Lightweight deep neural network models for electromyography signal recognition for prosthetic control

Abstract

In this paper, lightweight deep learning methods are proposed to recognize multichannel electromyography (EMG) signals against varying contraction levels. The classical machine learning, and signal processing methods namely, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), root mean square (RMS), and waveform length (WL) are adopted to convolutional neural network (CNN), and long short-term memory neural network (LSTM). Eight-channel recordings of nine amputees from a publicly available dataset are used for training and testing the proposed models considering prosthetic control strategies. Six class hand movements with three contraction levels are applied to WL and RMS-based feature extraction. After that, they are formed into appropriate input dimensions, and classified using the LDA, QDA, LDA-CNN, QDA-CNN, LSTM, and CNN. Depending on three prosthetic EMG validation approaches (Scheme 1-3), the accuracy rates of 41.68%, and 47.27% are yielded by LDA, and QDA with 32- dimensional RMS, and WL features while CNN with 2 × 16 input has 82.87% (up to 88.10%). The effect of the learnable filters of the DL approaches, and signal windowing on the success rate and delay time are discussed in the paper. The simulations show that 2D-CNN (accuracy of 82.87% with 1.7 ms delay) can be successfully adapted to prosthetic control devices.