Signal Slice and Tensor-Based Blind Separation of Underdetermined Mixtures

Abstract

Abstract A novel underdetermined blind separation method is proposed based on signal slice and tensor decomposition to explore effective statistical information and improve separation performance. Firstly, the whitening signal is partitioned into several slices, and the delay covariance matrix of each slice is calculated. These delay covariance matrices are then stacked into third-order tensors and compressed into low-dimensional core tensors using high-order singular value decomposition. Next, the third-order tensors are decomposed using canonical polyadic decomposition through weight nonlinear least square to obtain the mixed matrix. Finally, by leveraging signal independence, a matrix diagonalization method is employed to recover the source signals. Simulation results demonstrate that the proposed method effectively suppresses the influence of Gaussian noise and improves the estimation accuracy. Moreover, the proposed method achieves superior separation results compared to seven representative approaches.