부산대학교 도서관

The probability of defects increasing as the use of CRDM nozzle (Control Rod Drive Mechanism) was prolonged. The CIVA program, a non-destructive professional simulation tool, was used to determine whether defects occurring in the CRDM nozzle were detected when inspected using the TOFD(Time of Flight Diffraction) technique. In order to determine whether defects could be detected at five designated positions on the flat plate and welded parts, the results obtained when a simulation was performed according to the defect positions and various designated sizes were confirmed, and then data was collected to create a learning model. In order to preprocess the collected simulation data into a form similar to actual field data, Gaussian noise, blur, and trigonometric distortion were added to the simulation data. A deep learning model was created using the preprocessed simulation data using a CNN(Convolution Neural Network) neural network. A total of 7 learning models were created by creating a defect presence/absence judgment analysis model, a defect location judgment analysis model, and a defect size judgment analysis model. A defect analysis model was created by connecting the seven learning models created, and then the presence, location, and size of defects were analyzed by analyzing simulation test data and actual field data.