Using generative adversarial networks to improve the efficiency of crack detection in nuclear reactor inspection data

Tools

Branikas, E. and Murray, P. and West, G. (2022) Using generative adversarial networks to improve the efficiency of crack detection in nuclear reactor inspection data. In: 41st Annual Canadian Nuclear Society Conference, 2022-06-05 - 2022-06-08, Virtual.

[thumbnail of Branikas-etal-CNSC-2022-Using-generative-adversarial-networks-to-improve-the-efficiency-of-crack-detection-in-nuclear-reactor-inspection-data]
Preview
 Text. Filename: Branikas_etal_CNSC_2022_Using_generative_adversarial_networks_to_improve_the_efficiency_of_crack_detection_in_nuclear_reactor_inspection_data.pdf
Accepted Author Manuscript
License: Strathprints license 1.0
Download (616kB)| Preview

Abstract

In the nuclear sector, condition monitoring & inspection activities allow operators to assess and understand the health and condition of their plant. In the UK, where power stations typically belong to the Advanced Gas-cooled Reactor (AGR) family, inspections of the reactor cores are performed using various sensors including cameras which allow remote visual inspection (RVI) of key components. Specifically, RVI of the reactor core allows operators to view the inner surface of individual fuel channels selected for inspection. In this process, key reactor components are assessed, and any potential defects or anomalies can be detected and classified. Analysis of the resulting video footage is typically conducted manually, but recently, new approaches capable of identifying defects automatically, have been proposed to provide new decision support tools for analysing the data. However, current state-of-the-art techniques presented in literature have their limitations, particularly in terms of their ability to accurately detect and describe features such as cracks. This paper therefore proposes a new approach which addresses this challenge and is demonstrated through a case study using representative data gathered from AGR inspections. To achieve our goal, we implement a neural network that belongs to the family of Generative Adversarial Networks (GANs). Our approach provides a novel post-processing step which improves the pixel-level prediction of an existing image segmentation network designed to detect and segment cracks. We also provide a framework that can be incorporated into the output of other algorithms which produce binary crack predictions at their output. Our approach is demonstrated through application to Remote Visual Inspections of graphite core bricks in Advanced Gas-cooled Reactors and has the potential to be applicable to other inspection activities in the nuclear sector and other industries.

ORCID iDs

Branikas, E., Murray, P. ORCID logoORCID: https://orcid.org/0000-0002-6980-9276 and West, G. ORCID logoORCID: https://orcid.org/0000-0003-0884-6070;
CORE (COnnecting REpositories)