Blind Sources Separation of Non-stationary Signals Based on Adaptive Maximum Signal-to-Noise Ratio Method

In order to improve the blind separation performance of non-stationary signals, a new blind source separation algorithm named adaptive maximum signal-to-noise ratio algorithm was proposed. This algorithm uses the signal noise ratio function as the cost function parameter and an improved multinomial coefficient autoregressive model to estimate the best length of moving average window. Simulations showed that FastICA algorithm needs to assume the probability density function (PDF) of the sources to approximate their un-Gaussian features by choosing the appropriate nonlinear function. If the assumed PDF considerably deviates from the true one, the sources could not be separated correctly. In the case of the sources with identical kurtosis, the separation algorithm using cumulants failed to separate the sources. The comparison between the proposed method, the classical FastICA algorithm, and the separation algorithm using cumulants showed that the proposed method could retrieve the time-varying non-stationary source signals accurately, and the separation performance of the proposed method was not influenced by the PDF and the kurtosis of the source signals.