A nonlinear state space approach to nonlinear blind source separation

This paper addresses the prominent problem of separating noisy signals that have been convolutively mixed and nonlinearly distorted. The mixed signals are characterized by a nonlinear state space model which models both the statistical properties of the source signals and the overall nonlinear mixing process. A novel algorithm based on maximum likelihood framework has been rigorously developed for estimating the parameters in the model as well as inferring the source signals. In the proposed model, the nonlinear distortion function is modeled by using high order polynomials which enable the model to be formulated and optimized in a tractable manner. The strength of the proposed approach lies in the closed estimation of the source signals and the adaptive optimization procedure of the model parameters. This has resulted in high performance accuracy, fast convergence and efficient implementation of the estimation algorithm. Simulation has been conducted to verify the effectiveness of the proposed algorithm and the obtained results have shown 50% better accuracy than conventional nonlinear algorithms.