Data Mining-Based Prediction of Structural Deterioration Using Infrastructure Inspection Records and Environmental Exposure Variables

Structural deterioration threatens infrastructure reliability and increases maintenance demands, creating a need for predictive approaches that integrate historical condition information with environmental exposure. This study developed a data mining framework for predicting next-inspection structural deterioration using longitudinal infrastructure inspection records, asset characteristics, maintenance history, operational loading, and environmental variables. The analytical dataset comprised 5,035 inspection observations from 600 assets. Asset-grouped partitioning and stratified grouped cross-validation were applied to prevent information leakage across repeated inspections. Six machine-learning algorithms were evaluated, followed by hyperparameter optimization, probability calibration, threshold selection, and independent testing. Gradient Boosting provided the strongest overall performance and achieved an independent-test receiver operating characteristic area under the curve of 0.8090, precision–recall area under the curve of 0.9494, and balanced accuracy of 0.7258. Calibration reduced the Brier score from 0.1872 to 0.1235, while the optimized threshold produced precision of 0.9630 and specificity of 0.9082. Permutation analysis identified current and previous condition ratings as the dominant predictors, jointly accounting for 96.55% of positive predictive importance. Environmental exposure variables contributed to smaller incremental information after inspection history was considered. These findings demonstrate that calibrated, interpretable data-mining models can support deterioration forecasting and risk-informed infrastructure maintenance prioritization.