Physics-guided cross-scale temporal attention network for modeling and prediction of multi-axle heavy-duty vehicle dynamics

• A physics-guided cross-scale temporal attention network is proposed to address multi-scale coupling and strong nonlinearity in multi-axle heavy-duty vehicle dynamics modeling. • A hybrid BiLSTM-Transformer architecture with relative position encoding and multi-scale convolutional attention captures both local inter-axle dynamics and global coordination patterns. • A 14-DOF coupled dynamics model is embedded through hierarchical loss functions, ensuring physical consistency while maintaining data-driven learning capabilities. • The proposed method achieves superior prediction accuracy with RMSE of 0.0223 for longitudinal velocity and 0.0078 for yaw rate, significantly outperforming pure data-driven and standard PINN methods. • Comprehensive ablation studies validate the necessity of each architectural component, demonstrating the framework's effectiveness for multi-axle vehicle dynamics prediction in real-time applications. Multi-axle heavy-duty vehicles exhibit complex nonlinear dynamics and multi-scale temporal coupling that challenge real-time prediction. This paper proposes a Physics-Guided Cross-Scale Temporal Attention Network for five-axle vehicle state prediction. The framework establishes a 14-DOF coupled dynamics model and integrates BiLSTM to capture sequential inter-axle dependencies, Transformer with relative position encoding for global coordination patterns, and multi-scale convolutional attention (kernel sizes 3, 5, 7) to extract high-, mid-, and low-frequency temporal features corresponding to tire dynamics, suspension responses, and vehicle motion coordination. Adaptive gating fusion dynamically balances local and global temporal patterns, while hierarchical physics constraints ensure physically consistent predictions. Experiments demonstrate superior accuracy with RMSE values of 0.0223 (longitudinal velocity), 0.0075 (lateral velocity), 0.0396 (longitudinal acceleration), 0.0129 (lateral acceleration), 0.0063 (steering angle), and 0.0078 (yaw rate), significantly outperforming pure data-driven and standard PINN methods. Ablation studies confirm that all architectural components are essential, validating the cross-scale temporal attention design for multi-axle vehicle dynamics modeling.