Abstract

Decoding motor intent from surface electromyographic (sEMG) signals is essential for myoelectric prostheses and human-machine interfaces, yet remains challenged by the pronounced no stationarity of muscular signals, particularly in amputee users. These effects stem from physiological and structural changes that undermine the assumptions of conventional feature-based or decomposition- driven decoding methods.

This work presents a hybrid CNN-Transformer framework that directly processes multichannel sEMG signals without explicit motor unit decomposition. Convolutional layers extract local spatiotemporal features, while Transformer modules model long-range temporal and inter-channel dependencies, improving robustness to variability induced by force, posture, and fatigue.

Experimental results show stable multiclass performance with accuracies above 90% under both subject-dependent and subject- independent evaluations. Compared to decomposition-based approaches, the proposed method achieves comparable accuracy with lower computational complexity, supporting its suitability for scalable and near real-time myoelectric control

Keywords:Heavy metal; Physicochemical technology; Phytoremediation; Future scope