Deep Learning and Simulations for Nuclear Reactor Anomaly Detection and Localisation

Activity

Keynote

Description

The use of non-intrusive techniques for monitoring nuclear reactors is becoming more vital as western fleets age. As a consequence, the necessity to detect more frequently occurring operational anomalies is of upmost interest. In this talk, we will show how nuclear reactor core noise data can be employed to unfold from detector readings the nature and location of driving perturbations. Given that in-core instrumentation of certain western-type reactors are scarce in number of detectors, rendering formal inversion of the reactor transfer function impossible, this talk will describe the use of Deep Learning for the task of unfolding. We will discuss an approach to such a task doing so in the presence of multiple and simultaneously occurring perturbations or anomalies. A voxel-wise semantic segmentation network will be presented that can be used to determine the nature and source location of multiple and simultaneously occurring perturbations in the frequency domain. A nuclear reactor core simulation tool has been employed to provide simulated training data for two reactors. Lastly, we will talk about some pilot work towards the application of the aforementioned approach to real measurements, introducing a self-supervised domain adaptation procedure to align the representation distributions of simulated and real plant measurements

Period	7 Sept 2022
Event title	Computing, Engineering and Media Post Graduate Research Conference
Event type	Conference
Location	Leicester, United KingdomShow on map