基于多核学习的下肢肌电信号动作识别

为了提高下肢肌电控制系统中多运动模式识别的准确性,提出一种基于多核学习(MKL)和小波变换尺度间相关性特征提取的多类识别方法.根据多核学习理论,采用二叉树组合策略构造基于多核学习的多类分类器.对下肢4路表面肌电信号进行离散平稳小波变换,用小波系数尺度间的相关性提取特征向量输入构造的多类分类器,对水平行走时划分的支撑前期、支撑中期、支撑末期、摆动前期、摆动末期这5个细分运动状态进行分类.实验结果表明,所提的多模式识别方法能够以较高识别率区分多个细分运动状态,得到比标准的单核支持向量机(SVM)分类器更好的准确性.

Electromyography movement recognition of lower limb based on multiple kernel learning

In order to improve the precision of multi motion pattern recognition in lower limb myoelectric control system, a multi class recognition method was proposed based on the feature extraction using the inter scale dependency by the wavelet transform and the multiple kernel learning (MKL). A MKL based multi classifier was constructed by the binary tree combined strategy according to the MKL theory. Four channel surface electromyography signals of lower limb were decomposed by the stationary wavelet transform. Eigenvectors were extracted using the inter scale correlations between wavelet coefficients, and inputted into the MKL based multi classifier. Five subdividing patterns were identified in level ground walking, i.e. support prophase, support metaphase, support telophase, swing prophase and swing telophase. Experimental results show that the method can successfully identify these subdividing patterns with better accuracy than the standard single kernel support vector machine (SVM) classifier.