In the welding industry, radiographic inspection is commonly used to detect, classify, and assess various welding defects. However, this technique requires significant human resource allocation. Therefore, in this study, a convolutional neural network (CNN) was trained on radiographic images from the RIAWELC dataset to automate detection and classification of 4 different defect classes: Porosity (PO), Crack (CR), and Lack of Penetration (LP), and No Defect (ND). When presented with an input image from a radiograph patch in the RIAWELC data set, the model categorizes it into one of 4 defect types. The model was assessed on the final test set and achieved an impressive accuracy score of 95.62% and a low loss score of 0.1229. These results indicate robust generalization and minimal overfitting. Consequently, our proposed model is expected to perform effectively on new, unseen data, leading to improved efficiency, cost savings, and streamlined project schedules in the field of NDT inspections.
To read the full paper please click here.