Models
Models
Static Static
Nonlinearity LTI System LTI System Nonlinearity
uk yk uk yk
Fig. 1: Hammerstein Model (NL) Fig. 2: Wiener Model (LN)
Static Static Static

Nonlinearity LTI System Nonlinearity LTI System Nonlinearity LTI System
uk yk
uk yk
Fig. 3: Hammerstein-Wiener Model (NLN) Fig. 4: Hammerstein-Wiener Model (LNL)

Hammerstein
Hammerstein Model
Model Identification
Identification
Problem
ProblemFormulation
Formulation
ωk υk
vk
uk A B yk
N(.)
C D
Fig. 5: Hammerstein model
LTI subsystem Nonlinear subsystem
r
⎧ xk +1 = Axk + Bvk + ω k
⎨
(1) v k = N (u k ) = ∑ α g (ui i k ) (3)
⎩ y k = Cxk + Dvk + υ k (2) i =1
g i (• ) : ℜ p
→ ℜ , (i = 1, L , r ) known basis
p
yk ∈ ℜ m , xk ∈ ℜ n , vk ∈ ℜ p functions
ωk ∈ ℜn , υk ∈ ℜm α i ∈ ℜ p× p (i = 1,L, r ) unknown matrix
parameters
Identification
Identificationproblem
problem: : totoestimate
estimatethe
theunknown
unknownparameter
parametermatrices
matrices
α i ∈ ℜ p× p , (i = 1,L , r ) , and A, B, C, and D characterizing the nonlinear and the linear
, and A, B, C, and D characterizing the nonlinear and the linear
set {uk , yk }k =1 of
N
parts,
parts,respectively,
respectively,and andthethemodel
modelorder
ordern,n, from
froman
anN-point
N-pointdata
dataset of
observed
observedinput-output
input-outputmeasurements.
measurements.
Subspace
SubspaceIdentification
IdentificationAlgorithm
Algorithm
⎧ r
⎪ xk +1 = Axk +
⎪
∑
Bα i g i (u k ) + ω k
(3) → (1), (2) ⇒ ⎨ r
i =1
⎪
⎪ y k = Cxk + ∑
Dα i g i (u k ) + υ k
⎩ i =1
Normalization αi 2
=1 Identifiability
problem
Defining
~∆ ~∆ ∆
[
B = [B α 1 , L , B α r ] , D = [D α 1 , L , D α r ] , U k = g 1T (u k ), L , g r T (u k )]T
ωk υk
⎧⎪ xk +1 = Axk + B ~
Uk + ωk
⎨ ~ yk
~
⎪⎩ y k = Cx k + DU k + υ k
Uk A B
~
C D
Fig. 6: Equivalent LTI system

with input Uk
Linear Subspace Algorithms
(N4SID, MOESP,CVA)
~ˆ ~ˆ
Estimates Aˆ , B , Cˆ , D, model order n
~
α = [α1 , L , α r ]T , then B ~
Defining = Bα T , and D = Dα T , so that
~
⎡B ⎤ ⎡B⎤ T
∆
Θ BD = ⎢ ~ ⎥ = ⎢ ⎥α
⎣D⎦ ⎣D⎦
The problem then is how to compute estimates of matrices B, D, and α
~ ~
from the estimate of the matrices B , and D (i.e., from an estimate of Θ BD )
It is clear that the closest, in the 2-norm sense, estimates Bˆ , Dˆ , and α̂

