Knowledge-driven explainable AI for automated defect detection in nuclear reactor components

Critical assets in nuclear power plants, such as reactor vessels, coolant systems, and containment structures, must be inspected to ensure safe and reliable operations. However, these inspections are often complicated by the harsh environments in which they are conducted, with high radiation levels creating significant visual noise in inspection footage. Identifying defects, such as cracks or surface anomalies, is vital for preventing failures, but manual review by engineers can be time-consuming due to challenging visual conditions. To address these challenges, a novel, automated defect detection algorithm has been developed, integrating image processing techniques with a knowledge-driven framework. The approach uses frame differencing to detect temporal changes in video frames, thresholding to isolate potential defects, and morphological operations to eliminate noise. The main contribution of this work is a rule-based filtering process that incorporates domain-specific knowledge, including factors such as anomaly size, persistence across multiple frames, and proximity to critical surfaces. A key feature of this approach is the emphasis on explainability. Unlike black-box machine learning models, this method provides clear, rule-based justifications for each detected anomaly. Such transparency is crucial in the highly regulated nuclear industry, where every decision must be traceable and defensible. To validate the approach, it was applied to a case study involving Calandria Tubesheet Bore (CTSB) inspection videos, a particularly challenging dataset due to the visual noise caused by radiation. The knowledge-based rules tailored to this inspection process helped filter out irrelevant anomalies and generated detailed reports with visualisations to assist engineers in their final assessments.