| 孙鹏,宋杰,张克,卫文韬,李忠新.融合小波门控时序卷积网络的肌电手势识别研究[J].电子测量与仪器学报,2026,40(4):205-214 |
| 融合小波门控时序卷积网络的肌电手势识别研究 |
| Electromyographic signal gesture recognition methodbased on wavelet gated temporal convolution |
| |
| DOI： |
| 中文关键词: 肌电信号 手势识别 离散小波分解 时序卷积神经网络 门控单元 |
| 英文关键词:EMG signal Gesture recognition DWT TCN GLU |
| 基金项目:国家自然科学基金（62002171）资助项目 |
|
|
| Author | Institution |
| Sun Peng | School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China |
| Song Jie | School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China |
| Zhang Ke | The 208th Research Institute of China Ordnance Equipment Group, Beijing 102202, China |
| Wei Wentao | School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China |
| Li Zhongxin | School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China |
|
| 摘要点击次数: 340 |
| 全文下载次数: 138 |
| 中文摘要: |
| 在肌电信号手势识别领域,表面肌电信号固有的信号变异性是导致当前深度学习模型特征提取能力不足,识别准确率低的主要原因。针对这一问题,结合时序卷积神经网络的长期上下文建模优势与小波变换对非平稳信号的时频分析特性,提出一种新型小波门控时序卷积网络模型。首先通过小波卷积模块将输入的肌电信号进行多级离散小波分解,并将得到的各级分量分别进行一维卷积,小波域的卷积操作在实现多尺度提取时频特征的同时又自适应地增强关键的时频模式并抑制冗余分量,接着将卷积后的细节系数与近似系数进行离散逆小波变换以重构信号,最后将重构信号输入融合门控单元的时序卷积网络,利用时序卷积神经网络捕捉sEMG信号中的长期依赖关系,并使用门控单元对所提取的特征进行过滤。该网络结构在Ninapro DB1数据集上对52类手势分类实现了81.85%的准确率,相比于传统时序卷积网络,准确率提高了4.9%,与近年本领域主流深度模型(如DuNet、GengNet等)相比,该方法在保持模型参数量较小前提下,准确率相对提升幅度达4.0%~7.8%。 |
| 英文摘要: |
| This paper constructs a wavelet-gated temporal convolutional network model. First, the input electromyographic signals are subjected to multi-level discrete wavelet decomposition through a wavelet convolution module, and the components of each level are respectively subjected to one-dimensional convolution. Then, the detailed coefficients and approximate coefficients after convolution are reconstructed via discrete inverse wavelet transform. This process of multi-level decomposition, convolution, and step-by-step reconstruction enables the model to adaptively focus on key time-frequency features. The reconstructed signals are then input into a temporal convolutional network integrated with a gating unit. The proposed network structure achieves an accuracy of 81.85% for 52-class gesture classification on the Ninapro DB1 dataset, which is 4.9% higher than that of traditional temporal convolutional networks. Compared with recent mainstream deep models in this field (such as MSHilbNet, GengNet, etc.), this method achieves a relative accuracy improvement of 4.0%~7.8% while maintaining a smaller number of model parameters. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
