Corentin Ascone's thesis defense

Aiding human expertise of nuclear systems using digital twins: application to primary motor pump unit diagnosis

The diagnosis of nuclear equipment such as primary motor pump sets (GMPPs) requires a high level of expertise that is difficult to acquire and capitalize on within a company. This thesis, carried out in collaboration between the industrial company FRAMATOME (Base Installée de Jeumont, IB-J) and the LAMIH laboratory (UMR CNRS 8201), is an exploratory study aimed at developing an Artificial Intelligence-based help system to assist human experts in their diagnostic activities on this type of complex component. It is based on the Digital Twin (DN) concept of human-machine systems and applies it to FRAMATOME's needs.

An in-depth state of the art has made it possible to define a new holistic approach to JN design based on nine dimensions, which constitute the methodological framework applied to the development of the EXPERIA tool. The EXPERIA tool features a software architecture capable of managing several types of GMPP, integrating their data, and using Artificial Intelligence (AI) algorithmic models to provide diagnostic services to users. The major contributions of such a JN lie above all in the integration of data from multiple sources, but also in the consideration of a very small number of fault-characteristic data. As the safety requirements for nuclear systems are very high, failure scenarios are rare. This then poses a problem when creating algorithmic diagnostic models based on these data.

The proposed solution is based on the holistic approach of JN design relying on multiple modeling of the same system and on the complementarity between technical data and associated human knowledge. The EXPERIA tool was designed with this objective in mind, enabling experts to be assisted in their GMPP diagnostic tasks while capitalizing on their expert knowledge. EXPERIA integrates unsupervised and supervised machine learning models, and a module based on Dempster-Shafer theory to manage discrepancies between models and make diagnosis more reliable. Other models based on human knowledge are also implemented, enabling explicit diagnosis of good or bad operation. EXPERIA interfaces enable users to explore several models at will.

Preliminary results are encouraging, and prospects are proposed for the diagnosis and prognosis of normal and abnormal functioning of nuclear components for the optimization of their maintenance.