Research On Robust EMG Pattern Recognition With Multiple Confunding Factors

EMG control is largely influenced by many factors including electrode shifting,dynamic limb posture,dynamic muscle contraction strength,individual difference,longterm variation,and so on,which result in low classification accuracy(CA)in clinical practice.According to these influenced factors,this paper investigated the features extraction method,classifier generalization ability and adaptive learning strategy respectively.The contents of this paper include: optimizing the feature threshold method based on particle swarm optimization(PSO)algorithm,the features extraction method based on discrete Fourier transform(DFT),wavelet transform(WT)and wavelet packet transform(WPT),the learning strategy based on kernel function of support vector machine(SVM),online unsupervised learning strategy based on representative sample updating.Overviews about EMG pattern recognition(EMG-PR)at home and abroad were given in details.Then we found out some problems existing in the current research and determined the main research contents in this paper.In order to reduce the interference of confounding factors including electrode shiftting,limb posture changing and muscle contraction force changing,we firstly studied the features extraction method of EMG.We proposed a method for threshold optimization based on PSO.Compared with the traditional trail & error method,not only can it simplify the parameter selection process of features including zero crossing(ZC),myopulse percentage rate(MYOP),Willison amplitude(WAMP)and slope sign change(SSC),but it can also largely increase the CA by 10.2%.In addition,the hybrid features integrated traditional common features such as mean absolute value(MAV),root mean square(RMS)with DFT,WT,and WPT were proposed,which enhanced the robustness of EMG-PR and increased the CA by 30.5%,25.4%,22.9%,respectively.In view of the interference of confounding factors such as dominant hand/nondominant hand switching,limb posture changing and muscle contraction force changing,we investigated the generalization ability of classifier in this paper.Firstly,probabilistic neural networks(PNN)was used as classifier of EMG-PR,whose generalization ability is stronger than linear discriminant analysis(LDA).Then we studied the kernel function of SVM,and proposed a multi kernel learning methods to improve the generalization ability of SVM.The experiment result showed that the multiscale kernel Gauss kernel can reach the best CA,which increased by 1.5% comparing with the Gauss kernel.In view of the decline of CA resulting from confounding factors including long-term variation and limb posture changing,we proposed an online learning strategy based on representative samples,which can select some samples representing class information mostly from the training set.The experiment result showed that this method can enhance the CA in long-term wearing and other more complex confounding factors including electrode shifting and muscle force changing.