APSoC心音辅助诊断算法硬件加速方法

Hardware accelerating for classification of heart sound signals based on APSoC

针对云南边远山区低网络覆盖率和低传输速率下普通移动设备对神经网络处理速度慢、成本高、效率低的问题,提出一种基于APSoC的心音辅助诊断算法的硬件加速方法。在对5122例心音信号进行去噪、特征提取等预处理后,训练CNN网络模型用于心音样本分类。设计通用卷积电路与通用池化电路,将HLS优化后生成硬件电路部署至Zynq-7020 APSoC硬件平台,实现CNN算法的硬件加速。实验结果表明,相同条件下,其分类速度相比Intel-i7-8700提高了35倍,分类准确率仅损失了不到1%。该方法满足了高性能、低功耗、低成本等要求,为先心病初诊辅助诊断提供了一种离线解决方案。

To overcome poor network transmission in some remote mountain area of Yunnan and to get rid of the limitations of devices with low processing speed,high cost and low performance,a hardware accelerated method for classification of heart sound signals based on APSoC was proposed,which could work at offline state.5122 cases of heart sound signals were pre-processed,which included de-noising and feature extracting.The CNN network model was trained to be used for heart sound sample classification.A general circuit for convoluting and pooling was designed.A hardware and software co-design was employed on ZYNQ-7020 APSoC hardware platform according to its functions and resources.A hardware circuit was optimized using HLS,which was deployed to the hardware to realize the hardware acceleration of the CNN algorithm.Experimental results show that the computing time using this method was 35 times lesser than that using Intel-i7-8700,and it only cost 1%accuracy.This method meets the requirements of high performance,low power consumption,and low cost.It is hopeful to provide an off-line solution for the primary diagnosis of congenital heart disease.