Underdetermined mixing matrix estimation based on time-frequency single source points detection and eigenvalue decomposition

In this paper, a method of mixing matrix estimation based on time-frequency single source points detection and eigenvalue decomposition is proposed under the underdetermined blind source separation model. Firstly, short-time Fourier transform is used to transform non-sparse observed signals in time domain to sparse signals in time-frequency domain, and SSPs detection is employed to improve the sparsity. Secondly, the improved fuzzy C-means clustering algorithm is applied to estimate the number of sources. Then, we reselect the SSPs by the means of ordinary least squares to improve the estimation accuracy. Finally, the estimated mixing matrix is obtained by eigenvalue decomposition. Vast simulations illustrate that our method can estimate the mixing matrix accurately when the number of sources is unknown, and has strong robustness to noise.