Research On Convolutive Blind Source Separation Based On Swarm Intelligence

Blind source separation is the process of recovering source signals with the use of mixed signals merely,source signals and mix parameters are unknown,and it has found utility in many fields.Linear instantaneous mixture and convolutive mixture are the focus of scholars,because of the existence of time delay and attenuation in sound propagation,convolutive mixture is closer to the situation of daily life and it's more meaningful to have a study on it.The separation of convolutive mixed signals can be finished in time domain or frequency domain and we concentrate on the method in frequency domain because of larger computation and convolutive operation in time domain.Newton fixed-point iterative algorithm and gradient algorithm are widely used in traditional frequency methods to implement the separation of convolutive mixed signals,however they are all influenced by initial value of separation matrix,besides,gradient algorithm is influenced by step size as well,improper choices of these parameters may make algorithm deviate from the optimal value.Therefore,swarm intelligence algorithm with good optimization capacity is used to finish mixing process,and it avoids choosing initial value of separation matrix and step size,global convergence capacity is better.Through time-frequency conversion,convolutive mixtures are transformed to complex-valued linear instantaneous mixtures.In addition to order and proportion ambiguities,there is still phase ambiguity in complex domain and the problem should be considered when communication signals are tackled,therefore two kinds of methods are proposed.The first method extracts the real and imaginary parts of signals in order with the use of real Givens matrices,and bat algorithm is used to search the optimal values of rotation angles,the method can recover source signals and phase information simultaneously as long as the real and imaginary of source signals are independent.The second method eliminates phase ambiguity through post-processing of rotating angles.By experimental tests,the proposed methods can solve phase ambiguity effectively on the basis of ensuring separation accuracy.The real and imaginary parts of convolutive speech mixtures in frequency domain are not independent mutually,and there is no need to consider phase ambiguity.In order to finish the separation of convolutive mixtures,a method of constructing separation matrix of each frequency point by using complex Givens matrices is proposed,the number of parameters is reduced on the basis of guaranteeing orthogonality constraints.Then bat algorithm is used to optimize rotation angles,after eliminating order and scale ambiguities,time-domain signals are got through ISTFT.By experimental tests,the proposed method can realize separation of convolutive mixtures,and performs better than some other classical methods.Due to the disadvantage that independent vector analysis(IVA)algorithm is affected by the initial value of separation matrix because of the use of gradient algorithm,a frequency domain deconvolution method for finding the optimal initial value of separation matrix is proposed.Constructing initial value of separation matrix with complex Givens matrices,then get separation signals with the use of IVA,the mutual information of separation signals in each frequency point is fitness function and bat algorithm is used to search the optimal value.By experimental tests,the proposed method improves the performance of IVA on the basis of avoiding permutation ambiguity.