ѧ
JOURNAL OF SOFTWARE
1999 10 12  Vol.10 No.12 1999



ڴУӱ
챣
ժҪĲУϵͳԭͷ,һӱļСģ磨SCMACģ͡㷨ʵּ.ģ͡㷨ֱȨַӳ似,ѵΪַ,ȨصĹϵ.ΪȨַ,ҵӦȨ,岹ɻø߾.,ôļ洢ѰַȨصȷ,΢ڴ,ʹʵ.һʵ.ʵ,SCMACУ,ʹķٵΪ.
ؼʡСģ,,,У,.
ͼšTP18
A Novel CMAC Neural Network for Correcting the Sensor's Nonlinear Errors
ZHU Qing-bao
(Developing Department of Computer Technology Nanjing Normal University Nanjing 210097)
Abstract In this paper, the principles and the methods for correcting the nonlinear errors of the sensor system with a neural network are shown, and a novel simplified cerebella model articulation controller (SCMAC), which includes its model, algorithm and realized techniques, is proposed. The direct weight address mapping techniques are used in this model and algorithm, and the relation between the inputs and the weights is established by taking the inputs of training samples as their weight address, and the corresponding weights are found for any inputs by taking it as similar weight address, and the accurate outputs are obtained by the associable insertion algorithm. In addition, the weights are stored and addressed in a magnetic disk file, therefore, the overflow of internal memory of microcomputer are avoided, and the SCMAC is easily realized. Finally, a simulation experiment is given and the results show that the nonlinear errors of the sensor system are decreased to approximate zero after correcting with a SCMAC.
Key wordsCerebella model articulation controller (CMAC), sensor, nonlinear errors, correction, simulation.
ڼϵͳ,һôʰȡ,پӿ(׳Ʊ),һȵӦĵź.ǰѴͱϳΪϵͳ.Ŀǰ,ϵͳǵ롢ϵͳ,ɱʾΪ
y=f(u)(1)
yʾϵͳ,uʾϵͳ.
yźžA/Dתɱ΢ʶ,ǵĿǸݲõyδ֪u,
u=f1(y).(2)
ʵӦ,f(*)Ϊ,һΪԹϵ,һΪԺ.ν,ҲֻǰеһνΪԹϵ,Ƴ̶Ⱦ˸ഫϵͳԾ,һʷ,,ϸ˵,f(*)ǷԵ.f(*)һ򵥵֪Թϵ,ʽ(2)u.ǣںܶ,ʹ߲֪f(*)׼ȷʽ,ֻһνԶ(侫Ⱥ)һԵf(*)ӲķеһСβɽ,ַȫ̲,ҲӲƮƻӰ,˿ɿԺͲȵ,С,Ӧܵ.
Ϊ򲹳ϵͳķ,ʹyͨһ,ûڵԺΪ
p=F(y),
pΪ.
F(y)=f1(y),

p=f1(y)=u.(3)
Ȼ,F(y)ҲһԺ.ʽ(3)֪,p뱻ֵuһ,ʹĴԾ,ͼ1ʾ.

ͼ1
ʵӦ,ԺF(y)ıʽ׼ȷ.Ϊ,ǸСģ1(Сģ͹ؽڿcerebella model articulation controller,CMAC)бƽԺ,CMACѧϰƽF(y),ȡF(y).:CMAC뼴Ϊϵͳ,CMACΪp=u.CMACԸϵͳӦöƵ,ӳ,ռ䵽洢ӳ,һһһӳϵ,˸洢һġӴĴ洢ռ,΢ʵѵ.Ϊ,Ӹ洢洢ӳ˶ٵķӳ,Albus's CMAChashӳʵһ.hashӳпʹCMACĳЩѧϰвʹ̬仵,ʹ̱2.
,ڲͬпܼͬȨԪ,Albus's CMACѧµ˾ɵ3.hashӳ,Ȼ֤ѧϰ̵,ѧϰռõѧϰάָϵѸ,Ȩ洢ռʼ,Ϊ˽һ,ֱȨַӳƽ,ɴȨش洢Ԫ4.һȻ˽ϴĸĽ,ѧϰĺάϸʱ,ѧϰռõȨ洢ռȻӴ.һЧķǰһάֽɵά,ֱʱֱȨַӳƽĵάCMACȥƽ,ܱ֤ѧϰ̵,ܼؽʹ洢ռҪ,ʹһ΢Ҳʵ.һ,Ӧⶼһά,ʹάҲԷֽһά.磬Ŀǰһά,άԷֽһά.
,оһרӳһάԺCMACʮֱҪ,,CMACһֲһʵ58,ϱĴϵͳķУҪ,һֱֲȨַӳ似ļCMAC㷨.ΪCMAC,ǳ֮ΪSCMAC(simplified cerebella model articulation controller). 
1 SCMACĻԭṹ㷨
1.1 ԭṹ
SCMACļԭṹͼ2ʾ,XΪ״̬ռ(ͼ1еy),XпܵԪ,ͼеx1,x2,...,xi,xnΪѵԪ.ACMACAͬ,洢.ÿһѵڴ洢AһӦȨֵ.һxj,һyj=F(xj).Ϊ˶Ը״̬xjyj,Ҳ˶ӳ;f:XA,F:AY.
1.2 XAӳ
ѵ{(xi,pi),i=0,1,2,...,n}Ϊnɢļ,xiΪ,piΪ,x0<x1<x2<...<xn,xiʣ0,Rݵֵ.xiΪwi߼ַ,wi=f(xi),ȼڰpixiΪַ߼Ԫ.
1.3 ݵ
ʵʵXΪģ,X=(x1,x2,...,xn),ÿֵСĳһϵ.ģԭ,ǱĴ.(1) ֵƽ,任ɣ0,Rֵ.(2) ѧϰѵԪزܱ鲼ռ,ôδѧϰз,ʹӳ䵽ѵԪ뵥Ԫ.
ÿһԪxj, Ϊ0RΧڵһ,㷨