are such that
⎧⎪ 2
⎫⎪
( )
Bˆ , Dˆ , αˆ = argmin ⎨ Θ ˆ − ⎡ B ⎤α T
BD ⎢D⎥ ⎬
B , D ,α ⎪
⎩ ⎣ ⎦ 2⎪⎭
The solution to this optimization problem is provided by the SVD of Θ̂ BD .
Result
Result11
Let Θ ˆ ∈ ℜ ( n + m )×rp have rank s>p, and let its economy size SVD be
BD
partitioned as
s
⎡Σ1 0 ⎤ ⎡V1T ⎤
Θ BD = UΣV = ∑ σ i ui vi = [U1 U 2 ]⎢
ˆ T T
⎥⎢ T ⎥
(4)
i =1 ⎣ 0 Σ 2 ⎦ ⎣V2 ⎦
with U 1 ∈ ℜ ( n + m )× p , V1 ∈ ℜ rp × p , and Σ 1 = diag (σ 1 , σ 2 , L , σ p ) .
Then 2
⎛ ⎡ Bˆ ⎤ ⎞
⎜ ⎢ ⎥ , αˆ ⎟ = argmin Θ ˆ − ⎡ B ⎤α T = (U 1Σ 1 , V1 ),
⎜ Dˆ ⎟ BD ⎢D⎥
⎝⎣ ⎦ ⎠ B , D ,α ⎣ ⎦ 2
and the approximation error is given by Normalization

ˆ
2 in α provided by
ˆ − ⎡⎢ B ⎤⎥α T
Θ = σ p2 +1. the SVD
BD
ˆ
⎣D⎦ 2
Identification
Algorithm
The subspace algorithm can be summarized as follows.
~ ~
Step 1: Compute estimates of the system matrices A, B , C , D , and the ( )
model order n, using any available (linear) subspace algorithm, such as
N4SID, MOESP, CVA.
~ˆ ~ˆ
Step 2: Based on the estimates B and D compute an estimate Θ̂ BD of
matrix Θ BD .
Step 3: Compute the SVD of Θ̂ BD and its partition as in (4).
Step 4: Compute the estimates of the parameter matrices B, D, and α as
⎡ Bˆ ⎤
⎢ ˆ ⎥ = U1Σ1
⎣ D⎦
αˆ = V1
respectively.
Result
Result2:
2:Consistency
ConsistencyAnalysis
Analysis
Under some assumptions on persistency of excitation of the inputs, which

depend on the particular subspace method used in Step 1 of the algorithm,
⎛ ˆ ~ˆ ˆ ~ˆ ⎞
the estimates ⎜ A , B , C , D ⎟ are consistent in the sense that they converge
⎝ ⎠
to the true values when the number of data points N →∞ .
~ˆ ~ˆ
The consistency of B and D , implies that of B, D, and α .
Wiener
Wiener Model
Model Identification
Identification
Problem
ProblemFormulation
Formulation
ωk υk
uk vk yk
A B
N(.)
C D
Fig. 7: Wiener model
LTI subsystem Nonlinear subsystem

⎧ xk +1 = Axk + Bu k + ω k
r
⎨
(5) vk = N −1
( yk ) = ∑ α i g i ( yk ) (7)
⎩vk = Cxk + Duk + υ k (6) i =1
g i (• ) : ℜ m → ℜ m , (i = 1, L , r ) known basis
u k ∈ ℜ , xk ∈ ℜ , vk ∈ ℜ
p n m
functions
ωk ∈ ℜn , υk ∈ ℜm α i ∈ ℜ m× m (i = 1,L, r ) unknown matrix
parameters
Identification
Identificationproblem : totoestimate
problem: estimatethe
theunknown
unknownparameter
parametermatrices
matrices
α i ∈ ℜ m×m , (i = 1, L , r ), ,and
and A,
A,B,
B,C,
C,and
andDD characterizing
characterizingthe
thenonlinear
nonlinearand
andthe
the
set {uk , yk }k =1
N
linear
linearparts,
parts,respectively,
respectively,and
andthe
themodel
modelorder
ordern,n, from
froman
anN-point
N-pointdata
dataset
of
ofobserved
observedinput-output
input-outputmeasurements.
measurements.
Subspace
SubspaceIdentification
Algorithm
⎧ xk +1 = Axk + Buk + ω k
(7) → (6) ⇒ ⎪⎨ ∆ r ⇒ ⎨ ~ ~
⎪αYk = ∑ α i g i ( yk ) = Cxk + Duk + υ k ~
⎩Yk = Cxk + Duk + υ k
⎩ i =1
α = [α1 ,L ,α r ] , Yk = [g1T ( yk ),L, g rT ( yk )] ~∆ + ~∆ +

