Robust SEMG-based Gesture Identification Against Electrode Shift

The Surface Electromyography(sEMG)is a biological signal recorded via electrodes on skin surface.sEMG has been widely used in myoelectric control,due to richness of neural information,ease of acquisition and non-invasiveness.Although sEMG based gesture identification has achieved satisfying performance,accuracy may significantly drop since the distribution of sEMG patterns are influenced by electrode shift.As a result,a robust algorithm against electrode shift is a significant problem for sEMG based gesture identification.Although re-calibrating a recognition system by newly collected labeled samples after electrode shift is able to maintain the accuracy,collecting labeled samples is inconvenient and time-consuming for a user.Without the label information,the method may not work efficiently when the data distribution is changed significantly.This study firstly analyzes the influence of electrode shift.According to the distribution correlation and characteristic relationship of data,we propose the distribution mapping based ensemble method with weight fusion and the unsupervised domain adaptation method to against electrode shift with unlabeled samples.They aim to effectively learn on samples after electrode shift.The main contributions of study are:1)The online semi-supervised learning method which learns unlabeled test samples incrementally is proposed to adapt the distribution difference.The effectiveness of the new model is validated experimentally.2)The distributions of samples affected by different electrode shift are different but relevant.We propose a weighted multiple classifier system based on distribution mapping.The model uses different nonlinear functions to quantify the effect of electrode shift.The experimental results show that this method is robust to electrode shift.3)Based on the effectiveness of the new distribution information,an unsupervised domain adaptation method is proposed.It calibrates the classifiers with the estimated distribution parameters.The experimental results suggest that the proposed method enhances robustness of hand gesture identification against electrode shift.