Overdetermined Blind Source Separation Based on Singular Value Decomposition
Zhu Xiao-long①; Zhang Xian-da②
①Key Lab. for Radar Signal Processing Xidian University Xi’an 710071 China;②State Key Lab of Intelligent Tech. and Sys., Dept. of Automation Tsinghua University Beijing 100084
Abstract The problem of overdetermined Blind Source Separation (BSS) where there are more mixtures than sources is considered. Beginning with the Singular Value Decompo sition (SVD) of the separation matrix, a cost function is presented based on Independent Component Analysis (ICA), and then the ordinary gradient learning algorithm is developed. Secondly, resorting to the relative gradient, it is shown that the natural gradient learning algorithm for overdetermined BSS has the same form as that for usual complete BSS, which is verified by simulation results.