о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999 36 8 Vol.36 No.8 1999



СĹѹо
ʤ
ժҪнһֻСģ͡С.ʵѡȡṹСʵ˶Իѧʺݵѹͷ.ʵڴѹݵͬʱܺܺõػָԭʼס׼ȷطӳշλúǿ.ڷ෽иߵķֱʺȡ.ͨBPѵԱȣСӦǿٶȿ켰ŵ㣬رڷƽȡźŵķо.ˣСڹ״ŽϺõӦǰԽ.
ؼʡС磬ѹ࣬
ͼšTP399
STUDY OF SPECTRAL DATA REPRESENTATION AND CLASSIFICATION FOR WAVELET NEURAL NETWORK
ZHANG Yong-Sheng and YU Ke
(Department of Computer Science, Wuhan University of Technology, Wuhan 430070)
(Optical Fiber Sensing Research Center, Wuhan University of Technology, Wuhan 430070)
AbstractA model of neural network based on wavelet analysiswavelet neural network, is introduced in the paper here. The chemical substance infrared spectral compression representation and classification are realized with adaptive network structure and daughter wavelet.The experimental results show that the original spectra can be recovered well, the place and intensity of absorptive peaks can be defined exactly, and the data can be compressed greatly with the wavelet network.The resolution ratio and the characteristic collection ability of the wavelet network are better than that of other networks in classification. A comparison of the two training results of the wavelet and BP networks indicates that the wavelet network has better adaptability and faster convergence speed and can shield random noise. The wavelet network can be applied to the study and analysis of nonstationary and nonlinear signals, so it holds a bright future in spectra processing.
Key wordswavelet neural networkdata compressionclassification, infrared spectra
1
ǽл΢۽ṹԼҪ.׼ȷӳ⻯ĽṹͨҪӴĹ.ڱ֤Ҫǰ£ζԺ׽ѹϴؼݣƺ׵Ĵ洢ȴʽǺо.С12ݼС磬ǻС34ݶһµģ.С任ڴźʱнϺõĸƵʱ侫Ⱥ͵ƵƵʾȵŵ㣬ѹģʽʶȷŹ㷺Ӧ.ĽСں׵ѹͷ.ͨСȨֵڴѹݵͬʱܹܺõػָԭйף׼ȷطӳշλúǿȣӶﵽѹźԭʼźŵƥ.ڷ෽Сиߵķֱʼٵѵ.
2ԭ
2.1ݱʾ
С任ʵһֲͬռ֮Ļֱ任5

Ǿн֧ĺΪСͷֱΪnάmάռе.
һάźf(t),Сh(a,b,t)()԰ʽ

h(t)ΪСĸСΪһϵa,bֱΪh(a,b,t)Ӻƽӣźf(t)ֲṹķֱͨڲa,bСڵĴСλʵ.
СǻСģͣ÷СȡͨķSigmoidźűͨѡȡСԵʵֵģźs(t)Сh(a,b,t)ϣ

(3)ʽУΪźţwk,bkakֱΪȨֵСƽӺӣkΪСĸͼ1չʾ˽һһڵĵṹ.
wk,bkakͨСŻ

MΪݲ.ʽ(3)вMorletĸС(ͼ2ʾ)СҵƵĸ˹ʱƵͬʱֱʸߣ


ͼ1ѹʾС


ͼ2MorletĸС

t=(t-bk)/akEݶȷֱΪ



ùݶȷŻwkbkakֱΪiѭ

Tת.ͬ尴ʽеڣ



Ĳһά䲽6ݼѲÿѭʱʽӷֱwߣbߺaֱEСĳһ趨ѭΪֹ.Ϊ˺ʾÿѭEʽ

2.2ݷ
ıСṹԶźŽۺϷ.ѵΪPNڵ㣬Եpnڵʽʾ: 

УһSigmoid.ͼ3ʽ(14)ķСṹʾͼ.
wkbkakͨEŻ

Ϊpnڵ.ʽ(5)





ͼ3С
2.1ĹݶȷŻwk, bkakֱEСĳһ趨.
3
3.1ѹ
Ƕһ, 䡪ϩͪͪĻ(4000452cm-1Ϊ4cm-1888ݵ)ѹо׵ȫݵΪѵΪ룬͸ΪĿ30СѵwkakʼֵΪ(Χ-100+100)bkĳʼֵȡھȷֲΪ˼ӿٶȼֲСҪԲͲȵӦĲΪbkĳʼֵ,ѵΪ200ΣʼֱΪw=10-2a=b=10-7ѵEΪ0.000275ȫΪͼ˻ָEҲΪ0.000275.ͼ4˸ûԭʼͻָĺ.


