基于残差神经网络和表面肌电信号的肌肉疲劳监测研究  

Identification of Muscle Fatigue Based on SEMG Signal and ResNet

作　　者：方亚南[1,2] 汪晓红 丁一 FANG Yanan;WANG Xiaohong;DING Yi(College of International Education,Hefei College of Finance&Economics,Hefei 231299,Anhui Province;School of Economics and Management,Anhui Polytechnic University,Wuhu 241000,Anhui Province)

机构地区：[1]合肥财经职业学院国际教育学院,安徽合肥231299 [2]安徽工程大学经济与管理学院,安徽芜湖241000

出　　处：《沈阳工程学院学报(自然科学版)》2025年第1期71-77,共7页Journal of Shenyang Institute of Engineering:Natural Science

基　　金：教育部人文社科青年基金项目(23YJC630032);安徽省自然科学基金面上项目(2308085MG228);安徽省高校优秀青年基金(2023AH030023)。

摘　　要：对一种基于深度学习的肌肉电信号进行监测实验,共招募14名在校女大学生做肌电和心电两部分数据采集。首先,利用db小波对采集的肌电信号做两次小波变换,完成数据降噪;其次,通过快速傅里叶变换做数据过滤,然后进行时域和频域分析;最后,利用残差神经网络搭建数据处理模型,其中ResNet模型引入残差模块,能够有效地消除由于模型层数增加而导致的梯度弥散或梯度爆炸问题,提取肌肉收紧和放松状态时的数值特征。监测实验的数据处理结果显示:测试集识别准确率最终趋于0.766,训练集识别准确率最终趋于0.831,具有较高的准确率。In order to improve the accuracy of monitoring muscle fatigue,classify muscle fatigue status,and reduce the occurrence of overwork death,a deep learning-based muscle electrical signal monitoring experiment was carried out.The EMG data of 14 female college students were collected in the experiment.The d-b wavelet transformation was used to eliminate the high-frequency noise.Then,wavelet packet transformation was conducted to separate high-frequency components again for soft threshold processing to complete data denoising.Then FFT is used for data filtering,and time-domain and frequency-domain analysis is carried out.The valid data after analysis is marked with 0-1,and the ResNet is used to build the data processing model.The residual module is introduced into ResNet model,which can effectively eliminate the gradient dispersion or gradient explosion problems caused by the increase of model layers,and extract the numerical characteristics when the muscles are tightened and relaxed.The data processing results of the monitoring experiment show that the recognition accuracy of the test set tends to 0.766 and the recognition accuracy of the training set tends to 0.831,which has a high accuracy.

关 键 词：生理信号 残差神经网络 深度学习 特征提取 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]