Independent Component Analysis Of Low Spacing Frequency Domain

Cardiovascular and cerebrovascular diseases have become one of the most important killers of human beings,and the hearing dysfunction has become one of the important diseases that affect people’s normal learning and life.The application of blind source separation system to the auscultation system or hearing aid can greatly improve the medical level or quality of life.In speech signal processing,when the signal fragment is short enough,it is generally considered that the signal is steady state,and the blind source separation and correlation algorithm are applied.The blind source separation algorithm for speech signals generally includes two major categories: time domain method and frequency domain method.Because the time domain method is complex and computationally intensive,the frequency domain method is usually used.Based on the traditional independent component analysis algorithm to analyze the complexity of each step,the complexity of the mainly received signal length,number of frequency points with the ICA algorithm iterative algorithm convergence speed the influence of three aspects.In order to reduce the algorithm complexity,we can reduce the number of frequency points of iteration and accelerate the ICA iterative algorithm.In this paper,the number of frequency points for reducing iteration is studied.At the present stage,the independent component analysis of frequency domain is aimed at small spacing,which is not applicable to the spatial aliasing caused by distance similar to the distance between people and ears.According to the mathematical model mentioned in this paper,a general algorithm is given through mathematical derivation:The classical frequency domain independent component analysis algorithm is introduced.Will stubbornly in BSS algorithm in speech signal,because we need the time domain signal Fourier transform,so the frequency domain information contain imaginary part,namely in the frequency domain BSS algorithms need to try out the plural form of the ICA algorithm.This leads to the introduction of the complex ICA algorithm.Because the ICA algorithm of speech signal generates two uncertainties,the classical frequency domain BSS algorithm also needs to solve these two uncertainties.And the evaluation indexes of the speech quality are given,which are used as the standard for the evaluation of speech separation performance.ICA algorithm using FastICA and quantization algorithm combining the natural gradient algorithm,through a large number of experiments prove that the traditional frequency domain independent component analysis algorithm has good separation performance,but also found the relatively high complexity.Aiming at this problem,the analysis of the algorithm steps accordingly,determine the complexity of the algorithm are mainly concentrated in the ICA iterations,so this paper will research direction positioning on the direction of the reduced frequency.On the basis of traditional frequency-domain independent component analysis algorithm,the traditional algorithm is divided into two segments to choose the frequency point.The first frequency point chooses the determinant of covariance matrix as one of the classification frequency points,and the frequency point of the standard value is more than the primary point,and the remaining points are not selected.After a large number of experiments,it can be found that only after a frequency point selection,the selected primary points will have some frequency points that will deviate from the actual value,and have certain errors.Therefore,the second phase outlier algorithm is introduced to select the primary points for the first time.The selected points are selected as the final primary frequency points,and the remaining points are selected as the selected points.Through the above process,it is found that the complexity of the algorithm can be greatly reduced under the premise of optimizing the separation performance.In this paper,the concept of mutual information is introduced,and the standard of primary frequency selection in the first stage is replaced by mutual information,and the smaller the mutual information is,the smaller the mutual information is carried between the random variables.The experimental results show that when the mutual information is introduced as standard,the complexity is increased compared with the original value of the determinant,but the other separation indexes are improved greatly.Finally,two kinds of standards are analyzed,and two kinds of criteria are given.