Abstract:  To improve the correct recognition rate and efficiency of surface electromyogram signals（SEMGS） on the man-machine interface of smart wheelchair, feature representation and recognition are mainly studied based on wavelet packet multi-scale?decomposition. The collected SEMGS are decomposed at the specified scale and kernel functions according to the same set of orthogonal wavelet packet basis, and then this paper uses multi-scale decomposition coefficients of wavelet packet to construct the feature base vectors of SEMGS. Considering the possible feature information redundancy of multi-channel SEMGS, to eliminate the redundant information, reconstruction is done to the feature space of multi-channel SEMGS by orthogonal normalization, and it uses the dual coordinate vectors of refactoring feature vectors as the final feature representation of multi-channel SEMGS. It uses Nonlinear Autoregressive neural network（NARX） to realize the classification of four different action patterns of two channels of SEMGS. Experimental results show that the reconstructed dual coordinate vectors of wavelet packet multi-scale?decomposition can not only be used as the feature representation of SEMGS, but also simplify the structure of classifier effectively.

Key words: surface electromyogram（EMG） signals, man-machine interface, multi-scale decomposition of wavelet packet, feature representation, pattern recognition

摘要： 为提高智能轮椅人机接口中表面肌电信号的正确识别率和识别效率，主要研究了基于小波包多尺度分解的特征表示及识别。把采集的表面肌电信号在指定尺度及核函数的同一组正交小波包基下进行分解，用小波包多尺度分解的系数构造表面肌电信号的特征基向量。考虑到多通道表面肌电信号可能存在特征信息冗余，为消除这些冗余信息，对多通道表面肌电信号的特征空间通过正交规范化进行重构，并且用重构特征向量的对偶坐标向量作为表面肌电信号的最终特征表示。用非线性自回归神经网络实现了双通道表面肌电信号四种不同动作模式的分类。实验结果表明，小波包多尺度分解系数的重构对偶坐标向量不仅可作为表面肌电信号的特征表示，并能有效简化分类器的结构。

关键词: 表面肌电信号, 人机接口, 小波包多尺度分解, 特征表示, 模式识别