Independent component analysis in multi-channel forearm surface electromyography processing on armband approach for hand gesture classification

Abstract

Multi-channel surface electromyography acquisition is widely applied for gesture recognition in wearable armband devices and it may contain redundant information between the channels. An issue in this technique is the number of channels, which can improve the classification accuracy, but it requires more computational effort. Independent component analysis is a signal processing technique used in blind source separation problems and can be used to reduce dimension, separating linearly mixed sources. In this study, an analysis of the influence of independent component analysis in hand gesture classification with surface electromyography signals, acquired from the forearm, is proposed. Six gestures were acquired from 10 subjects, 4 time-domain features were extracted, and five classifiers were used in the evaluation. This work compares two approaches to extract the W matrix (the demixing matrix) in independent component analysis: one W matrix for each subject and one W for each gesture sample. Besides, the effect of increasing the number of channels in the classification is analyzed, aiming to find a statistically relevant number of independent components. The results showed that the two approaches of W matrix have no significant difference between them. Moreover, it was observed that the number of independent components affects the classifiers, but five components showed the same distribution of results compared with more components. Concerning the classifiers, extreme learning machine neural network and support vector machines presented the best results (over 90%).