基于时序二维化和卷积特征融合的表面肌电信号分类方法

针对传统模式识别方法在表面肌电信号(sEMG)分类时容易忽略非线性、时序性等特征的问题,文中提出基于时序二维化和卷积特征融合的分类方法.通过格拉姆角场转换实现时序二维化,保留sEMG原始时间序列的时间依赖性和相关性.为了在突出局部信息的同时充分保留细节信息,引入胶囊网络与卷积神经网络共同提取特征,并进行特征融合,实现不同条件下的手势识别.对比多种分类方法的实验表明,文中方法可以有效增强电极偏移情况和面向新对象时手部动作的整体识别水平,具有较强的鲁棒性.

Surface Electromyography Classification Method Based on Temporal Two-Dimensionalization and Convolution Feature Fusion

The traditional pattern recognition methods are prone to ignore characteristics of non-linearity and timing in the classification of surface electromyography(sEMG). Aiming at this problem, a sEMG signal classification method based on temporal two-dimensionalization and convolution feature fusion is proposed. Temporal two-dimensionalization is realized by Gramian angular field conversion to preserve the time dependence and correlation of original time series of sEMG. To highlight the local information and fully retain details simultaneously, a capsule network and a convolutional neural network are introduced to extract features together. In addition, the feature fusion is performed to realize the gesture recognition under different conditions. Experimental results show that the proposed method is more robust than other classification methods and it effectively enhances the electrode offset and the overall recognition level of hand movements facing new objects.