Classification of sEMG Biomedical Signals for Upper-Limb and Hand Rehabilitation Using a Hybrid CNN-SVM Architecture

Abstract

Electromyography (EMG) classification has been an important step to achieve the rehabilitation goal for lower/upper limbs and hands using robotic devices. To perform this step effectively, many researchers have adopted machine learning and deep learning algorithms. In this study, a hybrid CNN-SVM architecture was developed for the classification of surface EMG (sEMG) signals. The CNN part of the proposed architecture is used to extract relevant features from the data and the SVM part would use the extracted features for the classification task. This can be helpful as it will reduce human input and make results more consistent. For this work, we use the Ninapro DB2’s dataset, which contains 3 different Exercises B, C, and D. Thus, we obtained the following accuracy results: an accuracy of 78.56% for Exercise B, an accuracy of 72.84% for Exercise C, and an accuracy of 88.24% for Exercise D.