ͼ4, 䡪ϩͪͪͼ

ԿָĹܺܺõرԭй׵ԣرܹ׼ȷطӳշλúǿ.СѡȡһʮҪ⣬̫࣬Ȼʹٶȼӿ죬ҪӣѹҪͣ෴̫٣ѹȿ󣬵ʹٶȱ.ʵ飬ѡ30ǱȽϺʵ.׵ԭʼ͸ʵݿ30wk,bkak90ʾӶʵ׵ѹѹԼΪ101Чǿ.1ʵ׷ֵָ׷ֵ.
3.2׷
һЩʵĳЩеĲ֮ͬŹ׽ṹ(λǿΣĿȵ)ڸʽṹ佨ĳȷĶӦϵԲͬʽз.ǶOHNHCO廷4ʽг࣬ѡ4е125ʵΪѵ44ʵĺΪ⼯2ڵɣ4ƴʾ.ʼԲݽԤΪ˷ͻС˹̶Ⱦز10̶С(bkakk=1,2,...,10ȡĳЩ̶ֵ)ʼȨֵΪһֻͨȨֵŻ磬wȡֵ.155ѵδѵĺ׼4ʵʶɴ88׼ȷ.ٲöȨֵwСƽbkakʵȫŻ磬wabȡֵ150ѵ󣬷׼ȷʴﵽ94.ѡúʵС.
Сڶԭʼݷ¼ŵ㣺ýٵѵɴﵽϺõѵЧӶӿѵٶȣʶֱ.С԰ѳ߶ȡ任ȵĲԽкϲѵ.۾кܺõȡرڷƽȡźŵķ.
ǶBPСͬͳʼȨֵ5000εѵȽٶȣͼ5ʾѧϰߣɼСٶBP.


(a) BP磻(b) С
ͼ5ͼ
1ʵ׷ֵָ׷ֵ

ʵ׷ֵָ׷ֵ
(cm-1)͸(%)(cm-1)͸(%)
1324356.2324356.218130
2129918.5129818.215226
3128715.3128816.033068
4126810.112688.238141
5124218.0124119.981833
6122217.8122216.210110
762848.962646.037789
859755.559754.403103
958144.658144.159781


4 
ۿ֪СԺѹͷһӦüֵչʾڹ׷Ӧǰ.ⲨΣСӣѸӲҪҲҪߣ㷨ĽһŻۺܵһ.ҪָǣСĳʼֵѵкܴӰ.⣬òͬĸСпܻСѵԼӿٶ.
ĿõȻѧ(Ŀ5943206)
߼飺ʤУ19649˶ʿҪ硢ݿ⿪Ӧо.ɣУ19637ʿڣҪ¹˴мܲо.
ߵλʤ人ҵѧѧϵ人430070
ɡ人ҵѧġ人430070
ο
1Zhang Q, Benvenise A. Wavelet network. Proc IEEE Trans on Neural Network, 1992, 3(6): 889898
2Harold H S, Brain T, Shubha K, Neural network adaptive wavelet for signal representation and classification, Optical Engineering, 1992, 31(9): 19071916
3Bos M, Hoogendam E, Wavelet transform for the evaluation of peak intensities in flow-injection analysis. Analytica Chimica Acta, 1992, 267: 7380
4Bos M, Vrielink JAM. The wavelet transform for pre-processing IR spectra in the identification of mono and disubstituted benzens. Chemometrics and Intelligent Laboratory Systems, 1994, 23(2): 115122
5ǰ, ڿ. ʵС. : ӿƼѧ,1995
(Qin Qianqing, Yang Zongkai. Practical Wavelet Analysis. Xi'an: Xi'an University of Electron Science and Technology Press,1995)
6,,ĵ. Ӧü. ɳ: Ƽѧ,1993
(Hu Shouren, Shen Qing, Hu Dewen et al. Application Technology of Neural Network, Changsha: National Defence University of Science and Technology Press, 1993)
յڣ1997-11-18
޸ڣ1998-05-25
