Model Reduction For Identification of ARX Models

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

$5;

, 
E-mail: wangjianhong1980@yahoo.cn

:ARX
ARX
/

: $5;

Model Reduction for Identification of ARX Models


Jianhong Wang , Zhu Yong-hong
School of Mechanical and Electronic Engineering, Jingdezhen Ceramic Institute, Jingdezhen, 333403
E-mail: wangjianhong1980@yahoo.cn

Abstract: In this paper, we discuss the problem of model reduction in ARX system from the point of system identification.When
consider the process model represented by the linear regression form, based on the asymptotic analysis results of the unknown
parameters vector in the probability frame system, we derive the asymptotic variance matrix form of the unknown parameters
vector in ARX system. When obtain the identified parameters vector, we apply the most popular model reduction method L2
method and derive the identification strategy about the unknown parameters vector in the reduced model. Furthermore, we
analyse the asymptotic variance matrix form of the unknown parameters vector in the reduced model. Finally, the efficiency and
possibility of the proposed strategy can be confirmed by the simulation example results.
Key Words: ARX system; model reduction; asymptotic variance analysis

B (q) C (q)
1 (Introduction)  A ( q ) y (t ) = u (t ) + e (t )   
F (q) D (q)
y ( t ) u ( t )
//
e ( t ) 

 e ( t )

A ( q ) D ( q )
q B ( q ) 

 A ( q ) = F ( q ) = C ( q ) = D ( q ) = 1

// >@
A ( q ) = C ( q ) = D ( q ) = 1 >@

e ( t ) 
$5;
/ 
 A( q ) y (t ) = B ( q ) u (t ) + e (t )   

 F ( q ) = C ( q ) = D ( q ) = 1

/
B (q) 1
/  y (t ) = u (t ) + e (t )   
A(q) A(q)
 e ( t ) 1 A ( q ) 
2 (Problem Description) $5;
B ( q, ) 1
$5;  y (t ) = u (t ) + e (t )   
A ( q, ) A ( q, )
, :61164014

978-1-4577-2074-1/12/$26.00 2012
c IEEE 2093
A ( q, ) B ( q, )  ( t , ) = y ( t ) y ( t / t 1, )
A ( q, ) = 1 + a1 + " an q n , B ( q, ) = b1 + " bn q n
a
a
b
b
= y ( t ) H 1 ( ) G ( ) u ( t ) + 1 H 1 ( ) y ( t )  

= ( a1 " an b1 " bnb )


T
= H 1 ( ) y ( t ) G ( ) u ( t )
 a


N
0  Ef ( t ) = lim E f ( t )
N

( )
T t =1
= a1 " an a
b1 " bnb >@

0 = ( a 0
" a 0
b 0
" b 0
)
T
N * = arg min VN ( )
1 na 1 nb

= arg min E 2 ( t , ) N

0


 ( )
N N * ASN ( 0, P )   
( t ) = ( y ( t 1) " y ( t na ) u ( t 1) " u ( t nb ) )
T


 P = V ( ) 1 Q V ( ) 1
$5; 

{ }

Q = lim NE V ( ) V ( )
T
* *

y ( t ) = T ( t ) + e ( t )  N
  



N  1
 P = E ( t , o ) T ( t , o )
N
1 d
= arg min VN ( ) = arg min ( t , ) 
2
( t ,o ) = ( t , ) / =0
 N t =1   d
( t , ) = y ( t ) y ( t / ) = y ( t ) T ( t ) 
1
1 N

N = arg min VN ( )
(
P = N t ,N T t , N
N t =1
) ( )
N
1

1 N
1
1 N    N = 2 t ,N ( )
=
N
(t ) T (t ) N
(t ) y (t ) N t =1
t =1 t =1 $5;
$5; L2 L2
$5; 

 3 $5;
 (Accuracy analysis of the parameter
B (q) 1 estimation in ARX model)
G (q) = , H (q) =
A(q) A(q)
$5;  
y (t ) = G ( q ) u (t ) + H ( q ) e (t ) >@

 $5; ( t , )
y ( t ) = G ( q, ) u ( t ) + H ( q, ) e ( t ) 
 y ( t ) = G ( q, 0 ) u ( t ) + H ( q, 0 ) e ( t )    ( t , ) = H 1 ( ) y ( t ) G ( ) u ( t )

= G0 ( q ) u ( t ) + H 0 ( q ) e ( t ) = H 1 ( ) G ( ) u ( t ) + H 1 ( ) H 0 ( q ) e ( t )  
  G ( ) = G0 ( q ) G ( )
y ( t / t 1, ) = H 1 ( q, ) G ( q, ) u ( t ) 
   
