The 8th International Conference on System Reliability and Safety
Sicily, Italy - November 20-22, 2024

Zhiguo Zeng, CentraleSupélec, Université Paris-Saclay, France

 

Professor Zhiguo ZENG received the Ph.D. degree in reliability engineering from Beihang university in 2016. After receiving his PhD, he joined CentraleSupélec, Université Paris-Saclay, and became a full professor in 2023. His research focuses on the characterization and modeling of the failure/repair/maintenance behavior of components, complex systems and their reliability, maintainability, prognostics, safety, vulnerability and security. Dr. ZENG is an author/co-author of more than 150 papers in highly recognized international journals and conferences. He is recognized as Top Scholar by ScholarGPS based on the his strong publication records and impact of his research. His research has been funded by important government funding agencies like ANR and ERC, and also important industrial companies like EDF, SNCF, Orange and GE Healthcare. He is editorial board member of International Journal of Data Analysis Techniques and Strategies, and the leading guest editor of the special issue on “Dependent failure modeling” of the journal Applied Science. He is the co-head of the master program “Risk, Resilience and Engineering Management” in Universite Paris Saclay, and the engineering degree program “Operation Research and Risk Analytics”.

Speech Title: Empowering predictive maintenance with physics-informed machine learning and digital twins

Abstract: Predictive maintenance has become a key enabling technology for today’s industry. Recent advancement in deep learning has created great opportunities for data-driven predictive maintenance. However, the data-driven methods have to rely on large amount of failure data with labels to train deep learning models, which are often too expensive and time-consuming to collect in practice. Furthermore, most data-driven predictive maintenance models do not consider physical-knowledge during their training process and have no physical constraints in the model dynamics. As a result, the model lacks explainability and generalizes poorly on unseen data.
In this talk, we discuss some of our recent work aiming at addressing the limitations of data-driven approaches by enhancing them with physical knowledge and digital twins. In the first part of this talk, we present a two-phase physics-informed deep learning architecture to integrate physical knowledge for RUL prediction. The proposed architecture is illustrated on an open-source dataset of lithium prismatic batteries and the results show that considering the physical knowledge worsens the model’s capability to fit the degradation path, but improves its performance for RUL prediction. In the second part of this talk, we present a new framework for developing deep learning models for fault diagnosis based on digital twin. The developed fault diagnosis models are able to diagnose component-level failure based on system-level condition-monitoring data. The proposed framework is evaluated on a real-world robot system. The results showed that the deep learning model trained by digital twins is able to accurately diagnose the locations and modes of 13 faults/failure from 4 different motors.