Application of high-pressure water jets and U-Net based deep learning for the efficient decontamination of radioactive tanks

Abstract

An automated decontamination and monitoring system was developed by integrating a high-pressure rotating nozzle with a U-Net-based deep learning segmentation model. A high-pressure water jet (400 bar, 28 L/min) was applied inside radioactive tanks, and decontamination efficiency was evaluated by measuring residual Co concentrations using inductively coupled plasma-optical emission spectroscopy (ICP-OES). After four decontamination cycles, the overall residual contamination was reduced to 4.23 % +/- 8.10 %. Specifically, Groups 4 and 5 achieved complete decontamination, while Group 1 retained 19.3 %, highlighting the influence of nozzle positioning. The ICP-OES results confirmed that areas appearing visually decontaminated still contained 400-800 mg/kg of Co, thereby supporting the feasibility of image-based assessments. To enable real-time decontamination monitoring, deep learning models were trained to segment contaminated and decontaminated regions. Among the tested models, DeepLabv3 achieved the highest Dice coefficient (0.6988) and F0.5 score (0.7178), offering a balance between segmentation accuracy and false-positive reduction. The proposed system eliminated the need for manual contamination verification with smear paper, thereby reducing radiation exposure and enhancing decontamination efficiency. This study demonstrates that high-pressure jet decontamination, combined with real-time deep learning-based monitoring, offers a safer and more effective approach to radioactive waste management during nuclear decommissioning.