Algorithm for Nonlinear Blind Source Separation Based on Feature Vector Selection

A linear blind source separation algorithm based on generalized eigen-equation resolving is presented. Then a nonlinear blind source separation algorithm is proposed by extending the linear source separation algorithm to the nonlinear domain. The received mixing signals are first mapped to high-dimensional kernel feature space, and a feature vector basis given by the fitness function of the kernel feature space is constructed. Next, in the kernel feature space, the mixing signals are parameterized by the feature vector basis. Finally, the linear blind source separation algorithm based on signal variability is applied to the parameterized mixing signals. The proposed algorithm has simple computation and robustness, and is characterized by high accuracy. Simulation results illustrate well performance on the separation.