Why 300 Tray Points Multiply Per-Point False Calls
An EV battery tray is large and densely featured: up to 300 points across rivet nuts, wire thread inserts, cover mounting holes, ceramic composite tape and beam attachments. The tray carries the modules, so assembly completeness bears on pack structure safety. Walking the tray by hand is slow, and one rivet nut missed after fatigue is one latent risk.
300 points multiply the false calls
A 99 percent correct verdict per point sounds strong, but strung across 300 points the tray-level figure is not that number. The more points there are, the more opportunities a false call has to be amplified, and a full check does not permit a roughly-right verdict. That arithmetic is the design constraint: not pushing the single-point algorithm to its limit, but making each point's verdict certain enough and giving uncertain points an explicit route.
One shot covers the tray only if exposure is global
A 25-megapixel global-shutter area camera captures the tray while it is moving on the conveyor. Global shutter matters exactly here: every pixel is exposed at the same instant, so no rolling-shutter smear or geometric distortion appears in motion. If each of 300 points needed a stop, the cycle could not be held at all; covering every point in one capture is the precondition for the cycle time.
Why the models are split by feature class
CNNs are trained and run separately for bolts, wire thread inserts, machined openings and rivet nuts. The reason is not that more models are better, but that these classes look nothing alike: a bolt has thread and head, an insert is a fine embedded structure, an opening is only a contour, a rivet nut has a flange. One model learning all of them would compromise its threshold across incompatible appearances; split, each class is tuned for its own threshold and augmentation.
Point templates and model routing decide the changeover cost
At run time a point template routes each point to its class model, and feature extraction crops the region per point so the input is consistent for each model. Flexible changeover is software: switch the point template and the routing, without stopping the line. A changeover that does not stop the line still costs work: a new product class needs its model extended, and that belongs in the project schedule, not in an estimate for a parameter edit.
Where thresholds must stay strict
The tray verdict is aggregated from 300 point verdicts, and any abnormal key point blocks the whole tray. That rule sets the direction of the thresholds: key points should rather false-call than release, because a release is a pack structure risk, while false calls at non-key points can be absorbed in the recheck queue. Layering thresholds by point importance is more sensible than one tray-wide threshold, and it puts recheck capacity where it is actually needed.
196 seconds and the resolution limit
Matching pixel resolution to the smallest feature is the first gate in selection. When a feature is too small or too low in contrast, the first move is to improve the light and the imaging angle, not to buy more megapixels. Strobing does two jobs: a bright pulse keeps exposure short enough for a sharp capture, and it raises contrast on rivet nuts and thread inserts. The cycle is 196 seconds, with zero escapes and 99.5 percent process accuracy, point-level false calls absorbed by the recheck queue, subject to site acceptance results.
Questions about point count
How are 300 points guaranteed not to be skipped?
Each point is cropped and routed from the template and its verdict is written back by point number; any abnormal key point blocks the tray. The guarantee comes from the point list and the write-back, not from a person looking again.
Capture on the fly or stop and shoot?
With a continuously moving tray and a tight cycle, capture on the fly is the only option. If the station allows the tray to settle, a stopped capture is more forgiving optically but stretches the cycle.
Can false calls swamp the recheck station?
They can, which is why thresholds are layered by point importance and abnormal points are archived with their image.