Causative analysis of freight railway accident in specific scenes using a data-driven Bayesian network

Abstract

As the freight railway system is a typical complex system, freight railway accidents have various and complex accident scenes. A Data-Driven Bayesian Network (DDBN) with random variables representing scene elements, accident causes, and accident consequences as nodes, and conditional dependencies between nodes as edges is proposed to identify the most significant accident causes in various and complex specific accident scenes. First, an unsupervised-supervised method is designed to define the states of nodes in the DDBN, considering the characteristics of the data involving both continuous and discrete states. Second, a greedy algorithm is proposed to mine the causal sequence between nodes, and the direction of edges in DDBN is established accordingly. Then, an NB-K2-MLE approach is proposed to generate the structure and parameters of the DDBN from data. Finally, a risk calculation function based on DDBN is proposed to calculate the risk of various accident causes in specific scenes. In empirical analysis based on real accident data, the inference accuracy of DDBN reached 87.92 %, with precision and recall exceeding 70 %. More importantly, the research results indicate that the distribution of accident causes will concentrate with the refinement and specific of the scene, and the main accident causes at the macro level cannot be fully applicable to accident prevention in specific scenes. The DDBN constructed in this study can provide data support for the determination of the significant accident causes and the development of targeted accident prevention strategies in various and complex specific scenes.