Blind source separation based on subspace
Shangzhi Xu, Zhongfu Ye · 2006
The problem of blind source separation consists in recovering mutually statistically independent source signals from mixtures when nothing is known about the sources and the mixture structure. In general, we focus on the real linear instantaneous mixture of real mutually statistically independent sources. When the column vectors of mixing matrix are uncorrelated, it has the possibility that some noise subspaces and signal vectors can be defined. In this way, in this paper, a new simple method of blind source separation is proposed by using higher order statistic, which is based on those signal subspaces and the central limit theorem. This method searches directly the row vector of separating matrix at first, and then revise the cost function to continue to search. In our opinion, this method integrates the characteristics of the like fastICA methods and the like InforMax methods. Illustrative examples, which use different mixing matrices, are utilized to demonstrate the ability of the considered method.