电子测量与仪器学报

谢平,赵连洋,张艺滢,徐猛,江国乾,陈杰.多源域迁移学习的肌电-惯性特征融合及手势识别[J].电子测量与仪器学报,2024,38(7):187-195

多源域迁移学习的肌电-惯性特征融合及手势识别

Multi-source domain transfer learning for electromyography-inertial feature fusion and gesture recognition

DOI：

中文关键词: 肌电-惯性信号跨用户手势识别多源域迁移学习长短时记忆网络特征对齐

英文关键词:EMG inertial signal cross user gesture recognition multi-source transfer learning long short-term memory feature alignment

基金项目:国家自然科学基金（62076216）、河北省高等学校人文社会科学研究项目（SQ2021010）、教育部人文社科青年基金项目（21YJC890001）、河北省自然科学基金委员会重点项目（F2022203079）、河北省创新能力提升计划项目（22567619H）资助

作者	单位
谢平	1.燕山大学电气工程学院秦皇岛066000；2.河北省智能康复及神经调控重点实验室秦皇岛066000
赵连洋	燕山大学电气工程学院秦皇岛066000
张艺滢	北京理工大学设计与艺术学院北京100081
徐猛	燕山大学电气工程学院秦皇岛066000
江国乾	燕山大学电气工程学院秦皇岛066000
陈杰	燕山大学体育学院秦皇岛066000

Author	Institution
Xie Ping	1.College of Electrical Engineering，Yanshan University，Qin Huangdao 066000,China; 2.Hebei Key Laboratory of Intelligent Rehabilitation and Neuroregulation，Qin Huangdao 066000,China
Zhao Lianyang	College of Electrical Engineering，Yanshan University，Qin Huangdao 066000,China
Zhang Yiying	School of Design & Arts，Beijing Institute of Technology， Beijing 100081,China
Xu Meng	College of Electrical Engineering，Yanshan University，Qin Huangdao 066000,China
Jiang Guoqian	College of Electrical Engineering，Yanshan University，Qin Huangdao 066000,China
Chen Jie	College of Physical Education，Yanshan University，Qin Huangdao 066000,China

摘要点击次数: 50

全文下载次数: 300

中文摘要:

在跨用户手势识别研究中，针对单源域迁移学习存在的负迁移和模型泛化性能差的问题，本研究创新性地提出了一种基于肌电-惯性特征融合的多源域迁移学习策略，关键创新点在于整合多个源域的数据，并在此基础上采用域特有特征对齐与域分类器对齐的技术手段。这一方法旨在强化模型在不同用户间的手势识别性能，进而显著提升跨用户手势识别系统的准确性。首先，引入长短时记忆(long short-term memory, LSTM)网络模型，提取肌电惯性信息的平均绝对值、方差、峰值等时序特征；其次进行域特有特征对齐与域分类器对齐，利用多个源域数据完成对目标域的特征提取；最后融合分类损失、域特有特征差异损失和域分类器差异损失3个损失函数，协同优化整体损失。实验结果表明，所提方法与单源域、源域组合等多种传统方法相比，识别平均率有所提高，在NinaPro DB5数据集上，目标用户的手势识别平均准确率达到80%以上。

英文摘要:

In the realm of cross-user gesture recognition research, addressing the challenges of negative transfer and subpar model generalization observed in single-source domain transfer learning, this study presents a multi-source domain transfer learning strategy centered around the fusion of EMG and inertial features. The pivotal innovation lies in amalgamating data resources originating from diverse source domains, and subsequently employing techniques for domain-specific feature alignment and domain classifier alignment. The primary objective of this approach is to bolster model performance in gesture recognition across different users, thus significantly enhancing the accuracy of cross-user gesture recognition systems. Initially, the long short-term memory (LSTM) network model is introduced to extract time series features, encompassing metrics such as average absolute value, variance, and peak value derived from EMG and inertia data. Subsequently, domain-specific feature alignment and domain classifier alignment procedures are executed, facilitating feature extraction within the target domain utilizing data from multiple source domains. Lastly, the fusion of three loss functions—classification loss, domain-specific feature difference loss, and domain classifier difference loss—is undertaken to collectively optimize the overall loss. The experimental results demonstrate that the proposed method exhibits an improvement in average recognition rate compared to various traditional methods, such as single-source domain and source domain combination approaches. On the NinaPro DB5 dataset, the average gesture recognition accuracy for the target users exceeds 80%.

查看全文查看/发表评论下载PDF阅读器