EMG SIGNALS FOR FINGER MOVEMENT CLASSIFICATION BASED ON SHORT-TERM FOURIER TRANSFORM AND DEEP LEARNING

Abstract

An interface based on electromyographic (EMG) signals is considered one of the central fields in human-machine interface (HCI) research with broad practical use. This paper presents the recognition of 13 individual finger movements based on the time-frequency representation of EMG signals via spectrograms. A deep learning algorithm, namely a convolutional neural network (CNN), is used to extract features and classify them. Two approaches to EMG data representations are investigated: different window segmentation lengths and reduction of the measured channels. The overall highest accuracy of the classification reaches 95.5% for a segment length of 300 ms. The average accuracy attains more than 90% by reducing channels from four to three.