AFGLFNet: A novel multi-sensor fusion approach for predictive human lower-limb locomotion intention

Accurate prediction of lower-limb locomotions is pivotal for wearable rehabilitation devices, yet remains challenging due to the complexity of movement scenarios and sensor susceptibility to noise. For tasks requiring high responsiveness, algorithms must identify the intended motion before its onset to furnish control decisions in advance. To address this need, we propose a multi-sensor fusion approach for precise, real-time prediction of lower-limb motion patterns. The system employs a low-cost, lightweight sensor suite mounted on the lower limb of one side, integrating inertial measurement units (IMUs) and surface electromyography (sEMG). Building on these inputs, we develop an end-to-end Adaptive Frequency and Global–Local Fusion Network (AFGLFNet) designed to extract integrated features across temporal, spatial, and frequency dimensions from the fused signals. The model supports accurate and robust continuous prediction of seven common locomotion modes across prediction horizons from 150 to 500 ms. Experiments show that, when forecasting 250 ms ahead, AFGLFNet attains average recognition accuracies of 80.07% for sEMG-only, 92.05% for IMU-only, and 93.74% for sEMG+IMU, substantially outperforming five benchmark models. As noise levels increase, accuracy declines slowly, indicating strong anti-interference robustness. These findings substantiate the effectiveness of multi-sensor fusion and AFGLFNet for lower-limb motion prediction, providing insights into future applications in real-time control of lower-limb rehabilitation robot.