• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2022, Vol. 58 ›› Issue (7): 12-19.doi: 10.3901/JME.2022.07.012

• 机器人及机构学 • 上一篇    下一篇

扫码分享

基于迁移学习的手部离散动作识别方法研究

王文东, 李杰, 张俊博, 秦雷, 袁小庆   

  1. 西北工业大学机电学院 西安 710072
  • 收稿日期:2021-05-23 修回日期:2021-11-18 出版日期:2022-05-20 发布日期:2022-05-20
  • 通讯作者: 王文东(通信作者),男,1984年出生,博士,副教授。主要研究方向为人机交互与智能控制方法。E-mail:wdwang@nwpu.edu.cn
  • 作者简介:李杰,男,1998年出生,硕士研究生。主要研究方向为人体意图识别与人机交互。E-mail:li-jie@mail.nwpu.edu.cn;张俊博,男,1998年出生,硕士研究生。主要研究方向为人机交互与类脑混合控制。E-mail:jbzhang@nwpu.edu.cn;秦雷,男,1997年出生,硕士。主要研究方向为人机交互与智能机器人。E-mail:qinlei.chn@gmail.com;袁小庆,男,1979年出生,博士,教授,硕士研究生导师。主要研究方向为智能机器人与控制方法。E-mail:yuan@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(51605385)和陕西省自然科学基础研究(2020JM-131)资助项目。

Discrete Hand Motion Recognition Method Based on Transfer Learning

WANG Wendong, LI Jie, ZHANG Junbo, QIN Lei, YUAN Xiaoqing   

  1. School of Mechanical Engineering, Northwestern Polytechnical University, Xi'an 710072
  • Received:2021-05-23 Revised:2021-11-18 Online:2022-05-20 Published:2022-05-20

摘要: 表面肌电传感器作为新一代机器人人机交互接口设备,目前已在航空航天、军工应用、康复医疗、工业生产等多种环境中表现出巨大的应用潜力和价值。研究发现,表面肌电信号在识别手势动作时,若面临传感器移位、动作用户变化等问题,动作识别准确率将急剧下降,模型可复用能力变差。针对这一情况,提出一种基于小型辅助集的迁移学习建模方法。利用MMD算法对源领域数据集与目标领域数据集的高维距离进行评价,通过TCA算法缩小二者在全局特征上的边缘分布差异,引入小型辅助集对待测数据集创建伪标签,改进了迁移成分分析在数据条件分布相似性上的不足。以多名受试者作为研究对象,验证提出算法的适应性和合理性。肌电控制实验表明,新场景下受试者仅需进行小量训练(仅占源领域数据4%),迁移学习融合模型准确率可提高至80%以上。

关键词: 迁移学习, 表面肌电信号, 意图识别, 人机交互

Abstract: As a new generation of robot human-computer interaction interface equipment, surface EMG sensors have shown great application potential and value in various environments such as aerospace, military applications, rehabilitation medicine, and industrial production. It is found that when it faces problems such as sensor displacement and user changes, the accuracy of action recognition will drop sharply, and the reusability of the model will be poor. In view of this situation, a migration learning modeling method based on small auxiliary sets is proposed. The MMD algorithm is used to evaluate the high-dimensional distance between the source domain data set and the target domain data set, and their edge distribution difference is reduced by the TCA algorithm, and a small auxiliary set is introduced to create pseudo-labels for the data set to be tested, which improves the lack of similarity of data condition distribution in the analysis of migration components. Several subjects were used as research objects to verify the adaptability and rationality of the proposed algorithm. The electromyographic control experiments show that the subjects only need a small amount of training in the new scenario (only 4% of the source field data), and the accuracy of the migration learning fusion model can be increased to more than 80%.

Key words: transfer learning, sEMG, intention recognition, human-computer interaction

中图分类号: