A new algorithm for extracting specific signals with temporal structure

In this paper we develop a batch learning algorithm for semi-blind extraction of a desired source signal with temporal structure from linear mixtures. By studying the discrete time algorithm, an invariant set is obtained so that the non-divergence of the algorithm can be guaranteed. In the invariant set, the local convergence of the algorithm is analyzed. It is proven that the trajectories of the algorithm starting from the invariant set will converge to the desired source signal which is the most autocorrelated for an specific delay. The simulations verified the results.