Transformer-based digital twin modeling method for stress fields of gantry cranes

To meet the real-time structural safety monitoring requirements of hydropower-station gantry cranes, this study proposes a stress-field reconstruction method that integrates feature-point correlation analysis with a Transformer-based global attention mechanism. Using finite-element simulation data, correlations among 711,594 grid elements are computed to select representative measurement points and construct a feature subset, enabling effective input dimensionality reduction. The Transformer is then employed to capture the coupling relationships between the grids and the feature subset, allowing the full stress field to be reconstructed in real time from only a small number of measurement points. After training on 1,903 working-condition samples, the model achieves highly consistent results with finite-element simulations across multiple conditions, with an overall R² exceeding 0.999. The stress solution errors of the first 95% of finite elements are within [-7.10%, 7.10%].When fine-tuned using real monitoring data, the model freezes the base network and introduces a lightweight MLP calibration layer, enabling effective adaptation to measured signals while maintaining deviations within 5% FS on measurement points not used for calibration. This method exhibits strong reconstruction performance even under sensor-limited engineering conditions and provides essential technical support for ultimate strength assessment and fatigue-life prediction of hydropower-station gantry crane structures.