+ 1 H ( q, ) y ( t )
1

q 

2094 2012 24th Chinese Control and Decision Conference (CCDC)


N {
Q = lim NE V ( * ) V ( * )
T
} G ( )

H ( ) iw iwn
= e I " e I
T

N N
1
= lim
N N
E ( t , ) ( t , ) ( s, ) ( s, )   
t =1 s =1
* * * T *
B ( eiw , ) 1
2 iw
V ( * ) = E ( t , * ) T ( t , * ) E ( t , * ) ( t , * ) A ( e , ) A2 ( eiw , )

1
0
d A ( eiw , )
( t, ) = ( t, )    
d
G ( )

1 1
= G ( ) u ( t ) + u ( t )    G ( q, ) H ( q, ) u ( t )
H ( )
2
H ( ) ( t , ) = A ( q, )
e (t )
1 H ( )
+ H0 ( q ) e (t ) B ( eiw , ) 1
H 2 ( ) u ( t )
I A ( eiw , ) A ( e , )
iw iwn T
= e I " e iw

$5; e (t )
1 0
B (q) 1   
G (q) = , H (q) =    
A(q) A(q)
 1 N N

Q = lim
N
E (t,* )H1 (* ) G (* ) u(t) u( s) GT (* ) HT (* )T ( s,* )
N t =1 s=1
( t , ) B ( q, ) 1 N N
= A ( q, )
1
e ( t ) + A ( q, ) 0 + 2

u ( t ) + lim E (t,* )H1 (* ) H0 ( q) e
N t =1 s=1
( t
) e s) H0 ( q) HT (* ) T ( s,* )
(

A ( q, ) A ( q, )
N
a

N N
( t , )
= A ( q, ) 0
1
u ( t ) = u ( t )
= S + lim
N
E (t,* )H1 (* ) H0 ( q) H0 ( q) HT (* )T ( s,* )
N t =1 s=1
b A ( q, ) = S + E ( t,* ) H1 (* ) H0 ( q) H0 ( q) HT (* ) T ( s,* )
  1 N N
B ( ) jw
S = lim
N
E (t,* )H1 (* ) G (* ) u(t) u( s) GT (* ) HT (* )T ( s,* )
N t =1 s=1   
e
A ( ) H 1 ( * ) G ( * ) u ( t )
V ( * ) = E ( t, * ) T ( t, * ) E ( t, * )
# +H 1 ( * ) H0 ( q ) e ( t )

1 G ( ) B ( ) B ( ) jwn
= A ( ) == = E ( t, * ) T ( t, * ) E ( t, * ) H 1 ( * ) G ( * ) u ( t )
H ( )
e ,
A ( ) A ( ) E ( t, * ) H 1 ( * ) H0 ( q ) e ( t )
N
e jw 0

# = E ( t, * ) T ( t, * ) E ( t, * ) H 1 ( * ) G ( * ) u ( t )
  
e jwn
     
1 jw $5;
A ( ) e

P = E ( t, * ) T ( t, * ) E ( t, * ) H 1 ( * ) G ( * ) u ( t )
1
#
1

1 H ( ) 1 jwn
= A ( ) = e S + E ( t, * ) H 1 ( * ) H0 ( q) H0 ( q) H T ( * ) T ( s, * )
H ( ) A ( )
A ( )


E ( t, * ) T ( t, * ) E ( t, * ) H 1 ( * ) G ( * ) u ( t )
T
0
  
# 

0
P

d B ( q, ) d 1 4 $5;(Model reduction in
ARX model)
d k A ( q, ) d k A ( q, )
B ( q, ) 1 $5;
2
A ( q, ) A ( q, )
2

= q k na + nb
1
0
A ( q , ) 



 
( )
G q, H q, ( )
$5;

2012 24th Chinese Control and Decision Conference (CCDC) 2095


G ( q, ) H ( q, ) L2 e ( t ) 

 
J ( , ) G ( )
L2
>@>@ ( ()
= E G G ( ) u ( t )

)
   
H ( )
$5; L2

( ()
E H H ( ) e ( t )

)
L2 J ( , ) G ( )
 G ( q, ) ( ()
= E G G ( ) u ( t )
)
   
 H ( q, ) H ( )
H ( q, ) 1



( ()
+ E H H ( ) e ( t )
)
T
$5; 2 J ( , ) G ( ) G ( )
=E
H ( q, ) = 1 A ( q, ) L2 2
 H ( q, ) = 1 A ( q, ) 2 G ( )

( ()
)
E G G ( ) u ( t )
2
 

T
H ( ) H ( )
( G ( ) , H ( ) )
+ E

E H H ( ) ( () )
( G ( ) , H ( ) ) H 2 ( )

