Underdetermined mixing matrix estimation algorithm based on tensor analysis

Abstract

Aiming at the problems of difficult to extract effective feature information and the slow convergence speed of the underdetermined matrix estimation, an underdetermined matrix estimation algorithm of instantaneous mixtures based on tensor analysis was proposed to overcome the constraint of signal sparsity.In the proposed algorithm, the symmetric third-order tensor was constructed via the autocovariance matrix of segmentation sub-block, which was compressed into a kernel tensor to reduce the size of the data.An enhanced line search technology was applied to speed up the convergence of alternating least squares method, and the factor matrix was used as the measure of the mixing matrix estimation, but the selection of the number of segmentation sub-blocks was an open problem.Experimental results demonstrate that the proposed algorithm outperforms the sparse transformation method and the traditional high-order statistical method in handling the underdetermined mixing matrix estimation.

Keywords

• underdetermined matrix estimation;symmetric tensor;segmentation strategy;autocovariance matrix