| Title |
Integrating Safety Grade Prediction and RAG-Based Inspection Records for a Two-Stage National Highway Bridge Maintenance Decision Support Framework |
| Authors |
김다인(Kim, Dain) ; 정수민(Jung, Sumin) ; 박민건(Park, Mingeon) ; 허태민(Heo, Taemin) |
| DOI |
https://doi.org/10.12652/Ksce.2026.46.4.0335 |
| Keywords |
교량 유지관리, 안전등급 예측, 랜덤 포레스트, Retrieval-Augmented Generation, 점검 이력, 의사결정지원 Bridge maintenance, Safety grade prediction, Random Forest, Retrieval-Augmented Generation, Inspection history, Decision support |
| Abstract |
With the accelerated aging of bridge infrastructure in South Korea, there is an increasing need for a system to rationally prioritize inspection, maintenance, and reinforcement under limited personnel and budgets. However, deterioration prediction models in existing Bridge Management Systems (BMS) have primarily been utilized as independent analysis modules. This reveals a limitation in fully integrating the “evidence provision” and “context-based decision support” required in the actual decision-making stages. To address this, this study presents the applicability of a two-stage integrated decision support framework. This framework combines a machine learning classification model to screen high-risk bridges by predicting safety grades, and a Retrieval-Augmented Generation(RAG)-based question-answering system to provide contextual evidence by retrieving and summarizing past inspection records. A comparative analysis of algorithms (Decision Tree, Random Forest, and XGBoost) using data from approximately 8,890 national highway bridges demonstrated that while XGBoost showed higher overall accuracy, Random Forest was selected as the final prediction model for its superior performance in minimizing the omission of C-grade bridges requiring urgent inspection. Furthermore, the developed RAG system assists administrators' judgment by summarizing past damage recurrences and recommended actions, alongside the current and future (2036) prediction results from the Random Forest model. Across the five query categories Evaluation results demonstrated high faithfulness (93.4 %). By seamlessly connecting prediction, evidence provision, and decision support into a single continuous workflow rather than mere deterioration prediction, this study is expected to contribute to the advancement and practical applicability of future BMS. |