基于优化得分函数差分的频域盲解卷积算法  

A Frequency Blind Deconvolution Based on Optimal Score Function Difference

作　　者：许海翔[1] 黄克[1] 雷菊阳[1] 史习智[1] 

机构地区：[1]上海交通大学机械系统与振动国家重点实验室,上海200240

出　　处：《上海交通大学学报》2009年第3期361-366,共6页Journal of Shanghai Jiaotong University

基　　金：上海市科委重点基础项目(05JC14026)

摘　　要：基于完全互信息梯度表达式及得分函数差分的性质,提出一种在线频域盲解卷积块算法.本算法依赖于系统输出分布的优化多元得分函数,对源信号分布类型不作要求,为了避免幅值不确定引起算法收敛到平凡解,在算法的相对梯度中加入了单位方差约束;利用得分函数差分的性质,去除期望为零的矩阵元素,以保证相对梯度的正确方向.通过仿真比较,说明了该算法收敛速度快、分离效果较好且较稳定.在对观测信号进行预处理后该算法可用于语音源信号的盲解卷积.An online block algorithm for frequency domain blind deconvolution was proposed based on the complete mutual information gradient and the properties of the score function difference （SFD）. Contrary to simple nonlinear functions usually used in equivariant adaptive separations via independence （EASI）, the nonlinear score function of the proposed algorithm is neither fixed nor component-wise, it is an optimal nonlinear multivariate function which depends on the output distribution, so the proposed algorithm need not character the style of source signals. For overcoming the scale indeterminacy, the unit variance outputs are added to the relative gradient matrix. Some properties of SFD are also used to remove statistical zero elements, which guarantees the correct direction of the relative gradient matrix. The simulation shows the new algorithm performs more effectively in convergence and stability compared with natural gradient algorithm. With some preprocessing added to the observed signals, the proposed algorithm is more applicable to tackle speech language＇s blind deconvolution problem.

关 键 词：盲解卷积 得分函数差分 相对梯度 

分 类 号：TN911.3[电子电信—通信与信息系统]