Research On The Optimization Algorithm Of Convolutive Blind Source Separation Based On Joint Block Diagonalization

Blind source separation is a key technology in the field of signal processing.It is used to restore each source signal only according to the received signals when the known information of the source signal is very few or even completely absent.In recent years,considerable progress has been made on blind source separation technology,but the problem of noisy convolutive mixed speech signals' separation remains to be further explored.This paper focuses on blind source separation technology of noisy convolutive mixed signals.First of all,the speech enhancement algorithm based on signal subspace is used as denoising preprocessing part of blind source separation.It can reduce noise and make the mixed speech signals more pure.Then the author improves joint block diagonalization algorithm to avoid it converge to singular solutions.Finally,the improved algorithm is used to separate convolutive mixed speech signals.The main contributions of this paper are summarized as follows:(1)The mathematical models for blind source separation are introduced and some assumptions of source signals and noise signals are given in this paper.The author analyzes the evaluation indexes and the uncertainty of blind source separation that naturally exists.Three main categories of classic blind source separation algorithms are briefly introduced.The author chooses one algorithm of each category and uses them to separate noise-free speech signals and noisy signals.The results show that all of the three classic algorithms can not work well on noisy speech signals.(2)As for noisy speech signals,speech enhancement algorithms are mentioned to be as denoising preprocess before using blind source separation algorithms.The denoising of blind source separation has its own specialty.It not only has to suppress the influence of noise efficiently,but also has to minimize the loss of useful signals as possible.For this purpose,the author chooses the speech enhancement algorithm based on signal subspace as denoising part for blind source separation.The principle of signal subspace is introduced in detail and the simulation results of this algorithm verify its great removal effect of noise.(3)The noise-free convolutive mixed speech signals can be obtained after denoising preprocess.The author points out the disadvantage of classic nonorthogonal joint block diagonalization algorithm for this kind of signals.In order to overcome the algorithm's shortcomings,the author modifies its cost function and introduces the non-singularity constrain item to avoid singular solution.The author draws lessons from the thought of cyclic minimization to minimize cost function,obtains the optimal solution of the separation matrix and restores each source signal.The whole blind source separation procedure of noisy convolutive mixed speech signals is hackled.It is the combination of denoising preprocessing algorithm mentioned before and the improved joint block diagonalization algorithm.The numerical experiment compares the ability to suppress singular solutions of improved joint block diagonalization algorithm and the classic one.The improved algorithm can avoid singular solutions effectively.The author applies the whole improved blind source separation algorithm to noisy convolutive mixed speech signals.The results show the proposed algorithm can eliminate noise and restore each source signal accurately.Compare with the traditional algorithm,the proposed algorithm's 'PI and its MSE declines significantly.The performance of the proposed algorithm improves significantly.So it has application value to some extent.