CNN Classification of Finger Movements using Spectrum Analysis of sEMG Signals

Abstract

Electromyography (EMG) is a technique for measuring muscle responses or electrical activities in response to stimulation of nerves. EMG recorded by using electrodes attached to skin or surface electrodes is called surface electromyogram (sEMG) signals. sEMG signals contain information related to movements of muscles such as finger movements, arm movements, etc. Utilizing sEMG for a user to control a prosthesis is a non-invasive way. In this work, sEMG data is used to classify finger movements. sEMG signal is divided into small segments of either 200ms or 500ms. Each segment is converted to a two-dimensional array similar to an image using short-time Fourier transform (STFT) with different types of taper windows. STFT using Tukey window gives better accuracy than Hann window. The combination of Hann and Tukey windows gives similar results to Tukey window alone. Ten-fold cross-validation are performed on the proposed method. Accuracies of 85.29 (±0.53)% and 97.78 (±0.38)% are obtained for the window lengths of 200ms and 500ms, respectively.