The Study Of Heart Sounds Classification Based On Deep Learning Network

With the increasing morbidity and mortality of cardiovascular diseases,telemedicine and intelligent-assisted auscultation gradually entered into the crowd vision and was recognized as an important field of study.Affected by the surrounding environment and operating equipment,the collected heart sound signals are always mixed with certain noise.Besides,due to the flexibility of threshold setting methods and individual differences of the human body,there is always a large error in the periodical feature information obtained by segmenting the heart sound,which will affect the subsequent classification effect.In view of the above problems,this paper constructs a kind of biorthogonal wavelet base applied to the sound-noise denoising.Next,in the premise of avoiding heart sound segmentation and positioning,this paper builds a convolution neural network(CNN)model based on the constructed wavelet to classify the heart sounds,and the specific work is as follows:1.Heart Sound Denoising based on Wavelet Constructed.Based on the mathematical characteristics of wavelet bases,this paper analyzes its advantages in the application of denoising.On the basis of comprehensive and balanced consideration,this paper constructs a kind of double orthogonal wavelet base with compact support,symmetry and high order vanishing moment for the denoising of heart sound signals,and establishes the correlation coefficient and mean square error index system of signal and original signal after denoising.The denoising performance of the new constructed wavelet base is verified by comparison experiment with the traditional wavelets such as db5 wavelet,bior5.5 wavelet and sym5.The experimental results show that the mean square error and the cross-correlation coefficient of the heart sound wavelet base are higher in the noise treatment of normal heart sound,atrial vibrato and premature beats,and the overall denoising effect of the new wavelet base is stable for the different layers of wavelet decomposition.2.The classification and recognition of heart sound signals based on CNN.In this paper,the core idea of wavelet function as the activation function of neural network is discussed in detail,and the activation function of the convolution layer of CNN model is improved,and a CNN model based on constructed wavelet is built.This paper performed heart sound classification and recognition experiments on two data sets in turn,and compared this model with the improved LeNet-5 and improved Alex Net classification results of the related literature.The recognition rate of the improved CNN model based on wavelet in this paper was achieved 96.74%,which is 2.1% higher than the improved LeNet-5 with the same two-layer convolution layer,is slightly lower than the improved Alex Net by 0.31%,but the model training time is reduced by eight times.In a word,the heart sound wavelet constructed in this paper overcomes the difficulty of the traditional wavelet function in the orthogonality,symmetry,and high-order vanishing moments.The effectiveness of the wavelet function in the denoising of heart sounds also was verified through the correlation coefficient and the mean square error.At the same time,the superiority of the improved CNN model established in this paper in recognition accuracy and training time is also verified through various comparison tests.The research results in this paper show that the recognition of heart sound signals can be realized without deep CNN,which is of great significance for the development of telemedicine and intelligent auxiliary diagnosis.