(4)
ʽ(4),xjmaxxjֵ;xjminxjСֵ;xjֱ;iԪֵ.
1.4 ӳԤ
0,RÿHȡһѵԪ,ΪʹѵԪӳ䵽뵥Ԫ,ѵԪ´:
k=int(i/H), kʣ0,R/H,(5)
iǴǰѵԪֵ;KǴԪֵ.
,ȡH=4,{i=0,4,8,12,...},Ӧk={0,1,2,3,...}.
ֵj,ʽΪ:
k=int(j/H+1/2).(6)
ʽ(5)ʽ(6)ڶСĴͬ,ʽ(5)ȫ᷽ʽ,ʽ(6)45뷽ʽ.ʹѧϰʱֱΪH,ʵӦʱֱΪH/2.
,ӳ䵽뵥Ԫ,.
1.5 ѧϰ
辭ѵԪΪ{k1,k2,...,ki,...,kn},ΪP={p1,p2,...,pj,...,pn}.ѧϰ:
for (k=0;k<n,k++),
wk=pk,
wkΪkѵӦȨֵ.Ȼ,һѧϰ.
1.6 AYӳ
ͬʱ,ʽ(4)ʽ(5)任,õӦki,kiʣ0,R/H,
yi=F(ki)=wi.
ͬʱʽ(6)ӳ䵽ΪH/2̫.СHȻ߾ȣʹ̫࣬HҲȡСֵҪﵽ߾ѵ.νһ?ǵԾƲĹ,δѧϰϢʱ,ݹȥѧϰ֪ʶ׼ȷƶ.һ,ܷѵṩϢ,ͨ,ﵽɽ׼ȷƲӦ?ֲֵԭǡǡ.
һάy=f(x),nѵΪȾȡ,(x1,y1),(x2,y2),...,(xn,yn),xk=x0+kH;k=1,2,...,nΪڵ,HΪڵ򲽳.xi,xi>xk,²ֲֵʽ(ȥߴ):
(7)
ʽxi<xk,
(8)
ʽ
ʵ,CMACΪһֲѱڶоϿ,һǶȿ,ѵԪصȨؿɿб,ֻҪѵԪصȡֵͬ,ͿΪǵȾ,ôòֶʽв岹ʮֿе.
2 Ӧо
ñ㷨ȥУϵͳķ,ȡʮЧ.һӦʵ,˵Ӧ÷Ч.
2.1 ӦĿ
ùʽƳɵԲ΢,ЧΪ395m,ڣۣ100,+100(ֵ)Ϊ0.2m,100,200Ϊ1m,200,300Ϊ2.8m,300,395Ϊ4.9m(ͬ).Լ,ϵͳԵ˷ԱȽ.ñķУ.
2.2 淽
Ϊ˵Ⲣʧһ,ѵԪصȡֵR1,R2,R1=300,R2=395,һӦĴԽУ.ֻҪȡֵΪȫ,ʵȫУ.
ͼ1֪,У뼴Ǵϵͳ.ͼ1еy=x,ѵΪ(X*,P*),X*Ϊ,༴ϵͳپA/Dתͱȱ任ֵ(ʵӦ,ֱбȱ任.,Ϊ˵,ȡȱ任ֵ.P*Ϊ,ҲǶӦĴϵͳ.Ϊ˽һѵ,ΪH,ʹ΢״̬,һ׼΢̨΢,΢䴫,΢ĲֵΪx*ii*Hʱ,¼ֵ,Ϊp*i.iʣ0,M1,M=(R2R1)/H+1.
ΪӲ,Ĺ,ǶѧϰӦù̽˷.R1,R2ݷΧ,һʵֵ,С˷ϳ߷
x*=y=f(u)=bu+au2.(9)
x*c,
au2+bu-c=0,
(10)
ʽ(10),a=0.000 031 25,b=1.
x0=300,x*i=x0+i*H,i=0,1,2,...,M1,ʽ(10)ɵõp*j,j=(x*i-x0)/H,õһȨֵ{(j,wj),j=1,2,...,M1},wj=p*jΪjȨֵ.Ϊڴµ綪ʧ,һļб.
u,ʽ(9)Դϵͳз(A/Dתͱȱ任),õx,j=xx0(ƽ),ʽ(6),Kint(j/H+1/2),k0,1,2,...,M1.kΪȨļƫָ뼴ȡӦȨ.
xѧϰ()ʱ,ʽ(7)ʽ(8)в岹.
2.3 ʵ
Աķ1,еֵԴ㷨.
1Ƚ(Ϊֵ)

