Abstract
Accurate measurements of the neutron flux distribution within the reactor core are crucial for
ensuring reactor safety, achieving optimal performance, and understanding reactor operations.
This study aims to develop a Machine Learning (ML)-based anomaly detection technique to
identify anomalies in the neutron flux profile resulting from the improper insertion of copper
wires in the South African Fundamental Atomic Reactor Installationā1 (SAFARI-1). To achieve
this aim, the study first conducted data preprocessing steps, which involved handling missing
values, conducting decay correction, and normalisation. The physical characteristics of measured
thermal neutron flux profiles (key features) and the skewness of neutron flux shapes are extracted.
A statistical analysis was conducted to investigate whether these key features can be used to
estimate the anomalies mentioned above. The statistical analysis revealed that one of the key
features, i.e. the maximum point, can serve as an indicator for the control bank positions.
In contrast, the bottom minimum and top minimum points can be utilised as indicators for
anomalies caused by improperly inserted copper wires. Building on these insights, the Density-
Based Spatial Clustering of Applications with Noise (DBSCAN) clustering-based algorithm was
then employed to detect these anomalies using the key features of the neutron flux profiles.
The DBSCAN algorithm successfully grouped the neutron flux features with similar behaviour,
allowing for the detection of outliers. Finally, Convolutional Autoencoder (CAE) models were
developed to capture the normal neutron flux profiles and detect any deviations, such as anomalies
caused by improperly inserted copper wires. The CAEs for the four assemblies at different core
regions, such as near the centre (D5 and E6) and the periphery (B3 and H3), achieved promising
results. The CAE models demonstrated their capabilities in detecting the anomalies, obtaining
š¹1 scores of 0.621, 0.800, 0.696, and 0.579 for B3, D5, E6, and H3, respectively. These results
demonstrate that considering analytical statistical techniques and data-driven ML approaches
can enhance the quality of reactor monitoring and facilitate automated anomaly detection in the...