Contrastive Candidate-Proposal Graph Neural Network (CCP-GNN) for Automated Full-Network State Estimation in Civil Infrastructure Systems

Civil infrastructure systems are only sparsely monitored for state observations. Automatically estimating full-network, link-level states is challenging because, when inferring unobserved states from monitored links, information does not transfer reliably over long, heterogeneous network paths and not all monitored links serve as a reliable reference for an unmonitored link. This paper proposes a contrastive candidate-proposal graph neural network (CCP-GNN) method. It includes (1) context-aware, state-sensitive representation learning for reliable information transfer between monitored and unmonitored links, and (2) quality-aware relational inference to select high-quality monitored links as references per unmonitored link for reliable state estimation. The performance of CCP-GNN was evaluated on roadway systems from three cities – New York City, the City of Detroit, and the Town of Hempstead – representing large, medium, and small scales. Across the systems, the method achieved consistent accuracy, with MAPEs of 2.37%, 13.27%, and 11.51%, respectively, on testing links; and it also maintained balanced PICP and MPIW, indicating reliable and controlled uncertainty performance. Overall, the results show that CCP-GNN provides accurate and reliable state estimation, with effective generalizability across scales. The proposed CCP-GNN method offers full-network, link-level state estimates to support informed operation and maintenance decision-making in sparsely instrumented infrastructure systems.