Impact of Malfunctioning Sensors on Data-Driven Reduced Order Modelling: Application to Molten Salt Reactors

Abstract

Listen

Translate article

Over the years, the development of Data-Driven Reduced Order Modelling (DDROM) techniques has paved the way for novel approaches to combine the physical knowledge built in high-fidelity simulations with the physical observations from experimental measurements. On the one hand, these approaches allow updating and correcting the background information obtained from the physical model; on the other hand, they allow overcoming the sparsity of observations for a global state estimation. For these reasons, these approaches are of interest for applications where one of the two sources of information is incomplete: for example, for applications related to Circulating Fuel Reactors, such as the Molten Salt Fast Reactor. These reactors are characterised by a hostile and harsh environment and by the absence of solid structures inside the core, making the monitoring of the quantities of interest inside the core a challenging task. Many works of literature on DDROM assume that experimental data represent the truth, and although extensive research has been done on noisy sensors, few works of literature analyse what happens to the state estimation when one or more sensors malfunction. Then, the robust and reliable application of DDROM techniques, requires first investigating how their performance is affected by malfunctioning sensors. This work aims to investigate this aspect in the context of modelling and simulating the system response during an accidental transient occurring in a Molten Salt Fast Reactor, considering the impact of failed sensors on the performance of Data-Driven Reduced Order Modelling techniques. Quite importantly, this work also proposes a strategy based on Supervised Learning to compensate for the failed sensors.