ʱ 
ʱ
ƽδ岹
δ岹
ƽ岹
岹
ƽ
24.851 13.786 30.977 90.492 80.000 10.000 0
54.851 13.786 32.435 51.231 40.000 40.000 1
104.851 13.786 34.862 52.493 30.001 50.000 5
204.851 13.786 39.761 05.014 30.005 90.002 0

u=300,301,302,...,ui,...,395,
ei=fabs(piui),
aver=Q/s=ei/s,max=MAX{ei},min=MIN{ei},iʣ0,R2R1.
ʽ,uiǴϵͳʵֵ,piǾУʵֵ;eiľֵ;sĸ;averƽ.ݴϴ,Ϊ,CԱдһ麯һ.1ɸú.
1ݱ,(H)Ӵ󣨶Ӧ٣,δ岹ʱ,Ӵ,ԼΪH/2,CMACһ.Ĳ岹ʽȡ4,,岹,0.ʵ,˲㷨޳,Уϵڸ(0.05m),С.
,ǶԶһάԺ˷棬,ñӳһʵ仯ķԺʱ,üѧϰȡúܸߵıƽ.֮,ӳĺǷǵ仯Ļ仯ʽϴʱ,ѵҪһЩȡúܸߵľ.ڿϵͳеļⵥԪ͸ִϵͳķһ㶼ǰ,,ڷУرЧ.ΪĿǰѵһַ.
ֵע,ֲ岹Ӧ岹,֪ݵϢݵı仯,Ӷ׼ȷƲǰӦ.Խ,ҲԽ,֮С.,ֻб仯Ƶ,Ϊ.Լ,øģѧϰƽһԺʮ.ڽڴϵͳǱķУʱ,ֻҪעѵʱҪϴȡúܺõЧ.
3 
Ŀǰ,CMACоԴ̽ĽӦúͷչ׶,йѧ˸ָĸĽģ,HCMAC,TECMAC,CASCADE-CMAC ȵ.ͨ΢Ϣ±,רӳһάԺļģͺ㷨״,CMACУҲ״.岹,ʮ־ȷرƽԺ,ܴﵽ㷨ƽȵCMACҲδ.,άԺάԷֽһά⣬ñļģȥѧϰƽ.Ӧö֤˱ģͼ㷨ѧϰԺڴϵͳķУʮЧ,ʹ豸ķС.㷨ݡȸ,ڴ١ʹõƬҲʵֵص㣬й㷺Ӧǰͺܸߵƹֵ.
*оõʡίӦûĿ.
߼飺챣,1955,оԱ,ҪоΪԶƣܿ.
ͨѶϵ:챣Ͼ210097,Ͼʦѧ
ߵλϾʦѧϾ210097
E-mail: qb.zhu@jlonline.com
ο
1Albus J S. A new approach to manipulator control: the cerebella model articulation controller (CMAC). Journal of Dynsmic System, Measurement and Control, 1975,97:220227
2Lin Chun-Shin, Chiang Ching-Tsan. Learning convergence of CMAC technique. IEEE Transactions on Neural Networks, 1997,8(6):12811292
3Ker Jar-Shone, Kuo Yau-Hwang et al. Hardware implementation of CMAC neural network with reduced storage requirement. IEEE Transactions on Neural Networks, 1997,8(6):15451556
4,.СģĽ㷨о.Զѧ,1997,23(4):482487
(Liu Hui, Xu Xiao-ming et al. An improved CMAC neural network algorithm. Acta Automatica Sinica, 1997,23(4):482487)
5Leondes C T. CMAC-based techniques for adaptive learning control. Optimization Techniques. San Diego, CA: Academic Press, 1998. 227304
6Lin Chun-Shin, Chiang Ching-Tsan. Integration of CMAC technique and weighted regression for efficient learning and output differentiablity. IEEE Transactions on Systems, Man and Cybernetics, Part B, 1998,28(2):231237
7Hirashima Y, Iiguni Y, Adachi N. An adaptive control system design using a memory based learning system. International Journal of Control, 1997,68(5):10851102
8Lee Chau-Jhy, Lin Wei-Song. A method of clustering quantization for better training of CMAC. Journal of Engineers,1996,19(3):309320
1998-09-25յԭ,1998-12-23յ޸ĸ
