Analyzing Risk of Service Failures in Heavy Haul Rail Lines: A Hybrid Approach for Imbalanced Data

An incident in which a rail defect of size over a threshold value is noticed and the track is taken out of service is known as a service failure. This article aims at building accurate prediction models with binary outcome for risk of service failures on heavy haul rail segments. An analysis of the factors that influence the risk of a service failure is conducted and quantitative models are developed to predict locations where service failures are most likely to occur until the next inspection. To this end, data are collected from a Class I U.S. Railroads for six years from 2011 to 2016. Four prediction models (i.e., logistic regression, decision tree, multilayer perceptron, and gradient boosting classifier) are implemented and their results are compared. To account for the imbalanced classes between the normal operation and service failure, two treatments have been used including undersampling and oversampling. To improve the model performance, the parameters of each method are tuned using random search hyperparameter optimization. Later, bootstrap aggregation (or bagging) is incorporated into each method. The findings of the study show that the prediction performance is the highest when using bagging and oversampling as treatments with gradient boosting method. It was also identified that gross tonnage, presence of geometry defects, ambient temperature, segment length, and rail defect presence are the most important factors for predicting the risk of service failures. The results of this study are useful for railroads to develop effective strategies for rail inspections, preventive maintenance, and capital planning.