摘要
该文提出了一种多个均匀分布的源信号混叠的盲分离几何算法,该算法以矩阵的QR分解原理为分离的理论指导,并结合信号在各个阶段其scatter图所具有的特殊几何性质,首先将混叠信号进行白化,使其scatter图恢复为独立时的scatter图形状,然后将白化后的scatter图通过C2n次旋转变换,使其与各坐标轴平行,从而得到n个信号的分离.该方法第一次从代数上给出了几何算法的理论指导,从而真正得到了几何算法向多个信号混叠的推广.该算法不仅计算简单,同时有很好的仿真分离效果.在三个信号混叠的情况下,相对于Hyvarinen(2000)在分离时间上缩短了近30%.
Based on the idea of QR decomposition, a geometric blind separation algorithm is proposed for uniform distribution source signals in this paper. It is known that if source signals follow uniform distribution, the scatter plots of source signals and mixtures will show special geometrical property. Firstly, the mixture signals are whiten so that the scatter plot of which is changed into the scatter plot when they are independent, and then the scatter plot of whiten signals is rotated by Cn^2, times Givens transformation. Finally, the scatter plot is paralleled to the coordinate axis. Consequently, the mixture signals are separated. A theory guidance to implement the geometric algorithm is given from the viewpoint of algebra. And this paper firstly presents an algorithm to do blind source separation of more than three source signals. Compared with the algorithm of Hyvarinen（2000）, not only the separation accuracy of this algorithm is better, but also the separation time decreases. On the same condition that the mixture signals are mixed by three source signals, the experiment shows 30 percent decrease in the separation time.
出处
《计算机学报》
EI
CSCD
北大核心
2005年第9期1575-1581,共7页
Chinese Journal of Computers
基金
国家自然科学基金(60274006)
国家杰出青年自然科学基金(60325310)
国家教委跨世纪人才培养计划基金
广东省自然科学团队研究项目基金(04205783)
广东省自然科学基金重点项目(020826)资助
关键词
旋转变换
几何算法
盲分离
白化
scatter图
Givens transformation
geometric algorithm
blind separation
whiten
scatter plot
