Predicting Highway–Rail Grade Crossing Collision Risk by Neural Network Systems

Abstract

In this paper, train-vehicle crash risk at highway–rail grade crossings (HRGCs) is analyzed with a neural network (NN) model to return meaningful rankings of crash-contributory-variable importance based on different criteria, but also to produce dependent nonlinear contributor-crash curves with all other contributors considered for a specific contributor variable. Historical crash data for North Dakota public HRGCs from 1996 to 2014 were used for the study. Several principal findings were observed: (1) 22 input variables describing traffic characteristics and crossing characteristics are related to crashes at public HRGCs; (2) a mean-square error–based NN model and a connection weights–based NN model represent two relative contributory-variable importance lists for different application purposes; (3) the effect of different variables on crash likelihood is different when all other contributors are set at different levels, and the relationship between contributors and crash likelihood is dynamic nonlinear; and (4) in predictive and explanatory power, the neural network model outperforms the decision tree approach for the considered case study.