The Separation And Extraction Of Non-stationary Signals In Different Mixing Models

Blind source separation and extraction only use the observed signals measured by sensors to separate or extract source signals without knowing sources and how to mix.In signal processing field,blind source separation or extraction is a hot topic.It has been utilized in speech recognition,wireless communication,bio-medical and so on and owns wide prospect.Aiming at the problem,this paper puts the emphasis on the separation and extraction of non-stationary signals for different linear mixing models.The specific contents are as follows:First,for the linear standard mixing model,this paper proposes two algorithms for separating non-stationary signals by utilizing time-frequency analysis.The time varying features of non-stationary signals can be extracted by time-frequency analysis.By introducing time-frequency analysis,the blind source separation algorithms can utilize the time-frequency diversity to extract source signals.In traditional algorithms,the time-frequency concentration of the linear time-frequency transform is low and the bilinear time-frequency transform introduces the cross-term.This paper analyzes the disadvantages of the linear time-frequency transform and the bilinear time-frequency transform and then proposes a blind source separation algorithm by combining STFT and Wigner-Ville distribution.It wipes off the cross-term interference and retains good time-frequency concentration.It constructs a new strategy to choose time-frequency matrices and adopts joint approximate diagonalization to recover source signals.Besides,Combining with the Hough transform,this paper introduces local polynomial fourier transform that owns time-frequency concentration and restrains the cross-term to propose a blind source separation algorithm based on local polynomial Hough transform.In order to restrains noises and the cross-term,this algorithm transform the time-frequency plane that is formed by local polynomial fourier transform to the parameter plane.Meanwhile,this paper proposes another strategy to choose time-frequency matrices and lay a foundation for recovering source signals.The simulation results show that these two algorithms own good performance in separating non-stationary signals.Secondly,for the linear underdetermined mixing model,all kinds of blind source separation algorithms with single source points detection are proposed for the instantaneous mixing model and the time-delay mixing model.The methods based on single source points detection can be applied to solve the separation of non-stationary signals because there is no demandfor non-stationary.For the instantaneous mixed model,this paper proposes a real mixing matrix estimation algorithm.First,the observed signals are transformed to get better sparsity through time-frequency transform.Then,a method of detecting single source points are proposed.Finally,the improved clustering method can be utilized to get the mixing matrix.For the time-delay mixed model,this paper proposes a complex mixing matrix estimation algorithm.The observed signals for the time-delay mixing model can not get the clustering property.This paper constructs corresponding variables to form the clustering vector and utilizes the prior information of the receiving antenna array to detect single source points and estimate the mixing matrix.Besides,this paper proposes another complex matrix estimation method by detecting single source points.In this method,the detection method of the instantaneous mixing model is extended to the time-delay mixing model and the method of the instantaneous mixing model is used to estimate the mixing matrix.On the basis of estimating the mixing matrix,the method based on the subspace is adopted to recover the source signals.Experiment results prove that the proposed algorithms estimate the mixing matrix and sources well and own better performance than other algorithms.Finally,for the linear overdetermined mixing model,the blind source extraction algorithms based on the prior information are proposed for non-stationary signals.Some non-stationary signals own some prior information that can help recover source signals.There are many observed signals of the overdetermined model but few desired source signals,so blind source extraction is more efficient than blind source separation.This paper proposes three non-stationary signals extraction algorithms based on time structure and non-gaussianity for sources which own special time structure.According to the time structure and non-gaussianity of source signals,these algorithms construct the corresponding optimization model and get the unmixing vector to extract the desired source signal through iterations.What's more,the stability and the convergence property of the algorithms are analyzed.Simulation results show that three algorithms have different advantages in accuracy,robustness and convergence.According to the orientation information of the desired source signal,a blind source extraction algorithm is proposed.It utilizes the orientation information of the desired source signal and combines the minimum variance distortionless response and independent component analysis with reference to get the unmixing matrix,and then extracts the source signals with the FastICA algorithm.Simulation results show that the algorithmowns good performance.