  ѧ 
JOURNAL OF ASTRONAUTICS
1998  3   No.3 May   1998
Ƽڿ
ģ͵ŷϵͳ
벹о


(ѧԶϵ**100081)
硡ׯ
(ҵѧƹϵ**150001)
ժҪĴǶʹģŷϵͳĿƾȡ,ŶضԾŷϵͳӰ졣ʹ緽Ŷصѧģ,ȫԭƲ,ʵŶصĶ̬һŷϵͳ,оʹģʵֶ涯̬ģ͵߱ʶͿ,ʵ½ŷϵͳƵʵоµĲŷϵͳľʮЧ,ʵӦüֵ
ʡŷϵͳ硡Ŷءϵͳʵʱ
CONTROL AND COMPENSATION FOR NONLINEAR 
SERVO SYSTEMS BASED ON NEURAL NETWORKS

Xu Lixin
(Department of Automatic Control,Beijing Institute of Technology*Beijing*100081)
Wang ChanghongZhuang Xianyi
(Department of Control Engineering,Harbin Institute of Technology*Harbin*150001)
AbstractNeural network is used for servo systems design in two ways.The affection of disturbing moment in servo systems is discussed at first,we use feedforward neural network to build the mathematic model of disturbing moment,and design transfer function for compensation.Then,the schemes of training inverse dynamic model is provided.By means of neural networks model,the control properities of the servo system with unknown parameter are improved greatly after online learning course.Experimental results show that the new method is effective in identification and control of servo systems.
Key wordsServo systemsNeural networkDisturbing momentInverse systemsOn-line control
1
ŷϵͳѱ㷺Ӧں,Ե豸еĲת̨,ʵеķת̨,Ŀ걳ģװеĹѧ涯ϵͳȵȡ,ŷϵͳĿԴһЩ:,ŷϵͳȵܵĦΪŶصơŶŷϵͳ,ʹǷֱʼظȽ,Ӱŷϵͳĵԡ,ض,رǵһַģͲδ֪ʱ,Ҫ߾ȡƵ,Ŀ㷨ܵһľޡ,оŷϵͳܵ·Ҫۺʵ塣
,ۼӦõоչʾڱϵͳṹ,Լ֯Ӧѧϰص㡣ģ͵ĿƷʾϺõ,ڻˡڵȡ˳ɹӦã1ݡģΪ,Ӧ缼Ŷģ,ʹòʵŶصĶ̬һŷϵͳ,оʹģʵֶ涯̬ģ͵߱ʶͿơʵ½ŷϵͳƵʵо
2׼
2.1Ŷط
ŷϵͳеŶҪĦ,⻹иתءĦصĴСṹ,,شСйءתҪ븺йءŶŷϵͳȵҪϰ,ѧһ˴,ȡһЩɹ2,3ݡ
1)ֶģ͡

(1)
MSMCֱΪĦغͿĦ,MFΪת̨йжӦʦõĦء
2)䳣ֵģ,ΪĦһΧǳֵ

(2)
ף3ݸĦģ״̬۲,ʹòԭǰ,ʵת̨ϵͳľ̬پȡ,ף2н˳ֵתΪŶءף3ݸĦصķֶģ͡ڽģͱȽϴֲ,ʵӳĦصʵ,Ħشڹ㸽Ĺɹ״̬ʱ,ԻЧ,ܷͬʱĦغ͸תزΪȷĦغ͸תģŷϵͳȵĹؼ⡣
2.2涯̬ģʹ
һԶ̬ϵͳ
(3)
״̬xRn,uRm,yRr,f(*,*)h(*,*)ΪӦάԺ
14ݶʽ(3)ʾϵͳ:uy,һӦϵͳ:rwʹõr(t)=y(t)ʱ,еʽw(t)=u(t),ϵͳϵͳƵϵͳ,ϵͳ,ƿ档
24ݶڵ롢ϵͳ,ʽ(3)m=1,r=1,ϵͳԽa()ʹʽС

14ݶڵ롢ϵͳ,m=1,r=1,ϵͳΪĳֱҪԽa()n
,Եõ
1:

(4)
ĵ롢ŷϵͳ,R(*)0,涯̬ϵͳһڡ
֤:x1=,x2=
,ʽ(4)״̬Ϊ

Сa()=2=n
1Ƴϵͳʽ(4)档
֤ϡ
3ŶصĶ̬
3.1Ŷؽģ
ƽΪ򵥵ǰ,ƽԺ,Ŷصѧģ͡ŶҪĦغͺֵת,ŶӰ,Ϊϵͳ͸һʱ,ŶF()()ķԺȻһ̡̬ʹǰԪ罨һģ͡Ϊʹѵ,ֱת(أ0)ͷת(أ0)ģ͡ڷֱƽΪ򵥵ķԺ,Ȩֵѵѡöǰʵ,ÿ纬һڵ,һڵ,,ѡ8ڵ㡣ڵ㼤Ϊ˫кtanh(x)=(1-exp(-2x))/(1+exp(-2x)),ڵ㼤ΪԺṹԺ,Ϊ,ΪŶصĳֵʦ,I֪ƽϵʽ:

