Machine Learning-based Anomaly Detection and Perturbation Analysis in Nuclear Reactors

Machine Learning-based Anomaly Detection and Perturbation Analysis in Nuclear Reactors

This is a past event

This talk will provide an overview and main results of a recently finished 4-year £5M EU-H2020 project (Sep 2017 – Aug 2021) entitled “Core Monitoring Techniques and Experimental Validation and Demonstration (Cortex -” that involved 20 partners from 11 countries (EU, Japan and US). Georgios was leading activities in a work package related to machine learning and signal processing a large part of which was supported by Aiden. Machine Learning in the context of nuclear reactor perturbation analysis is a relatively unexplored area, due to many reasons that concern both open data availability (highly confidential and sensitive data) and existence of real-life rare or extreme events (few incidents). Based on analysis of the core reactor neutron flux, it is possible to derive useful information for building fault/anomaly detection systems, but this requires simulating various scenarios with very accurate simulation models (e.g. fuel assembly vibrations), in both the frequency- and time- domain. In this project we developed various deep learning approaches to detect the various perturbations taking place within the reactor core, and localise them, i.e. where those perturbations originated from (I, j, k). However, there are several open challenges that will be discussed during the talk, e.g. validation in the absence of real data, simulating TBs of data (that takes weeks) and training models at that scale. Some of these open challenges have been part of a bid that was recently submitted in Horizon Europe (2022-2025), where if funded, Aberdeen team will be leading activities in domain adaptation, self-supervised, and uncertainty estimation.

Aiden Durrant and Georgios Leontidis
Meston 4 and Online

For more information, contact Ehud Reiter  (