A robust H ∞ learning approach to blind separation of slowly time-varying mixture of acoustic electromechanical signals

Although many techniques have been developed for solving the blind source separation (BSS) problem, some issues related to robustness of BSS algorithms are yet to be addressed. Most of the BSS algorithms developed assume the mixing system to be stationary. In this paper, we present a robust approach based on H ∞ learning to address the instantaneous BSS problem in a non-stationary mixing environment. The motivation behind applying H ∞ filter is that these are robust to errors arising out of model uncertainties, parameter variations and additive noise. Acoustic electromechanical signals have been considered for simulation purpose. Simulation results demonstrate that the H ∞ filter performs superior to Kalman filter and VS-NGA algorithm. To ensure practicability of the proposed approach, the H ∞ learning algorithm has been implemented and tested on Texas Instrument's TMS320C6713 floating point DSP platform successfully.