A second-order statistics based algorithm for blind separation of signals

We propose an information theory based objective function for measuring the statistics independent of source signals. Then, we develop a learning algorithm for blind separation of nonstationary signals by minimizing the objective function, in which the property of nonstationary and direct architecture neural network is applied. The analysis demonstrates the equivalence of two neural architectures in some special cases. The computer simulation shows the validity of the proposed algorithm. We give the performance surface of the object function at the last of the paper.