A novel iterative conditional maximization method for post-nonlinear underdetermined blind source separation

An iterative conditional maximization method originated from Bayesian statistics is proposed in this paper to offer a solution for blind source separation under a post-nonlinear underdetermined environment. The proposed algorithm estimate the sources and mixing matrix through their individual marginal probabilities instead of join probability. A Generalized Gaussian Distribution model is applied to approximate the prior information of probability distributions. The unknown nonlinear function is also estimated and modeled by a Multilayer Perceptron (MLP) neural network. All parameters are updated iteratively until convergence to a fixed state has been achieved. The proposed algorithm is tested on real audio wave and the performance is measured by modified Mean Square Error (MSE). The obtained results show that the proposed algorithm gains substantial improvements compared with the conventional linear algorithm.