Join
Programs
About
Automatic Welding Inspection (AWI) by Canam is an AI vision system, in production on Canam's steel joist lines, inspecting every weld in real time. Using four cameras per line, a deep-learning model catches missing welds within seconds, gives inspectors a second set of eyes, and learns from their feedback.
Open-web steel joists are assembled by manual welding of web members and spacers to the top and bottom chords. Each connection point is a node. A joist carries dozens of nodes, and a single line presents several thousand node welds per shift. While welder performance is near-perfect, the acceptable number of missed welds for a load-bearing member is zero. Quality control is then essential.
To ensure this, the process relies on several successive visual inspection steps. At the line exit, a checker inspects every node on bundles of stacked joists, under welding glare and spatter. The task is demanding: slow, requiring full attention across thousands of near-identical welds, and, like any visual check, prone to error. Shop logs also record missed welds without location or trend data.
No commercial system addressed this need. Machine-vision and laser-scanning products inspect weld geometry on single, fixed parts. None detected the absence of a weld on stacked joists of varying sizes moving on a conveyor, where one joist partly hides another. Published research targeted porosity, undercut and penetration, not missing welds, and no dataset existed.
Canam therefore developed the complete solution in-house, from hardware to AI: a capture station robust to spatter, heat and vibration deep-learning-powered software producing one located verdict per weld from a moving, multi-camera video stream and a checker interface delivering that verdict within seconds.
AWI is in daily production, installed between the welding and checking stations, where it performs one of the inspection steps automatically on every bundle of stacked joists. Each line has four industrial cameras, two per side, viewing the top and bottom chords under LED lighting. Cameras and lights sit in protective housings on Canam-built stands and feed a GPU server in the plant.
As the bundle moves past, a You Only Look Once (YOLO) deep-learning detector, optimized with TensorRT for real-time inference, analyses every frame and classifies each node, spacer and seat as welded, defective or hidden from view. No decision rests on a single image. Each weld appears in many frames from two cameras, so the software tracks it through the sequence, reconciles both viewpoints, and issues one verdict per weld with its position on the joist. When another joist hides a node, AWI reports it as blocked instead of guessing, so the checker knows which nodes to inspect manually.
The result reaches the checker's touchscreen within seconds, before the bundle arrives at the checking station: a drawing of each joist, with every missing weld highlighted in its section. Now guided, the checker verifies the bundle, repairs any missing weld, and uses the screen to report any case where AWI was wrong, a false alarm or a missed weld. Every reported case becomes a training example: the images are labelled, the model retrained, and the improved model goes into service once it passes accuracy checks.
