Զѧ
ACTA AUTOMATICA SINICA
1997ꡡ236ڡVol.23No.61997




ɫ·ϵͳPNN˲
ΰ
ժҪͨоϵͳɫͳԣ 䣬һԤ⣬ Ϊģʽģʽ ʵѵ. ɴɫ·ϵͳ˲·. 
ؼʡϵͳ PNN˲ɫ.
PNN FILTERING FOR NONLINEAR SYSTEM
WITH COLORED NOISE
LIU JIANHUAXU XIAOMINGZHANG WEIJIANG
(Department of Automation, Shanghai Jiaotong University, Shanghai200030)
AbstractIn this paper, by dividing the interval of the system noise and realizing on-line training of the probabilistic neural network, a new filtering method for nonlinear systems driven by colored noise is proposed. 
Key wordsNonlinear systems, PNN filtering, colored noise.
1
ϵͳ״̬ƷĿǰƹ㿨˲(EKF)Ϊ ڵҪڹƾȺ˲ĸٶȷ治. Ӱ쾫ȵԭһڷԺľֲԻɵһ棬ϵͳķʹ˲ļܴ󣬲òһЩƼ㣬ʹܵ˽ϴӰ죻ߣEKFϢӦòϸ£ֻͨʵֵԤֵĲԵزԤֵ. ڸٶȷ棬EKFĸԤֵʵֵĲĻȷ ļȥصģ ֽϴȵ״̬ͻʱȻ˽ӷЧԲʮ. ɫΣ ÷ֽⷽ һЩ͵ɫзֽ⣬ʹΪɷֵϣͬʱĳɷӦΪϵͳ״̬. ַȱҪõͽɷֽļ֣ ״̬άEKFļ. ڼѧˮƽߣԼۺӦüõشչΪṩõ. 1988Dr.Specht˸(PNN) ͻŵѵٶȿ죬ʵ. 
ĲDr.SpechtPNNԭͣɫ·ϵͳ״̬ƵPNN ڹƾȺ͸ٶؼָ귽нϴĸĽ. 
2ɫ·ϵͳ״̬ƵPNN
ѵ̵ҪģʽиģʽĲ࣬Ҫͨģʽиģʽϳ̶ȵóڸĸʣͳԽۺϣԵõBayesŽ. ڷ˲⣬ͨŶ䣬õһԤ⣬ΪPNNģʽƷϵͳ״̬. 
Ƿɢϵͳ
xk=(xk-1,k-1)+(xk-1, k-1)wk,(1a)
zk=h(xkk)+vk. (1b)
Ц,,hֱΪnn,nm,1nάӳ䣬xkRn1Ϊ״̬wkRm1Ϊϵͳɫ vkRl1ΪɫzkRl1 Ϊϵͳ. ϵͳ(1)ǿɹ۲. 
ٶ״̬Ҫ;ǦǣҶڦǣģ0ʹã(xk-1,k-1)ģǳ.Ϊ㣬m=1Ϊֻƹ㼴. infwk=bk1,supwk=bk2ϵͳwkŶ[bk1,bk2]еĸʷֲܶȺǦ(k). 
[bk1,bk2]wkĸܶȺ(k)лֵ֣ΪL0=bk1,L1,,Lj-1, Lj=bk2;ʹwk(L0,L1],(L1,L2],,(Lj-1, Lj]еĸǾȵ. 
Ϊʵ˲Ԥڸٹ̣ Ŷ֮ѡֵ㣬maxLi-Li-1=ġ(i=1,2,,j),Bk=bk2-bk1,
Lj+1-Lj+2=Lj+2-Lj+3==Lj+p-1-Lj+p=-,
Lj+pbk2+Bk/2, Lj+p-1bk2+Bk/2,
L-1-L-2=L-2-L-3==L-p+1-L-p=,
L-pbk1-Bk/2, L-p+1bk1-Bk/2.
ǣLi(i=-p, -p+1, , j+p)ΪŶ[bk1-Bk/2, bk2+Bk/2]ķֵ. 
һĹk-1-xk-1ǣk-1Ϊ(k-1,)򵱦(k-1-,k-1)(k-1+,k-1)ʱ(k-1,)ƹһĴΪ((k-1,))=((k-1-,k-1),(k-1+,k-1))(k-1-,k-1)(k-1+,k-1)ʱ(k-1,)ƹһĴΪ((k-1,))=(((k-1+),k-1),((k-1-),k-1)).
