Bridge Failure Risk Prediction Using Geospatial Data Processing via Multi-Head Attention Deep Learning Model

Bridge failures are a significant threat to infrastructure safety and public security, which demand cost-effective and scalable monitoring systems. This work proposes a novel data-driven framework for early warning of bridge collapse risk, leveraging Interferometric Synthetic Aperture Radar (InSAR) displacement time series and Deep Learning. A mathematical formulation of a bridge-collapse risk index is introduced, allowing quantitative estimation of the probability of failure from displacement data. To overcome data scarcity, a synthetic bridge dataset is generated through a combination of geometrical transformations and stochastic perturbations applied to real InSAR observations. Then, a Multi-Head Attention-based Neural Network is trained to predict the risk increment over time windows, using both on-bridge and surrounding geospatial points as input. The model is validated on the historical collapse of the Tadcaster bridge and stress tested on the Cantiano bridge, which failed during the 2022 Marche flood in Italy. The results show that the proposed approach effectively captures the temporal evolution of structural instability, with a conservative (risk-overestimating) but consistent prediction trend. These findings demonstrate the potential of combining InSAR data and attention-based Deep Learning models for scalable, non-invasive bridge health monitoring.