Abstract
While automatic welding process has been introduced at the manufacturing site today to improve the welding efficiency and weldment quality, there are still some joint which is difficult to be automatically welded. Horizontal penetration bead welding in Shipyard, for instance, where weld pool shape varies easily and tracing technique for its variation is required, is manually welded by skilled welder. In order to automate such skillful welding, our research team works on development of weld pool recognition technique with visual sensor and control robot system. In this research, feature points of weld pool are recognized by using CNNs based learning model in real time during CO2 welding on V-groove joint with gap variation. The chemical composition of the flux cored wire is specially designed for bridge performance and back bead quality. It is adopted the straight stepped weaving to adapt a weld pool shape with gap variation. In order to reduce work processes of ceramic backing attachment, with and without ceramic backing welding has been studied in this research. From the images by a CMOS camera, it is confirmed that the pool lead length and width (PLL, PLW) which are calculated by feature points are recognized with high accuracy by CNNs learning model. On the other hand, it is also found that a large corpus of labeled images is required to obtain the high performance of learning model. In order to reduce costly expert annotation, we propose a self-training method which uses unlabeled images. As a result, it is confirmed that the PLL and PLW are recognized accurately by the self-training method proposed. Finally, results of demonstration of automatic welding with real time image recognition and robot control are described. These results show that horizontal penetration bead welding with and without ceramic backing is possible to be automated by robot system proposed.
