Académique Documents
Professionnel Documents
Culture Documents
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
Outline
Outline
References
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.
LSD - September 2005 J.C. Gómez 7
⎧ D , l=0
hl = ⎨ l−1
⎩CA B , l > 0
⎡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
X = [x1 , x2 , L , x N ]
⎡ C ⎤
⎢ CA ⎥
Γα = ⎢ ⎥ Extended (α > n)
⎢ M ⎥ Observability Matrix State Sequence Matrix
⎢ α −1 ⎥
⎣CA ⎦
LSD - September 2005 J.C. Gómez 13
⎡ 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 ⎥⎦
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
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.
LSD - September 2005 J.C. Gómez 15
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.
LSD - September 2005 J.C. Gómez 16
⎡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
−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
⎠
Γα = Γα Â shift-invariance property
• Bˆ and D
ˆ : solving a system of linear equations
Yα = Γα X + H α Uα + Nα
and
Yα Uα⊥ = Γα X + Nα Uα⊥
Block-oriented Nonlinear
Block-oriented Nonlinear Models
Models
Static Static
Nonlinearity LTI System LTI System Nonlinearity
uk yk uk yk
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
LSD - September 2005 J.C. Gómez 21
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
LSD - September 2005 J.C. Gómez 22
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
~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 )
Identification
IdentificationAlgorithm
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.
LSD - September 2005 J.C. Gómez 26
Result
Result2:
2:Consistency
ConsistencyAnalysis
Analysis
Wiener
Wiener Model
Model Identification
Identification
Problem
ProblemFormulation
Formulation
ωk υk
uk vk yk
A B
N(.)
C D
Fig. 7: Wiener model
⎨
(5) vk = N −1
( yk ) = ∑ α i g i ( yk ) (7)
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
LSD - September 2005 J.C. Gómez 28
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
IdentificationAlgorithm
Algorithm
⎧ xk +1 = Axk + Buk + ω k
⎧ 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
Normalization α
+
=1 Identifiability
2
problem
LSD - September 2005 J.C. Gómez 29
ωk υk
⎧ xk +1 = Axk + Buk + ω k
⎨ ~ ~ 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
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⎭
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
[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
0.64 × 10−8
Φν (ω ) = Spectrum of the zero
1.2 − 0.4 cos(ω )
mean coloured noise
LSD - September 2005 J.C. Gómez 34
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
Validation results
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.
LSD - September 2005 J.C. Gómez 38
Fig. 13: True (blue) and Estimated (red) Outputs (validation data)
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
LSD - September 2005 J.C. Gómez 42
Estimation and Validation data
Fig. 18: True (blue) and estimated (red) Output (Estimation/Validation data).
LSD - September 2005 J.C. Gómez 45
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.
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