| Abstract Scope |
In gas metal arc welding, weld quality is affected by many factors, including welding conditions and the external environment. Welding process monitoring is vital to validating the planned process. Vision-based monitoring systems have been used for arc welding automation. Mainly, a laser vision sensor is widely used and researched in the field of welding automation because it can acquire weld bead shape and joint geometry for various weld joints with high precision. In order to enhance the result of vision monitoring, image information of the molten pool is required in addition to the profile information by the laser vision sensor. This study proposes a method to capture both the molten pool and a laser line on solidified weld at the back of the molten pool using a vision camera coupled with a stripe laser. In order to improve the original image captured using the camera, radiation emitted during gas metal arc welding was analyzed. The spectral characteristics of the radiation emitted were analyzed using a spectrometer, and a suitable wavelength band proper to the camera was determined to acquire both the laser line and the molten pool. The structure of the stripe laser and vision camera was designed considering the distance and angle so that the weld pool and laser line concurrently could be captured in one image. To clearly observe the captured image, image processing, such as image detection algorithm and curve fitting, was performed to extract the molten pool and laser line, and the improved image was used for welding quality using a convolutional neural network-based algorithm. Keywords: Gas metal arc welding, Spectral characteristics, Laser vision sensor, Vision camera, Image processing, Convolutional neural network |