Interpretable accident prediction at highway-rail grade crossings: a deep learning approach

Abstract

Accidents at highway-rail grade crossings (HRGCs) pose significant risks to life and property, leading to substantial losses each year in the United States. Accurate and interpretable accident prediction provides a viable solution for improving the safety of HRGCs. Although encouraging processes have been achieved, existing studies either exhibit insufficient predictive performance or lack inherent interpretability, hindering efforts to further enhance the safety of HRGCs. To fill this gap, a well-designed deep learning model for accurate and interpretable accident prediction at HRGCs is proposed in this study. First, a word embedding approach is employed to generate vector representations of the category characteristics of HRGCs, effectively capturing the semantic information inherent in these characteristics. Second, the attention mechanism is used to separately aggregate the category characteristics and numerical characteristics, which can dynamically identify the key contributing characteristics of the accidents at HRGCs. The HRGCs data from Louisiana, Texas, and Washington were employed for a comparative analysis with the baseline model, demonstrating and validating the superiority and practicality of the proposed deep learning model. Finally, an interpretive analysis of the prediction process and prediction results of the proposed deep learning model is conducted. Eventually, this study explores the causative factors of accidents at HRGCs in a data-driven manner, providing valuable insights for further improving the safety performance of HRGCs.