Cold Plate: Line Scan Plus Robot Re-Shoots at Rivets
A cold plate is a large, complex sheet-metal workpiece. It needs scratches, dents and dirt detected across the whole face, and rivet quality checked along folded edges. The two jobs conflict: the full face wants speed and breadth, while a rivet wants proximity and a straight-on view. Sending inspectors back and forth is heavy work, and the judgment scale changes with the person.
One plate carries two different imaging jobs
Surface defects are an area problem needing high-resolution coverage; rivet quality is a point problem needing frontal imaging. One setup for both forces a compromise: a line scan covers wide but cannot see the side edge, while an area camera is flexible but needs multiple frames. The division follows: a fixed line-scan station scans the whole plate, and a robot carrying a high-resolution area camera handles local re-shoots.
The line-scan blind spot sits on the folded edge
A line-scan camera images row by row as the part moves and builds the full-plate image, with a linear light keeping brightness even between rows. Its weakness is not plate-face sharpness but side visibility. Rivets sit on folded or upright structures angled away from the scan direction, so the line-scan view is blocked or severely oblique. Rivet quality goes to the robot area camera, which squares up to each rivet and removes oblique distortion.
Re-shoot points are computed, not preset
Re-shoot positions are generated from the line-scan result: the scan localizes suspect regions and the robot visits those coordinates. Side-edge rivet points are laid out at fixed positions in plate coordinates. With no anomaly, the re-shoot is skipped and the plate is judged within takt. The dynamic part matters: suspect regions differ per plate, and fixed points would miss or waste shots.
Two coordinate systems have to become one verdict
The line scan and area camera are two systems in two frames, and merging them naively does not line up. Station calibration maps each image to plate coordinates, the suspect region goes to the area channel for confirmation, and both results are written back in one frame and fused by region. Without it the fused conclusion splits: the line scan reports a defect while the area camera photographed next door.
Why a single-channel call is not scrapped outright
The fusion rule is: a region reported by both channels gains confidence and is graded directly; a region reported by one channel loses confidence and goes to manual review. This is the trade between false accepts and false rejects. A single-channel report may be real or an artifact; scrapping would pull good plates, passing would let defects through. Routing to review hands the uncertainty to a person at controlled cost.
How two stages fit into 45 seconds
The plate cycle is 45 seconds, with the two stages pipelined, subject to site acceptance results. The line scan completes while the plate is conveyed, taking almost no extra time. The robot re-shoots only after suspect regions appear. Robot path time for re-shoots belongs in the cycle budget, and suspect regions must be prioritized when there are many, or the takt will swing with defect count. Workpieces with complex edges and many re-shoot points are better served by an extra robot or an optimised shooting order, leaving margin in the cycle.
The two-level architecture, coarse scan then precision confirmation, fits sheet-metal parts with folded structures: a main channel screens broadly, an auxiliary confirms locally, and fusion produces the verdict. Selection should set line-scan resolution from the smallest defect and re-shoot point count from structural complexity, budgeting the two separately.
Where line scan and re-shoot divide
Can a line scan alone check rivets?
Rivets sit on side and folded edges, where a fixed line-scan view cannot square up; an angle-adjustable area re-shoot is required.
Can re-shooting break the takt?
It can if points have no priority. When suspects are numerous they must be ranked, high-risk first with the rest sent to review, or the cycle will not hold.
When the two channels disagree, which wins?
Confidence decides, not arrival order. Two-channel agreement raises confidence; a single-channel report lowers it and routes to review. The conflict is made explicit rather than letting one channel silently override the other.