Machine learning based life prediction of rail tracks using environmental and operational factors

Abstract

This study presents an integrated method using machine learning approaches for rail track life prediction, including environmental and operational aspects. Comprehensive data were obtained by analyzing metro rail replacement and maintenance data spanning over two decades from South Korea. Multiple regression-based machine learning models were used in the proposed framework to forecast rail life, including categorical boosting (CATB), extreme gradient boosting (XGB), random forest, and decision trees. The average temperature, maintenance count, and passenger volume were revealed from the feature importance evaluated using permutation and Shapley value analyses after Bayesian optimization. The results demonstrate that the XGB and CATB models obtain a coefficient of determination of approximately 0.81 for the test set under actual conditions despite minimal outlier removal. Moreover, this research demonstrates useful applications by mapping forecasts to particular rail segments, thus enabling data-informed maintenance scheduling and proactive decision-making in asset management systems.