FPGA Design And Implementation Of Arrhythmia Classification System Based On CNN

As the incidence of cardiovascular disease increases year by year,the diagnosis and prevention of cardiovascular disease has attracted more and more attention.Electrocardiogram(ECG)is an important basis for doctors to diagnose heart disease.The traditional clinical diagnosis method is to make a diagnosis by manual analysis of the patient’s ECG.Such a diagnosis may lead to misdiagnosis and missed diagnosis.Research on arrhythmia classification algorithms can assist doctors’ diagnosis on the one hand,improve the diagnosis efficiency of doctors,and be conducive to the reasonable allocation of medical resources;meanwhile,it can also transplant algorithms to wearable devices to provide users with disease warning and health monitoring Functions.In recent years,with the advancement of neural network research,neural networks have shown more powerful classification performance in many research areas.The automatic recognition of ECG signals by applying convolutional neural networks can better extract ECG signals.Multi-classification process with higher accuracy.In order to make the convolutional neural network better process the one-dimensional time series such as the ECG signal,some researchers have proposed to use the one-hot encoding technology to image the ECG signal and combine it with the CNN model to study the arrhythmia classification and achieved good results.The disadvantage of the classification effect is the high redundancy of the image data.In view of the shortcomings of the above methods,this paper uses a combination model of GASF(Gramian Angular Summation Field)imagery algorithm based on time series time correlation characteristics combined with a convolutional neural network model.Sexual features,on the other hand,greatly reduce the redundancy of imaged data.The main results of this paper are as follows:1.A five-class model of arrhythmia based on GASF ECG signal imaging algorithm and convolutional neural network was trained.First,the three-layer wavelet threshold denoising algorithm is used to denoise the ECG signal in the MIT-BIH arrhythmia database;the differential threshold method is used to perform R wave detection and heart beat segmentation on the ECG signal;secondly,the GASF matrix with a size of 40 * 40 is selected Extract the time-dependent features of the ECG signal to generate an ECG signal feature matrix data set;use the ECG signal feature matrix data set to train the arrhythmia five-class convolutional neural network model,and the final model’s average classification accuracy on the test set is 97.51%.2.Designed and optimized the main module IP core of FPGA implementation process of arrhythmia classification system.This includes writing Verilog HDL code to implement a three-layer wavelet threshold denoising process on the FPGA and generating a denoising IP core;using Xilinx’s HLS(High-Level Synthesis)tool to design R-wave detection IP core and GASF imaging respectively IP core,convolutional layer IP core,pooling layer IP core,fully connected layer IP core in convolutional neural network module.And by improving the algorithm structure,adding reasonable optimization instructions and adding AXI4 high-speed bus interface,further improve the computing performance of the above IP core.The generated high-performance IP core will be used to build an FPGA-based arrhythmia classification system.3.Implemented FPGA-based arrhythmia classification system.Using the generated high-performance algorithm module IP core,through software and hardware co-simulation design,an arrhythmia classification system with high classification accuracy,fast operation speed,and low power consumption is finally realized on the Xilinx ZC706 development platform,and its calculation speed is about ARM 5.73 times the calculation speed.