Exploiting temporal dynamics to improve the robustness of continuous myoelectric control
Loading...
Date
2025-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of New Brunswick
Abstract
Myoelectric control based on Surface electromyography pattern recognition (sEMG-PR) offers intuitive and dexterous control of powered prostheses for people with limb differences. However, conventional sEMG-PR systems often struggle with transitions between movements, impacting online usability. In this thesis, we investigated these transition-specific challenges and proposed novel approaches to enhance the performance and user experience of sEMG-PR systems.
We first established a comprehensive framework for evaluating classifier performance during transitions, incorporating transition-specific metrics and continuous dynamic datasets. This framework represents an improvement over conventional evaluation methods, which often focus primarily on steady-state performance and neglect transitions. Our analysis, utilizing this enhanced framework, revealed that classifiers, even with similar steady-state performance, can differ substantially in their ability to handle transitions. This finding underscores the limitations of conventional evaluation methods.
Next, we systematically investigated various error-mitigation strategies, including existing and novel post-processing techniques. While some techniques showed promise, particularly those based on rejection, our findings suggest that relying solely on post-hoc error correction may not be sufficient to address the challenges of transitions effectively.
Finally, we explored incorporating continuous dynamic data, inclusive of transitions, into the training process. Our results demonstrated the advantages of leveraging Long Short-Term Memory (LSTM) networks, which can effectively capture the dynamic nature of transitions. Furthermore, we pioneered the use of self-supervised learning for sEMG-PR, and demonstrated its effectiveness in learning meaningful and robust representations from unlabeled continuous dynamic data, leading to enhanced performance both offline and online.
Our findings underscore the crucial role of temporal information, dynamic training data, and appropriate model selection, particularly temporal models like LSTMs, in achieving robust and reliable sEMG-PR based myoelectric control. The proposed approaches have the potential to significantly enhance the usability and effectiveness of these systems, paving the way for more intuitive and user-friendly prosthetic devices.