Image segmentation in friction stir welding using convolutional neural networks

Friction stir welding is an advanced joining process widely used for its superior weld quality, high seam strength, and minimal distortion. Therefore, it is widely applied in industries that require these characteristics, such as the aerospace and the automotive sector (battery trays for electric vehicles). Nevertheless, process irregularities may occur, leading to defects that are often visible on the weld seam surface. For this reason, the weld seam surface is typically inspected after welding by means of human visual inspection. This requires well-trained personnel and involves challenges such as detailed documentation and time-consuming evaluation, particularly when multiple or long weld seams are examined. To address the limitations of post-process inspection, inline quality monitoring systems are gaining increasing importance. Several studies have investigated surface quality monitoring approaches, often in combination with image classification models. The present work contributes to automated surface quality monitoring by investigating the use of modern convolutional neural networks in a pixel-level approach (semantic segmentation). Segmentation enables the identification, localization, and quantification of surface defects. In this study, promising results were obtained for the segmentation of flash, surface galling, and surface lack of fill, with a global accuracy of 97.69% and a mean intersection over union (mIoU) of 83.78%. An IoU of 79.18% was achieved for flash, while the IoUs for surface lack of fill and surface galling were 80.22% and 66.87%, respectively. Future studies will compare automated visual inspection with process monitoring approaches based on process responses, such as process forces and process temperatures.