Underdetermined Blind Source Separation Based On Improved K-means Clustering

Blind source separation technology is an important branch in the area of themodern digital signal processing. Blind source separation refers to the source signaland mixed system features which are unknown, the process only through observedsignals to estimate mixed matrix and separate source signal. Because of its importantapplication in wireless communication, biomedicine and so on, so these applicationscatch scientific workers more and more attentions to blind source separation.According to the more or less amount of source signals and the observed signals,blind source separation problem can be classified into three situations:overdetermined blind source separation, positive definite blind source separation andunderdetermined blind source separation. Generally we group togetheroverdetermined blind source separation and positive definite blind source separationas one kind, namely the number of observed signals are equal or greater than thenumber of source signals, in this case, we usually use independent componentanalysis (ICA) to separate blind source separation,which is simple and smart. If whenunderdetermined situation, namely the number of observed signals are fewer than thenumber of source signals, in this kind of situation, we often use the sparsity of thesignals at one moment to make the signals have sparse representation, again on thebasis of the existing algorithm, to restore the source signals.According to blind source separation problemof sparse component analysis, thekey work and innovations of this paper mainly include:(1)From the positive definite blind source separation, the paper analyses thebasic problem of blind source separation, easier to understand underdetermined blindsource separation, which provides a very good transition.(2)Some important algorithms of underdetermined blind source separation areintroduced, and pointes out that the concrete implementation process and theadvantages and disadvantages of these algorithms.(3)An improved K-means clustering algorithm to estimate the mixing matrix isproposed. First the algorithm based on this algorithm, the data is pre-processed andthen initial cluster centers ares elected by the zoning law and then cluster to estimatethe mixing matrix.This paper mainly improves data preprocessing, the zoning methodto estimate the initial cluster centers. The proposed algorithm improves the inadequacies of the k-means algorithm which is very sensitive to the sample inputsequence and initial cluster centers.then we used improved clustering algorithm tocluster sparse observed mixed-signal, in order to estimate the mixing matrixaccurately for underdetermined blind separation.(4)On the recovery of the source signals, if the source is not fully sparse, and notmore than two non-zero source signals in the sample interval. We propose the sparsedecomposition standards (SSDP). Relatively classic linear programming algorithm,this standards improve recovery accuracy of source signals.