|
Title |
Prediction of Bridge Remaining Useful Life Using Inspection-Derived Deterioration Rates and Machine Learning
|
|
Authors |
오승철(Seung-Cheol Oh) ; 이동현(Dong-Hyun Lee) ; 배성호(Sung-Ho Bae) ; 이승진(Seung-Jin Lee) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.5.100 |
|
Keywords |
교량 결함도 점수; 교량 단위 열화율; 잔존수명; 기계학습; 교량 자산관리 Bridge Deficiency Index; Bridge-level deterioration rate; Remaining useful life; Machine learning; Bridge asset management |
|
Abstract |
This study proposes a bridge-level deterioration modeling framework using deterioration rates derived from the Bridge Deficiency Index (BDI). The analysis was conducted using 9,482 inspection records from 3,961 highway bridges in South Korea, of which 3,928 bridges with valid deterioration segments were used for modeling. The inspection records were concentrated in intermediate condition grades, with 91.4% classified as Grade B and 8.4% as Grade C. This distribution limits the direct prediction of BDI or condition grades, particularly in high-deficiency ranges near the D-grade threshold. To address this limitation, segment-wise BDI change rates were calculated from adjacent inspection records, and recovery segments with decreasing BDI values were excluded. The valid deterioration segments were then aggregated into bridge-level observed deterioration rates and used as the target variable for machine-learning regression models. Random Forest, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) were trained using service life, bridge geometry, and traffic-related variables. The models achieved R2 values of 0.741-0.765, Root Mean Squared Error (RMSE) values of 0.00335-0.00352/year, and Mean Absolute Error (MAE) values of 0.00213-0.00231/year.
Random Forest showed the lowest error and was adopted for the life estimation. The predicted deterioration rates were converted into remaining useful life and total service life estimates using the D-grade thresholds. The proposed framework provides a quantitative procedure for converting maintenance-affected inspection records into bridge-level indicators for network-level bridge management.
|