Abstract: In order to meet the needs of many human hand action pattern recognitions in complex applications such as exercise rehabilitation training and human-machine interaction, influences of sEMG signals sampling channel layouts, training sample preparations, feature extraction modes, classifier structure parameters, and so on were analysed herein. Optimal signal sampling scheme was established, and 3 BP neural network classifiers were designed based on time domain statistic, self-regression model coefficient, and wavelet packet decomposition coefficient. Experimental results show that average recognition rate of classifiers for 6 kinds of single finger actions, 13 kinds of multi finger actions, 20 kinds of hand movements are 98.5%, 92.4%, and 90.9% respectively. Output delay is less than 190 ms, which verifies effectiveness and practicability of the proposed method.

Key words: surface electromyography(sEMG), hand action, neutral network, pattern recognition

摘要： 为了满足手部运动功能康复器的主动康复训练对多种人手动作模式识别的需求，分析了表面肌电信号采样通道设置布局、训练样本制作、特征提取方式、模式分类器结构参数等因素对手部动作识别的影响，设计了针对前臂的表面肌电信号采集方案，分别基于时域统计量、自回归模型系数、小波包分解系数特征设计了BP神经网络分类器。实验结果表明：对6种单指动作、13种多指动作、20种手部动作的最佳平均识别率分别为98.5%、92.4%、90.9%，计算时间小于190 ms，验证了所提出方法的有效性和实用性。

关键词: 表面肌电, 手部动作, 神经网络, 模式识别