Semi-Supervised Gan-Based Defect Detection on Radiographic Images of Friction Stir Welded Dissimilar Joints (AA6082:AA5083)

The latest green welding technology, friction stir welding (FSW), was believed to manufacture exceptional welds with minimal defects. It is evident that the amount and type of defects determine the performance and the application of any weldment fabricated using any welding technique. The world of artificial intelligence is setting its footprints in almost every research domain. This work focuses on the implementation of the machine learning concept for classifying defective and non-defective weldments using an unsupervised learning technique called anomaly detection. Subsurface defects of weldments were captured on the films using conventional radiography testing. The developed semi-supervised generative adversarial network model built for the classification turned out to be a robust model for classifying the defective and non-defective radiographic images. This developed model for detecting anomalies in radiography weld images may be utilized in a various industrial applications where digital radiography is used to examine welds, such as the aerospace and automobile sectors. The model may be used to automate the identification of anomalies in radiography images, resulting in a faster, more accurate, and less prone to human error inspection process. It may also be used to spot defects that human inspectors may overlook, resulting in better safety and reliability of the inspected structures.