Développement d’un outil de pronostic pour la maintenance des systèmes mécaniques.

Abstract

Maintenance is becoming increasingly important in companies and tends to evolve for reactivity and cost needs. A particular evolution concerns the way to apprehend the phenomena of failure: little by little the industrialists tend, not only to anticipate them by the recourse to preventive actions, but in addition to do it in the most just possible way with a goal reducing costs and risks. This evolution has given a growing share to the prognosis process. The activity of fault prognosis is today considered as a key process in industrial maintenance strategies. However, in practice, prognostic tools are still rare. Today's stabilized approaches rely on a history of significant incidents to be representative of potentially predictable events The purpose of this thesis is to propose a tool to predict the degradation of equipment without prior knowledge of its behavior, and to generate prognostic indicators to optimize maintenance strategies. Various techniques, of vibratory signal processing, have been explored and tested, on a test bench designed and realized as part of the research axes of this work. Two techniques of artificial intelligence have been exploited in the diagnosis and prognosis of defects in rotating machines, where indicator selection techniques have been explored. The combination of vibration signal processing techniques and artificial intelligence by neural networks has made it possible to provide an efficient prognostic tool and to quantify the relevance of the sources of information used and proposed.