Adaptive fast algorithm based on natural gradient for instantaneous blind source separation

Abstract

A fast adaptive algorithm based on natural gradient is proposed in this paper to address the convergence issue in instantaneous blind source separation BSS problem in a noisy and non-stationary mixing scenario. The natural gradient technique overcomes slow convergence properties of the gradient adaptation technique when the slope of the cost function varies widely for small changes in the parameters. To speed up the convergence further we adopt one of the heuristic methods by incorporating a momentum term. Synthetically generated data as well as real world data have been considered for validation purpose. Numerical experiments on sinusoidal signals and acoustic electromechanical signals confirm the superior performance of the proposed algorithm over the conventional natural gradient algorithm NGA in both noisy and noiseless situation as well as in stationary and non-stationary mixing scenario.