Predicting urban rail transit safety via artificial neural networks

• Published in: Safety science Vol. 167; p. 106282
• Main Authors: Awad, Farah A., Graham, Daniel J., Singh, Ramandeep, AitBihiOuali, Laila
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2023
• Subjects: Artificial neural networks; Rail injuries; Safety benchmarking; Safety performance; Urban rail safety; Safety benchmarking; Rail injuries; Urban rail safety; Artificial neural networks; Safety performance
• ISSN: 0925-7535; 1879-1042
• DOI: 10.1016/j.ssci.2023.106282

•This study presents the first macro-level cross-country analysis of urban rail transit safety.•The study utilizes an international dataset from 31 systems to analyze the operational safety of urban rail.•This study applies supervised machine learning models which consider the socio-technical nature and non-linear interactions.•This study provides a benchmarking model for urban rail transit safety outputs which accommodates heterogeneity. This paper studies the operational safety of urban rail transit (URT) systems through Artificial Neural Networks. While recent safety literature adopting systematic models of analysis consider the complexity of URT operations, they focus on single systems or single components of the operational process. Our study contributes to the URT safety literature by having a macro perspective, while considering that such complex socio-technical systems involve multiple non-linear interactions among their components. To our knowledge, we present the first cross-country analysis of URT safety through machine learning models in the literature, using a unique international dataset from 31 URT systems which comprises annual system-level data. Two models are estimated to predict the annual URT injuries. The first model includes safety-related incidents as inputs, while the second includes operational characteristics of the system. Additionally, a closed-form formula is presented to predict the annual number of injuries based on operational features of the URT system along with an illustrative example to demonstrate benchmarking applications. The results are promising and indicate good generalizability. The models proposed in this study could be useful for operators and policy makers as they aid in prioritizing improvements, predicting future safety performance based on changes in operational features, and as a benchmarking tool.