KMΪתسΪֵ,(d)/(dt)=0,KMI=F()ɵŶF()͵õһ(,F())ccccccccΪʦѵǰȨֵΪ֤ѵЧ,ѵӦǽϴΧ
3.2װ
Ȩֵѵ,Ŷصģ͡Ħطֶģ:ٶαΪ⻬ķ,ٶνӽԹϵ
ŷϵͳͼ1ʾʵʱٷź,ͨģͼŶصĴС,ΪʵŶصȫ,㲹ڴΪ:



ͼ1Ŷزϵͳṹ
ʹӲ·ʵ,ԭһнСʱ䳣ĹԻڡ,ģͼǰڼʵ,(s)ɱʵ֡
4ƶ涯̬ģͽ
4.1ģ
öǰϵͳ涯̬ģ,Ľλæ,ٶ,ǼٶΪ,UƽUd
ϵͳʶĽǶ,ѧϰɰϵͳ,ΪĹģ(ԪĿ)ͼѧϰʱ,ѡöͬƵźิϵķʽΪѧϰԤλźš
ǵһʽ(4)ʾʵŷϵͳ,вδ֪,14HzΪҪƵ,ԭϽ,ԤλźӦƵΧ,ͬΪ˾ʵֵķ,ֻѡźƵʸϳԤλź,ڸӵԤλź,ֻӵṹ͸ѵʱ伴ʵ֡ԤλźƵΪf1=2Hz,f2=3.33Hz,ֵΪ18,Ԥλźſд:
r=18sin(2*2*t)+18sin(2*3.33*t)
rźż뵽ջѵṹ,ɻϵͳʵź:λźŦ,ź,źuy,v,aʾ,,,ѵɶλźy,ٶźv,ٶźaźu,y,v,uֱӲáaֱӶٶź΢ֻ,ڲٻʵźܴ,ٶȼʧ,,ֳҲȱٶȼƵ豸,ʹá-΢5ȡٶźšҪָʹá-΢ȡٶź,߽,Ҳߵʵ,һֽϺõĻȡٶȷ֮,ͬҪ֤Rͦġ
4.2涯̬ϵͳɽṹ


ͼ2(a)ϵͳɽṹ


ͼ2(b)ϵͳf=2Hz
źӦ
ƶһʵϵͳ,ٻڵٶȻ,ʵ鲻Ҫȷ֪ģͽṹͲ,ֻ漴ɡ
ϵͳͼ2(a)ʾ
λôõλ,λôΪ5V,Ӧ180λ,СѹֱΪ0.01V,źźλôźžHY-1232ݲɼӿڰṩ32·ն˵ݲɼͨ,ܴﵽ0.03%ľȺ12λķֱʡ,˰廹һ12λֱʵģͨ,ɲչ,ֱPC߽ݡʵоʹжݲɼʵʱơΪɵļ,˱뱣֤һʱɴļ,,ѡüΪ486/DX-66,вʱΪ10msʶŷģ,ֱʹPIDϵͳ,ϵͳf=2Hz,ֵ18(0.5V)źӦͼ2(b)ʾ
4.3ѧϰԭĽ
ѧϰһֻӦƷkawato6ӦáʹñջϵͳķȨֵ,ĿǼС,Ӷ߱ʶ涯̬ģ,ʵָ߾ȵλÿơѧϰƽṹͼ3ʾ


ͼ3ѧϰƽṹ
涯̬ǰ,ƶ,淴ɡ淴ΪһķԱ֤ѧϰڼϵͳĽȶ,ΪźȨֵ

(5)
ѧϰһ߿Ʒѧϰ,ɷϵǰȡ淴Ҫ,ʱʵȫ涯̬ơ
ѧϰڿƵͬʱֽѧϰ,ѧϰõϢһʵŻơģͽȶʱ,ѧϰƿɽһ߱ʶ涯̬ģ;;㶯ϵͳܵŶʱ,ѧϰܹЧȨֵ,Ӧʶ涯̬ģͲǶ,ѧϰһȽġǿ³ԵĿƷ
ʽ(5)Ȩֵ,ʵз㲻㡣,ݶȷľֲСʱ,ʹѧϰֹͣ״̬,ÿεȨֵļƫ,ʱϳ,ռùϵͳԴΪ,ʽѵ㷨7ݸĽȨֵʽ(5)
WijΪԪiԪjȨ,ڵk:



(6)
ΪһСPijk)Ϊ

(7)
Cij(k)ԭ㷨ͬ,ڴ,

(8)


