改进的高阶收敛FastICA算法.PDF

第32 卷第10期 东 北 大 学 学 报 ( 自 然 科 学 版 ) V o l. 32, No. 10 2 0 1 1 年 10 月 Journal of Nort heastern U n vers ty( N atural Sc ence) O c t. 2 0 1 1 FastICA 1 1 2 2 季 策 , 胡祥楠 , 朱丽春 , 张志伟 ( 1. , 110819; 2. , 1000 12) : FastICA , , , . , FastICA . , , FastI CA . : , . : ; ; FastICA; ; : T N 911. 6 : A : 1005- 3026( 2011) 10- 1390- 04 Improved Higher Order Convergent FastICA Algorithm JI Ce, H U X iang-nan, ZH U L i-chun , ZH A N G Zhi-w ei ( 1. School of Infor mat o n Sc ence Eng neer ng, No rtheastern U n vers ty , Shenyang 110 819, Ch na; 2 . Nat onal Astronom cal Observator es, Ch nese Academy of Sc ences, Be j ng 1000 12 , Ch na. Correspond ng author: JI Ce, E- ma l: hux angnan 121 @ 163. com) Abstract: H gh order F ast ICA algor thms have the advantages of s mple f orm and f ast convergence rate. However, they are sens t ve to t he r n t al values aff ect ng convergence effect and even result ng n nconvergence f the n t al values are not chosen appropr ately. T o solve t he problem, the F ast ICA algor t hms of the t h rd and f f th order converg ence w ere mproved w t h t he st eepest descent met hod. F rst , the n t al v alues w ere calculat ed w t h t he st eepest descent met hod. T hen , the opt m al solut on w as calculat ed w t h t he h gh order convergence F ast ICA algor t hm . Speech s gnal separat on ex per ments show ed that t he m proved algor thm can separate m x ed s g nal and overcome t he n t al value sens t v ty problems eff ect vely. Key words: ndependent component analys s; N ew ton s terat on method; Fast ICA; st eepest descent method; n t al value sens t v ty