Research On Underdetermined Blind Source Separation And Extraction For Anechoic Mixtures

In recent years,the blind source separation and extraction have received more and more attention in the signal processing field.And they have been widely used in biomedicine,wireless communication technology,image processing,speech signal processing and many other research fields.Traditional blind separation algorithms are based on independents component analysis,which can only solve the overdetermined or normal mixing models.However,due to the limited number of receiving devices and unknown number of source signals,the underdetermined case is more common in the actual situation.Most of the existing underdetermined blind source separation algorithms are aimed at the linear instantaneous mixtures,ignoring the propagation delay between the source signal and the receiving array.In this paper,we mainly study on the underdetermined anechoic mixtures,including the source signal separation and the interested source signal extraction:First of all,in order to solve the problem of undetermined blind source separation under anechoic model.A mixing matrix estimation algorithm based on single source domain detection is proposed under the single source time-frequency neighborhood assumption.A new single source domain pre-extraction algorithm is used to filter out the single source time-frequency windows.Next,the single source time-frequency windows are clustered to estimate the attenuation parameters.Then the single source domains will be further screened and grouped by the estimated attenuation parameters.Finally,the clustering analysis of each single source domain group is performed to estimate the delay parameters.The estimation of the mixing matrix has laid a great foundation for the source signals recovery.As we can see from the experiments that our algorithm improves the single source domain detection accuracy.And it has a great estimation performance even when there is a large number of source signals.Then,in order to recover the weaker sparsity signals from the underdetermined anechoic mixtures,we mainly study the characteristics of the single source points in the spatial time-frequency distribution of the source signals.And a new single source points detection criterion is proposed which relaxes the sparsity hypothesis of the source signals.First,the proposed single source points detection criterion is used to detect the single source points.To further enhance the estimation accuracy,the clustering analysis of the selected single source time-frequency points is performed.Then the estimated mixing parameters are obtained as well as the mixing matrix.Finally,based on the estimated mixing matrix,the recovery signals can be obtained.Simulation results show that our algorithm can greatly separate the source signals from the underdetermined anechoic model.And compared with the existing algorithms,our algorithm has higher estimation accuracy and stronger noise robustness in the aspect of the mixing matrix estimation.Finally,to solve the blind signal extraction problem which the number of interested signals is far less than the number of source signals,a new ICA with reference signal algorithm is proposed in this paper.A closeness measurement function is obtained by the prior azimuth information of the target signal of interest.Next,it combines with the ICA algorithm to get a new optimization model.Then,iterate and optimize until the optimal separation weight vector obtained.Thus realize the extraction of interested source signal.The simulation experiment achieves the accurate extraction of the interest target.Even if there is a great deviation of the prior azimuth information,the target signal can be extracted accurately.