Convolution Neural Network for EMG-Based Finger Gesture Classification for Novel and Trained Gestures

Surface Electromyography (sEMG) has been extensively investigated for its applications with machine learning algorithms in creating functional and intuitive controls for prosthetic arms. Most proposed sEMG classification systems focus on using various types of classification algorithms to predict the movement classes (labels) of the sEMG data collected from isometric gestures based on discrete pre-assigned labels. In general, when a class is not used in training or calibration of the classifier, it cannot be classified. The objective of the proposed study was to investigate whether a Convolution Neural Network (CNN) could be used to classify sEMG signals from novel gestures that the CNN had not been trained with. Results indicate that the proposed CNN can predict some novel gestures that it has not seen in training. This ability of generalization varied depending on the datasets and subjects.