A Density Based Clustering Approach to Underdetermined Blind Source Separation
Yuduo Wang, Qingyi Wang · 2021 China Automation Congress (CAC) · 2021
Blind source separation(BSS) is the process recovering the source signals unknown. Once the number of observed signal is less than the source, it can be referred to underdetermined blind source separation(UBSS). In this paper, one mathematic model based on Sparse Component Analysis(SCA) applied to UBSS is introduced, which can be summarized in two steps after preprocess of signal. What makes the method in this paper different from the classic is the ability of estimating the number of source signals. In this paper, mixture signals are preprocessed by short time Fourier transform(STFT) to strengthen the sparsity, after which some points called single signal point(SSP) are selected for further process. Then based on Density Peak Clustering(DPC) clustering, the method in this paper can precisely get the source signal number while figuring out the channel matrix also called mixture matrix in step one by plotting the decision graph of SSP. After acquiring the estimated matrix, in step two using L1 norm the source signals are finally recovered. With experiments for vocal signal, it has been proven that the method in this paper has a better performance than some other methods which is further demonstrated in the rest of paper.