Զѧ
ACTA AUTOMATICA SINICA
1997 23 5 Vol.23 No.5 1997



ú̬ǰʶ
ϵͳ1)
졡Ӣ
ժҪú̬ǰԷϵͳȫֳģֱ֤ͣʶṹĿԣѧϰ㷨Ϊ̬BP.
ؼʡǰʶ̬BP㷨.
NONLINEAR SYSTEM IDENTIFICATION USING FEEDFORWARD 
NEURAL NETWORKS WITH RECURRENT HIDDEN LAYER
BAO XIAOHONGJIA YINGMIN
(The Seventh Research Division, Beijing University
of Aeronautics and Astronautics, Beijing 100083)
AbstractThis paper presents a feedforward neural network which contains a dynamic hidden layer (RHNN) for the modeling of a nonlinear system and the feasibility of this approach is proved. The learning algorithm is the dynamic BP. In the simulation a comparison between RHNN and RFNN for modeling a nonlinear system is made.
Key wordsFeedforward neural networks, system identification, dynamic BP algorithm.
1
ʶϵͳ˴Щ󶼽ڷԶ̬ϵͳģϣ1ݣʱӷĶǰ(RFNN)ƽЧ.ֵעģǴɢ״̬Ƴľֲģͣ1ݣԭ״̬еȫϲһд˹ϵ⣬ú̬ǰ(RHNN)ֱӶ״̬ķϵͳϵģͣѧϰ㷨Ϊ̬BPڷжԴϵͳϵֱRFNNRHNN˱ʶȽϣܸ׼ȷرʶϵͳ.
2Էϵͳģ
һ㷺ӦõķϵͳڴΧ(ȫ)ģͱʾ
x(k+1)=f(x(k),u(k)),y(k)=h(x(k),u(k)).
(1),(2)
Ϊϵͳڲ״̬Ϊϵͳ룻ΪfhΪ.
ϵͳ(1)(2)¼裺
H1)޵M()Сx(M)+y(M)ޣϵͳȶģ
H2)f,hLipschitzϵͳĽΨһ.
1.H1),H2)ķϵͳ(1)(2)һ̬Ĳǰ(ͼ1ʾ)k[0,M]⾫ȱƽԣʹ

(3)

ͼ1RHNNʾͼ
֤.ǰNNf,NNhֱƽf(.),h(.).NNfWf={1Wf,2Wf,1b},1WfRn1(n+r),2WfRnn1,1bRn1.NNhWh={1Wh,2Wh,2b},1WhRn2(n+r),2WhRmn2,2bRn2,

(4)
x=(x1,,xn)T,g(x)=(g(x1),,g(xn))T,



(5)


(6)


(7a)

(7b)

(7c)
ף2ж1ӦάȨֵW*f,W*hʹ(3)ʽƵͶ壬1W*,2W*,3W*,HW*,uW*,1b*,2b*ʹ(3)ʽ.
3ѧϰ㷨
覨ʾдȨֵΪ򵥼ƣֻMISOϵͳƵMIMOϵͳ.ԤĿ꺯

(8)
Nݳ.

ʽЦΪȡԸȨֵĵ.
Ѷ̬дΪʽ

ע⵽Ķ̬ԣԦеĲͬ




(k,)Ƕֽ̬ж̬BP.
4
.

ֱRFNNĴͱRHNNĲԴϵͳϵʶ֪ypu1,2ֱӹϵuW=0.1ͼ2ͼ3ʾ.
1

RFNN(ʶ)RHNN(ʶ)
㷨BP=0.05̬BP=0.25
ṹ4-45-12-10-8-1
ѵ
ѵk=20,000k=20,000
Ԥ
Ԥͼ2ͼ3
1)2)



ͼ2RFNNԤ

ͼ3RHNNԤ
ɵ½ۣ
1)RFNN״̵̬ķϵͳϵȫֱƽǿ.
2)RHNNϵȽǿ.Ҵ˹ϵıƽѵƵйأԤѵƵԤ⾫ȸ.
ú̬ĲǰԷϵͳбʶ֤˴˷Ŀ.ѧϰ㷨Ϊ̬BP.ֽṹͨѧϰ״̵̬ķϵͳϵܽнϺõıʶ.
1)ȻѧĿ.
ߵλպѧоҡ100083
ο
[1]Leontaritis I J, Billings S A. Input-output parametric models for nonlinear systems, part I:Deterministic non-linear systems. Int.J.Control, 1985, 41(2):303328.
[2]꣬ 봿.зɢ̬ϵͳʶĿ. Ӧã199411(4)413420.
ոڡ1995-08-17
