基于迁移学习的肌电信号手势识别方法研究  

Research on EMG Signals Gesture Recognition Based on Migration Learning Strategy

作　　者：游昌欣 韩晶[1] 常皓 任兵 李玉菡 韩衍 YOU Changxin;HAN Jing;CHANG Hao;REN Bing;LI Yuhan;HAN Yan(School of Mechanical and Electrical Engineering,North University of China,Taiyuan 030051,China;Unit 32382 of PLA,Beijing 100072,China)

机构地区：[1]中北大学机电工程学院,太原030051 [2]中国人民解放军32382部队,北京100072

出　　处：《中北大学学报（自然科学版）》2024年第4期531-538,共8页Journal of North University of China(Natural Science Edition)

基　　金：中央引导地方科技发展基金项目(YDZJSX2022A024);中央引导地方科技发展基金项目(YDZJSX2023A026);中国航天伺服驱动与传动实验室开放基金(LASAT-20210302);2023年度山西省研究生教育创新项目(2023KY595)。

摘　　要：在利用肌电信号进行手势识别的过程中,受行为习惯、肌肉组织以及佩戴方式不同等影响,往往存在手势训练时间长、数据量大、识别准确度低、实时性差等难题。为此,本文运用迁移学习理论对LSTM算法模型进行改进,选取不同状态的手势组成源任务,以训练好的LSTM网络作为源网络模型,对设定的6种手势完成两种迁移策略与非迁移策略下的手势识别对比实验。结果表明:采用预训练方式的迁移学习策略识别效果优于固定值方式的迁移策略。当采用预训练迁移策略的方式改进LSTM手势识别算法时,其训练时间仅为使用LSTM识别算法所需训练时间的1/16,当每个动作仅重复20次时,准确率就可以达到80.2%,比仅使用源网络中的LSTM手势识别方法平均高出22%。因此,采用迁移学习方法在减轻训练量的同时,也可以提高手势识别的准确率。In the process of gesture recognition using EMG signals,affected by behavioral habits,muscle tissues,and different wearing styles,there were often difficulties such as long gesture training time,large data volume,low recognition accuracy,and poor real-time performance.For this reason,this paper used the migration learning theory to improve the LSTM algorithm model,selected different states of gestures to form the source task,took the trained LSTM network as the source network model,and completed the gesture recognition comparison experiments under the two migration strategies and non-migration strategies for the six types of gestures set. The results show that the recognition effect of the migration learningstrategy using the pre-training approach is better than that of the migration strategy using the fixed-valueapproach. When the LSTM gesture recognition algorithm is improved by using the pre-training migrationstrategy, the average training time is only one-sixteenth of the training time required to use the LSTM recognitionalgorithm, and when each action is repeated only 20 times, the accuracy rate can reach 80. 2%,which is 22% higher on average than that of using only the LSTM gesture recognition method in thesource network. Therefore, the gesture recognition accuracy can be improved while reducing the trainingamount by using the migration learning method.

关 键 词：肌电传感技术 迁移学习 神经网络 无人系统 手势识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]