Research On The Mixing Matrix Estimation Of Underdetermined Blind Separation Based On Mixing Clustering And Meshes Density

Blind source separation is a newly hot topic in the field of signal processing. It estimatessources using mixed signals without prior knowledge such as sources number, location andmixing process. Blind source separation is widely used in biomedical engineering, speechenhancement, digital communication systems, image processing, remote sensing, radar andsonar and other fields.According to the number of relations of mixed-signals and source signals, blind sourceseparation can be divided into underdetermined blind source separation, positive blind sourceseparation and overdetermined blind source separation. Most previous studies focus onoverdetermined blind source separation, in which the number of source signals must be lessthan the number of mixed-signals. However, these conditions are not completely satisfied inreality. Considering the most practical situations, this thesis focuses on the underdeterminedblind source separation, in which the number of source signals is greater than that of mixedsignals.Now the main method of solving underdetermined blind separation is the two-stepmethod based on sparse signal. The two-step method firstly estimates the mixing matrix, thenobtains source signals using corresponding separation technology. In the two-step method,underdetermined blind mixing matrix estimation algorithm plays a key role. The estimationaccuracy of the algorithm directly affects that of the source signals. Now the main method ofsolving the mixing matrix estimation is the clustering methods, such as K-means and theHard-Lost algorithm. The primary contributions of the dissertation are summarized blow.1. The traditional K-means clustering algorithm, although with high rapid convergence,is too dependent on the initial clustering centers. With regards to this, this thesis, proposes amixed clustering method based on the improvement artificial colony algorithm and theK-means algorithm. The new method combines the advantages of regulating ability of globaloptimization and local optimization with rapid convergence of K-means clustering algorithmto improve the robustness of the algorithm. Experiments show that the clustering effect of thenew method is significantly improved, not only the stability.2. Cluster centers have a greater offset with real situations due to the noise points andoutliers. Although the stability of mixed cluster method is higher than the traditional methods, the accuracy is not high due to the offset.With regards to the real situation, this thesisproposes a new algorithm based on mixing clustering and mesh density. Experiments showthat compared to the traditional underdetermined blind mixing matrix estimation algorithm,the new algorithm has advantages of high stability and estimation accuracy. Through theabove two aspects of improvement, the subject finally realize a new algorithm of blind matrixestimating based on hybrid clustering and mesh density. This algorithm gets higher robustnessand precision than the traditional algorithm.