һ覵(k-1-,k-1)(k-1+,k-1),Ҧ(k-1,k-1)0, ڦ(k-1-,k-1)(k-1+,k-1)(k-1,k-1)0ȫ. 
1=min(k-1,k-1)(Li-Li+1)(i=0, 1, , j), Ĺʽ
  (2)
Ҧ(k-1-,k-1)+i*1(k-1+,k-1)ͦ(k-1-,k-1)+(i*-1)1(k-1+,k-1). ɴˣɵõһϵͳ״̬x*i(i=-p,-p+1,,j+p+i).
ɼx*i(i=-p,-p+1,,j+p+i*)Ԥ״ֵ̬ķ. 
x*i(i=-p,-p+1,,j+p+i*)(1b)ʽõһԤi(i=-p,-p+1,,j+p+i*). ϵͳzk,{i-k}(i=-p,-p+1,,j+p+i*-1,j+p+i*)һԤֵͨBayesȫʹʽü{P(i-k)}(i=-p,-p+1,,j+p+i*-1,j+p+i*),i[-p,p+1,,j+p+i*-1,j+p+i*],ʹ
P(i-zk)=max{P(i-zk)},
i{-p,-p+1,,j+p+i*-1,j+p+i*}.(3)
ǣx*i伴Ϊ״̬ͻ״̬ƵBayesŽ. ɴ˿Եõ½. 
1.x*kΪBayesŹֵϵͳ(1)״̬ͻ䣬ɷֵL-p=(bk2-3bk1)/2,L-p+1,,Lj+p-1,Lj+p=(3bk2-bk1)/2wkĸܶȦ[bk1-Bk/2,bk2+Bk/2] ͨ(2)ʽõһϵͳܴڵ״̬(3)ʽõ״ֵ̬Сk-x*kޣ.
ڽϴֵƫϵͳ仯ԼŶ ΪǿŶɵ״̬ͻ[bk1-Bk/2,bk2+Bk/2]ѡֵ㣬ֵΪ
L*-1=bk1-Bk,L*-2=bk1-2Bk,L*-3=bk1-4Bk,,L*-M=bk1-2M-1Bk,
L*1=bk2+Bk,L*2=bk2+2Bk,L*3=bk2+4Bk,,L*M=bk2+2M-1Bk.
ЩֵԭеŶŶڷֵΪһ壬ʹ˲ĸΪǿ. ԴˣԵõ½. 
2.[bk1-Bk/2,bk2+Bk/2]ֵL*i(i=-M,-M-1,,-1, 1,,M-1,M)(1a)ʽеwkڳŶ״̬Ŷxkȡֵ࣬˲ĹΪ[2i䦣(k-1,k-1)+(k,k)]/3 . 2i䦣(k-1,k-1);xkȡֵڲ࣬˲ĹΪ[2i-1(k-1,k-1)+(k,k)]/2i+1(k-1,k-1).رأ(x,k)Ϊģ ǰʵֱֵΪ1/31/4 ٹÿʹΧΪԭе1/31/4.
ɴ˿ɼ PNN˲ĸȫȡŶŶֵ״ͨӦıʽٹ̣ʹ˲ݻ ķѾ֤ƪ ڴ˲. 
оɫ·ϵͳPNN˲ص㣺
1) ˲ĹƾBayesŹƼŶĻȷ
2) ˲ĸٶŶֵȷ ٹ̿ͨʽ. 
ˣ PNN˲иݻص. ţ÷ʵӦойķչǰ. 
1) õҽίʿ.
ߵλϺͨѧԶоϺ200030
ο
[1] ΰ֮ӿ.һֻ˹ķ˲. Ϣƣ1996 25(1)1-7.
[2]Specht D F. Probabilistic neural networks and the polyomial adaline as complementary techniques for classification. IEEE Trans. NN, 1990,1(1):111-121.
[3]Specht D F. Probabilistic neural networks. Neural Network, 1990, 3(1):109-118.
[4]Ruck D, Rogers S, Kabrisky M, Oxley M, Suter B. The multilayer perception as an approximation to a Bayes optimal discriminant function. IEEE Trans., N N, 1992, 3(1):296-298.
ո 1995-05-06
