Upper Limb Motion Recognition Based on Two-stream Convolutional Neural Network for sEMG Signals

doi:10.3969/j.issn.1004-132X.2026.03.019

Abstract

Abstract:

In order to enhance the accuracy of upper limb motion recognition based on sEMG signals and to validate the applications of the intent recognition model in real rehabilitation robots， a upper limb motion recognition method was proposed using a two-stream convolutional neural network for sEMG signals. The approach began by applying wavelet threshold denoising， bandpass filtering， full-wave rectification， and envelope smoothing， followed by sample construction using a sliding window. The original EMG signals were then processed with variational mode decomposition and discrete wavelet packet transform. Key intrinsic mode functions and wavelet packet transform coefficients were extracted as inputs for the two branches of the model to enable high-level feature learning. A temporal convolutional network was employed to capture temporal dynamics and global dependencies within the features. The feature fusion module then integrated the high-level feature information. The proposed method achieves average recognition accuracies of 93.43%， 92.37%， and 97.54% on the public Ninapro DB4/DB5 datasets respectively and self-collected data for 6 upper limb movements. The average recognition accuracy reaches 87% for the 6 upper limb movements of 5 participants.

Key words: upper extremity motion recognition, two-stream convolutional neural network, surface electromyographic（sEMG） signal, variational modal decomposition, discrete wavelet packet transform, upper extremity motion recognition experiment

摘要：

为提高基于表面肌电信号的上肢动作识别精度，验证意图识别模型在实际康复机器人上的应用，提出了一种基于双流卷积神经网络的表面肌电信号上肢动作识别方法。采用小波阈值去噪、带通滤波、全波整流与包络平滑，并以滑动窗口进行样本构建。对原始肌电信号进行变分模态分解和离散小波包变换，同时提取突出的本征模态函数和离散小波包变换系数作为模型两个分支的输入进行高层特征的学习。采用时间卷积网络捕捉特征中的时间动态信息和全局依赖关系，最终通过特征融合模块实现高层特征信息的融合。所提方法在公开数据集Ninapro DB4/DB5和自采的6类上肢动作数据中，平均识别准确率分别达到了93.43%、92.37%和97.54%，并且在上肢动作识别实验中5名实验人员的6类上肢动作的平均识别准确率达到了87%。

关键词: 上肢动作识别, 双流卷积神经网络, 表面肌电信号, 变分模态分解, 离散小波包变换, 上肢动作识别实验

CLC Number:

TP911.7

LI Xianhua, YIN Sheng, QIU Xun, DU Pengfei, SONG Tao. Upper Limb Motion Recognition Based on Two-stream Convolutional Neural Network for sEMG Signals[J]. China Mechanical Engineering, 2026, 37(3): 697-707.

李宪华, 尹胜, 邱洵, 杜鹏飞, 宋韬. 基于双流卷积神经网络的表面肌电信号上肢动作识别[J]. 中国机械工程, 2026, 37(3): 697-707.

