PatchTST In-Sensor Computing for Real-Time Knee Joint Torque Prediction in Walking Based on sEMG

Knee joint torque prediction holds critical applications in rehabilitation medicine, human-machine interaction control, and assistive device development. However, data transmission delays and dependence on remote computation make it challenging for traditional methods to achieve high-precision, real-time knee joint torque prediction during walking. This paper proposes an in-sensor computation method based on surface electromyography (sEMG) and PatchTST, leveraging the model's multi-head self-attention mechanism and temporal feature extraction capabilities to capture subtle changes in muscle activation throughout the gait cycle. All signal acquisition, data processing, and pattern recognition occur within a device specifically designed for in-sensor computation. Experimental results show that this method achieves an RMSE of 0.032±0.08 Nm/kg (p=0.032), significantly outperforming traditional models like LSTM, CNN-LSTM and Transformer. Furthermore, the model consistently maintained high accuracy across different walking states for all participants. The proposed method demonstrates the potential for real-time, natural human-machine interaction in biomedical applications.