T
C =α C , D =α D
Normalization α
+
=1 Identifiability
2
problem
ωk υk
⎨ ~ ~ uk A B Yk
Y
⎩ k = C x k + D uk + υ~k ~ ~
C D
Fig. 8: Equivalent LTI model
with output Yk
Linear Subspace Algorithms
(N4SID, MOESP,CVA)
~ˆ ~ˆ
Estimates Aˆ , Bˆ , C , D, model order n

The problem is how to compute estimates of matrices C, D, and α+
~ ~
from the estimates of the matrices C , and D
Similarly to what was done for the Hammerstein model the closest, in the
2-norm sense, estimates Cˆ , Dˆ , and α̂ + are such that
(Cˆ , Dˆ ,αˆ ) ⎧ ~ˆ ⎫
2
~ˆ
+
= argmin ⎨ ⎡C D ⎤ − α + [C D] ⎬
⎢
C , D ,α + ⎩ ⎣
⎥⎦ 2⎭
The solution to this optimization problem is provided by the SVD of

the matrix ⎡C~ˆ D~ˆ ⎤
⎢⎣ ⎥⎦
Result
Result33
⎡ ~ˆ ~ˆ ⎤ mr ×( n + p )
Let ⎢C D ⎥ ∈ ℜ have rank s>m, and let its economy size SVD
⎣ ⎦
be partitioned as
⎡C~ˆ ~ˆ ⎤ s
⎡Σ1 0 ⎤ ⎡V1T ⎤
D = UΣV = ∑ σ i ui vi = [U1 U 2 ]⎢
T T
⎥⎢ T ⎥
(8)
⎢⎣ ⎥⎦ i =1 ⎣ 0 Σ 2 ⎦ ⎣V2 ⎦
with U 1 ∈ ℜ mr × m , V1 ∈ ℜ ( n + p )× m , and Σ 1 = diag (σ 1 , σ 2 , L , σ m ) .
Then
(αˆ , [Cˆ ])
2
+
Dˆ = argmin ⎡C
⎢⎣
~ˆ ~ˆ
D ⎤ − α + [C
⎥⎦
(
D ] = U 1 , Σ 1V1T , )
C , D ,α + 2
and the approximation error is given by Normalization

in α+ provided
[ ]
2
⎡C~ˆ ~ˆ
D ⎤ − αˆ + Cˆ Dˆ = σ m2 +1. by the SVD
⎢⎣ ⎥⎦ 2

Identification
Algorithm
The subspace algorithm can be summarized as follows.
~ ~
(
Step 1: Compute estimates of the system matrices A, B, C , D , and the )
model order n, using any available (linear) subspace algorithm, such as
N4SID, MOESP, CVA.
~ˆ ~ˆ
Step 2: Compute the SVD of ⎡⎢C D ⎤⎥ and its partition as in (8).
⎣ ⎦
Step 3: Compute the estimates of the parameter matrices C, D, and α+ as
[Cˆ Dˆ ] = Σ V 1 1
T
αˆ = U1+
respectively.
Simulation
Simulation Examples
Examples
Example
Example1:
1:Hammerstein
HammersteinModel
ModelID
ID((“academic”)
“academic”)
The True System
z 2 + 0 .7 z − 1 .5 linear subsystem
G( z) =
z 3 + 0.9 z 2 + 0.15 z + 0.002
N (u k ) = 0.8589 u k + 0.0149 u k2 − 0.5113 u k3 − 0.0263 u k4 nonlinear subsystem
The input and noise

uk = sin (0.0005πk ) + 0.5 sin (0.0015πk ) + input
+ 0.3 sin (0.0025πk ) + 0.1 sin (0.0035πk ) + γ k ( γ k white noise with
variance 10 −6 )
0.64 × 10−8
Φν (ω ) = Spectrum of the zero
1.2 − 0.4 cos(ω )
mean coloured noise
The Estimated Nonlinear Subsystem
Nˆ (u k ) = 0.8589 u k + 0.0142 u k2 − 0.5113 u k3 − 0.0260 u k4 Estimated nonlinear subsystem
Fig.9: True (blue) and Estimated (green)

nonlinear characteristic.
The Estimated Linear Subsystem

0.9986 z 2 + 0.6997 z − 1.4984
Gˆ (z ) = Estimated linear subsystem
z 3 + 0.9002 z 2 + 0.1495 z + 0.0014
Validation results
Fig. 10: True (green) and Estimated (blue) Output.

