Research On The Algorithm Of Frequency Domain Blind Source Separation Based On ICA

The main idea of blind source separation(BSS)is the process of separating and extracting various source-signals which can not be observed directly from the observation signals received from a number of sensors.According to the difference of mixed signals,blind source separation can be divided into linear or nonlinear mixed models.At present,the main research is linear mixed model,including linear instantaneous mixtures and linear convolution mixtures.Linear instantaneous mixed model has been studied very thoroughly,and it has been used in many fields.Independent component analysis(ICA)is the most commonly method of blind source separation.However,in the real environment,the transmission of signals is often influenced by many factors,and the mathematical model of the source signal is closer to the convolutive model.This thesis will study blind source separation algorithm in frequency domain based on ICA,and mainly include the following aspects:First,the basic theory of blind source separation was introduced in detail,and the basic principle and properties of independent component analysis also was systematically introduced,including the ICA model,the uncertainty of constraint conditions and special features.Furthermore,several typical objective functions and optimization algorithms of independent component analysis are studied.Second,the basic principles of blind source separation algorithm in frequency domain was introduced,and blind source separation algorithm in the complex domain was introduced in detail,and the influence of ICA's inherent amplitude uncertainty and permutation uncertainty on blind source separation frequency domain algorithm were thoroughly discussed.Thirdly,for the permutation uncertainty of blind source separation frequency domain algorithm,the characteristics of the separation signals after using the short time Fourier transform of the speech signal were studied,and the traditional sorting algorithm based on the amplitude correlation in frequency domain was carefully analyzed.Then,an improved amplitude correlation sorting algorithm is proposed based on the short time mean amplitude function and determining the undesired separation frequency points.Simulation results show the effectiveness of the proposed algorithm.Fourth,direction of arrival(DOA)sorting algorithm was studied,and the advantages and disadvantages between the amplitude correlation sorting algorithm and DOA sorting algorithm also were analyzed,then DOA sorting algorithm is combined with the improved amplitude correlation sorting algorithm.A dividing frequency sorting algorithm is proposed based on the improved amplitude correlation sorting algorithm and DOA sorting algorithm.Simulation results show the effectiveness of the proposed algorithm.Fifth,the basic principle of constraint independent component analysis(CICA)is introduced,and the power density spectrum and statistical difference of the modulus in frequency domain between the target speech signal and the diffuse background noise were introduced in detail.Because the speech has a sparser modulus than the diffuse background noise,a blind speech signal extraction algorithm is proposed based on the modulus diversity constraint.Simulation results show the effectiveness of the proposed algorithm.