Continuous knee joint angle prediction for exoskeleton applications using sEMG–IMU fusion and PGQNet

Continuous and accurate prediction of lower-limb joint motion is important for natural assistance and stable human–robot interaction in lower-limb exoskeletons. However, exoskeleton assistance alters neuromuscular activation and limb kinematics, complicating continuous knee joint angle prediction. This study proposes a Phase-Guided Query Network (PGQNet) for future knee joint angle prediction based on surface electromyography (sEMG) and inertial measurement unit (IMU) fusion. PGQNet uses dual-branch temporal convolutional networks to encode modality-specific features and constructs phase-conditioned queries by combining time-varying gait-phase descriptors with modality identity embeddings. These queries perform intra-modal temporal readout and feature reinjection before multimodal fusion and bidirectional long short-term memory regression. Experiments involving ten participants were conducted during low-speed level walking with and without exoskeleton assistance at prediction horizons of 50–200 ms. The bimodal configuration consistently outperformed sEMG-only input. Under exoskeleton assistance, its whole-cycle RMSE was reduced by 31.0%–36.9% relative to sEMG-only input and by 6.3%–8.9% relative to IMU-only input. PGQNet also achieved the best mean RMSE, mean absolute error, and correlation among the compared models under both conditions. Although whole-cycle performance was comparable between conditions, assistance reduced swing-phase RMSE but increased support-phase RMSE across all prediction horizons. Toe-off was the most challenging gait region, with RMSE approximately 3.0–3.2 times that around heel strike without the exoskeleton and 2.5–2.6 times with assistance. Additional experiments demonstrated partial transferability to unseen participants, low average sensitivity to gait-phase perturbations, and software-level feasibility for quasi-online implementation.