基于SE与XceptionTime肌电信号手势识别  

sEMG Gesture Recognition based on SE Attention and Xception Time

作　　者：刘聪 许婷婷 胡胜 LIU Cong;XU Tingting;HU Sheng(School of Electrical and Electronic Engin.,Hubei Univ.of Tech.,Wuhan 430068,China;Wuhan Hua’an Science and Tech.Co.,Ltd.,Postdoctoral Workstation,Wuhan 430068,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430068 [2]武汉华安科技股份有限公司博士后科研工作站,湖北武汉430068

出　　处：《湖北工业大学学报》2025年第2期29-35,共7页Journal of Hubei University of Technology

基　　金：湖北省重点研发计划(2021BGD013);武汉市科技局知识创新专项(2022010801010255)。

摘　　要：传统基于卷积神经网络的表面肌电信号手势识别算法普遍识别精度不高,究其原因:一是常规卷积对输入通道做卷积并求和,进而造成sEMG信号的通道特征和时序特征混杂一起;二是这些算法对sEMG信号中不同通道间的特征提取不够充分。针对这一问题,提出了一种基于SE注意力与XceptionTime的肌电信号手势识别模型。首先,对原始的表面肌电信号分别进行滤波、归一化和滑动窗口的预处理操作;然后本网络中的SE-XceptionTime模块通过使用深度可分离卷积可将通道特征与时序特征分开映射,并设置3个不同卷积核大小滤波器提取长或短的时序信息;最后在该卷积后加入SE注意力机制,以此来关注通道中的有效特征,弱化无关信息的干扰。实验结果表明,在NinaproDB1数据集上52类手势识别准确度达到了88.72%,均高于目前主流的深度学习手势识别网络模型,充分验证了网络模型的有效性。Conventional convolutional neural network-based surface electromyography signal gesture recognition algorithms generally have poor recognition accuracy.The reasons for this are as follows:first conventional convolution does convolution and summation on the input channels,which in turn causes channel features and timing features of the sEMG signal to be mixed;second,these algorithms do not sufficiently extract the features between different channels in the sEMG signal.To address this problem,this paper innovatively proposes sEMG gesture recognition model based on SE attention and XceptionTime.First,the original sEMG signal is preprocessed with filtering,normalisation and sliding window operations respectively;Then the SE-XceptionTime module in the network then maps channel features to temporal features separately by using depthwise separable convolutions and sets three different convolution kernel size filters to extract long or short temporal information;finally,the SE attention mechanism is added after this convolution,a SE attention mechanism is added after this convolution to focus on the effective features in the channel and weaken the interference of irrelevant information.The experimental results show that the accuracy of 52 types of gesture recognition on the NinaproDB1 dataset reaches 88.72%,which is higher than the current mainstream deep learning gesture recognition network models,fully validating the effectiveness of the network model in this paper.

关 键 词：表面肌电信号 手势识别 深度可分离卷积 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]