Systemic approach to the operational support of machine learning systems for predictive maintenance in the nuclear industry

The rise of digital technologies, connected systems, and artificial intelligence is opening up new perspectives for the evolution of maintenance practices, particularly in the nuclear industry. These advances make it possible to implement more continuous, intelligent, and anticipatory equipment monitoring. In line with this dynamic, this industrial PhD, conducted in collaboration with INEO Nucléaire, aims to leverage the opportunities offered by Industry 4.0 while taking into account the gradual evolution of equipment over time, whose lifespan spans several decades. It also builds upon the expertise of maintenance operators, who possess in-depth knowledge of the systems they operate daily. Within this context, our central research question is as follows: How can we ensure the robustness and reliability of a machine learning system for predictive maintenance in dynamic and constantly evolving industrial environments, particularly in the nuclear sector?

To address this question, a conceptual framework was developed, covering the entire lifecycle of machine learning systems in a critical industrial environment. This framework places particular emphasis on model maintenance, especially the detection and management of drifts caused by the non-stationarity of the environment in which the equipment operates.
Adopting a systemic approach, this work distinguishes itself from purely algorithmic approaches by incorporating software-related challenges associated with implementing an online drift management mechanism. It results in several complementary contributions.
The first introduces an architectural pattern focused on concept drift management, formalised as an encapsulated and reusable software solution. This pattern provides a reference framework that clarifies a dense body of literature and highlights the technical implications stemming from the choice of a drift management strategy.
The second contribution underscores the importance of high-quality labelling, made possible through the integration of active learning strategies, while also exploring the use of incremental learning as a relevant alternative in constrained environments, particularly on edge computing infrastructures.
Finally, the third contribution presents a lightweight, modular, and scalable software architecture designed for machine learning systems dedicated to predictive maintenance. Based on a microservices approach, it clearly separates the system’s operational and maintenance functions, enhancing its performance and maintainability compared to a monolithic architecture. This directly operational solution strengthens the industrial relevance of the thesis and its concrete applicability for INEO Nucléaire.