Research On Denoising And Segmentation Of Heart Sound Signal

Heart sound is one of the common physiological signals in the human body.It contains a lot of cardiac pathological information and can reflect the health of the heart.Cardiovascular disease has always been one of the main causes of human death.Early extraction of physiological and pathological information of the heart is of great value in the treatment of cardiovascular disease.Successful automated auscultation of heart sounds can be used as an effective auxiliary diagnostic tool to help ordinary medical staff Determine if expert diagnosis is needed,especially in areas where clinicians are scarce and medical care is limited.Therefore,computer-aided analysis of heart sound signals is of great significance for the preliminary diagnosis of cardiovascular disease.According to the non-linear and non-stationary characteristics of heart sound,this paper mainly studies the heart sound denoising method and heart sound segmentation method of heart sound signal.The main research contents are as follows:1.Noise reduction processing for heart sound signals.In this paper,an improved adaptive hybrid interval threshold de-noising algorithm for heart sound signals based on ICEEMDAN was proposed.Aiming at problems of traditional heart sound signals de-noising method being easy to eliminate parts of high frequency useful information and cause loss of information,Empirical Mode Decomposition(EMD)methods have problems such as modal aliasing,Firstly,heart sound signals were decomposed into different Intrinsic Mode Functions(IMFs)with ICEEMDAN,and use a joint strategy to find two critical modal component K values,then the noise dominant modal components and aliasing ones are denoised by using Wavelet Packets(WP)and the ICEEMDAN algorithm.Reconstruction with the remaining components to obtain the denoised heart sound signal.The results show that under different noise level conditions,the algorithm can effectively remove the normal,abnormal heart sound noise component,and retain the useful information of the high frequency,Improved signal-to-noise ratio by 9dB.2.Adaptive threshold segmentation of heart sound signals.This paper proposes a method of heart sound segmentation based on adaptive threshold.For the characteristics of low frequency,susceptible to interference and a large amount of noise.Firstly,a method for extracting the continuous average energy envelope of the heart sound signal after noise reduction is adopted.Finally,based on the continuous average energy envelope and autocorrelation function,An adaptive threshold heart sound segmentation method.The simulation results show that the envelope characteristics of the heart sound signal extracted by this method are more robust.It's more accurate that the proposed segmentation algorithm is compared with the heart sound segmentation algorithm based on short-time energy entropy ratio method and short-time autocorrelation function method.3.Classification of fundamental heart sound based on machine learning.,In this paper,a classification method based on improving artificial bees colony algorithm is proposed tooptimize BP neural network.For the problems of BP neural network relying on initial weights,slowing convergence and easily falling into local extremum,the development ability of standard Artificial Bees Colony algorithm is weak,local search ability is poor,etc,The adaptive and global optimal strategies are introduced to improve the global search and probability selection algorithm of honey source in the Artificial Bees Colony algorithm,and the optimal solution of the current iteration is used to improve the development ability.In addition,chaotic systems are used to generate initial populations in order to enhance the global convergence of the bee colony algorithm.Finally,The proposed algorithm is applied to heart sound recognition.Experimental results show,the classification accuracy of the proposed algorithm is better than that of the classical classification algorithm.Under the condition of Mel-scale Frequency Cepstral Coefficients characteristic parameters,the classification accuracy rate is over 94%.