Synthesizing Ontology and Graph Neural Network to Unveil the Implicit Rules for US Bridge Preservation Decisions

Abstract

Bridges are essential portions of a nation’s infrastructure systems. Although general rules and guidelines are available for bridge preservation activity prediction, due to the intricate interdependencies among bridge elements, defects, and preservation activities, departments of transportation (DOTs) rely heavily on bridge engineers’ experience to determine preservation needs. Hence, identifying and organizing the unwritten and experience-based domain knowledge is essential to automate bridge preservation planning. This research collected 13,994 defects for 442 bridges in North Carolina. A graph neural network (GNN) model was developed to predict preservation activities using a defect dependency graph. This research created a bridge preservation ontology to further leverage experience-based domain knowledge to derive 80 unwritten activity-triggering rules via ontology axioms. A heterogeneous graph was constructed considering the semantics related to bridge defects and elements from the axioms. Tests revealed that, with a domain ontology, the GNN model improved prediction accuracy, precision, recall, and F1 score by 4.78%, 4.03%, 15.03%, and 11.62%, respectively. This research contributes to the body of knowledge by proposing a new graph theory–based bridge inspection database to enable GNN learning considering spatial and logical dependencies. Construction practitioners can instantly access and clearly comprehend bridge maintenance contextual information using the ontology database and machine learning models. The framework provides a systematic model for bridge preservation activity planning and enhances the robustness and reliability of bridge preservation decision-making. This research will assist DOT engineers and managers in improving knowledge sharing and automatic planning in bridge management.