An Adaptive Dynamic Mixing Model for sEMG Real-Time ICA on an Ultra-Low Power Processor

Blind Source Separation (BSS) has shown promise in enhancing the interpretability and usability of surface electromyography (sEMG) signals for Human-Machine Interfaces (HMIs). However, existing BSS algorithms often rely on offline processing with isometric movements, limiting their applicability in real-time human-machine interfaces. This study proposes a dynamic mixing approach for sEMG signal BSS implemented on an embedded platform, enabling real-time processing. The algorithm is based on dynamically updating BSS matrices and adaptively adjusting the signal separation process based on the changes in the sEMG sources during contractions. We implemented our solution end-to-end, leveraging an ultra-low power processor specialized for computation-intensive edge applications. Experimental results demonstrate the effectiveness of the proposed approach in separating the sEMG sources in real-time, with 91.8% accuracy and a power envelope ≤ 83.4 mW. Real-time operation is proven as 128 ms of data are processed with a latency ≤ 97.6 ms. This work demonstrates the performance of the BSS algorithm on wearable sEMG-based systems for HMIs, opening doors for advanced applications in rehabilitation, prosthetics, and human-computer interaction.