FAST-TCN: A Frequency-Aware Spatio-Temporal Network for sEMG-Based Continuous Motion Prediction in Upper Limb Rehabilitation Robotics

Abstract

Practical sEMG-based motion prediction for rehabilitation robotics faces two critical challenges: low robustness caused by inter-subject variability and signal non-stationarity, and mechanical safety hazards due to jittery control signals. Conventional Temporal Convolutional Network(TCN)-based methods often prioritize temporal patterns while neglecting spectral characteristics and physical constraints. To address these, FAST-TCN is proposed, a Frequency-Aware Spatio-Temporal network based on a dual-stream architecture. Prediction robustness is enhanced by integrating multi-scale temporal convolutions with a spectral gating mechanism for dynamic noise suppression. Furthermore, domain shifts across individuals are mitigated via instance normalization, supplemented by a minimal one-shot linear calibration to ensure biomechanical consistency. Trajectory safety is guaranteed by incorporating a physics-informed smoothness constraint into the loss function. Evaluated on a 24-subject dataset, FAST-TCN achieves a Mean Absolute Error (MAE) of 10.58◦ and a coefficient of determination (R2) of 0.9068, outperforming standard TCN baselines by 16.8%. These results demonstrate a robust and smooth solution for continuous motion prediction in rehabilitation robotics.