,ƽȨֵ仯ΪCij(k),ʵĽȨֵһַǳЧѵ㷨,ܹʵϵͳ,ڽ϶̵ʱʵʱ㡣
5ʵ
5.1Ŷز
ŶҪӰŷϵͳĵٸԼλ,ص㿼첹ǰָı仯ʹϵͳٺ㶨ٶź,䶨λȡԺĿ걳ģװеĹѧ涯ϵͳΪ,öԱȷ,ŶǰбȽϡѡ-0.5,0.5ڵ20ȷֲ(λrad/s)ѵȨֵ,źΪ0.1rad/sĵбź,Ƚϲǰϵͳĸپ,ڸٹٸλź,ͼ4(a)ͼ4(b)ʾ,,ͼ4(a)ǲǰ,ͼ4(b)ǲߡÿ,ǰ¼˶бźŵĸ,Ϊλ,нϴԾ״̬лġʾ:


ͼ4(a)ǰ


ͼ4(b)
1.0.3mradС0.1mrad
2.̬0.1mard͵ӽ޾̬
,ģͲΧڵź,ƾԽ͸,ϵͳѹĲʵŷϵͳĵٸپȺͶλȡ
5.2涯̬ѧϰʵ
5.2.1ģͱƽ
涯̬ģ㺬3ڵ,ֱʾλæ,ٶȺͽǼٶȡֻһڵ,ʾϵͳu,㺬15ڵ,ڵ㼤Ϊ

ѻõнȡ200Ϊѵ,ݽйһ,һʾΪ{1,1,1,u},



ͼ5涯̬ģͱƽ
ڿuȽС,ڡ2Χ䶯,ʲһһȨֵѵǳҪ,ܹЧѧϰٶ,ѵʱ䡣ѵһ,ʱѧϰ,ıƽͼ5ʾ,u1(k)Ϊʵʿ,õ߱ʾu2(k)Ϊƽ,ʵ߱ʾ
5.2.2һƵźŵѧϰ
źf=2Hz,ֵ0.5V,÷ѧϰ,ѧϰ1000κĿƽͼ6ʾ


ͼ6(a)ͼ6(b)

]
ͼ6(c)ͼ6(d)


ͼ6(e)
ͼ6f=2Hzźѧϰƽ
ͼ6(a)ΪѧϰϵͳӦ,yr(k)Ϊ,y3(k)ΪϵͳӦͼ6(b)Ϊѧϰĸ,ͼ6(a)ͼ2(b)Ƚ,Է־ѧϰϵͳԸ뼸ûλͺ,ɷֵ0.5VС0.23V(ͼ6(b))
ڼӵĿ,涯̬ǰͳ淴,ѵǰ߱Ƚͼ6(c)ͼ6(d)ʾun1(k)un2(k)ʾѵǰǰ,up1(k)up2(k)ʾȽͼ6(c)6(d)ɼ,ѧϰ,ǰȡ淴,ʵ涯̬ƵĿġ
ʹʽѵ㷨ѵͼ6(e)ʾȻ㷨˽ΪѧϰЧ
5.2.3Ƶźŵѧϰ
ѡλź
r=9sin(2*2*t)+9sin(2*3*t)


ͼ7Ƶźѧϰƽ
2000ѧϰͼ7ʾ
ͼ7з,Ȼѧϰ2000,ٽϵһƵźҪܶ,˵ƵźѧϰǱȽϸӵ,Ҫѧϰѵʱ䡣,ȨֵҪźŽе,Ȩֵ仯ĸԡ
6ᡡ
ĴǶʹģŷϵͳĿƾȡŷϵͳӦõʵҽ,緽Ŷصѧģ,ǰķֶģͺͳֵ,ܹʵӳŶصı仯̡ȻȫԭƲڴݺ,ڽṹ,ʵ֡µĲǿкЧġ䲹ܵŷϵͳ,ЧͬŶ(縺ƫ)Ķ̬,ŷϵͳģδ֪,˻ǰ̬ģѵŷϵͳ涯̬ʵʱ,ӦõʵҽоѧϰһȽ,еĿƷ,ܹЧչϵͳƵ,ƿϵͳܡ,δŶصӰ졣ʵŷϵͳ,ŶҲǽλúʱϴӰŷϵͳľȡ,żźź͸ӿʵõѧϰ㷨ҲҪһо⡣
ο
1Kawato M,Furukawa K,Suzuli R.A Hierarchical Neural Network Model for Control and Learning of Voluntary Moment.Biological Cybernetics.1987,57(2):169-185
2,ܹ,ȫ.״̬۲ڵϵͳеӦ.廪ѧѧ,1980,20(4):33-46
3Ԫ,Ӣ.ת̨ĦصĶ̬.Զѧ,1983,9(4):248-252
4,Ԫ.ԿƵϵͳ.廪ѧ,1991,pp30-94
5,ΰ.Ը١΢.ϵͳѧѧ.1994,14(2):177-183
6Kawato M,Gomi H.Neural Metwork Control for A Closed-Loop System Using Feedback-Error-Learning.Neural Networks.1993,6(6):933-946
7,,ׯ.ѵһ㷨.ҵѧѧ.1997,29(3):85-87
