Research On Fast Blind Source Separation Algorithm Based On Multi-standard Fusion

Blind Source Separation(BSS)can restore source signal well only depending on the observed mixed data without any prior information of source signal and transmission channel.The method can be applied not only to image and medical signal processing,but also widely used in the processing of acoustic signals,especially in binaural hearing aids.BSS is mainly used for speech enhancement in hearing aids.It can use the characteristics of its own algorithm to maintain the interaural clues of the target sound source without knowing the distribution of microphone arrays,channel transmission characteristics and destroying the interaural clues of other sound sources.However,the complexity of BSS algorithm is relatively high,which is inconsistent with the requirement of hearing aids for low power consumption and low latency of chips.In order to make BSS play a greater role in hearing aids,the first thing to do is to find a way to reduce the complexity of BSS.In order to better solve the problem of high complexity of BSS,a fast BSS algorithm based on multi-standard fusion was proposed from the frequency domain.Firstly,the whole process of recovering the source signal from the mixed signal by BSS is introduced,and the method and complexity of each stage are described in detail.Secondly,the fast BSS algorithm based on multi-standard fusion is discussed in detail.The basic idea of the frequency bin screening is that the separation matrix of a small part of the frequency bins is obtained by iteratively updating by the ICA algorithm,and the separation matrix of the residual frequency bins is obtained by other low complexity methods.The new algorithm proposed in this paper takes into account the energy and independence of the signal when selecting the frequency bin of the separation matrix by the ICA algorithm.In the paper,the multi-standard fusion frequency screening method is adopted innovatively:the inner product and the parametric centered correntropy are used as the criteria for frequency bin selection.The specific methods are as follows:(1)Calculate the inner product and the parametric centered correntropy of the frequency bins in the set range;(2)Set the threshold range of the two standards according to the change of the number of frequency bins with the threshold;(3)In order to determine the joint threshold,the separation performance and the running time of the algorithm are simulated in the range of the joint threshold;(4)Normalize the two criteria,then the normalized values are compared with the thresholds.If the values meet the two threshold requirements,the separation matrix of the frequency bin is obtained updated iteratively,otherwise it is classified as the residual frequency bin.Compared with the current single screening standard,the number of frequency bins that need to be iterated by the multi-standard method is greatly reduced,and the complexity of the algorithm will also be reduced.Moreover,the obtained primary frequency bins will have a greater probability of being those the frequency of good separation performance.The new algorithm proposed in this paper can better balance the relationship between separation performance and complexity,so that when the complexity is greatly reduced,it can also have good separation performance.Considering that the separation matrix of the unselected frequency bins depends on the relevant information of the selected frequency bins,this paper proposes a clustering algorithm to further remove the frequency bins with poor separation performance and improve the accuracy of the estimation parameters.Specific methods are:(1)Calculating the distance between the attenuation parameters corresponding to the frequency bins selected by the multi-standard fusion method;(2)Determining whether the attenuation parameter of each frequency bin is the core bin according to the DBSCAN algorithm.If as long as one of the four attenuation parameters at the frequency bin is not a core point,the frequency bin is divided into a set of residual frequency bins,and the separation matrix is solved by a non-iterative method.For the selected frequency bins,the separation matrix is solved by the iterative update method using the same separation method as the traditional BSS algorithm..For unselected frequency bins,the proposed method uses the mixed information of the estimated selected intermediate frequency to complete the separation.It does not need iterative updating and has low complexity.The experimental results show that the proposed multi-standard fusion frequency bin selection algorithm is superior to the traditional BSS in performance and complexity,While SDR,SIR and PESQ are increased by 10.93dB,7.87dB.and 0.080 respectively,the running time of the proposed algorithm is only 8%of the conventional method,and the complexity is only 27.3%of the traditional FDICA.