Hardware Design Of Lightweight Network For Heart Sound Classification

Cardiac auscultation has long been the main method of screening for congenital heart disease in primary care hospitals in China,but it has many drawbacks in terms of difficulty of screening and high screening requirements.In recent years,artificial intelligence has been widely used in many fields such as intelligent driving and intelligent medical care.It is important to combine artificial intelligence technology,digital signal processing technology and cardiac auscultation technology to research and implement a more efficient and accurate congenital heart disease assisted diagnosis system.Neural networks are an important research direction in artificial intelligence,and in recent years,with the improvement of computer computing power,the research of neural networks has made rapid progress.FPGA is a programmable hardware platform with many advantages,such as low power consumption,high stability and parallel computing,etc.The implementation and hardware acceleration of a heart sound classification system through FPGAs can significantly improve the portability and real-time performance of the heart sound classification system.A kind of design of the heart sound classification system based on the improved lightweight neural network Mobile Net was put forward in this paper,which implements the core layer of the system in hardware through high-level synthesis tools.The following works are included in this paper.1.Improving the heartsound classification system: traditional convolutional neural networks are computationally intensive and have many weight parameters,which are not suitable for deployment in embedded devices with limited hardware resources.In this paper,we use Mobile Net,a lightweight network,to build a lightweight heartsound classification network consisting of 8 layers of convolution based on the idea of deep separable convolution proposed in it.The fully connected layers in Mobile Net are replaced by point-by-point convolutional layers to reduce the number of different types of convolutional layers and reduce hardware resource consumption.2.The parallelism characteristics of the traditional convolutional layer,deep convolutional layer and point-by-point convolutional layer are compared,and a parallel speedup of 9 times is achieved for the traditional convolutional layer,9 times for the deep convolutional layer and up to 144 times for the point-by-point convolutional layer through intra-convolutional window parallelism,inter-convolutional window parallelism,interinput channel parallelism and inter-output channel parallelism.3.The memory scheme of input feature map data and weight parameter data is optimized,and their storage in memory is re-arranged,using line buffer,sliding window technology,and reasonable allocation of hardware resources such as BRAM and registers to ensure the concurrent read capability of hardware modules.A classification accuracy of 96.27% was achieved on the test set of 324 cases in laboratory heart sound database.In terms of computational efficiency,the heart sound classifier achieves about 19 times hardware acceleration on FPGAs compared to generalpurpose CPUs,and reduces the computing time by about 34% compared to other lowparallelism lightweight heart sound classification systems.