Mechanical feedback and artificial intelligence-based optimization in robotic welding

Robotic welding systems play a critical role in modern manufacturing due to their ability to provide high precision, repeatability, and efficiency. However, one of the main issues that reduces production quality is the inability to instantly compensate for deviations occurring in the weld seam. These deviations directly lead to decreased quality and increased costs. Numerous factors, such as thermal deformations, variations in part tolerances, programming errors, and additional manual quality control steps, also reduce overall manufacturing efficiency. Therefore, the development of automatic, fast, and reliable optimization systems that can address these multifactorial challenges of industrial automation has critical importance. In this thesis, two primary solutions are proposed: a mechanical feedback-based seam tracking system and an AI-assisted defect and seam deviation detection system. Both systems were independently modeled, tested, and their performances analyzed. The mechanical feedback system, utilizing integrated analog angle sensors, employs a contact-based approach to detect seam deviations in real time and utilizes this data for trajectory correction. This approach, validated with a 96–98% consistency between simulation and field data, demonstrated high trajectory tracking performance. In field experiments, the system, controlled with a hybrid FCM-PID algorithm, compensated an initial trajectory deviation of 25 mm within a 4–5 second settling time and reduced the final positional error to the order of ±500 µm. By incorporating oscillation levels into the fuzzy rule base, this hybrid approach shortened the settling time by more than 14% compared to conventional PID control, proving its effectiveness in industrial applications. In the AI-based system, YOLOv5-s and YOLOv8-m models were trained on a custom dataset to detect weld seam deviations and defects. Experimental results showed that model performance is directly related to the size of the training dataset. Particularly in defect detection, increasing the dataset size by 30% and the number of epochs by 25% led to a significant improvement in both F1 score and confidence score of the YOLOv8-m model. This finding demonstrates the system's strong capability for learning and generalization through data augmentation. In conclusion, this thesis presents solutions to the frequently encountered issues of seam deviation and quality control in robotic welding automation, both as independent modules and as an integrated framework. The developed systems provide flexible, cost-effective, and high-performance alternatives tailored to the diverse requirements of industrial applications.