桂彬彬,
陈胜勇
1.天津理工大学计算机科学与工程学院 天津 300384
2.计算机视觉与系统教育部重点实验室 天津 300384
基金项目:国家自然科学基金(U1509207)，天津市虚拟仿真实验教学建设项目基金(津教政办[2019]69)
详细信息
作者简介:杨淑莹：女，1964年生，博士，教授，硕士生导师，研究方向为模式识别、图像处理
桂彬彬：男，1995年生，硕士生，研究方向为医学信号处理
陈胜勇：男，1973年生，博士，教授，博士生导师，研究方向为计算机视觉、图像处理
通讯作者:杨淑莹 yangshuying@email.tjut.edu.cn
中图分类号:R540.41; TP391计量
文章访问数:309
HTML全文浏览量:149
PDF下载量:62
被引次数:0
出版历程
收稿日期:2020-08-31
修回日期:2021-03-23
网络出版日期:2021-04-14
刊出日期:2021-10-18
Arrhythmia Detection Based on Wavelet Decomposition and 1D-GoogLeNet
Shuying YANG,,Binbin GUI,
Shengyong CHEN
1. School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China
2. Key Laboratory of Computer Vision and Systems, Ministry of Education, Tianjin 300384, China
Funds:The National Natural Science Foundation of China(U1509207), Tianjin Virtual Simulation Experiment Teaching Construction Project Fund(JMEC[2019]69)
摘要
摘要:心电图(ECG)信号的准确分类对于心脏病的自动诊断非常重要。为了实现对心律失常的智能分类，该文提出一种基于小波分解和1D-GoogLeNet的精确分类方法。在该方法中，利用Db6小波对ECG信号进行8级分解，得到既含时域信息又有频域信息的多维数据。随后，分解的样本用作1D-GoogLeNet的输入训练该模型。在提出的1D-GoogLeNet模型中，借鉴Inception在图像特征提取中的优异性能，将2维卷积变换为1维卷积学习ECG的特征，并且简化各个Inception的结构，降低模型参数。该文提出的神经网络分类器能够有效缓解计算效率低、收敛困难和模型退化的问题。在实验中，选用MIT-BIH心律失常数据集测试所提模型的性能，对比了信号的不同分解分量组合作为输入的检测结果，当输入数据由{d2-d7}组合时，所提1D-GoogLeNet模型可以达到96.58%的平均准确率。此外，还对比了该模型与未经结构优化的简单1维GoogLeNet在数据集上的表现，前者在准确率上比后者提高了4.7%，训练效率提高了118%。
关键词:心律失常分类/
小波分解/
Inception/
卷积神经网络
Abstract:The accurate classification of ElectroCardioGram (ECG) signals is essential for the automatic diagnosis of heart disease. In order to realize the intelligent classification of arrhythmia, an accurate classification method based on wavelet decomposition and 1D-GoogLeNet is proposed. In this method, Db6 wavelet is used to decompose the ECG signal in eight levels to obtain multi-dimensional data containing both time domain information and frequency domain information. Subsequently, Decomposed samples are used as input to 1D-GoogLeNet to train the model. In the proposed 1D-GoogLeNet model, using Inception's excellent performance in image feature extraction, the two-dimensional convolution is transformed into one-dimensional convolution to learn the features of ECG, and the structure of each Inception is simplified, and the model parameters are reduced. The deep learning classifier proposed in this paper can effectively alleviate the problems of low computational efficiency, difficulty in convergence and model degradation. In the experiment, the MIT-BIH arrhythmia dataset is used to test the performance of the proposed model. The experiment compares the detection results when different decomposition component combinations are used as input. When the input data is combined by {d2-d7}, the proposed 1D-GoogLeNet model can achieve an average accuracy of 96.58%. In addition, the performance of the model and the simple one-dimensional GoogLeNet without structural optimization on the data set is compared. The accuracy of the former is 4.7% higher than the latter, and the training efficiency is increased by 118%.
Key words:Classification of arrhythmia/
Wavelet decomposition/
Inception/
Convolutional Neural Network(CNN)
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=a451a864-7615-47c8-9811-f0c52efeeadb