2
T
2 J ( , ) G ( ) G ( )
= E

1 2 2    
J ( , ) = G ( eiw , ) G ( eiw , ) u ( w ) dw
T
H ( ) H ( )
4 0
E
 
1 2 2
+ H ( e , ) H ( e , ) e ( w ) dw
iw iw

4 0
   

 
1 2 
( ) ( )
2
J , = G eiw , G ( eiw , ) u ( w ) dw
4 0
 
1 2
( )
2
+ H eiw , H ( eiw , ) e ( w ) dw
4 0 
( G ( ) , H ( ) )
( ) ( )   
1
2 J N ,N( j ) J N ,N( )
j
( j +1) ( j) ( j)
 N = N
N( ) 2 ( j)
j
N
( )
= arg min J ,   

 + 
( ) J N
j
( )
,
N N 
* = lim
N >@>@


( * , * ) J
( , ) P
1
Cov J * *
( ) 1
( () )
2
J , = E G G ( ) u ( t )    
( * , * ) J
( , )
T T
2 J * *
1
( () )
2
+ E H H ( ) e ( t )     
2
     $5;
1
( () )
2

= E G G ( ) u ( t )
L2
2
 $5;
+ E ( H ( ) H ( ) )
1 2

2096 2012 24th Chinese Control and Decision Conference (CCDC)



Bode Diagram
50




Magnitude (dB)
0

( * , * ) J
( , )
1
Cov J * *
-50

( * , * ) J
( * , * )
T T
P J
1 -100
G ( ) G ( ) T 2 G ( )
( () )
90

E E G G ( ) u ( t )
2
0

=

Phase (deg)
T -90

H ( ) H ( ) 2
( )


+ E



( ()
E H H ( )
H
2

) -180

-270

T T -360
G ( ) G ( ) H ( ) H ( ) 10
-2 -1
10 10
0 1
10
2
10
E E Frequency (rad/sec)



E ( t , ) ( t , ) E ( t , ) H ( ) G ( ) u ( t )
1 Bode Diagram
* T *
* 1 *  * 50

S + E ( t , * ) H 1 ( * ) H 0 ( q ) H 0 ( q ) H T ( * ) T ( s, * )
0

Magnitude (dB)
-50

E ( t , * ) T ( t , * ) E ( t , * ) H 1 ( * ) G ( * ) u ( t )
T
-100

T T
G ( ) G ( ) H ( ) H ( ) -150

E E
-200
-180

T -270
G ( ) G ( ) T 2 G ( )
( ()
E G G ( ) u ( t ) )

Phase (deg)
E -360

2 (32)
T
-450

H ( ) H ( ) H 2 ( )

( () )
-540
+ E E H H ( )


2 -630
10
-1
10
0
10
1
10
2

Frequency (rad/sec)

5 (Example Analysis) 

6 (Conclusion)
L2

$5;
ARX
B0 ( q ) 1
y (t ) = u (t ) + e (t ) L2
A0 ( q ) A0 ( q )

A0 ( q ) = 0.35q5 + 0.28q 4 + 0.24q 3 + 0.13q 2 0.05 
B0 ( q ) = 0.12q 3 + 0.05q 2 0.02q 0.06
e ( t ) 
 {u ( t ) , y ( t )}t =1,2"1000 [1] F Tjarnstrom. L2 model reduction and variance
reduction[J]. Automatica, 2002, 38(3):1517-1530.

[2] F Tjarnstrom. Variance analysis of L2 model reduction

when undermodeling[J]. Automatica, 2003,


39(5):1809-1815.
 [3] Ljung L. System identification: Theory for the user[M]:.
 L2 New Jersey: Prentice Hall, 1999.
[4] C R Rojas. The cost of complexity in system
 identification: frequency function estimation of finite
impulse response systems[J]. IEEE Transactions of
Automatic Control, 2010, 55(10): 2298-2309.
[5] Michel Gevers. Identification and the information matrix:
how to get just sufficiently rish[J]. IEEE Transactions of
B ( q ) A ( q ) 1 A ( q )  Automatic Control, 2009, 54(2): 2828-2839.
%RGHSORW [6] Lennart Ljung. Estimating linear time invariant models of
non-linear time-varying system[J]. European Journal of
Control, 2001.7 (2):203-219.

[7] H Hjalmarssion. A geometric approach to variance


analysis in system identification[J]. IEEE Transactions of
 Automatic Control, 2011, 56(5): 983-997.

2012 24th Chinese Control and Decision Conference (CCDC) 2097


[8] B Ninness. Variance error quantifications that are Automatic Control, 2004, 49(8): 1275-1291.
exact for finite model order[J]. IEEE Transactions of

2098 2012 24th Chinese Control and Decision Conference (CCDC)

You might also like