Influence of EMG Data from Different Days on Feature Selection Methods

Abstract

Feature selection is an method which can effectively reduce the dimension of electromyography (EMG) feature vector. Previous studies only focused on the performance of feature selection method in the same day. However, the distribution of EMG data from different days is different. And few studies focused on the influence of different day’s EMG data on feature selection. We mainly discussed the influence of different days’ EMG data on feature selection methods in this study. We adopt three feature selection methods (none feature selection (NFS), sequential forward selection (SFS) and particle swam optimization (PSO)) in this study. Six subjects performed eleven upper limb motions, and the EMG signal was collect in consecutive five days. First, we use the data from the first day for feature selection. After feature selection, we used the EMG feature vectors after feature selection to train a support vector machine (SVM). We performed the same feature selection as the first day on the EMG data of remaining four days. Then, we used the SVM trained in the first day to classify the EMG data of each four days. Finally, we discussed the performance of SVM under three different feature selection methods in different days. The results indicated that EMG data from different days have insignificant influence on feature selection methods (p=0.218). SVM with PSO had a shorter time consumption when compare to NFS (p < 0.05) and SFS(p < 0.05). PSO feature selection method is more suitable to improve the real-time performance of EMG pattern recognition under different days.