Deep parallel spatial-temporal cross-attention network for continuous joint angle prediction using sEMG signals across multiple terrains

Accurate decoding of user motion intention from surface electromyography (sEMG) signals is a critical challenge in the control of lower limb exoskeletons, particularly under non-stationary conditions and across complex terrains. Existing deep learning methods often struggle to effectively capture multi-granular signal features and decouple complex spatiotemporal dependencies, limiting their robustness. To address these issues, this paper proposes a Deep Parallel Spatial-Temporal Cross-Attention Network (PST-CAN) for continuous prediction of lower limb joint angles. First, a Multi-Scale Xception (MS-Xception) encoding module is designed to adaptively extract cross-scale spatiotemporal features from raw sEMG signals using parallel depthwise separable convolutions. Second, a Cascaded Parallel Spatiotemporal Dual Cross-Attention (PST-DCA) mechanism is introduced. By employing a parallel dual-branch structure, this mechanism explicitly decouples and models spatial muscle synergies and temporal evolution patterns, mitigating information loss associated with traditional serial fusion strategies. Experimental validation on the multi-terrain ENABL3S dataset demonstrates that PST-CAN achieves an average MAE of 2.59 \(^{\circ }\) , an RMSE of 4.19 \(^{\circ }\) , an \(R^2\) of 0.961, and a CC of 0.980 under the revised chronological training/validation/testing protocol. The proposed framework achieves the best overall performance among the evaluated baseline models and maintains robust trajectory tracking and phase alignment across complex locomotion conditions.