Research On Heart Sound Period Characteristics And Construction Method Of Heart Sound Neural Network

Heart sound signal is a biological signal that exists in humans.By analyzing the characteristics of heart sound signal,it can help people detect and monitor the health of the heart.Similarly,as a biological signal that can judge disease,people have been studying it.Therefore,the research in this thesis is based on this.Among the traditional algorithms for heart sound research,many algorithms have appeared,including preprocessing algorithms,segmentation algorithms,and so on.For the purpose of heart sound classification,this thesis uses the current mainstream convolutional neural network(CNN)for image and biological signals on the classifier,and combines CNN with heart sound feature extraction method to propose a design method of heart sound neural network.In the research of heart sound signals,there are two basic methods for processing heart sounds: extracting heart sound components and extracting heart sound cycles.In the current research,most researchers use the method of extracting heart sound components.In the method of extracting heart sound components,the effect of manually labeling the components is the best.However,manual tagging has the disadvantages of heavy workload and high cost.In the method of automatically extracting heart sound components,due to the difference between abnormal and normal heart sounds,it is difficult to design a method that can accurately extract heart sound components for all heart sound signals.For the method of extracting the heart sound cycle,the heart sound cycle is studied as a whole,because the normal heart sound is a quasi-periodic signal,and generally the complete heart sound cycle is used as the characteristic representation of the heart sound.However,if the heart sound period is to be extracted,there is a certain requirement for the length of the heart sound data,which can reflect the periodic characteristics of the heart sound without making the data volume too large.Combining the two basic methods mentioned above,in order to analyze and solve the existing problems,this article adopts single-cycle and multi-cycle methods respectively.On the premise of a single cycle,the heart sound signal is periodically segmented by the periodic feature extraction method of the heart sound,and the length of a cycle is evaluated according to the correlation of the heart sound.Therefore,this thesis uses an envelope autocorrelation function to extract the cycle of the heart sound..At the same time,the heart sound is also a signal in the timefrequency domain.Therefore,the spectrogram can be used as a two-dimensional representation of the characteristics of the heart sound.Based on the idea of signal losslessness,a heart sound sound spectrum designed with the heart sound period as the content is proposed.Under the premise of multiple cycles,as an autocorrelation function that can provide the basis for extracting heart sound cycles,it also contains the periodic characteristics of heart sounds,and the periodicity of normal heart sounds is strong,and the periodicity of abnormal heart sounds is weak,so it can be passed Observe the autocorrelation function of the entire heart sound recording to determine the nature of the heart sound signal.Then based on the above two characteristics of heart sound characteristics,a heart sound neural network is designed to summarize the characteristics of heart sounds.The heart sound spectrum and the heart sound autocorrelation function are used as the input of the heart sound neural network.Through training,testing,comparison and analysis,a neural network structure suitable for the heart sound sound spectrum map is obtained.To sum up,this thesis takes the heart sound period feature extraction as the research basis of this thesis,and designs the heart sound neural network to solve the problem of heart sound recognition and classification.This thesis studies the pretreatment,feature extraction methods of heart sound signals,and the structural design of heart sound neural networks.Starting from the mathematical definition to the derivation of experimental methods,the heart sound cycle extraction method,heart sound neural network design method,and heart sound neural network Operation mechanism.Through the periodic extraction of heart sounds,a 97.33% periodic extraction accuracy rate can be obtained;the heart sound phonogram has obtained 96.16% accuracy on the test set on the heart sound neural network,with a specificity of 0.8800,a sensitivity of 0.9100,and a revised accuracy rate It is 89.50%;the heart sound autocorrelation function graph obtained 100% recognition rate of the verification set on the heart sound neural network,the specificity is 0.8667 on the test set,the sensitivity is 0.9375,and the correction accuracy rate is 90.21%.The correction accuracy obtained by combining these two characterization methods with the heart sound neural network is 3.48% and 4.19% higher than the best results of Physio Net / Cin C Challenge 2016,respectively.