Data-driven methods to predict track degradation: A case study

Abstract

Predicting track degradation is extremely important due to its beneficial effects on increasing track safety and scheduling preventive maintenance. Logistic regression and Gradient Boosting Machine (GBM) were employed in this paper for predicting Yellow-tag Defect in the Next Inspection (YDNI), which is an indicator of maintenance requirement in the track. The geometry data of this paper were collected from passenger tracks with a length of 155 miles from 2011 to 2020. In addition to geometry data, Ground Penetrating Radar (GPR) data were used as additional explanatory variables to improve the performance of models. As there are a limited number of YDNI observations, Synthetic Minority Oversampling Technique (SMOTE) is employed to resolve the issue of imbalance dataset. Results indicate that the GBM outperforms logistic regression and can successfully predict YDNI with a very high sensitivity of 0.94. Parametric analysis indicated the direct correlation between the Ballast Fouling Index (BFI) and the probability of defect occurrence. Finally, the model can be integrated into a decision-making framework which generates predictions based on a given probability threshold for classification. Thus, the railway companies can select a specific threshold according to their maintenance budget and desired level of track safety for scheduling preventive maintenance interventions.