sEMG Classication with Convolutional Neural Networks: A Multi-Label Approach for Prosthetic Hand Control

Sammanfattning: In myoelectric prosthesis design, there is often a trade-off between control robustness and range of executable movements. As a low movement error rate is necessary in any real application, this often results in a quite severe limitation on the dexterity of the user. One possible remedy for this could come from the use of multi-label machine learning methods, where complex hand movements can be expressed as the sum of several simple movements. I investigate how effective state of the art deep learning methods are at classifying HD-sEMG signals. Notable weight is put on extracting multilabel information from both the spatial and temporal signal domain of EMG signals by use of convolutional neural networks (CNN). In addition, to investigate the feasibility of reducing the number of necessary electrodes, a novel method for quantifying channel importance is proposed. I show that multi-label classication performance can rival that of classical single-label methods, even with a large set of labels. Despite the general stochasticity of sEMG signals, no manual feature engineering is necessary and a very short time window is suficient for accurate classication.