Robust Algorithm Research For Electrode Offset Problem In EMG Gesture Recognition

Prosthetic limb control based on surface electromyography(s EMG)on the residual limbs of amputees is an important human-computer interaction technique.The non-stationarity of the signal caused by the s EMG electrode shift will cause a sharp decrease in the recognition rate of motion,which will seriously impede the daily use of myoelectric prosthetic control.Most of the existing robust algorithms for solving electrode shift are based on high-density s EMG signals,and can achieve good performance.However,high-density s EMG prosthetics do not meet the requirements for clinical practicality.The low-density s EMG prosthesis are convenient and with high practicality for users’ daily use.The problem is the lack of effective and robust motion recognition methods for the quantitative characterization of low-density s EMG electrode shift and for processing the electrode shift effects.In this study,we focused on the non-stationarity of s EMG caused by electrode shift.We designed an electrode shift experiment and collected 8-channel s EMG signals of 10 types of hand gestures.Twenty-five healthy subjects and 4 amputees were enrolled.New algorithms that can robustly and accurately decode hand gestures from s EMG signals were developed and validated on this dataset.Specifically,our major works are summarized as follows.Firstly,we proposed a new method to visualize the changes of feature space caused by electrode shift.The t-distribution random neighborhood embedding(t-SNE)method was used to perform nonlinear reduction of high-dimensional s EMG features so that the feature space changes caused by electrode shift can be visualized.A new index of spatial distance ratio was proposed to characterize the feature space changes caused by electrode shift quantitatively.By comparing the results of various s EMG recognition algorithms,we found that the electrode shift caused a significant change in the time domain(TD)feature distribution.The non-negative matrix factorization(NMF)algorithm can effectively reduce the extent of the feature space change,while the robust algorithms developed for high-density electrode environments have limited effectiveness in our dataset(low-density electrodes).Secondly,we developed a new feature space invariant algorithm based on NMF.We improved the existing NMF algorithm through the row sum constraints of the coefficient matrix to improve the consistency of the signal feature space in the presence of electrode shifts.A new sliding window nonnegative matrix factorization(SW-NMF)method is further developed and combined with Self-Enhancing Linear Discriminant Analysis(SE-LDA)to combat the changes of s EMG feature space.The results showed that in the case of electrode shift,the recognition rate of SW-NMF is 82.09 ± 10.15%,which is significantly higher than the recognition rate of 55.04 ± 12.10% of the traditional TD + Linear Discriminant Analysis(LDA)classification method.Thirdly,we further developed an active compensation algorithm for electrode shift.Different from the above SW-NMF algorithm which finds a constant projection space,another idea is to compensate for the feature space change caused by electrode shift.To this end,we proposed a tracing algorithm(TA),which first compares the root-mean-square(RMS)characteristics of the 2-second data of the up-cut action before and after the electrode shift to find the smallest difference in the location interval,and then uses the electrode configuration at both endpoints to train the classifier.The classification results showed that the classification result of the TA algorithm(85.37 ± 9.84%)is significantly higher than the TD feature and the SW-NMF algorithm.Non-stationarity of signals is a common phenomenon in biomedical signal processing and related human-computer interaction research,and it is also one of the major challenges encountered in current research.Aiming at the problem of electrode shift in the prostheses control based on s EMG,this paper proposed two different robust recognition algorithms by the observation and characterization of feature space changes,along with actively compensating for the feature space changes.Thus,we reduced the spatial projection invariance and improved the effect of s EMG prosthesis in daily life and clinical rehabilitation.This study is conducive to the popularization of human-computer interaction technology in daily life.