Figures/Tables 19

References 19

[1]	王亚楠，吴思缈，刘鸣. 中国脑卒中15年变化趋势和特点［J］. 华西医学， 2021， 36（6）： 803-807.
	WANG Yanan， WU Simiao， LIU Ming. Temporal Trends and Characteristics of Stroke in China in the Past 15 Years［J］. West China Medical Journal， 2021， 36（6）： 803-807.
[2]	CHOCKALINGAM M， VASANTHAN L T， BALASUBRAMANIAN S， et al. Experiences of Patients Who Had a Stroke and Rehabilitation Professionals with Upper Limb Rehabilitation Robots： a Qualitative Systematic Review Protocol［J］. BMJ Open， 2022， 12（9）： e065177.
[3]	FU Rongrong， ZHANG Baozhong， LIANG Haifeng， et al. Gesture Recognition of sEMG Signal Based on GASF-LDA Feature Enhancement and Adaptive ABC Optimized SVM［J］. Biomedical Signal Processing and Control， 2023， 85： 105104.
[4]	HYE N M， HANY U， CHAKRAVARTY S， et al. Artificial Intelligence for sEMG-based Muscular Movement Recognition for Hand Prosthesis［J］. IEEE Access， 2023， 11： 38850-38863.
[5]	PRABHAVATHY T， ELUMALAI V K， BALAJI E， et al. A Surface Electromyography Based Hand Gesture Recognition Framework Leveraging Variational Mode Decomposition Technique and Deep Learning Classifier［J］. Engineering Applications of Artificial Intelligence， 2024， 130： 107669.
[6]	CHEN Qingzheng， TAO Qing， ZHAO Muchao， et al. CNN-based Gesture Recognition Using Raw Numerical Gray-scale Images of Surface Electromyography［J］. Biomedical Signal Processing and Control， 2025， 101： 107176.
[7]	LIU Xiaoguang， ZHANG Mingjin， WANG Jiawei， et al. Gesture Recognition of Continuous Wavelet Transform and Deep Convolution Attention Network［J］. Mathematical Biosciences and Engineering， 2023， 20（6）： 11139-11154.
[8]	XIONG Baoping， CHEN Wensheng， NIU Yinxi， et al. A Global and Local Feature Fused CNN Architecture for the SEMG-based Hand Gesture Recognition［J］. Computers in Biology and Medicine， 2023， 166： 107497.
[9]	ATZORI M， MÜLLER H. The Ninapro Database： a Resource for sEMG Naturally Controlled Robotic Hand Prosthetics［C］∥2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society （EMBC）. Milano， 2015： 7151-7154.
[10]	SHEN Shu， GU Kang， CHEN Xinrong， et al. Gesture Recognition through sEMG with Wearable Device Based on Deep Learning［J］. Mobile Networks and Applications， 2020， 25（6）： 2447-2458.
[11]	CHAI Yuanyuan， LIU Keping， LI Chunxu， et al. A Novel Method Based on Long Short Term Memory Network and Discrete-time Zeroing Neural Algorithm for Upper-limb Continuous Estimation Using sEMG Signals［J］. Biomedical Signal Processing and Control， 2021， 67： 102416.
[12]	姜海燕，许先静，钟凌珺，等. 采用变分模态分解与领域自适应的表面肌电信号手势识别［J］. 西安交通大学学报， 2024， 58（5）： 75-87.
	JIANG Haiyan， XU Xianjing， ZHONG Lingjun， et al. Gesture Recognition of Surface Electromyography Based on Variational Mode Decomposition and Domain Adaptation［J］. Journal of Xi’an Jiaotong University， 2024， 58（5）： 75-87.
[13]	NGUYEN P T， KUO C H. A Novel Surface Electromyographic Gesture Recognition Using Discrete Cosine Transform-based Attention Network［J］. IEEE Signal Processing Letters， 2024， 31： 266-270.
[14]	PENG Xiangdong， ZHOU Xiao， ZHU Huaqiang， et al. MSFF-Net： Multi-stream Feature Fusion Network for Surface Electromyography Gesture Recognition［J］. PLoS One， 2022， 17（11）： e0276436.
[15]	WU Yuheng， ZHENG Bin， ZHAO Yongting. Dynamic Gesture Recognition Based on LSTM-CNN［C］∥2018 Chinese Automation Congress （CAC）. IEEE， 2018： 2446-2450.
[16]	JOSEPHS D， DRAKE C， HEROY A， et al. sEMG Gesture Recognition with a Simple Model of Attention［J］. Proceedings of Machine Learning Research， 2020， 136： 126-138.
[17]	XU Zhengyuan， YU Junxiao， XIANG Wentao， et al. A Novel SE-CNN Attention Architecture for sEMG-based Hand Gesture Recognition［J］. Computer Modeling in Engineering & Sciences， 2023， 134（1）： 157-177.
[18]	WANG Zihao， WAN Huiying， MENG Long， et al. Optimization of Inter-subject sEMG-based Hand Gesture Recognition Tasks Using Unsupervised Domain Adaptation Techniques［J］. Biomedical Signal Processing and Control， 2024， 92： 106086.
[19]	PENG Fulai， CHEN Cai， Danyang LYU， et al. Gesture Recognition by Ensemble Extreme Learning Machine Based on Surface Electromyography Signals［J］. Frontiers in Human Neuroscience， 2022， 16： 911204.

数据集	VMD	DWPT	VMD+DWPT	本文
DB4_12	90.35	87.97	92.57	93.43
DB4_52	79.07	79.57	84.21	86.97
DB5_12	84.56	89.68	91.03	92.37
DB5_52	65.96	80.27	84.25	84.93
NI_6	93.87	88.82	96.78	97.54

数据集	VMD	DWPT	VMD+DWPT	本文
DB4_12	90.35	87.97	92.57	93.43
DB4_52	79.07	79.57	84.21	86.97
DB5_12	84.56	89.68	91.03	92.37
DB5_52	65.96	80.27	84.25	84.93
NI_6	93.87	88.82	96.78	97.54

数据集	VMD	DWPT	VMD+DWPT	本文
DB4_12	93.36	89.68	93.58	94.98
DB4_52	89.78	79.57	92.27	94.12
DB5_12	91.71	91.47	93.65	94.24
DB5_52	81.50	85.67	89.59	90.59
NI_6	96.86	91.90	97.89	98.70

数据集	VMD	DWPT	VMD+DWPT	本文
DB4_12	93.36	89.68	93.58	94.98
DB4_52	89.78	79.57	92.27	94.12
DB5_12	91.71	91.47	93.65	94.24
DB5_52	81.50	85.67	89.59	90.59
NI_6	96.86	91.90	97.89	98.70

数据集（手势数量）	方法	模型类型	识别准确率（所有实验人员）	识别准确率（单独实验人员）
NinaPro_DB4（52）	文献［13］	FANet	78.70	未提供
	文献［8］	GLF-CNN	82.2	未提供
	文献［14］	CNN	未提供	84.87
	本文模型	CNN	86.97	94.12
NinaPro_DB5（12）	文献［15］	LSTM-CNN	未提供	71.66
	文献［16］	CNN	89.18	未提供
	文献［17］	SE-CNN	89.54	未提供
	本文模型	CNN	92.37	94.24
NinaPro_DB5（52）	文献［10］	CNN	未提供	74.51
	文献［18］	CDEM	84	未提供
	文献［19］	EELM	77.9	未提供
	本文模型	CNN	84.93	90.59