A robust H∞ learning approach to blind separation of signals

A robust estimation technique based on the H∞ filter (learning) is proposed in this paper to address the instantaneous Blind source separation (BSS) problem in a non-stationary mixing environment. It is assumed that the variations in the mixing system are small. The learning algorithm is obtained by applying H∞ filter to the BSS model with state-space representation. The motivation behind applying H∞ filter is its robustness towards errors arising out of model uncertainties, parameter variations and noise. The proposed algorithm is applied to both synthetically generated signals and practical sound signals. A performance comparison between the H∞ filter, Kalman filter, ICA based on mutual information and Nonlinear PCA establishes the robustness of the proposed H∞ approach.