A system has been developed for automatic MAG welding with ceramic backing. This system comprises a camera to capture the images of the molten pool for recognizing feature points to control the torch. A regression-based deep convolutional neural network (DCNN), which outputs continuous values from image inputs, was used to recognize feature points such as arc center and the leading end of the molten pool. This has enabled the accurate recognition of the distance from the arc center to the leading end of the molten pool, as well as the width of the molten pool, with an average error of 0.44 mm or less. The formation of a proper back bead has been confirmed in a welding experiment on a test piece with a tapered gap (from 3 to 10 mm).
展开▼