东北大学学报:自然科学版 ›› 2014, Vol. 35 ›› Issue (2): 204-207.DOI: 10.12068/j.issn.1005-3026.2014.02.012

• 信息与控制 • 上一篇    下一篇

引入松弛因子的高阶收敛FastICA算法

季策1,王艳茹1,沙明博2,杨正义3   

  1. (1.东北大学 信息科学与工程学院, 辽宁 沈阳110819; 2.奥维通信股份有限公司, 辽宁 沈阳110179;3.中国人民解放军94816部队, 福建 福州350002)
  • 收稿日期:2013-01-08 修回日期:2013-01-08 出版日期:2014-02-15 发布日期:2013-11-22
  • 通讯作者: 季策
  • 作者简介:季策(1969-),女,辽宁沈阳人,东北大学副教授.
  • 基金资助:
    国家自然科学基金资助项目(11273001,61074073,61273164);教育部新世纪优秀人才支持计划项目(NCET-10-0306).

Relaxation Factorbased FastICA with Higher Order Convergence

JI Ce1, WANG Yanru1, SHA Mingbo2, YANG Zhengyi3   

  1. 1. School of Information Science & Engineering, Northeastern University, Shenyang 110819, China; 2. Allwin Telecommunication Co., Ltd., Shenyang 110179, China; 3. 94816 PLA Troops, Fuzhou 350002, China.
  • Received:2013-01-08 Revised:2013-01-08 Online:2014-02-15 Published:2013-11-22
  • Contact: JI Ce
  • About author:-
  • Supported by:
    -

摘要: 高阶收敛的FastICA算法对初始值的选择较为敏感,如果初始值选择不当不仅会影响算法的收敛效果,甚至可能导致不收敛的结果.针对这一问题,将松弛因子引入高阶收敛的牛顿迭代法中,通过适当的修正,获得了既能保证一定收敛速度,又能有效克服初值敏感性的改进三阶、五阶FastICA算法.仿真工具采用Matlab软件,应用3种算法对语音信号进行分离;结果表明,对比基本FastICA算法,改进后的算法有效地分离了混合信号,并且降低了算法对初始权值的依赖性.

关键词: 独立分量分析, FastICA, 松弛因子, 初值敏感性

Abstract: High order FastICA(fast independent component analysis)algorithm has the characteristics of simple form and quick convergence. However, the algorithm is sensitive to its initial value which affects the convergence effect and even results in nonconvergence if it is not chosen appropriately. In order to solve this problem, a relaxation factor is introduced into high order Newton iterative method. Through the appropriate correction, the improved high order FastICA algorithm can be obtained, which can not only guarantee the convergence speed, but also effectively overcome the initial value sensitivity problem. Applying the algorithm to the separation experiment of speech signals, the result shows that the proposed algorithm effectively separates the mixed signal,and reduces the dependence on the initial value.

Key words: independent component analysis(ICA), FastICA, relaxation factor, initial value sensitivity

中图分类号: