Weld path detection in robotic welding processes with deep learning-assisted image processing

With the Industry 4.0 revolution, digitalization and automation have led to radical changes in production processes, and artificial intelligence and robot technologies have made significant contributions to critical processes such as welding processes. Especially in the manufacturing sector, since the welding process directly affects the durability and reliability of the products, it is of great importance to determine the welding path accurately and precisely. In this thesis, an artificial intelligence-supported welding robot prototype was designed and implemented in order to minimize welding errors and optimize production processes. The robot is equipped with a passive vision system due to its simplicity and cost-effectiveness. However, since there is no external laser source in the passive vision system, sensitivity problems may occur. In order to solve these problems, the welding path was determined with artificial intelligence-based deep learning methods. In the study, the workpiece images obtained with the camera were converted into a segmentation problem and the welding path was determined using widely used structures such as deep learning-based TransUNet and UNet. In order for the models to give more successful results in different environmental conditions, data augmentation methods were applied and the best estimation performance was targeted by making various changes in the model parameters. In addition, studies were conducted on the data set created with raw images during model training and performance evaluation was performed. After the detection of the welding path, in some cases, gaps occurred between the lines and the workpiece profiles were estimated to eliminate these gaps. After the predictions made using transfer learning-based structures, the gaps were completed with line or curve fitting methods. In order to increase the prediction performance, VGG16, VGG19 and DenseNet121 architectures were used and the obtained results were compared. The models created in the study gave successful results with high accuracy rates and made a significant contribution to increasing the welding path tracking precision of welding robots.