sEMG-based hand gestures classification using a semi-supervised multi-layer neural networks with Autoencoder

This work presents a semi-supervised multilayer neural network (MLNN) with an Autoencoder to develop a classification model for recognizing hand gestures from electromyographic (EMG) signals. Using a Myo armband equipped with eight non-invasive surface-mounted biosensors, raw surface EMG (sEMG) sensor data were captured corresponding to five hand gestures: Fist, Open hand, Wave in, Wave out, and Double tap. The sensor collected data underwent preprocessing, feature extraction, label assignment, and dataset organization for classification tasks. The model implementation, validation, and testing demonstrated its efficacy after incorporating synthetic sEMG data generated by an Autoencoder. In comparison to the state-of-the-art techniques from the literature, the proposed model exhibited strong performance, achieving accuracy of 99.68%, 100%, and 99.26% during training, validation, and testing, respectively. Comparatively, the proposed MLNN with Autoencoder model outperformed a K-Nearest Neighbors model established for comparative evaluation.