Neural Network with Momentum for Dynamic Source Separation and its Convergence Analysis

This paper addresses the problem of blind source separation (BSS) of n independent sources from their m linear mixtures in the over-determined cases ( m n > ) with unknown and dynamically changing number of sources. The system architecture including an on-line source number estimator and an auto-adjust separation mechanism is considered based on the feed-forward neural network (FNN). To speed up and stabilize the iteration procedure, we propose to modify the FNN by adding a momentum term, and convergence analysis for the new algorithm is also presented, provided that the learning rate is set as a constant and the momentum factor an adaptive variable. Computer simulation results confirm that our approach is feasible for dynamic BSS cases and has satisfied convergence speed and steady-state error performance. Moreover, the proposed algorithm can ensure the separation of weak or badly scaled signals.