Approach to Assist in the Discovery of Railway Accident Scenarios Based on Supervised Learning

Abstract

The European Community has developed a real turning point in the common rail transport policy by defining new ambitions to rebalance sustainably the sharing between modes of transport, develop intermodality, fight congestion, and finally place safety at the heart of European action. To consolidate the usual methods of railway safety analysis, this chapter proposes two complementary railway safety assessment methods based on AI techniques and in particular on machine learning (ML). The study seeks to exploit, by machine learning, the lessons resulting from “Experience Feedback” (REX) in order to help and assist safety experts, technical investigators, and certification bodies to assess the level of safety of a new rail transport system. Unfortunately, safety in rail transport improves essentially on the basis of in-depth knowledge of accidents and incidents resulting from “experience feedback”. As stipulated by European regulations, all players in rail transport and in particular infrastructure managers and railway undertakings are obliged to set up a system of “experience feedback” in order to understand the causes and the seriousness of the consequences engendered by rail accidents and incidents. So, the knowledge of accidents and incidents results essentially from the contribution of lessons learned and experiences acquired. To explain and understand the causes and circumstances of accident risks and therefore at least avoid the reproduction of similar accidents, we have oriented our study toward the use of approaches derived from AI and machine learning. From experience feedback, the main objective is to exploit a set of insecurity events in order to anticipate and prevent the reproduction of the risks of accidents or similar incidents and possibly to discover and identify new scenarios of potential accidents liable to jeopardize safety. This chapter proposes a new hybrid method based on three machine learning algorithms. The first stage of acquiring knowledge led to the development of two accident scenario databases. The first base relates to the analysis of “functional safety”, and the second base relates to the analysis of the “security of critical software”. The second step, which is based on a concept classification algorithm, makes it possible to group the accident scenarios into coherent classes such as the class relating to train collision or derailment problems. For each class of accident or incident scenarios, the third step implements a learning technique based on production rules in order to identify some relevant safety rules. In the fourth step, the previously generated production rules are transferred to an expert system in order to deduce the potential accident risks. Finally, a case-based reasoning (CBR) system makes it possible to search, on the basis of “experience feedback”, for the cases closest to this new risk of accident and proposes the most appropriate prevention or protection measures.