EOL Engine Checks: Why Judgment and Archive Cameras Split
Before an engine assembly leaves the line, more than a dozen assembly states must be confirmed: nameplate number against the barcode, filter element, charge-air hose, intake pipe, flywheel bolts, inspection hole cover, protective caps, NOx sensor strap, parts box, water pump connector and the direction its outlet faces. They sit on the top, sides, front, rear and underside, and an operator circling the unit with that list misses the far side first.
The list is written by part; the cameras are mounted by position
Detection items are written by part, cameras by position. The first engineering step is translating the parts list into a view list: which items fall on top, which on the side and end faces, which only appear when looking up. Only then can the robot count and the zone each robot owns be answered. Skip it and added cameras buy cost, not coverage.
The station measures about 4922 mm x 5039 mm x 3008 mm and runs the assembly through on a slat or roller conveyor. Floor space and traffic routes belong at proposal stage, not in a site discussion after the frame arrives.
Two 20 MP cameras per robot, and neither is a spare
Each robot carries two 20-megapixel area-scan cameras, one judging and one archiving, plus a third archive camera over the top. The split is practical. A judgment image wants crisp features and stable parameters; an archive frame has to show the whole unit as it stood. One exposure for both fails both: cut it for noise and the detail an engineer needs at rework is gone, raise gain for a presentable picture and every threshold has to loosen.
Direction is a different problem from presence
The hardest item is not whether a bolt is tight but which way the water pump outlet points. Presence detection cannot answer that. Orientation items need their own viewing angle designed at planning stage, plus pass and fail samples of the same feature for training. Their cost is paid while deciding where to stand the camera, not while training the model.
Where no angle works, a person signs off
A few positions stay hidden behind structure whatever the layout. The scheme does not force a verdict there: those items are declared not visually inspectable and completed through manual entry. That looks like a concession, but it protects the criterion. Guessing at something the camera cannot see costs more in false calls than stating plainly that the item belongs to a person.
Barcode first, then the checklist follows the product
The assembly is scanned on entry, the model identified and its inspection list loaded: which items apply this cycle and what counts as a pass. AI recognition and template comparison then verify configuration, presence and appearance, and results are bound to the barcode and uploaded to MES/QMS, so a rework ticket shows which item failed and what the camera saw.
What the 70-second cycle actually contains
The cycle is about 70 seconds, customized to the workpiece and subject to site acceptance results. Inside it sit the barcode read, multi-angle capture by each robot, dual-engine judgment, exception routing and data upload. Spreading the work across robots is what keeps it there: view zones run in parallel, so cycle time does not grow linearly with the checkpoint count. Manual assembly errors fell by more than 90 percent, subject to site acceptance results.
Questions asked at selection stage
Why not one high-resolution camera for the whole unit?
From any single position a large part of the surface is occluded. A panorama looks complete, but only the face turned toward the lens can be judged. Multiple robots spread the occlusion; they do not add megapixels.
Is the archive camera redundant?
No. The judgment image is for the algorithm, the archive image for people. When a verdict is disputed, only the full frame shows what the part looked like at that moment.
What changes when a checkpoint is added?
If it falls inside an existing view zone, only the list and parameters change. If not, the camera layout has to move. That is why the view list comes before the layout.