Knowledge Graph-Based Structural Safety Risk Diagnosis and Control of Prestressed Concrete Bridges

Reliable structural safety risk diagnosis and control of prestressed concrete bridges is essential for safe operation. Unfortunately, current practice relies heavily on inspectors’ engineering knowledge, field experience, and subjective judgments. In addition, existing general-purpose large language models (LLMs) often underperform in bridge defect diagnosis because of missing domain terminology and hallucinated technical statements. Therefore, there is a need to establish a trustworthy and explainable method for structural safety risk diagnosis and control. This study develops a domain knowledge-graph-enhanced framework, the prestressed concrete bridge defect knowledge-graph-enhanced LLM (PCBDK-LLM), to support defect diagnosis and treatment recommendations for prestressed concrete bridges. First, a prestressed concrete bridge defect knowledge graph is constructed using a hybrid approach that combines direct text-driven extraction from standards, peer-reviewed literature, and inspection reports with an ontology-based schema and consistency axioms. Then, the authors propose a retrieval module (REM) that performs ontology-aware chunking and hybrid similarity search to ground a locally deployed dialogue model (DeepSeek-R1) on verified domain knowledge. Eight real rehabilitation cases (eight bridges) are used to evaluate model recommendations against a reference solution documented in the project materials. Results indicate that the proposed PCBDK-LLM generates treatment suggestions that are more consistent with the reference plan than the baseline LLM and the ablation variants.