Real-Time Hip Joint Torque Estimation Using CNN–LSTM–Attention and IMUs for Exoskeleton Assistance

Abstract

To overcome the limitations of traditional control strategies that depend on explicit gait event detection or complex biomechanical models, this work introduces a deep learning framework for real-time torque estimation in a hip exoskeleton. The core of this method is a data-driven convolutional neural network (CNN)–LSTM–attention hybrid network that directly regresses standardized biomechanical torque profiles from kinematic data collected by wearable inertial measurement units (IMUs). We generated the supervisory torque labels for training by performing an inverse dynamics analysis using opensim on a public dataset. This approach enables the model to learn a standardized, template-based assistive strategy rather than subject-specific biological joint torques. The method’s effectiveness was validated on five healthy subjects walking at three speeds (2.7, 4.5, and 6.3 km/h). Quantitative analysis shows the model’s prediction accuracy (R2 = 0.9209) is significantly superior to baseline models. Our assistive strategy yielded significant physiological benefits, lowering the wearer’s net metabolic cost by 17.5% on average. Although the results validate the potential of this data-driven framework, its reliance on steady-state laboratory data for training remains a limitation.