Contributions and limitations of AI and machine learning in railway operations, maintenance, and safety: A literature review

This study provides a non-exhaustive overview of AI methods, algorithms, and applications, particularly machine learning (ML) in rail transport. Drawing inspiration from the two European directives relating to the development (Directive 2012/34/EU) and interoperability (Directive (EU) 2016/797) of the railway system, this review proposes to classify AI applications according to (1) the “structural elements†of the railway system (infrastructure, rolling stock, energy, control and command, and signaling), (2) the “functional elements†(operation, maintenance, and telematics applications), and (3) both “structural and functional†elements with a view to identifying AI approaches aimed at improving railway safety, in particular the analysis of railway accidents and incidents based on investigation reports. Several “classic†AI techniques are implemented, including ML (supervised, semi-supervised, unsupervised), deep learning such as artificial neural networks (ANN), natural language processing (NLP), case-based reasoning (CBR), etc. However, the lack of interoperability between “classic†AI tools and the inadequacy of these approaches to capitalize, share, and reuse this knowledge have driven research toward the development of new approaches based on ontology’s and knowledge graphs. A study of all AI applications shows that the stages of acquisition, retrieval, analysis, structuring, formalization, modeling, processing, and interpretation of the data produced pose a crucial problem in the rail transport sector. Moreover, with complex models called “black boxes,†it is difficult to understand how and why the internal reasoning mechanisms of the AI system impact the solution and predictions. The new approach of explainable AI (XAI) or explainable machine learning (XML) can provide an element of response to this problem, particularly when it comes to a crucial issue such as railway safety.