Advanced thermal monitoring in scaled reactors using deep learning-enhanced drone IR imaging

Abstract

This paper proposes a novel remote and regional thermal anomaly detection system for nuclear power plants, addressing critical challenges in physically inaccessible and high-radiation environments where traditional monitoring methods expose personnel to safety risks. The system integrates drone-based infrared and visible-light imaging with advanced deep learning techniques to achieve regional, intuitive thermal monitoring of complex plant components. We evaluated six deep-learning models including Mask R-CNN, Swin Transformer, Mask2Former, and the YOLOv8 series on a dataset of thermal-visual drone images captured at the URI-LO facility, a 1/8-scale integral effect test reactor. Experiments under three operational scenarios - normal, steam generator fault, and reactor coolant pump failure - demonstrate the system's robust detection capabilities. The YOLOv8m model achieved optimal performance with mAP50 scores of 0.947 for object detection and 0.964 for instance segmentation, while maintaining precision and recall above 93%. Our dual-level analysis combining bounding box detection with pixel-level segmentation provides higher spatial resolution, enabling precise localization of thermal anomalies. Overall, the proposed system offers a promising solution to improve monitoring safety, reliability, and operational efficiency in nuclear power plants, with potential applications in other high-risk environments where rapid fault detection is critical.