sEMG-Driven Hand Dynamics Estimation With Incremental Online Learning on a Parallel Ultra-Low-Power Microcontroller.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) is a key technology for non-invasive human-machine interfaces.
• Inherent variability in sEMG signals, particularly across sessions, hinders the generalization of machine learning models.
• Advancements are shifting sEMG control from classifying static positions to regressing dynamic hand movements for more fluid control.

Purpose of the Study:

• To develop an incremental online-training strategy for sEMG-based estimation of simultaneous multi-finger forces.
• To create a compact Temporal Convolutional Network (TCN) suitable for embedded, on-device learning.
• To validate the method's performance and efficiency in real-world scenarios.

Main Methods:

• An incremental online-training strategy was implemented using a small Temporal Convolutional Network (TCN).
• The method was validated on the HYSER dataset for cross-day performance evaluation.
• The approach was deployed on an ultra-low power GAP9 microcontroller to assess latency and energy consumption.

Main Results:

• The incremental online training achieved a cross-day Mean Absolute Error (MAE) of 9.58 ± 3.89% of Maximum Voluntary Contraction on a challenging, improvised force sequence dataset.
• Performance was comparable to non-embeddable, accuracy-oriented offline training methods.
• Deployment on the GAP9 microcontroller resulted in a low latency of 1.49 ms and energy consumption of 40.4 uJ per update step.

Conclusions:

• The proposed incremental online-training strategy enables accurate, real-time, on-device estimation of multi-finger forces from sEMG signals.
• The solution addresses the limitations of sEMG variability and meets the requirements for embedded systems.
• This approach facilitates versatile and fluid control for human-machine interfaces in various applications.