Railway operational hazard prediction and control based on knowledge graph embedding and topological analysis

Abstract

Railway operational accidents usually result from the domino effects of a series of interrelated hazards. Predicting and controlling potential hazards in advance are valuable for ensuring safe railway operations. A variety of hazards form a heterogeneous hazard relationship network because of their complex interactions. The potential hazards can be predicted and controlled by use of such a relationship network structure. In this paper, a new knowledge graph-based hazard prediction and control approach is proposed, aiming to prevent railway operational accidents using the relationship network of hazards. Its originality is to leverage knowledge graph embedding and topological analysis to predict and control hazards, by means of both a novel convolutional architecture on hyperplanes and some tailored topological indicators. The outcomes of the proposed approach can offer railway operators the decision basis of accident prevention, in the form of potential hazards and their corresponding control measures. An application to the UK's railway accident data shows that 13.25 % and 4.38 % of hazard prediction accuracy gains in Hit@3 and Hit@10 evaluation metrics are respectively achieved by the proposed method over the best baseline methods. Furthermore, it also demonstrates the effectiveness of the proposed method in formulating targeted hazard control measures.