Detection of Weld Defects in Patch-based Radiographic Images Using Deep Learning

Industry 4.0, the fourth industrial revolution, is characterized by the integration of digital technologies into manufacturing processes aimed at improving production efficiency, flexibility, and customization. Welding, a pivotal process within this industry, can result in defects and deformations in welded structures. To mitigate these issues, non-destructive testing via radiography is employed. Nevertheless, the manual analysis of images proves to be time-consuming and complex due to the varied shapes, sizes, and positions of defects.In this research, we propose a promising approach utilizing deep learning, specifically segmentation techniques, to automate the detection of welding defects in radiographic images. The image dataset was curated from three datasets, the most comprehensive of which was obtained through collaboration with the Algerian company Eurl TESTIAL, with input from domain experts. After constructing a new patch-based dataset, we sequentially utilized two models: one for binary classification and the other for multi classification, aiming to localize defects in welds. The system is built on deep convolutional neural network architectures, namely VGG, ResNet, DenseNet, and Inception. Evaluation of the proposed models reveals that the combination of Inception V3 and DenseNet121 achieved promising results, reaching an accuracy rate of 99%. This automated detection system has the potential to revolutionize non-destructive testing by reducing time and cost, while also enabling the detection of previously invisible defects, thereby enhancing the safety and quality of welded products. Consequently, the application of deep learning in this field proves to be a promising approach with the strong potential to transform the non-destructive testing process.