Fusing Multi-modal Statistics For Joint Blind Source Separation

Joint Blind Source Separation(J-BSS)is an emerging signal processing technique in recent years,which has wide application potential in multi-set fusion.J-BSS recovers the source signal by using the statistical characteristics of the source signal,such as inter-group correlation and intra-group independence.J-BSS based on multi-modal statistics aims at the fusion of second-order covariance,second-order pseudo covariance,fourth-order cumulant and other statistics to solve the problem of J-BSS with unknown statistical characteristics of source signals.This paper mainly studies the J-BSS method of multi-modal statistical fusion processing.The main results are as follows:·We presents a joint tensor diagonalization algorithm based on fourth-order cumulants.Based on the Jacobian successive rotation strategy,tensor joint diagonalization,a highly nonlinear optimization problem,is transformed into a series of simple sub-optimization problems that can receive a closed solution by linear operations.By alternating iterations,the factor matrices of the tensor are updated,and then the mixed matrices of different datasets are estimated to realize J-BSS.The results of computer simulation show that the proposed algorithm has faster convergence and higher accuracy than the existing BSS and J-BSS algorithms.·We presents a joint tensor/matrix diagonalization algorithm for multi-modal statistical fusion processing.The algorithm can fuse three types of statistics,namely the second-order covariance,the second-order pseudo covariance,and the fourth-order cumulant,within a tensor/matrix diagonalization framework.Through the Jacobin successive rotation fitting factor matrices,the target tensors are updated and tends to diagonalize.Then the mixed matrices of different datasets are estimated to realize J-BSS.The simulation results show that the proposed algorithm has good performance for a variety of signals with different statistical characteristics,and can flexibly adjust the type of statistics involved in the operation.The proposed method can also flexibly select or emphasize datasets that have strong source dependence in the calculation according to prior knowledge,in order to obtain more accurate results.In addition,both of the above algorithms have shown good performance in processing fetal ECG(Fetal ECG,FECG).