Conference paper Open Access

A Deep Learning Approach to Anomaly Detection in Nuclear Reactors

Francesco Caliva; Fabio De Sousa Ribeiro; Antonios Mylonakis; Christophe Demazière; Georgios Leontidis; Stefanos Kollias


Citation Style Language JSON Export

{
  "publisher": "Zenodo", 
  "DOI": "10.5281/zenodo.1410084", 
  "language": "eng", 
  "title": "A Deep Learning Approach to Anomaly Detection in Nuclear Reactors", 
  "issued": {
    "date-parts": [
      [
        2018, 
        9, 
        6
      ]
    ]
  }, 
  "abstract": "<p>In this work, a novel deep learning approach to<br>\nunfold nuclear power reactor signals is proposed. It includes a<br>\ncombination of convolutional neural networks (CNN), denoising<br>\nautoencoders (DAE) and k-means clustering of representations.<br>\nMonitoring nuclear reactors while running at nominal conditions<br>\nis critical. Based on analysis of the core reactor neutron flux, it is<br>\npossible to derive useful information for building fault/anomaly<br>\ndetection systems. By leveraging signal and image pre-processing<br>\ntechniques, the high and low energy spectra of the signals were<br>\nappropriated into a compatible format for CNN training. Firstly,<br>\na CNN was employed to unfold the signal into either twelve or<br>\nforty-eight perturbation location sources, followed by a k-means<br>\nclustering and k-Nearest Neighbour coarse-to-fine procedure,<br>\nwhich significantly increases the unfolding resolution. Secondly, a<br>\nDAE was utilised to denoise and reconstruct power reactor signals<br>\nat varying levels of noise and/or corruption. The reconstructed<br>\nsignals were evaluated w.r.t. their original counter parts, by way<br>\nof normalised cross correlation and unfolding metrics. The results<br>\nillustrate that the origin of perturbations can be localised with<br>\nhigh accuracy, despite limited training data and obscured/noisy<br>\nsignals, across various levels of granularity.</p>", 
  "author": [
    {
      "family": "Francesco Caliva"
    }, 
    {
      "family": "Fabio De Sousa Ribeiro"
    }, 
    {
      "family": "Antonios Mylonakis"
    }, 
    {
      "family": "Christophe Demazi\u00e8re"
    }, 
    {
      "family": "Georgios Leontidis"
    }, 
    {
      "family": "Stefanos Kollias"
    }
  ], 
  "id": "1410084", 
  "event-place": "Rio de Janeiro", 
  "type": "paper-conference", 
  "event": "Joint Conference on Neural Networks (IJCNN)"
}
52
138
views
downloads
All versions This version
Views 5252
Downloads 138138
Data volume 391.7 MB391.7 MB
Unique views 5252
Unique downloads 134134

Share

Cite as