Study On Blind Source Separation Algorithm

Blind source separation (BSS) aims to extract independent signals from their linear mixtures captured by a number of sensors without knowing the channel information. The separation is carried out only with the assumption of mutual statistical independence among the source signals and is hence called a"blind"method. It has broad application in speech recognition, multi-user communication, array signal processing and biomedicine etc. In recent years, BSS has drawn lots of attention in signal processing community and neural networks community. The main works of the dissertation can be summarized as follows:·Discussing the basic theory of blind source separation in detail and introducing the natural gradient algorithm and the back-propagation algorithm. Inspired by the back-propagation algorithm, we give an improved natural gradient algorithm, which incorporates a momentum term into the natural gradient for increasing the convergence speed. Simulation results show this method accelerates the convergence speed efficiently.·Although the improved algorithm can accelerate the convergence speed, the performance of the algorithm is sensitive to the step-size and the momentum factor. For achieving faster convergence speed and higher tracking capability, variable step-size and momentum factor must be used. So we present an adaptive improved natural gradient algorithm, which use an appropriate estimation function to control the step-size and the momentum factor. The ability of the algorithm to improve the convergence in a time-invariant mixing environment and to track a changing environment, such as time-varying mixing matrix or signal power, is demonstrated by extensive simulation results. Moreover, its good performance of separating extremely weak or badly scaled stationary signals is illustrated too.