Automated fillet weld inspection based on deep learning from 2D images

This work presents an automated weld inspection system based on a one-stage Neural Network trained through a series of 2D weld seam images obtained in the same study. Our method uses You Only Look Once in version 8 (YOLOv8) for object detection. Several models have been trained so that the system can predict: the type of Flux-Cored Arc Welding (FCAW)/Gas Metal Arc Welding (GMAW) weld, two of the common welding processes that have been used in most industries: shipbuilding, automotive, aeronautics among others; detect if a weld is well manufactured or defective; and a third experiment where four classes are attempted to be detected: a correctly manufactured weld, if it presents a lack of penetration defect, an undercut defect or other manufacturing problems. The presented system does not stop at determining the correct welding parameters. The study is based on finding a robust and reliable system to support this difficult and critical task of weld inspection. High performance was achieved in all three experiments carried out in this study, both in those that established a binary classification (the first two) and in the one that established a multiclass classification (the third experiment). In all of them, an average prediction success rate of over 97\% was achieved.