Title:Fast Hip Joint Moment Estimation with A General Moment Feature Generation Method

Abstract:The hip joint moment during walking is a crucial basis for hip exoskeleton control. Compared to generating assistive torque profiles based on gait estimation, estimating hip joint moment directly using hip joint angles offers advantages such as simplified sensing and adaptability to variable walking speeds. Existing methods that directly estimate moment from hip joint angles are mainly used for offline biomechanical estimation. However, they suffer from long computation time and lack of personalization, rendering them unsuitable for personalized control of hip exoskeletons. To address these challenges, this paper proposes a fast hip joint moment estimation method based on generalized moment features (GMF). The method first employs a GMF generator to learn a feature representation of joint moment, namely the proposed GMF, which is independent of individual differences. Subsequently, a GRU-based neural network with fast computational performance is trained to learn the mapping from the joint kinematics to the GMF. Finally, the predicted GMF is decoded into the joint moment with a GMF decoder. The joint estimation model is trained and tested on a dataset comprising 20 subjects under 28 walking speed conditions. Results show that the proposed method achieves a root mean square error of 0.1180 $\pm$ 0.0021 Nm/kg for subjects in test dataset, and the computation time per estimation using the employed GRU-based estimator is 1.3420 $\pm$ 0.0031 ms, significantly faster than mainstream neural network architectures, while maintaining comparable network accuracy. These promising results demonstrate that the proposed method enhances the accuracy and computational speed of joint moment estimation neural networks, with potential for guiding exoskeleton control.