ENABLING MYOELECTRIC CONTROL TRAINING USING CONTINUOUS DATA THROUGH SELF-SUPERVISED REPRESENTATION LEARNING

Abstract

In this work, we explore the potential of integrating continuous transition data into the training process for pattern recognition-based myoelectric control. We use a set of steady-state and continuous transition performance metrics to compare the performance of classifiers trained with continuous data versus the traditional ramp contraction approach. We further compare the performance of the popular LDA classifier with that of a deep gated recurrent unit (GRU) classifier capable of leveraging the temporal dynamics. We also introduce a novel self-supervised contrastive representation learning approach with augmentations that significantly improves the offline steady-state and transition performance. This work provides compelling early evidence of the potential for semi-supervised learning approaches to leverage temporal dynamics in continuous training data to improve the performance of pattern recognition-based myoelectric control.