Example
Example2:
2:Hammerstein
HammersteinModel
ModelID
ID((Binary
Binary Distillation
DistillationColumn
Column))
Input: reflux ratio (u)
Outputs: overhead flow rate (y1)

overhead methanol concentration (y2)
bottom flow rate (y3)
bottom methanol concentration (y4)
Fig. 11: Schematic representation of the

distillation column
(Weischedel & McAvoy, 1980)
Fig. 12: Left Plot: Estimation (first 1000 points), and validation (remaining
1000 points) Input Data. Right Plot: Estimation (first 1000 points) and
Validation (remaining 1000 points) Output Data.
Fig. 13: True (blue) and Estimated (red) Outputs (validation data)

Third order model with eigenvalues at {0.4916, 0.9557 , 0.9726}
Third order polynomial
Fig. 14: Estimated Nonlinear

Characteristic
Example
Example3:
3:Wiener
WienerModel
ModelID
ID((pH
pH Neutralization
NeutralizationProcess
Process))
• base: NaOH acid: HNO3

(u 3 )
buffer: NaHCO3
• Manipulated variable: base
flow rate (u1)
• Disturbances: buffer flow rate
effluent solution (u2) and acid flow rate (u3)
• Output: pH of the effluent
Fig. 15: Schematic representation of the pH solution (y)
Neutralization Process
(Henson & Seborg, 92, 94, 97)
Simulation Model based on first principles (introducing two reaction

invariants for each inlet stream)
x& = f ( x) + g ( x)u1 + p ( x)u2
h ( x, y ) = 0
where ∆
x =[x1 , x2 ] = [Wa , Wb ]
T T
T
⎡u u ⎤
f ( x) = ⎢ 3 (Wa 3 − x1 ), 3 (Wb 3 − x2 )⎥
⎣V V ⎦
T
⎡1 1 ⎤
g ( x) = ⎢ (Wa1 − x1 ), (Wb1 − x2 )⎥
⎣V V ⎦
T
⎡1 1 ⎤
p ( x) = ⎢ (Wa 2 − x1 ), (Wb 2 − x2 )⎥
⎣V V ⎦
y −14 −y 1 + 2 ×10 y − pK 2
h( x, y ) = x1 + 10 − 10 + x2
1 + 10 pK1 − y + 10 y − pK 2
Estimation and Validation data
Fig. 16: Estimation (first 1000 points) and validation

(remaining 600 points) input-output data.

0.0062 z 2 − 0.0122 z + 0.006
Third order model Gˆ (z ) =
z 3 − 2.9466 z 2 + 2.8940 z − 0.9474
Third order polynomial
Nˆ −1 ( yk ) = 0.0319 yk3 + 0.0358 yk2 + 0.9989 yk
Fig. 17: Estimated Nonlinear Characteristic.

Validation results
Fig. 18: True (blue) and estimated (red) Output (Estimation/Validation data).
Conclusions
Conclusions
• New subspace methods for the simultaneous identification of the
linear and nonlinear parts of multivariable Hammerstein and
Wiener models have been presented.
• The proposed methods make use of a standard (linear) subspace
method followed by a 2-norm minimization problem which is
solved via an SVD.
• The proposed methods generalize all the families of linear
subspace methods to this class of nonlinear models.
• The method provides consistent estimates under the same
conditions on persistency of excitation required by the (linear)
subspace method used as the first step of the algorithm.
• The estimated models are in a format which is suitable for their
use in standard (linear) Model Predictive Control schemes.

