Abstract:

A method for blind estimation of underdetermined mixing matrix based on improved K-means clustering is proposed when the source number is unknown. First, the density parameter of the projection points of the mixing signals on half of the unit ultra sphere is calculated. Then, the projection points with low density are removed and the initial clustering centers are chosen from the projection points with high density. Finally, cluster the remaining points, use the DaviesBoudin index to estimate the source number, and estimate the mixing matrix. The simulation results show that the proposed algorithm’s complexity is lower and its running time is only about 1% to 3% of that of the Laplace mixed model potential function algorithm; its source number estimation accuracy is much higher than that of the robust competitive agglomeration algorithm; when the signal to noise ratio is greater than 13 dB, its accuracy is higher than 96.6% and its estimated mixing matrix error is small. When SNR is higher, it can relax the sparsity requirement of the sources.