Myoelectric Signal Classification Using Convolutional Neural Networks with Pre-Extracted Features

Abstract

This paper presents a model in which time and frequency domain data are pre-extracted from electromyographic (EMG) signals and used as input to a Convolutional Neural Network (CNN) to classify a patient's limb movements. EMG is a source of information for the development of prostheses, but its classification still represents a challenge due to its inherent variability and non-stationary nature. The proposed approach provides representative information about the EMG signal to CNN, which then selects the relevant features and performs the classification more accurately. The investigation evaluated the work using six data sets, with upper and lower limb movements (p < 0.05). The comparison with other similar approaches demonstrates its potential, reaching 98.84% of accuracy, superior to the traditional algorithms. The study suggests the feasibility of employing CNN-type networks in the case of the EMG signal, combined with a pre-processing technique.