Blind Separation Of Underdetermined Mixtures Of Sources

BSS (Blind Source Separation) is a fundamental and challenging researchtopic in signal processing field. Benefiting from the promising applications inspeech recognition, denoising, wireless communication, sonar problem, biologicalmedical signal processing. fiber communication et al, BSS has become one of thehottest spots in signal processing field and neural network field, In the nearly20years, the theories and algorithms about BSS have got great developments. Manyeffective algorithms have been presented, and their performances are different fromthe ability to separate source signal. However, the BSS theory is very profound andthe BSS algorithms are very difficult to implement, to this day, the study on BSS isstill far away from mature. Many theory problems and BSS techniques are expectedto continue to be discussed. the main contributions of this thesis are as follows:As for sparse source signals, this thesis analyzes the important property ofobserved signals. The scatter plot of observed signals shows straight line formalong the column vector of mixing matrix. The modulus of observed signal point islarger, it is important in the procedure of estimate mixing model. The conventionalK-means BSS algorithm deals with all observed points with different modulus bythe same way, so it is not accurate. And we present K-PCA algorithm to make upthe lack of conventional K-means BSS algorithm.Blind separation of underdetermined delay mixtures is studied, An attenuationand delay matrix recovery in signal source intervals is proposed. Our algorithmfirstly estimates the attenuation after the detection of the samples in singal sourceintervals and then recovers the delay by the optimum of a cost function. In sourcerecovery for the model with delay, a new algorithm based on sparsity is proposed.In the underdetermined case, estimates that the number of source signals andanalyzes the performance of the signal sparse solution. Presents a new judge signalsparse solution criteria.This paper put forward a new kind of two-step approach toseparate underdetermined blind signals. The simulating results illustrate the betterperformance of the method.