Aimming to the blind separation of non-stationary sources, a new on-line algorithm base on self-organizing
network is proposed. The separation structure and cost function are established by utilizing the multilayer networks. The
learning rule for the network’s parameters is derived from the nature gradient descent minimization of the cost function that
takes the minimum only when the network outputs are uncorrelated with each other. The multilayer networks and nature
gradient principal are combined, which improves the performance in non-stationary signals separation. Finally, simulation
results show that the proposed algorithm has good convergence precision and stability.