基于数据驱动可迁移网络的肌电模式识别研究

Electromyography (EMG) pattern recognition based myoelectric control technology using traditional machine learning algorithms, is faced with heavy training burden, high user dependence, and poor robustness etc. Aiming at solving these common problems, a systematic study on the EMG pattern recognition based on deep learning is proposed. The major contents of this study could be summarized as follows: (1) To meet the needs of deep learning for large sample data sets, the hand gestures in Chinese sign language are selected to form the target gesture set and high density array electrode is used for EMG collection to build a standard database for the study of EMG pattern recognition based on deep learning; (2) To overcome the influence of electrode deviation and individual differences among users, blind source separation technology is used to decompose high density array EMG signals, and channel optimization scheme for myoelectric pattern recognition is explored based on muscle activation area. (3) To enhance the robustness of the classifier and reduce user dependent gestures, large scale China sign language gesture set is used as the source task, small scale EMG control gesture set is designed as the target task, and the research of EMG pattern recognition based on large data driven migration network is carried out. The launching of this study is expected to solve some key problems that impede the development of myoelectric control technology and promote the application of this technology in rehabilitation engineering and human-machine interaction etc.

旨在解决基于模式识别的肌电控制技术面临的训练负担重、用户依赖度高、算法鲁棒性差等由传统机器学习算法带来的共性问题，本课题拟对基于深度学习的手势肌电模式识别展开系统研究，主要研究内容包括：（1）为满足深度学习对大样本数据集的需求，以中国手语动作为目标手势集，采用高密度阵列电极采集肌电信号，构建适用于基于深度学习的肌电模式识别研究的标准数据库。（2）为克服电极偏移及用户间个体差异等因素的影响，采用盲源分离技术分解高密度阵列肌电信号，探索在肌肉激活区域基础上优选用于肌电模式识别的通道组合方案。（3）为提升分类器的鲁棒性并减少用户依赖，将中国手语包含的大规模手势集作为源任务、小规模待识别的肌电控制手势集作为目标任务，探索基于大数据驱动可迁移网络的肌电模式识别方案。本课题研究工作的开展，有望通过深度学习分类器的泛化能力，解决阻碍肌电控制技术发展的关键难题，推动该技术在康复工程及人机交互等领域的应用。

旨在解决基于模式识别的肌电控制技术面临的训练负担重、用户依赖度高、算法鲁棒性差等由传统机器学习算法带来的共性问题，本课题对基于深度学习的手势肌电模式识别展开了系统研究，主要研究工作和取得的成果可概述如下：（1）研制了128通道高密度阵列肌电采集系统，以中国手语动作为目标手势集，构建了包含163种手势、65位受试者的高密度阵列手势肌电数据库，为基于深度学习的肌电模式识别打下了数据基础；（2）采用主分量分析、独立分量分析、非负矩阵分解等盲源分离技术分解高密度阵列肌电信号，开展了基于肌肉主要激活区域的电极校准研究以及基于空间激活模式的肌电模式识别研究，为克服电极偏移及用户间个体差异等因素的影响提出了有效的解决方案；（3）探索了基于数据驱动可迁移网络的肌电模式识别方案，并将脉冲循环神经网络（SRNN）引入肌电模式识别中，验证了利用深度学习方法实现高鲁棒性、低训练负担和低功耗肌电模式识别的可行性；（4）面向肌电控制实际应用的需求，扩展了同步肌电模式识别和肌力估计、基于深度学习算法的腕戴式手势识别设备研制、以及基于PPG信号的手势识别等研究内容。本课题取得的研究成果，对于通过深度学习分类器的泛化能力解决阻碍肌电控制技术发展的关键难题、推动该技术在康复工程及人机交互等领域的应用具有潜在的应用价值。

内容获取失败，请点击重试