Research On Underdetermined Blind Source Separation Algorithm Based On Compressed Sensing And Its Application

Blind source separation technology is a kind of signal processing technology with great challenges and application value.It can recover the source signal only by analyzing the received signal in the absence of prior information of the source signal and channel parameters.Among them,underdetermined blind source separation has more source signals than receiving sources,which is more in line with the actual application scenario.It has great value in practical engineering,so it has become a research hotspot.This paper studies the underdetermined blind source separation of speech signals,and proposes the following improvements to address some of the shortcomings of underdetermined blind source separation of speech signals:First,for the case where the source signal is not sufficiently sparse,a single source point detection algorithm is used to obtain more sparse observation data.Then,the number of clusters and cluster centers are estimated on a Density-based spatial clustering of applications with noise(DBSCAN)algorithm.The cluster centers represent the column vectors of the mixed matrix and the number of clusters.Represents the number of source signals.In order to prevent the selection of the initial value of the cluster center and the situation of being trapped in a local optimal solution,a particle swarm optimization(Particle Swarm Optimization(PSO))is used to optimize the estimated cluster center to obtain a better mixture matrix estimate.The PSO-DBSCAN algorithm is proposed to overcome the disadvantages of traditional clustering algorithms that need to know the number of clusters in advance,and it can correct the clustering data based on the density connected domain during clustering iteration,and remove some interference points to improve the clustering accuracy.It can be learned from simulation experiments that the proposed algorithm can effectively improve the estimation accuracy of the hybrid matrix under high signal-to-noise ratio,and can automatically estimate the number of source signals.Secondly,for the signal sparse reconstruction algorithm,the dictionary sparse representation algorithm based on dictionary learning is studied,and the K-SVD algorithm is used to perform dictionary learning on known specific signals to construct a complete dictionary.For the difference in the sparseness representation of each frame of speech signal,a variable step size adaptive threshold sparsity adaptive matching pursuit reconstruction algorithm based on inner product distribution(VT-SAMP)is proposed to solve the candidate atoms in the SAMP algorithm iteration process.The precision and computational pressure are increased exponentially,and the adaptive transformation threshold improves the accuracy of atom selection.Compared with the traditional signal reconstruction algorithm,the algorithm uses the signal inner product distribution as the control method,which can improve the accuracy of sparsity estimation and the accuracy of signal reconstruction.The feasibility of this algorithm is proved by comparing other dictionary learning algorithms with the adaptive sparsity dictionary learning and reconstruction algorithm based on VT-SAMP and K-SVD.Finally,the source signal is restored based on the signal sparse representation model of compressed sensing theory.Compared with the original signal,the reconstructed signal can basically restore the characteristics of the original signal.