Research On Hardware Acceleration Of Heart Sound Classification Algorithm Of Congenital Heart Disease Based On APSOC

The heart sound can well show the working state of the heart.Yunnan is a high-risk area of congenital heart disease（CHD）.The initial diagnosis of CHD mainly relies on cardiac auscultation by experienced auscultation experts.This method is inefficient.It also consumes a lot of manpower and material resources.In this study,the APSOC technology was used to integrate digital signal processing,deep learning,and edge computing to realize hardware acceleration for CHD heard sound classification algorithm on APSOC.The research work of this thesis is divided into the following four parts.1)How heart sound signal is pre-treated and its feature is extracted.A complete cardiac cycle is about 0.8 seconds long.A complete cardiac cycle of abnormal signals contains pathological murmurs of congenital heart disease.In this article,the average phase of the heart sound signal is 2 s,which contains at least one complete cardiac cycle.The STFT is performed on it.The convolutional neural network model is designed.The local feature analysis method of the convolutional neural network is used to perform the time-frequency map analysis.2)How hardware implementation is made for the designed convolutional neural network.The computing modules are designed into a transplantable IP core by using the Vivado HLS tool.In order to make hardware acceleration for computing process of the IP core,the logic design,data flow design,and parallelism design were done.The resource consumption of the HLS IP core is analyzed to ensure that there is no waste.3)How the IP core and Xilinx official IP core are designed.First data trend analysis is conducted.Then board-level design is done by using integrated design tools.Next synthesis are conducted and implemented into hardware platform.Finally resource analysis is made.4)All functions including IP core initialization,IP core configuration,DDR data transmission,logic function design with IP core,etc are designed on ARM core in Zynq to establish the hardware platform.The results show that the 120 time-frequency spectra maps of heart sound can be classified at accuracy of 0.9.A single heart sound map proccesed within 3ms.The power consumption of chip is 2.207 w.Compared with the existing literature,the Convolution Operations amount in this method is higher 10.04 times than one in the literature but time consumtion is just higher 1.28 times than one in the literature.The accuracy in test set by using this method is higher 0.04 than the accuracy described in the literature.Meanwhile,the average resource consumption ratio is also lower than the consumption ratio described in the literature by about 36.3%.The results show the system has low power consumption,high efficiency,resource saving,and good accuracy by using the hardware acceleration method.It is expected to be used for screening congenital heart disease at villages.