sEMG多特征融合的自适应神经网络下肢运动意图识别研究  

Study of adaptive neural network based on sEMG signal multi-feature fusion for lower limb motion intention recognition

作　　者：刘瑞恒 张峻霞[1,2] 钱芊橙 LIU Ruiheng;ZHANG Junxia;QIAN Qiancheng(College of Mechanical Engineering,Tianjin University of Science and Technology,Tianjin 300222,China;Tianjin Key Laboratory of Light Industry and Food Engineering Machinery Equipment Integrated Design and Online Monitoring,Tianjin 300222,China)

机构地区：[1]天津科技大学机械工程学院,天津300222 [2]天津市轻工与食品工程机械装备集成设计与在线监控重点实验室,天津300222

出　　处：《现代电子技术》2022年第7期33-40,共8页Modern Electronics Technique

基　　金：2019天津市研究生科研创新项目(2019YJSB014);天津市科技支撑计划资助项目(14ZCZDSY00010)。

摘　　要：针对表面肌电信号单一特征进行动作意图识别准确率低的问题,提出一种利用表面肌电信号多特征融合的动态自适应神经网络算法,实现8种下肢运动意图的准确识别。采集8种下肢动作的表面肌电信号,利用小波基函数对原始信号进行降噪处理,提取时域、小波变换和样本熵的原始特征参数。对原始特征进行主成分分析,降低特征维度,使用改进的差分进化算法优化各个特征的权重值;针对传统BP神经网络梯度下降法收敛速度慢的问题,使用动态自适应学习率的神经网络算法代替传统BP神经网络识别方法,既提升了模型的收敛速度,又提高了运动意图识别的准确率。实验结果表明,采用多特征融合的自适应神经网络模型识别8种下肢运动意图,平均识别准确率达到94.89%,比单特征的BP神经网络方法识别准确率提高10%以上,动作的识别时间只需要280 ms。该方法在300 ms内可实现对下肢动作的识别,能够达到运动意图识别的要求。In view of the low accuracy of action intention recognition based on single feature of sEMG(surface electronomyography)signals,a dynamic adaptive neural network algorithm based on multi-feature fusion of sEMG signals is proposed to realize accurate recognition of eight kinds of lower limb motion intentions.The sEMG signals of eight kinds of lower limb motions are collected.The original signals are subjected to denoising processing by wavelet basis function.The original feature parameters of time domain,wavelet transform and sample entropy(SE)are extracted.The principal component analysis(PCA)is performed on the original features to reduce the feature dimension,and the improved differential evolution(DE)algorithm is used to optimize the weight value of each feature.In view of the slow convergence speed of the traditional back propagation neural network(BPNN)gradient descent method,the dynamic adaptive learning rate neural network algorithm is used to replace the traditional BPNN recognition method,which not only improves the convergence speed of the model,but also improves its accuracy of motion intention recognition.The experimental results show that,when the proposed model is used to identify the eight kinds of lower limb motion intentions,its average recognition accuracy can reach 94.89%,which is 10%(at least)higher than that of the model based on BP neural network method with single feature,and its duration of the motion recognition is only 280 ms.The proposed method can realize the recognition of lower limb motions within 300 ms.Therefore,it can meet the requirements of motion intention recognition.

关 键 词：下肢运动意图识别 多特征融合 动态自适应神经网络 特征提取 下肢表面肌电信号 差分进化算法 小波分析 主成分分析 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]