The Research Of Blind Source Separation Algorithm Based On Noisy Model

As the research hotspots in signal processing area, Blind source separation(BSS)has been successfully applied to the fields of image and the speech signal processing,biomedical signal analysis and processing, the antenna array signal processing and so on.The main purpose of BSS is only using the observation data received by the sensors toresume the source signals, while under the situation with unknown transmit channel andunknown sources.This paper has systematically explained the main solution of the blind sourceseparation problem----Independent Component Analysis (ICA). The basic assumptions,basic characters, optimization criterion,performances and evaluation index of ICAalgorithms are summarized. FastICA and natural gradient algorithms are analyzed basedon the model of instantaneous linear mixture blind source separation.Most blind source separation algorithms are established on noisy-free model.However, when the signal involves noise, the performance of these algorithms will dropeven defeat. In order to solve this problem, two kinds of noise separation model are setup, according to the different superposition of noise. On this basis, it summarizes fourmethods to solve noisy BBS. The classical BSS algorithms were improved to eliminatenoise caused by the deviation, such as Gaussian moments of FastICA algorithm and thebias removal technology of natural gradient algorithm. Wavelet analysis denoising isalso studied for BSS.The main innovations of this paper are proposed as follows:(1)FastICA algorithm is optimized by the quadratic convergence of Newtoniteration method and is widely applied in BBS. In order to accelerate the convergencespeed of the algorithm, this article gives two improved FastICA algorithms which use thefifth-order convergence of Newton iterative method. The experimental results show theperformance of the improved algorithms are better than the original algorithm.(2)A new wavelet threshold function is proposed by analyzing the main factorswhich affect the performance of wavelet threshold denoising.(3)The bias removal technology of natural gradient algorithm is used under thenoisy circumstances to estimate the demixing matrix. Then the new wavelet thresholdfunction is applied to separated signals to remove noise. This method is robust andimproves the SNR of the ultimate results obviously.Though BSS has matured to a point, there are still many problems and difficult in practical applications, so the direction of next research is briefly discussed at the end ofthis paper.