Blind Source Separation Algorithm Research With Communication Signals

With the rapid development of computer technology, digital signal processing technology in communication, medicine and other fields has been widely used. Blind Source Separation (BSS) technique as a sophisticated signal processing method is researched by many scholars.BSS recover unknown signals only based on independent statistical characters of the original signals without any prior knowledge of the signal transmission channel and source signals. The BSS can be divided into linear blind source separation and nonlinear blind source separation by transmission mode. Under linear map, BSS can be resolved only use the independent statistical character between the source signals. Nonlinear map is a sick problem and hard to make. It need more work on this subject.This paper researched the linear and nonlinear blind source separation.For the linear mapping BSS problem, we systematically studied the algorithms based on information theory, joint approximate diagonalization and the negative -entropy, in which negative-entropy-based FastICA algorithm has the advantage of fast convergence, suitably applied in the engineering environment, but its drawback is also exist. The answer is sensitive to the initial of separating matrix, inappropriate initialization of the separating matrix will come to wrong solutions. In this paper, FastICA algorithm is improved by combing Newton's method and Shamarskii method to change the iteration mode. This will reduce the sensitivity on the initialization of separating matrix. Real signal simulation and complex signal simulation showed that the improved FastICA algorithm is not sensitive to the separating matrix which randomly initialed, and the separation efficiency and convergence rate are also improved. Compared with algorithms based on information theory, joint approximate diagonalization, improved FastICA is better. For the case of non-linear mapping, this paper studied Markov Post-Nonlinear separation (Markov-PNL) algorithm and the mutual information Post-Nonlinear separation (MIM-PNL). Firstly for the Markov-PNL, this paper discussed the effect which Markov order make on. The stimulation shows the weak instance of the Markov-PNL algorithm; as traditional MIM-PNL and Markov-PNL both calculate the score function, the convergence speed is slow. Base on parametric of the score function, this paper use multilayer perception for nonlinear blind source separation and add a damping term to the iteration which speed up the convergence. Simulation result shows that the improved MIM-PNL algorithm increased separating efficiency and convergent speed in some extend.