TimesNet for Real-Time Gait Phase Estimation in Exoskeleton Rehabilitation

Real-time gait phase estimation is essential for synchronized control of lower limb exoskeletons in stroke and spinal cord injury rehabilitation. Accurate phase estimation enables properly timed assistive torque, directly impacting patient safety and rehabilitation outcomes. The control loop operates at ۲۰۰ Hz, requiring inference below ۵ ms for clinical deployment. This paper compares ۱۵ deep learning architectures for gait phase prediction using hip joint angles. Evaluated models include LSTM, BiLSTM, TCN, SSM, Dense, and TimesNet. Models were trained on data from ۸ healthy participants walking under multiple conditions (normal, force-applied, simulated abnormal gaits), totaling ۵۲ motion samples. Performance was evaluated using MSE, RMSE, MAE, parameter count, and inference time. Results show two-layer BiLSTM with polar output achieves the highest accuracy (MSE of ۵.۸۵ right, ۷.۸۶ left), but its ۶.۳۸ ms inference time exceeds the ۵ ms constraint, making it unsuitable for safe patient use. In contrast, TimesNet achieves an optimal clinical trade-off with MSE of ۷.۰۷ (right) and ۹.۵۷ (left), RMSE of ۲.۶۶° (right) and ۳.۰۹° (left), and inference time of ۴.۳۳ ms, fully satisfying real-time requirements. TimesNet provides the best balance between accuracy and efficiency for safe, effective exoskeleton-based gait rehabilitation.

کلیدواژه ها:

Gait Phase Estimation ، TimesNet ، LSTM ، BiLSTM ، TCN ، Lower Limb Exoskeleton ، Real-Time Inference ، Deep Learning ، Time Series Analysis