A Robust Feature Set for Wearable Multichannel Myoelectric Devices in Practice

A wearable myoelectric device is essentially a surface electromyography(sEMG) based human machine interaction (HMI) system. The non-stationary property of sEMG could be one of the obstacles that degrade the user experience of wearable myoelectric devices because they need to be put on and taken off frequently, which brings in the time-variation effects specified in this paper. In order to reduce the influence cause by time-variation on wearable myoelectric devices, the common spatial pattern(CSP) was employed to extract relatively robust features from multichannel sEMG signals. As comparation, the conventional time domain(TD) features together with time domain auto regressive(TDAR) features were investigated. To implement the experiment, 16 electrodes were utilized to simulate a wearable myoelectric device, and an absolute value based Teager-Kaiser energy operator(TKEO) strategy was proposed to realize onset detection. The experiment was divided into 2 sessions, corresponding to the situations before and after time-variation happened respectively. The support vector machine (SVM) and linear discrimination analysis(LDA) were used to build the pattern recognition models. In session2, the average classification accuracy of CSP(95.2±2.3%) is superior to TD(90.3±4.4%) and TDAR(92.4±4.2%) under LDA. The result suggests that the CSP feature set is relatively robust against accuracy reduction resulted from time-variation in motion recognition.