Autorisé à présenter ses travaux en vue de l’obtention du Doctorat de l’Université Paris-Saclay, préparé à l’Université Paris-Saclay GS Sciences de l’ingénierie et des systèmes en :
Traitement du signal et des images
« Diagnostic et Pronostic des défauts dans les systèmes complexes multivariés »
le JEUDI 5 JANVIER 2023 à 10h00
à
Amphi I, Bâtiment Gustave Eiffel
CentraleSupélec, 3 rue Joliot-Curie, 91190 Gif-sur-Yvette
Membres du jury :
M. Mohamed BENBOUZID, Professeur des universités, Université de Bretagne Occidentale, FRANCE – Rapporteur
M. Vincent COCQUEMPOT, Professeur des universités, Université de Lille, FRANCE – Rapporteur
M. Demba DIALLO, Professeur des universités, Université Paris Saclay, FRANCE – Examinateur
M. Didier THEILLIOL, Professeur des universités, Université de Lorraine, FRANCE – Examinateur
Mme Audine SUBIAS, Professeure des universités, INSA Toulouse – Université de Toulouse, FRANCE – Examinateur
M. Antoine PICOT, Maître de conférences (HDR), INP Toulouse – Université de Toulouse, FRANCE – Examinateur
M. Claude DELPHA, Maître de conférences (HDR), Université Paris-Saclay, FRANCE – Directeur de thèse
“Fault Diagnosis and Prognosis in multivariate complex systems”
Abstract :
Fault diagnosis and prognosis have attracted huge attention in industry and academia for the increasing requirements on reliability, availability, maintainability, and safety. Despite the significant progress, the existing fault diagnosis methodologies still suffer from challenges, such as the lack of sufficient faulty data for training, ineffectiveness to complex distributed data, low sensitivity to incipient faults, and the interference of noise and outliers. Therefore, this work proposes a new one-class classification method implemented by generating anchors and selecting the region margin to determine a healthy region as a decision area. Then a particular distance measurement called local Mahalanobis distance is then defined to indicate the distance between a sample and the healthy region. Based on the proposed one-class classification method and the LMD index, this work first develops an incipient fault detection approach by combining the LMD index and the empirical probability density cumulative sum technique. This work also discusses the efficiency of LMD as a representative feature for fault detection. Secondly, this work proposes the faulty variable isolation method for single fault cases by combining the LMD technique with the contribution plot idea. Thirdly, an analytical expression of fault increasing rate is derived from the LMD index for the fault severity estimation task. Finally, we further develop a new reconstruction-based approach using the local Mahalanobis distance as a detection index to improve the isolation and estimation performance. The improved method can accurately isolate multiple faulty variables and estimate their fault amplitudes simultaneously. The case study based on the Continuous-flow Stirred Tank Reactor process data shows that the LMD technique has significant benefits for the fault diagnosis problem, such as high sensitivity to incipient faults, robustness to noise and outliers, and no distribution assumption. The fault diagnosis methods developed on LMD significantly outperform state-of-the-art solutions. The comparative study on the Case Western Reserve University bearing data indicates that the LMD technique can be used as a feature extraction approach and is more effective and robust than the other statistical techniques.
