Development of deep learning-based error detection model for repeated measurement data of pipe wall-thinning in nuclear secondary system

A practical deep learning (DL)–based technique is presented to automatically detect human and measurement errors in repeated ultrasonic testing (UT) measurement data used for pipe wall-thinning evaluation in nuclear power plants. In Korean plants, repeated UT data are stored as two-dimensional matrices in CHECWORKS program and are used for wear evaluation methods. However, various error patterns are frequently embedded in the data, and manual review is rarely performed due to the large data volume and limited program functionality. In this study, UT data collected from Korean nuclear power plants were labeled for error and used to train DL models. A Deep Neural Network (DNN) and Convolutional Neural Networks (CNNs) were developed and comparatively evaluated. The best-performing CNN models achieved F1-scores exceeding 97%, demonstrating high reliability in error detection. When the detected erroneous data were removed prior to wear evaluation, UT data became better aligned and negative wear-rate artifacts were eliminated, leading to more stable and realistic wear trends. The proposed technique is being prepared for integration into ToSPACE program, providing a practical tool to enhance the reliability of UT data and to support field maintenance decision-making.