TOWARD THE USE OF MACHINE LEARNING MODELS IN THE AID OF PREDICTIVE MAINTENANCE OF NUCLEAR REACTORS

Nuclear microreactors may be important in decarbonizing our energy portfolio. Due to their reduced size and the harsh environment created by their operational conditions, implementing the necessary instrumentation to perform real-time monitoring of a given quantity within the reactor vessel may be challenging. In this work, we will explore using a combination of machine learning models and probe measurements to reconstruct the temperature field distribution over a section of the vessel wall of a High-Temperature Gas Reactor (HTGR), accounting for normal and off-normal operational conditions. This approach aims to demonstrate the effectiveness of this methodology in the predictive maintenance of nuclear reactor components. We performed the temperature field reconstruction using a Convolutional Neural Network (CNN) model capable of solving the boundary element method through the resolution of the Kirchhoff-Helmholtz integral equation with no source. This model was trained using data from a computational model and tested against experimental and computational data. We obtained the experimental data from a Texas A&M facility while we built the computational model through the Multiphysics Object-Oriented Simulation Environment (MOOSE) framework. We validated our computational models against experimental data for normal and off-normal operational conditions. Our preliminary results show that the CNN model fails to accurately reconstruct temperature distributions under off-normal conditions when trained exclusively on normal operational cases and under the assumption that we only have eight available thermocouple readings over the vessel wall. However, adding off-normal cases to the training dataset significantly improved the CNN accuracy in reconstructing temperature fields for off-normal operational conditions.