Research On Permutation Algorithm Of Convolutive Blind Source Separation In Frequency Domain

Blind source separation is a technique that only uses the received mixed signals to separate and recover the source signals without knowing the source signals and the propagation process.Convolutive blind source separation model has become a research hotspot in this field because it is most appropriate to the actual situation.The core of the most commonly used frequency domain convolutive blind separation algorithm is to use the time-frequency conversion tool to convert the time domain convolutive mixed signals into the frequency domain instantaneous mixed signals,and then separate the signals at each frequency bin through the instantaneous mixed blind separation algorithm.However,due to the existence of amplitude ambiguity and permutation ambiguity,people often need to transform back to the time domain after eliminating these two types of ambiguities,so as to obtain the more accurate separated signals in the time domain.Among them,the permutation ambiguity has a greater impact on the performance and is more difficult to eliminate.This thesis focuses on how to effectively solve the problem of permutation ambiguity in the process of convolutive blind source separation in frequency domain under the environment of speech signal or radar signal as the source signal.The specific research contents and results of this thesis are summarized as follows:(1)In this thesis,the theory of convolutive blind source separation is studied.Firstly,two typical mathematical models of blind source separation are deeply studied,and then the time domain method and frequency domain method of convolutive blind separation are introduced in detail,and their respective advantages and disadvantages are also analyzed.The inherent amplitude ambiguity and permutation ambiguity of frequency domain convolutive blind source separation are deeply explored,and the causes of ambiguities are explained mathematically.At the same time,the two most commonly used instantaneous blind source separation algorithms are also introduced,paving the way for their subsequent applications.(2)Aiming at the poor robustness of the traditional amplitude correlation permutation method,we propose an energy correlation permutation algorithm based on frequency bins correction,which takes into account both universality and high performance.At present,there are two types of methods commonly used to solve the problem of permutation ambiguity.They are the permutation algorithms based on amplitude correlation(including Anemuller algorithm and Murata algorithm)and the permutation algorithm based on DOA(DOA algorithm).Murata algorithm has strong universality,but its performance is unstable.Based on this,we propose an improved permutation algorithm.It is found that the energy correlation performance of signals among adjacent frequency bins better reflects the relationship between distance and correlation.Therefore,the improved algorithm utilizes the energy correlation of signals on adjacent frequency bins.In addition,the reliability of signal permutation on each frequency bin is measured,and the frequency bins with unreliable sorting results are corrected in time,which effectively avoids the occurrence of continuity errors and improves the robustness of the algorithm significantly.Simulation results show that the improved algorithm has obvious advantages in separation performance compared with the traditional amplitude correlation permutation algorithm in the case of different convolutive mixture models and different source signals.(3)This thesis proposes a frequency domain blind separation permutation algorithm based on performance weight clustering.The correlation permutation algorithm based on binary mask uses the characteristic that the estimated signals obtained by binary mask do not have the uncertainty of the order,and they are used as the reference for permutation.The limitation of the algorithm itself makes it difficult to guarantee its performance when the source signals are not sufficiently sparse.Therefore,we want to select a more appropriate permutation reference object,and the frequency domain blind separation permutation algorithm based on performance weight clustering come into being.The algorithm uses clustering to obtain the permutation reference,and calculates the accuracy of the separated signals on each frequency bin.According to the accuracy of the separation results,different frequency bins are given different clustering weights,thereby improving the reliability of the clustering results.By segmenting the frequency bins,the propagation of errors can be effectively suppressed and the performance of the algorithm is improved.Finally,several groups of simulation experiments verify the universality and performance superiority of the frequency domain blind separation permutation algorithm based on performance weight clustering,and the influence of the number of receivers on the performance of the algorithm is explored in the radar signal environment.