Using Machine Learning Methods for Modeling Freight Train Derailment Severity

This paper focuses on identifying factors affecting the severity of freight train derailment. To examine the train derailment, it is necessary to study the point of derailment and the number of cars derailing. Previous studies have used truncated geometric distributions with two key assumptions: (1) cars in a train get involved in a derailment independently of one another, and (2) probabilities of cars involved in derailments are all the same along the train length. The underlying assumptions are clearly violated in the real world. Therefore, in this study, different classification approaches, including decision tree, random forest, support vector machine, and AdaBoost techniques, have been used to avoid fixed assumptions. The results show that the decision tree is the best classifier to predict the severity of train derailment for the US accident database, and the two-level severity scenario (one car derailed or more) presents better results to classify derailment severity. The research also shows that freight train derailment severity has been affected mainly by (1) train speed, (2) cause of the accident, and (3) train weight-to-train length ratio. Among these features, cause of accident is the most important feature in classifying accident severity; also, the causes of one-car derailments are mostly related to mechanical and electrical failures. In mechanical and electrical failure, train speed plays a significant role in determining the severity of accidents. The factor of train weight to length comes into account when an accident's cause is related to human factors.