Deep Graph Learning for Spatial-Temporal Group-Correlated In-Service Performance Prediction of Transportation Infrastructure System

The rapid advancement of urbanization has led to increasing attention on the in-service performance prediction of transportation infrastructure systems, including bridges, tunnels, and roads. Traditional data-driven methods for structural performance prediction often separate the spatial topology from the temporal evolution, which fails to fully utilize spatial–temporal group-correlations among individual structures inside an infrastructure system or different components within a single structure. Therefore, it is difficult to effectively capture complex dependencies of multiple spatial nodes and long-range time series, which leads to limited prediction accuracy and efficiency. To address the above challenge, deep graph learning has emerged as a powerful tool for capturing complex spatial–temporal dependencies embedded inside the monitoring data of transportation infrastructure systems. This article systematically summarizes recent advances in spatial–temporal deep graph learning methodologies and the corresponding applications on in-service performance prediction of transportation infrastructure systems. First, a unified mathematical framework of spatial–temporal graph learning is abstracted and formulated to establish a standardized pipeline of multivariate group-correlation prediction tasks for transportation infrastructure systems. Then, a taxonomy of spatial–temporal graph modeling approaches is summarized according to intrinsic learning mechanisms, including topology-based, distance-based, similarity-based, mobility-based and adaptive methods. Afterwards, deep graph learning for group correlation modeling and spatial–temporal prediction is mainly categorized into STGRNNs, STGCNNs and STGSATs, according to distinct modelling strategies of intrinsic spatial–temporal evolutive relationships. Furthermore, various applications for in-service performance prediction under real-world scenes are reviewed, including a global level of transportation infrastructure system and a local level of individual structure. Finally, future directions are prospected to address challenges of low-quality data, dynamic graph modeling, cross-scene transferability, and lightweight deployment.