Analysis and prediction of railway accident risks using machine learning

International audienceThe harmful consequences of rail accidents, which sometimes lead to loss of life and the destruction of the system and its environment, are at the basis of the implementation of a "experience feedback" (REX) system considered as the essential means to promote the improvement of safety. REX seeks to identify adverse events with a view to highlighting all the causes that contributed to the occurrence of a particular accident and therefore to avoid at least the reproduction of new accidents and similar incidents. Accident and incident investigation reports provide a wealth of informative information for accident prevention. It would be appropriate to exploit these reports in order to extract the relevant information and suggest ways to avoid the reproduction of adverse events. In this context, knowledge of the causes of accidents results mainly from the contribution of lessons learned and experiences gained, whether positive or negative. However, the exploitation of information and the search for lessons from past events is a crucial step in the REX process. This process of analyzing and using data from experience can be facilitated if there are methods and tools available to technical investigators. It seems advisable to consider the use of artificial intelligence (AI) techniques and in particular automatic learning methods in order to understand the origins and circumstances of accidents and therefore propose solutions to avoid the reproduction of similar insecurity events. The fact that the lessons one can learn from a REX depends on the experiences of the situations experienced in the past, constitutes in itself a key argument in favor of machine learning. Thus, the identification of knowledge about rail accidents and incidents and share them among REX actors constitute a process of learning sequences of undesirable events. The approach proposed in this manuscript for the prevention of railway accidents is a hybrid method built around several algorithms and uses several methods of automatic learning: Learning by classification of concepts, Rule-based machine learning (RBML) and Case-based reasoning (CBR)