Underdetermined Mixed Matrix Estimation Based On DPC-KFCM Algorithm

The task of blind source separation is extracted from the aliasing signal separate source signals and,with the continuous development of social science and technology,the blind source separation technology has also been improved,and can be in more than one in the field of technology research to provide technical support for separation of signals aliasing,greatly promote the common development in many areas,however,more in line with the actual application of underdetermined blind source separation technology,Although the research started late,it has become a research focus in the field of blind source separation due to its high research value and high challenge.At present,the main method to solve the problem of underdetermined blind source separation is the "two-step method" based on sparse component analysis,that is,under the premise that the signal is sparse enough,the estimation of the mixed matrix and the reconstruction of the source signal are solved one after another.In this dissertation,the underdetermined blind source separation algorithm under linear instantaneous mixing model is studied based on this method,and an improved scheme is proposed.The main contents can be summarized as the following two aspects:(1)In the mixed matrix estimation stage,the Fuzzy C-means clustering(FCM)algorithm represented by soft clustering is improved in this dissertation to realize the estimation of the mixed matrix.Aiming at the problems of low estimation accuracy and poor robustness of FCM algorithm in underdetermined mixed matrix estimation,An improved kernel-based Fuzzy Cmeans(KFCM)Clustering algorithm based on Density Peak Clustering(DPC)is proposed.KFCM algorithm based on Gaussian kernel function is constructed by introducing kernel function into FCM algorithm,which can effectively overcome the influence of noise points and outliers on clustering results and improve the estimation accuracy of mixed matrix.To improve the traditional algorithms of DPC and and the integration of KFCM algorithm,by introducing the K neighbor in DPC algorithm thought,through the determination of the unified local density measure function to realize the unified measurement of any size data set,and the local density and high density distance setting threshold to realize the KFCM algorithm the initial clustering center and clustering center automatically determine the number,To improve the robustness of the algorithm.Experimental results show that the proposed algorithm improves the estimation accuracy and robustness of the underdetermined mixed matrix effectively compared with the comparison algorithm.(2)In the source signal reconstruction stage,in order to realize the effective separation of speech signals in the underdetermined blind source environment,a source signal reconstruction algorithm based on Double Sparsity SGK(DSSGK)was proposed.The underdetermined blind source separation theory and compressed sensing theory were compared from three aspects of theory,mathematical model and research objective,and it can be concluded that they are similar in the above three aspects.In this dissertation,the problem of source signal reconstruction under underdetermined blind source separation was solved by using compressed sensing theory.First of all,by constructing a complete dictionary to realize the voice signal sparse representation,which,for some typical dictionary learning algorithm computational complexity is higher,the dictionary size limited and long run defects,this dissertation adopts DSSGK dictionary learning algorithm for training the dictionary,the algorithm USES double sparse dictionary structure combines analytical dictionary and effective learning dictionary,The dictionary training has both speed guarantee and quality guarantee,and the structure is integrated with SGK algorithm to effectively improve the execution efficiency of signal sparse representation.Then the underdetermined blind source separation of speech signals is realized by using the mixed matrix estimation results and the orthogonal matching pursuit algorithm based on the compressed sensing theory.The experimental results show that the proposed algorithm has great potential in processing large-scale data,and the speed of dictionary training increases the efficiency of source signal reconstruction.