Automated defects detection of AA 6063-MgAZ31B TIG welding using radiographic images and deep learning

Detection of welding defects using radiographic images is the most common method of nondestructive testing. It offers better quality of the welding in aircraft, chemical, and shipbuilding industries. Numerous image-processing techniques have been developed to identify weld defects using radiographic images. Weld defects can appear in radiographic images in different shapes, sizes, locations, and contrasts. In this work, the weld specimens of AA 6063 and MgAZ31B dissimilar alloy joints were prepared with different current levels using Tungsten Inert Gas (TIG) welding. The novel test datasets were developed from radiographic images of welded specimens. The various deep learning models such Convolutional Neural Network (CNN), Residual Network (ResNet) and Inception V3 were also used for weld defects detection. In experimentatal study, the available dataset for the training consists of weld defects such as porosity, crack, lack of fusion, and slag infusion were used for investigation. The effectiveness of three deep learning models were compared and found ResNet has the highest classification accuracy of 99.67% for training the images and 99.52% for testing images of AA 6063 and MgAZ31B weld joint.