Physics-Informed Deep Transfer Learning for sEMG-Based Multiple Joint Angle and Torque Estimation

Accurately estimating joint angles and torques is essential for human-machine interfaces, such as exoskeleton control and human-robot collaboration. Musculoskeletal models driven by surface electromyographic (sEMG) signals incorporate human physiological knowledge but are computationally complex, whereas data-driven methods offer faster inference but lack interpretability. To bridge this gap, a physics-informed deep transfer learning framework is proposed in this paper, which integrates the physical knowledge as a soft constraint into the data-driven method. Additionally, a novel neural network called CNN-Mamba is proposed to implement the framework. It combines the local feature extraction capability of convolutional neural network (CNN) with the global modeling ability of the state space model (SSM) known as Mamba. To demonstrate its capability to simultaneously estimate the angles and torques for multiple joints, the method was validated on two datasets: wrist-hand coupled motion (10 subjects, a total of 30 trials) and walking motion (8 subjects, a total of 32 trials). The average Pearson correlation coefficient between estimated values and ground truth for all joints exceeded 0.96, reaching a maximum of 0.99. Ablation and comparative experiments further confirmed the advantages of the proposed method. Moreover, our approach demonstrates comparable performance to model-based methods while achieving at least 14× faster inference speed. Specifically, estimating the joint angle and torque for a motion duration of 1 second takes only 161.8 ms. Overall, our innovative approach provides an effective solution for estimating multi-physiological parameters across multiple joints in human motion, with promising applications in human-machine interaction.