Automated generation of training dataset for crack detection in nuclear power plant components

Inspection for crack-like features in nuclear power plant components is vital to maintain safe continued operation. The traditional manual-based inspection can discover such features but could be time-consuming and repetitive. Deep learning technique such as convolutional neural network (CNN) offers improved efficiency by automating the inspection of cracks in images and videos. However, a significant overhead of the CNN implementation is the preparation of the large training dataset for training the classification system. The traditional manual-based labelling process costs intensive labour and could become prone to inconsistent labelling standard to annotate cropped patches from inspection videos. As a result, the CNN system may learn irrelevant features for decision-making. This paper presents an automated labelling technique to efficiently generate crack training dataset with consistent labelling standard. The proposed labelling technique is based on binary masks and can generate sufficient labelled patches with customised resolution. A showcase study of automated patch labelling of crack-like features in superheater tube plate upper radius is provided. The result shows that the proposed labelling technique is suitable for the fast creation of training dataset to build autonomous crack detection systems. The only overhead is the generation of binary crack masks which weights much less than the manual labelling burden.