In V groove butt welding with a ceramic backing material, the molten pool should penetrate well. Regardless of the gap fluctuations, the molten pool must maintain a good penetration shape. For this purpose, the molten pool state is detected using ResNet50 as one of the deep learning. The molten pool using a CMOS camera is taken. Fundamental experiments are performed, and images are collected for learning of ResNet50. Good estimation results are obtained for untraining data. The gap and its center are detected processing the molten pool images. The seam tracking is carried out using PI controller, with inputs being the difference between the wire tip and the gap center, and the output is Y axis position. The weaving width is adjusted to fit the gap. The molten pool state is controlled adjusting the travel speed of the welding torch, to keep constant the arc position, because the state of the molten pool depends on the arc position. If there is gap fluctuation as the disturbance and the reference of the arc position is same, it is difficult to get the same penetration of the molten pool. Therefore, the reference according to the output of ResNet50 is adjusted. Molten pool control is based on PI controller with the input is the difference between the arc position and its reference, and the output is the travel speed. The control performance is verified in a case where the gap varies from 7 to 3 mm, and good results are obtained.
Plasma arc welding (PAW) was employed in joining thick materials with groove. Due to the high-density plasma arc, keyhole welding was used in the butt welding. The gap might be taken in places due to the heat distortion during the welding. To achieve a high-quality welding, the adaptive control is required according to the gap. The authors tried to apply CMOS camera to obtain information from the top surface and achieve synchronization between the camera shutter and welding current. Thus, a clear image of the weld pool with the keyhole and the gap was taken. The brightness distribution in the front of the weld pool becomes uneven due to the plasma arc. To classify the gap and select the welding conditions, such as the welding current and the plasma gas flow rate, the authors used one of the deep learning algorithms, convolutional neural network (CNN). In the training and testing data, the performance of the CNN was found to be satisfactory. However, if the plasma torch does not trace the welding line, the pool image will be different from the training image. Hence, the CNN will not be able to estimate the correct gap. To avoid this, the authors detected the welding line from the pool image and selected the image area near the welding line as the input of the CNN, i.e., the input area is selected based on the image obtained during the welding. The validity of the proposed method was verified by the welding experiments.
Plasma arc welding can achieve high-quality welding joints in high-strength manufacturing fields, such as aviation and automotive, and improve production efficiency. It is important to observe the weld pool state in real-time in robot automatic welding. However, the electrodes of the plasma welding torch cannot be observed from the outside. Teaching the weld line to torch in real-time to be observable to humans will be difficult. Also, it is difficult to process the image to obtain the position of the weld line in K-PAW. In this study, a camera was utilized to observe the weld pool. The authors estimate the weld line position in real time by image processing based on U-Net prediction. The U-Net model demonstrates sufficient prediction where the accuracy reached 99.5% for the training data and 96.5% for the test data recognition. Moreover, a control method utilized weld line position estimated from the boundary area to verify the effectiveness of this prediction model from 3 mm within the deviation of 1 mm, which is within the range of permissible welding errors. It could reduce image processing errors in the weld pool image and provide higher recognition accuracy than image processing. Combining vision sensing technologies and deep learning methods will provide new technologies to enable higher welding precision and improve welding quality. It could also accelerate the development of welding technology in the intelligent manufacturing field.
While modern power sources have improved process stability, real-time monitoring and feedback control remain essential for ensuring consistent weld quality under dynamic conditions. To address this need, a vision-based closed-loop control system was developed for pulsed Metal-Active Gas (MAG) welding. The system dynamically adjusts the welding speed based on real-time visual feedback in the welding process. Otsu thresholding combined with morphological operations was applied to molten pool images for brightness-based feature extraction. These features, representing the dynamic behavior of the molten pool, were incorporated into a feedback loop for real-time control. Without relying on complex model-based prediction or sensor fusion, the proposed method reduces fluctuations in weld bead geometry and lowers the occurrence of defects. The experimental results showed that, under optimized control conditions and after a steady welding state was achieved, the weld bead’s height deviation exhibited an average standard deviation of 0.08 mm, and a process stability rate of 92%. The combination of conventional hardware and straightforward image processing makes the proposed approach practical for industrial implementation.