Study On Blind Source Separation With Dynamically Changing Source Number

From the view of information theory, the main methods of Blind Source Separation (BSS) are stated briefly, and the Independent Component Analysis (ICA) is discussed in details. In most approaches of BSS, the number of source signals is typically assumed to be known a priori, but this is not true in practical applications. The divergence reason of the natural gradient algorithm which used to solve over-determined BSS problem with unknown and dynamical changing source number is analyzed. A new adaptive blind separation algorithm is given. With this algorithm, the BSS can converge stably to correct point. It also can simplify the complexity without computing the correction among output components. Finally, using the BSS algorithm based on dynamical neural network (DNN), the application of BSS in radar signal sorting is studied. The number of input radar signals can be determined by increasing or deleting the output number of DNN adaptively, and the good separation performance is obtained.