Data mining of temporal sequences for the prediction of infrequent failure events : application on floating train data for predictive maintenance.

Summary Nowadays, in order to meet economic and social requirements, rail transport systems need to be operated with a high level of safety and reliability. In particular, there is a growing need for monitoring and maintenance support tools to anticipate failures of railway rolling stock components. To develop such tools, commercial trains are equipped with intelligent sensors that send real-time information on the status of various subsystems. This information is in the form of long time sequences consisting of a succession of events. The development of automatic analysis tools for these sequences will allow the identification of significant associations between events in order to predict the occurrence of serious failures. This thesis addresses the problem of time sequence mining for the prediction of rare events and is part of a global context of development of decision support tools. We aim at studying and developing various methods to discover association rules between events on the one hand and to build classification models on the other hand. These rules and/or classifiers can then be used to analyze online a stream of incoming events in order to predict the occurrence of target events corresponding to failures. Two methodologies are considered in this thesis: The first one is based on association rule mining, which is a temporal approach and a pattern recognition based approach. The main challenges faced by this work are mainly related to the rarity of the target events to predict, the important redundancy of some events and the very frequent presence of "bursts". The results obtained on real data collected by sensors onboard a fleet of commercial trains show the effectiveness of the proposed approaches.