Accurate assessment technology of BIM bridge structure status using spatiotemporal transformation graph neural network

Traditional bridge structural condition assessments struggle to integrate the physical topological information of BIM models with the spatiotemporal evolution of long-term monitoring data, resulting in low assessment accuracy and a lack of damage localization capabilities. To address this, this paper proposes a precise assessment framework based on a spatiotemporal transformer graph neural network (ST-Transformer GNN). Component attributes and connectivity relationships are automatically extracted from the IFC (Industry Foundation Classes) model of BIM (Building Information Modeling), constructing a dynamic graph structure with physical weights. Multi-source sensor data is mapped to corresponding nodes based on spatial location, forming a temporal input sequence. A physics-guided spatiotemporal attention module based on the Transformer architecture is then designed. This module, through a stiffness-constrained spatial propagation mechanism and temporal dynamic weighting, enables joint modeling of inter-component load responses. Finally, by combining multi-task learning with gradient interpretability analysis, component-level health status is output, and a damage heatmap is generated. Experimental results demonstrate that this method achieves an average component-level health status classification accuracy of 96.3%, and achieves a peak error of 0.8με in temporal response prediction, significantly outperforming competing models such as ST-GCN and DCRNN. The conclusions suggest that the deep integration of BIM and the physically enhanced spatiotemporal transform graph neural network (ST-Transformer GNN) can effectively enhance the intelligence level of bridge condition assessment.