Research On Frequency Domain Algorithm For Convolutive Blind Source Separation

Blind source separation refers to the process of separating source signals from mixed signals when source signals and mixture model are unknown.In reality,the observed signals usually have convolutive mixture due to reflection,delay,and interference.At present,the frequency domain method is a research hotspot of convolutive blind source separation.The theory of frequency domain method is to convert the time domain convolutive mixed signal into the frequency domain instantaneous mixed signal by the STFT,and then use the instantaneous blind source separation algorithm to separate the mixed signal on each frequency point.However,due to the inherent scaling and permutation ambiguity of the blind source separation algorithm,these two ambiguities problems must be solved before the ISTFT.The problem of permutation ambiguity is more serious.For the frequency domain algorithm in convolutive blind source separation,the research contents and results of this thesis are summarized as follows:(1)For the scaling and permutation ambiguities of the frequency domain method,this thesis takes the speech signal as an example to study the minimum distortion method for solving the scaling ambiguity and three methods for solving the permutation ambiguity.The three methods are the amplitude correlation method,the direction of arrival method and the independent vector analysis(IVA)method.This thesis simulates and compares the three algorithms.Afterwards,this thesis proposes an IVA convolutive blind separation algorithm based on step-size adaptive to solve the order ambiguity problem.The algorithm uses the JADE algorithm to initialize the separation matrices,which makes the initial value more reasonable and avoids local convergence.The step size parameters are adaptively optimized to improve the convergence speed of the algorithm.The simulation results show that the proposed algorithm improves the performance and significantly shortens the computation time.(2)This thesis studies the convolutive mixture in sparse.The sparse case in this thesis is divided into two types,which are source signal time-frequency sparseness and convolutive filter sparseness.In the case of source signal time-frequency sparseness,this thesis studies a binary time-frequency mask method which uses this feature for blind separation.and proposes a correlation sorting algorithm based on the binary mask for the disadvantages that the method is easy to cause signal distortion and cannot make deconvolution.The algorithm takes the binary time-frequency mask separation signal as a reference,and correlates the signals which are separated on each frequency point,so that the correct order of each frequency point can be obtained,and the distortion and deconvolution problems can be solved.The simulation results verify the performance improvement of the method.(3)In the case of convolutive filter sparseness,this thesis explores how to use sparsity and finds the L1 norm feature of sparse filters,and then proposes a sorting and restoration algorithm based on L1 norm minimization.The simulation results show that the method can recover the frequency domain permutation of sparse filters when there is no scaling problem.Then this thesis studies a single source sparsely convolutive channel estimation method from the field of blind channel estimation.The method uses the L1 norm feature of sparse channel to estimate the sparse channel from the single source convolutive signal using convex optimization.The simulation verifies that the method can make effective estimation even when the time-frequency information of the observed signal is incomplete.Finally,this thesis considers two sparse cases and proposes a multi-source sparse convolutive blind separation algorithm based on convex optimization,which provides a new solution for the convolutive blind separation in sparse case.Simulation experiments verify the effectiveness of the proposed method.