Research On The Algorithms Of Blind Source Separation

This dissertation primarily discusses the linear mixture model of the sources which is most common in BSS, and reviews the independent component analysis (ICA) which is used to resolve BSS problems. The uppermost contributions of this dissertation are as follows:1. In BSS, the general methods firstly assume the probability density function (pdf) of sources to obtain the important activation function (AF), and then separate the source signal from mixture signals. If the assumed pdf is different from the true pdf considerably, the sources will not be separated correctly. Aiming at this problem, we brings forward a kind of switching algorithm based on kurtosis, which is used to adaptively learn activation function of the ICA without assuming the pdf of sources. Computer simulation shows that this algorithm can separate sources effectively.2. Standard stochastic gradient algorithm of blind source separation is only adapted to Euclidean space. In general Riemannian space, natural gradient algorithm should be used, which, however, leads to numerical instability when the source signals are non-stationary or their magnitudes change rapidly. The nonholonomic constraints can avoid the instability. Using generalized Gaussian distribution model to simulate the nonlinear activation function, the nonholonomic natural gradient algorithm can be adapted to the source signals of arbitrary distribution by choosing different Gaussian exponents. Computer simulations show the validity of proposed method.