NONPARAMETRIC APPROACH TO IDENTIFICATION
OF BLOCK-ORIENTED SYSTEMS
Grzegorz Mzyk, Zygmunt Hasiewicz
The Institute of Computer Engineering, Control and Robotics
Wrocław University of Technology
Wrocław, Poland
Statement of the problem
k
u
k
y
k
z
k
w
( )
µ
{ }
∞
=
0
i
i
γ
k
v
Fig.1. Hammerstein system
( )
{
|
}
( )
k
k
R u
y
u
u
c u
s
µ
=
=
=
+
E
∑
∞
=
−
+
=
0
)
(
i
k
i
k
i
k
z
u
y
µ
γ
NONPARAMETRIC METHODS
Kernel regression estimation
1
1
ˆ ( )
( )
( )
M
M
i
i
M
i
i
i
u u
u u
u
y K
K
h M
h M
µ
=
=
−
−
=
∑
∑
Theorem 1. If
( )
0
h M
→
and
( )
Mh M
→ ∞
as
M
→ ∞
then
ˆ ( )
( )
M
u
u
µ
µ
→
in probability as
M
→ ∞
in each continuity point of ()
µ
and the input probability density.
Theorem 2. If
()
µ
is twice differentiable in the point
u
then for
1/ 5
( )
(
)
h M
O M
−
=
it holds that
2/ 5
ˆ ( )
( )
(
)
M
u
u
O M
µ
µ
−
−
=
Orthogonal expansion
( )
( )
( )
u
g u
f u
µ
=
, where
( )
( )
( )
u
g u
f u
µ
=
0
0
( )
( ),
( )
( )
i i
i i
i
i
g u
a
u
f u
b
u
φ
φ
∞
∞
=
=
=
=
∑
∑
( ),
( )
i
k
i
k
i
i
k
a
y
u
b
u
φ
φ
=
=
E
E
1
1
1
1
ˆ
ˆ
( ),
( )
M
M
i
k i
k
i
i
k
k
k
a
y
u
b
u
M
M
φ
φ
=
=
=
=
∑
∑
(
)
0
( )
0
ˆ
( )
ˆ ( )
ˆ ( )
q M
i i
i
M
q M
i i
i
u
a
u
b
u
µ
φ
φ
=
=
=
∑
∑
The convergence conditions:
trigonometric series
2
( )
lim
0
M
q M
M
→∞
=
Legendre series
2
( )
lim
0
M
q M
M
→∞
=
Laguerre series
6
( )
lim
0
M
q M
M
→∞
=
Hermite series
5/ 3
( )
lim
0
M
q
M
M
→∞
=
Daubechies wavelets
2 ( ) 2
2
lim
0
q M
M
M
+
→∞
=
COMBINED PARAMETRIC-
NONPARAMETRIC ALGORITHMS
Estimation of the static nonlinearity
( )
( , )
k
k
u
u c
µ
µ
=
1
2
( )
sin(
)
c u
k
k
u
e
c u
µ
=
+
(
)
2
,
,
1
ˆ
ˆ
arg min
( , )
N
N M
c
k M
k
k
c
w
u c
µ
=
=
−
∑
(
)
1
,
,
ˆ
ˆ
T
T
N M
N
N
N
N M
c
W
−
= Φ Φ
Φ
1
2
( ( ), ( ),...,
( ))
T
N
k
k
N
k
u
u
u
φ
φ
φ
Φ =
(
)
1
2
( )
( ), ( ),...,
( )
T
k
k
k
m
k
u
f u
f u
f u
φ
=
Identification of the linear dynamics
(
)
1
,
,
,
,
ˆ
ˆ
ˆ
ˆ
T
T
N M
N M
N M
N M N
Y
θ
−
= Ψ
Θ
Ψ
0
1
1
2
(
, ,...,
, ,
,...,
)
T
s
p
θ
α α
α β β
β
=
T
k
k
k
y
z
ϑ θ
=
+
1
1
2
(
,
,...,
,
,
,...,
)
T
k
k
k
k s
k
k
k p
w w
w
y
y
y
ϑ
−
−
−
−
−
=
1
2
( ,
,...,
)
T
N
N
Y
y y
y
=
,
1
2
ˆ
( ,
,...,
)
T
N M
N
ϑ ϑ
ϑ
Θ
=
,
,
0,
ˆ
ˆ
ˆ
/
i M
i M
M
γ
χ
χ
=
,
1
1
ˆ
(
)(
)
M i
i M
k i
k
k
y
y u
u
M
χ
−
+
=
=
−
−
∑
1
1
1
1
,
M
M
k
k
k
k
y
y
u
u
M
M
=
=
=
=
∑
∑
Nonparametric instrumental variables
*
1
(
,...,
,
,...,
)
T
k
k
k s
k
k p
w
w
y
y
ψ
−
−
−
=
#
1
(
,...,
,
,...,
)
T
k
k
k s
k
k p
w
w
y
y
ψ
−
−
−
=
,
,
0
ˆ
ˆ
APR
k
i M
k i M
i
y
w
γ
−
=
=
∑
-0,4
-0,2
0,0
0,2
0,4
0,6
0,8
1,0
1,2
1,4
-1,2
-0,8
-0,4
0,0
0,4
0,8
1,2
( )
u
µ
nonparametric
kernel estimate
parametric-nonparametric
estimate
true
Example
-0,2
0,0
0,2
0,4
0,6
0,8
1,0
1,2
-1
1
3
5
7
9
11
impulse response
true
nonparametric
estimate
Conclusions
• Each part is identified separately
• The convergence is strictly proved
Hasiewicz, Z. and G. Mzyk (2004). Combined parametric-nonparametric identification
of Hammerstein systems. IEEE Transactions on Automatic Control, vol. 49, pp. 1370-
1376.
Hasiewicz, Z. and G. Mzyk (2006). Nonparametric instrumental variables for
Hammerstein system identification. IEEE Transactions on Automatic Control
(submitted to)
• The method works under existence of correlated random
noise