Knee Joint Angle Estimation Based on EMG-FMG Fusion Temporal Convolutional Network

Abstract

Accurate perception of human motion intent and continuous joint kinematics estimation are essential for the control of lower-limb rehabilitation exoskeletons. To address the limitations of unnatural human-machine interaction in lower-limb exoskeletons and the vulnerability of surface electromyography (EMG) to sweat, fatigue, and environmental noise, this paper proposes a continuous knee joint angle prediction framework based on the bimodal fusion of EMG and force myography (FMG). Leveraging a self-designed flexible bimodal sensor, this work achieves hardware-level synchronized acquisition of neural drive signals and physical muscle deformation signals at the same anatomical location, effectively eliminating temporal drift between heterogeneous signals. A bimodal prediction model integrating a dual-level attention mechanism with a Temporal Convolutional Network (TCN) is developed. Specifically, a Local Cross-modal Attention (LCA) module is first employed to dynamically recalibrate EMG amplitudes in real-time using stable FMG features. Subsequently, a Global Spatial Self-Attention (GSA) mechanism quantifies the spatial synergistic weights across different muscle groups. Finally, a TCN architecture with dilated convolutions captures long-range temporal dependencies within gait cycles to achieve precise mapping of knee joint angles. Experimental results demonstrate that the proposed EMG-FMG_TCN model exhibits superior performance in knee joint tracking, achieving a coefficient of determination ( R2 ) of 0.984 and a root mean square error (RMSE) as low as 2.43°. Compared with the EMG-only unimodal model, the Bimodal fusion approach reduces the prediction error by 41.7%. Furthermore, the proposed model significantly outperforms conventional LSTM and CNN architectures in terms of dynamic response and tracking smoothness, while demonstrating superior accuracy and real-time efficiency compared to state-of-the-art models such as IDB-EMGNet and Conv-LSTM. This study provides a robust solution for joint angle estimation in lower-limb rehabilitation exoskeletons, showing significant promise for clinical applications.