Variable Step-Size Online Algorithm for Blind Separation Based on the Extended Infomax

A novel variable step-size online algorithm for mixed signals with sub- and super-Gaussian source distributions based on the extended infomax is present. The extended infomax algorithm usually separates the sources by batch processing and it requires adequate samples to estimate the kurtosis of the output signals so the algorithm will be invalid when the channel matrix is changed. An improved online estimation model of the kurtosis is introduced in this paper, we interpose a detection machine-made to judge whether the channel matrix is changed or not in separation process. In order to solve the ambivalent tradeoff between convergence rate and steady-state error, a variable step-size online algorithm is proposed. The step-size updating regulation is controlled by the kurtosis variance of the signals because the kurtosis fluctuation can describe the state of separation. This online algorithm accelerated the convergence rate and reduced the steady-state error efficiently.