Research Seminar
Subspace Identification of
Hammerstein and Wiener Models
Speaker
Juan C. Gómez
Laboratory for System Dynamics and Signal Processing
FCEIA, Universidad Nacional de Rosario
ARGENTINA
jcgomez@fceia.unr.edu.ar

Seminar LSD 2005 BW

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Seminar LSD 2005 BW

Transféré par

Droits d'auteur :

Formats disponibles

Research Seminar

LSD - September 2005 J.C. Gómez 1

 Introduction: Motivation, New results

LSD - September 2005 J.C. Gómez 2

LSD - September 2005 J.C. Gómez 3

• Subspace Methods have been very successful for the identification of

LSD - September 2005 J.C. Gómez 4

1. Gómez, J.C. and Baeyens, E.. Subspace Identification of

LSD - September 2005 J.C. Gómez 6

 They don’t require a particular (canonical) state-space realization Æ

LSD - September 2005 J.C. Gómez 8

xk +1 = Axk + Buk + Kek State-space model in

 The identification problem

LSD - September 2005 J.C. Gómez 9

 Realization-based 4SID Methods

where the impulse response coefficients hl are related to the system

LSD - September 2005 J.C. Gómez 10

An estimate of the extended observability matrix can be computed by a

LSD - September 2005 J.C. Gómez 11

LSD - September 2005 J.C. Gómez 12

Yα = Γα X + H α Uα + Nα fundamental equation (1)

⎡ y1 y2 L yN ⎤ Output block Hankel matrix

 The main idea of Direct 4SID methods

LSD - September 2005 J.C. Gómez 14

such that Uα Uα⊥ = I

Yα Uα⊥ = Γα XUα⊥ (3)

change of coordinates in state-space

• MOESP (Verhaegen, 1994): Wr = I , Wc = ⎛⎜ ΦΠ U⊥ T Φ T ⎞⎟ ΦΠ U⊥ T

• N4SID (Van Overschee and de Moor, 1994):

 Computation of the system matrices

Given an estimate Γ̂α of the extended observability matrix, estimates

• Ĉ : first row block of Γ̂α

LSD - September 2005 J.C. Gómez 18

noise term needs to be

LSD - September 2005 J.C. Gómez 19

Fig. 1: Hammerstein Model (NL) Fig. 2: Wiener Model (LN)

Static Static Static

Fig. 3: Hammerstein-Wiener Model (NLN) Fig. 4: Hammerstein-Wiener Model (LNL)

LSD - September 2005 J.C. Gómez 20

Fig. 6: Equivalent LTI system

LSD - September 2005 J.C. Gómez 23

It is clear that the closest, in the 2-norm sense, estimates Bˆ , Dˆ , and α̂

and the approximation error is given by Normalization

LSD - September 2005 J.C. Gómez 25

Under some assumptions on persistency of excitation of the inputs, which

LSD - September 2005 J.C. Gómez 27

LTI subsystem Nonlinear subsystem

⎩vk = Cxk + Duk + υ k (6) i =1

α = [α1 ,L ,α r ] , Yk = [g1T ( yk ),L, g rT ( yk )] ~∆ + ~∆ +

LSD - September 2005 J.C. Gómez 30

The solution to this optimization problem is provided by the SVD of

LSD - September 2005 J.C. Gómez 31

and the approximation error is given by Normalization

LSD - September 2005 J.C. Gómez 32

LSD - September 2005 J.C. Gómez 33

N (u k ) = 0.8589 u k + 0.0149 u k2 − 0.5113 u k3 − 0.0263 u k4 nonlinear subsystem

 The input and noise

Fig.9: True (blue) and Estimated (green)

 The Estimated Linear Subsystem

Fig. 10: True (green) and Estimated (blue) Output.

Input: reflux ratio (u)

Outputs: overhead flow rate (y1)

Introduction: Motivation, New results

They don’t require a particular (canonical) state-space realization Æ

The identification problem

Realization-based 4SID Methods

The main idea of Direct 4SID methods

Computation of the system matrices

The input and noise

The Estimated Linear Subsystem

The Estimated Linear Subsystem

Simulation Model based on first principles (introducing two reaction

The Estimated Linear Subsystem

The Estimated Nonlinear Subsystem