Machine learning and knowledge-driven approaches for automated Calandria Tubesheet Bore Inspection in CANDU reactors

Bruce Power launched its Major Component Replacement (MCR) programme in 2020, focusing on renewing critical components within CANDU reactors, including steam generators, pressure tubes, calandria tubes, and feeder tubes. As part of this programme, the inspection of Calandria Tubesheet Bores (CTSB) presents significant opportunity to deploy AI techniques to reduce time and subjectivity and lengthy and laborious process by providing a set of tools which assist analysts in identifying and classify anomalies within visual inspection footage.

During the MCR outage of Bruce A Reactor 3 (MCR3), a collaborative team from Bruce Power, the University of Strathclyde, Prolucid, and ATS Corporation deployed automated inspection technology, generating a substantial ground-truth dataset. Engineers provided expert labelling of indications across the full inspection series, creating a valuable resource for advancing automated detection through machine learning and knowledge-driven approaches.

Building on this dataset, the team developed and trained a Convolutional Neural Network (CNN) anomaly detection model specifically developed for CTSB inspection videos. The CNN can recognise subtle indication features that rule-based techniques may struggle to detect reliably. In parallel, a knowledge-driven methodology was developed, incorporating insights from the MCR3 labelled dataset. This combines the interpretability of rule-based systems with the pattern-recognition capabilities of machine learning, ensuring both transparency and accuracy in indication detection.

The enhanced system meets the explainability requirements of nuclear applications while improving detection accuracy and reducing false positives. By leveraging operational inspection data, the solution achieves robustness and reliability critical for deployment in high-consequence environments. Integrating CNN-based and knowledge-driven approaches ensures comprehensive coverage of indication types while maintaining the transparency essential to regulatory and industry acceptance. Future work will focus on integrating engineer feedback, expanding labelled datasets for quantitative evaluation, refining algorithms to handle varying light levels and defect types, and providing insights into defect severity.