A Research On Compressed Sensing Algorithm Of Heart Sound Signal Based On Deep Learning

The wearable heart sound signal acquisition device is generally powered by batteries,the energy consumption of the acquisition device needs to be strictly limited.Reducing the amount of data transmitted through compressed sensing can reduce the energy consumption in the signal transmission stage,increase the battery life of the wearable device,and prolong the life of the battery.The traditional heart sound signal compressive sensing algorithm has problems that the performance of the measurement matrix is not good enough and the effect of the reconstruction needs to be improved.Based on deep learning,this thesis designs a network to build heart sound signal compressive sensing measurement matrix,a heart sound signal reconstruction network and a heart sound signal denoising network.The main research contents are as follows:1)According to the limitations of the existing rule-based designed Gaussian measurement matrix and the measurement matrix constructed by the Dr2Net network,this thesis uses the Inception module and additional convolutional layers as a new decoder,and make use of multiple channels to learn different dimensions features,construct the compressed sensing measurement matrix of heart sound signal.After the signal were reconstructed by the BP algorithm,the measurement matrix generated by the proposed network achieved an improvement rate of more than 5%in the indicators PSNR,RMSE and PRD compared with the comparison measurement matrix.2)A heart sound signal reconstruction network based on a multi-channel and multi-scale feature learning module is proposed.The basic module uses 3 channels to extract local spatial features of different scales,uses the Non-local block structure to calculate long-distance dependencies and then learn non-local temporal features,uses LSTM to learn the local timing features of the heart sound signal,and finally stacks the features of each channel,using 1x1 Convolution kernel to fuse multi-scale features.Compared with the reconstruction network CSNet and BP reconstruction algorithm,the reconstruction network proposed has achieved an improvement rate of more than 4%in the three indicators of PRD,RMSE and PSNR;compared with the BP reconstruction algorithm,the reconstruction time is improved about 1-2 orders of magnitude.3)This time we abandon the traditional Encode-Decoder denoising network design pattern,utilize modular network design pattern,and every module brings in attention mechanism.The denoising network consists of 4 SKA(SK and Attention)modules,and every SKA module consists of a Sk Block and a Attention Block.The Sk block utilizes the Selective Kernel Convolution structure,which adaptive adjust the receptive field sizes of neurons;the Attention Block is based on the Squeeze-and-Excitation module,which implements a spatial attention mechanism.The denoising capabilities of different networks are tested in 7 cases including lung sound noise,motion artifact,Gaussian white,pink,red,purple noise and a mixture of the above 6 noises.The denoising ability of different networks is verified,and the network proposed has achieved more than 14%improvement in indicators of SNR₄8)),PRD and RMSE.4)The testing heart signal acquisition hardware is based on the STM32F103T8 and MAX9812 chip,and the host computer uses Python for signal reconstruction and signal denoising.The UI is designed using Py Qt5.Then in real situations,we verifies the signal reconstruction and denoising effects for different sampling frequencies and different signal frame lengths.