| Acoustic signal separation technology is a basic problem in the field of signal processing,and it is also a signal processing problem that needs to be solved urgently in complex environmental acoustic sensing applications such as sensor networks.Acoustic signal separation technology can be divided into single-channel and multi-channel separation from the perspective of the number of channels.In this paper,the space-time frequency mode in multi-channel acoustic signal separation is studied,and optimizes the Multichannel Non-negative Matrix Factorization(MNMF)model.The non-negative matrix factorization technology is widely used in the research of single-channel and multi-channel acoustic signal separation technology because of its simple implementation advantages.Usually,the amplitude spectrum matrix of the mixed signal is factorized into the form of multiplying two non-negative matrices.In this study,while fully considering the signal characteristics,it also fully mines the multi-channel spatial information,so as to separate the acoustic signal of the MNMF model.The main researches are as follows:Aiming at the problem that the sparseness of basis matrix and coefficient matrix in MNMF model is difficult to control,a sparse orthogonal joint constraint multi-channel non-negative matrix decomposition acoustic signal separation method is proposed.Firstly,the cost function is constructed by sparse orthogonal joint constraint terms based on multi-channel extended Itakura-Saito(IS)divergence,and the auxiliary functions of signal sparse and signal orthogonal constraints are given on the basis of the cost function to realize the cost function minimization solution.Then,by designing iterative update rules,the multi-channel non-negative matrix factorization basis matrix and coefficient matrix of sparse orthogonal optimization are obtained.Finally,based on the spatial characteristics of multi-channel signals,non-negative matrix factorization basis clustering is carried out.The model parameters are brought into the multi-channel Wiener filter to solve the separated signal,and the influence of the sparse orthogonal constraint on the sparsity and continuity of the basis matrix and coefficient matrix is discussed.In order to further improve the performance of acoustic signal separation,a sparse orthogonal joint constrained MNMF acoustic signal separation method based on Direction of Arrival(DOA)initialization is proposed to solve the problem of lack of signal prior information in the multi-channel non-negative matrix factorization acoustic signal separation algorithm model.Firstly,the incident angle of the acoustic signal is estimated by the multiple signal classification algorithm,and then the steering vector containing the direction information of the acoustic signal is brought into the expression of the spatial characteristic matrix,thus the spatial characteristic matrix initialized based on DOA is obtained,and it is used as the initial value of spatial characteristic matrix,it is brought into the algorithm for iterative update.The test data of two-channel audio data and four-channel acoustic target separation show that,for audio data,the proposed sparse orthogonal joint constrained MNMF acoustic signal separation algorithm and the sparse orthogonal joint constrained MNMF algorithm based on DOA initialization are in the performance index of signal distortion ratio(Signal to Distortion Ratio,SDR)are improved by 0.84 d B and0.98 d B respectively.For helicopter sound source data,the proposed two algorithms are respectively improved by 4.53 d B and 5.16 d B in SDR.The simulation verifies that adding sparse orthogonal joint constraints and the method based on DOA initialization improves the performance of acoustic signal separation. |
PDF Full Text Request