卷积神经网络结合子域适应的低采样率肌电手势识别  

Convolutional neural network combined with subdomain adaptation for low sampling rate EMG-based gesture recognition

作　　者：周雕 熊馨[1] 周建华[1] 宗静 张琪[1] ZHOU Diao;XIONG Xin;ZHOU Jianhua;ZONG Jing;ZHANG Qi(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《浙江大学学报（工学版）》2024年第10期2011-2019,共9页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(82060329).

摘　　要：为了提升模型识别低采样率肌电手势的性能,提出新的识别方法.通过信息扩展层对预处理后的低采样率肌电信号信息进行扩展,增强关键特征的表示.在特征提取网络中,利用子域适应网络提取源域与目标域中的域不变特征后进行域不变特征分类.使用NinaPro数据库中的DB1和DB5子数据库对所提方法进行评估.实验结果表明,所提方法识别53种和52种手势的最高准确率分别为90.89%(DB1)、89.90%(DB5)和82.01%(DB1)、77.07%(DB5),能够降低电极移位、肌肉疲劳、皮肤阻抗的变化和肌肉相对电极的相对运动等因素对低采样率肌电手势识别的影响.A new recognition method was proposed to improve the performance of low sampling rate electromyography(EMG)-based gesture recognition.The information of the pre-processed low sampling rate EMG signal was extended by an information extension layer,and the representation of key features was enhanced.In the feature extraction network,domain invariant features in the source and target domains were extracted by the subdomain adaptation network,then the domain invariant features were classified.The proposed method was evaluated using the DB1 and DB5 sub-databases of the NinaPro database.Experimental results showed that the proposed method recognized 53 and 52 gestures with the highest accuracy of 90.89%(DB1),89.90%(DB5)and 82.01%(DB1),77.07%(DB5),respectively.The effects of factors on low sampling rate EMG-based gesture recognition are reduced by the proposed method,factors that include electrode shift,muscle fatigue,changes in skin impedance,and the relative movement of the muscle relative to the electrodes.

关 键 词：低采样率表面肌电 手势识别 子域适应 信息扩展 挤压与激励注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]