Localizing Perturbations in Pressurized Water Reactors Using One-Dimensional Deep Convolutional Neural Networks

Laurent Pantera, Petr Stulík, Antoni Vidal-Ferràndiz, Amanda Carreño, Damián Ginestar, George Ioannou, Thanos Tasakos, Georgios Alexandridis, Andreas StafylopatisSensors 2022 · DOI · OpenReview

TL;DR

This paper introduces a novel approach using 1D Convolutional Neural Networks (CNNs) to pinpoint the location of anomalies within a nuclear reactor core by analyzing "neutron noise" signals. By training on simulated data and validating on real-world measurements from a VVER-1000 reactor, the method demonstrates promising accuracy in identifying perturbation locations, particularly at a frequency of 10 Hz.

What Problem Does This Paper Address?

Nuclear power plants (NPPs) rely on numerous sensors to monitor their operational status and ensure safety. While these sensors typically report steady-state values, small fluctuations, known as "neutron noise," can occur. These fluctuations, though seemingly minor, can carry crucial information about internal reactor phenomena and potential anomalies. Identifying the precise location of these perturbations is vital for effective diagnostics and maintenance. Traditional methods for analyzing neutron noise often struggle with complex reactor geometries and the inverse problem of inferring perturbation locations from limited sensor data. This paper aims to bridge this gap by developing a machine learning-based solution for accurate perturbation localization in pressurized water reactors (PWRs), specifically addressing the hexagonal geometry of VVER-1000 reactors.

Methods

The core of this work lies in employing a one-dimensional Deep Convolutional Neural Network (CNN) for perturbation localization. The methodology involves three main stages:

  1. Simulation: Perturbation scenarios are simulated using the FEMFFUSION code, which was extended to handle hexagonal reactor geometries and analyze neutron noise in both time and frequency domains. This process generates datasets of neutron noise signals corresponding to known perturbation locations.
  2. Training: The simulated datasets are used to train the 1D CNN models. The CNN architecture consists of multiple convolutional layers with increasing numbers of feature maps (128, 256, and 512), followed by normalization and pooling layers. The final layers are dense layers that output the predicted location of the perturbation.
  3. Validation: The trained models are then tested on actual plant measurements from the Temelín VVER-1000 power plant.

A key aspect of the CNN architecture is its input processing. Neutron noise signals are transformed from the frequency domain to the time domain using a specific formula (Equation 8). For plant measurements, which are already in the time domain, the data is downsampled to reduce computation time while retaining essential frequency content. The models are trained to identify perturbations at the fuel assembly level, outputting a radial position prediction.

Datasets

  • Simulated Data: Generated using the FEMFFUSION code for VVER-1000 reactor models. These simulations cover scenarios with generic absorbers of variable strength at three specific frequencies: 0.1 Hz, 1 Hz, and 10 Hz. The simulations are performed on a discretized hexagonal grid representing the reactor core.
  • Plant Measurements: Acquired from the Temelín VVER-1000 Power Plant in the Czech Republic. These measurements are from Self-Powered Neutron Detectors (SPNDs) and are used to validate the performance of the trained models.

Key Insights and Ideas

  • Neutron Noise as a Diagnostic Tool: The paper reinforces the idea that subtle fluctuations in neutron flux ("neutron noise") contain valuable information about the state of a nuclear reactor core, including the location of anomalies.
  • CNNs for Spatial Localization: The application of 1D CNNs, typically used for sequential data, is extended here to infer spatial locations within a reactor core by analyzing time-series neutron noise signals. This demonstrates the versatility of CNNs beyond traditional image processing.
  • Frequency-Specific Perturbation Signatures: The study highlights that different perturbation frequencies (0.1 Hz, 1 Hz, 10 Hz) yield distinct signatures in neutron noise, and the CNN models are trained to recognize these.
  • Hexagonal Geometry Handling: The extension of the FEMFFUSION code to handle hexagonal reactor geometries is crucial, as many modern reactors (like VVER-1000) utilize this design, which differs significantly from the more commonly studied Cartesian geometries.
  • Trade-off between Frequency Resolution and Training Time: The paper explicitly discusses the compromise made by decimating signals and using a limited number of samples for training, which improves computational efficiency but limits the frequency resolution, impacting performance at lower frequencies.

Key Results

  • Training Accuracy: The trained machine learning models achieved high accuracy on the simulated data, with accuracies ranging from 93.25% to 93.95% across different detector groups, indicating a strong ability to identify the correct fuel assembly.
  • Localization Performance: The proposed methodology demonstrated promising localization results, particularly at a frequency of 10 Hz, where the prediction accuracy for the radial location was described as "well described."
  • Frequency Dependence: Performance was less optimal at lower frequencies (0.1 Hz and 1 Hz) due to limitations in frequency resolution imposed by signal processing choices for faster training. The paper notes that at 1 Hz, the interpretation of results was less obvious, and the models predicted only one location.

Limitations and Future Work

  • Frequency Resolution Limitation: The paper acknowledges that the signal processing choices (decimation and using 200 samples) limited the frequency resolution to approximately 0.5 Hz. This exclusion of perturbations below 0.1 Hz and the less precise characterization at 1 Hz are noted as limitations.
  • Focus on Radial Location: The current model primarily predicts the radial position of perturbations, with axial location not being the primary focus.
  • Limited Frequencies: Simulated data was generated for only three specific frequencies (0.1 Hz, 1 Hz, 10 Hz).
  • Future Work:
    • Generate simulated data for a wider range of frequencies (greater than 5 Hz) to improve analysis of low-noise frequencies.
  • Train models for each fuel cycle and investigate specific phenomena like Incompatible Rod Insertion (IRI) if they can be linked to specific frequencies.
  • Improve frequency resolution by increasing the sample size used for model training, particularly for very low frequencies (less than 2 Hz).
  • Extend the methodology to consider the axial location of perturbations in addition to the radial one.

Summary updated

Abstract

This work outlines an approach for localizing anomalies in nuclear reactor cores during their steady state operation, employing deep, one-dimensional, convolutional neural networks. Anomalies are characterized by the application of perturbation diagnostic techniques, based on the analysis of the so-called “neutron-noise” signals: that is, fluctuations of the neutron flux around the mean value observed in a steady-state power level. The proposed methodology is comprised of three steps: initially, certain reactor core perturbations scenarios are simulated in software, creating the respective perturbation datasets, which are specific to a given reactor geometry; then, the said datasets are used to train deep learning models that learn to identify and locate the given perturbations within the nuclear reactor core; lastly, the models are tested on actual plant measurements. The overall methodology is validated on hexagonal, pre-Konvoi, pressurized water, and VVER-1000 type nuclear reactors. The simulated data are generated by the FEMFFUSION code, which is extended in order to deal with the hexagonal geometry in the time and frequency domains. The examined perturbations are absorbers of variable strength, and the trained models are tested on actual plant data acquired by the in-core detectors of the Temelín VVER-1000 Power Plant in the Czech Republic. The whole approach is realized in the framework of Euratom’s CORTEX project.
BibTeX
@article{pantera2022localizing,
  author = {Laurent Pantera and Petr Stulík and Antoni Vidal-Ferràndiz and Amanda Carreño and Damián Ginestar and George Ioannou and Thanos Tasakos and Georgios Alexandridis and Andreas Stafylopatis},
  title = {{Localizing Perturbations in Pressurized Water Reactors Using One-Dimensional Deep Convolutional Neural Networks}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22010113},
}
Figure