|
Title |
Long-Term Performance Prediction of Concrete Bridge Decks through Machine Learning and Survival Analysis Approaches
|
|
Authors |
김우혁(Woo-Hyeok Kim) ; 양지수(Ji-Su Yang) ; 민근형(Geun-Hyeong Min) ; 김우석(Woo-Seok Kim) |
|
DOI |
https://doi.org/10.4334/JKCI.2026.38.2.147 |
|
Keywords |
기계학습; 교량 바닥판 열화; 결함도 지수; 생존 분석 machine learning; deterioration; defect index; survival analysis |
|
Abstract |
Our goal was to predict the long-term performance of concrete bridge decks in Korea by integrating inspection records from the Facility Management System (FMS) with regional climatic data. Based on a dataset of 5,353 bridges, several preprocessing procedures were applied, including imbalance correction (SMOTE-Tomek), categorical variable transformation (One-Hot Encoding), and normalization (StandardScaler). Subsequently, four long-term prediction models were developed: Weibull, CoxPH, Lasso?Cox, and PINN-Cox. Model performance was evaluated using the Concordance Index (C-index), Integrated Brier Score (IBS), Integrated Mean Absolute Error (IMAE), and Royston?Sauerbrei’s . Among the models, the Weibull model achieved the best performance, followed by CoxPH, Lasso?Cox, and PINN-Cox. Furthermore, by converting the survival function into a cumulative distribution function (CDF) and applying monotonic regression to the average defect index (DI) with respect to service years, we propose a procedure to align probability-based predictions with actual deterioration indices. The results confirm the feasibility of estimating the long-term performance of individual bridges by applying parallel shifts to the average deterioration curve, thereby reflecting information from inspection histories.
|