Abstract
In weld-based manufacturing processes such as welding and metal deposition additive manufacturing (AM), the weld bead is a direct indicator of manufacturing quality. For example, the geometry of the weld bead was optimized to a net shape which outperformed conventional geometries. Automatic monitoring of weld bead is thus of prime importance for welding process control and quality assurance. This paper develops a general-purpose convolutional neural network (CNN) for pixel-level detection and monitoring of beads, regardless of welding materials, machine, manufacturing conditions, etc. To achieve the generality, we collected a great variety of welding images containing 2677 single-line beads from 231 research articles, followed by pixel-wise hand-annotation. Consequently, the trained CNN can recognize different beads from various backgrounds at a pixel level. Case studies show that compared to the image-level classification in prior research, its pixel-level labeling permits real-time, complete characterization of weld beads (e.g., detailed morphology, discontinuity, spatter, and uniformity) for more informed process control. This research represents a significant step towards developing a truly human-like monitoring system with low-level scene understanding ability and general applicability.