A feedback approach to over complete BSS and its learning algorithm

When the number of sensors is less than that of the signal sources, this problem is called´ Over Complete BSS´ (OC-BSS), which is a difficult problem for lack of information about signal sources and a mixing process. A feedback approach has been proposed for the OC-BSS. In this paper, new learning algorithms used in the feedforward separation and the feedback cancelation are proposed. The separated single source is fed back to the inputs of the separation block, and is subtracted from the observations, in order to reduce the number of equivalent signal sources. Signal distortion, which is caused in the subtraction process, is suppressed by a spectral suppression technique. The learning of the feedforward separation is accelerated by using the constraints derived from the estimated mixing block. The proposed method can improve a signal to interference ratio by 6 ~ 10 dB compared to the conventional methods.