Edge ejects the defective closure before bagging

The problem

Sampling checks a fraction while the press makes dozens of closures per shot.

A high-cavity press produces dozens of pieces per shot with cycle times under 2 seconds. Manual sampling can only check a fraction. Lines without an automated camera leave control gaps right after a mold or color change — exactly when rejects tend to spike.

• A high-cavity closure press makes dozens of pieces per shot with cycles under 2 seconds. Hourly manual sampling can only look at a small fraction of that.
• Lines without an automated camera depend on 1-5% sampling, which statistically misses intermittent defects coming from a single cavity. A cavity that fails one shot in fifty can run for hours before a sample catches it.
• The control gap is widest right after a mold or color change, exactly when rejects tend to spike: residual color, flash on a worn parting line, a short shot from a cold hot-runner tip.
• A defective closure that reaches the bag reaches the bottler's capper. A damaged thread or a broken tamper band becomes a leak or a torque complaint at their filling line.

How it fits the IRIS system

Edge — piece-by-piece verdict in milliseconds, at the line, nothing sent to the cloud.

ILEAN Edge places a camera over the line with a CNN trained on good and bad examples of that specific closure — thread, short-shot, flash, stain. The defective piece is ejected in milliseconds, before reaching the outbound bag or octabin, with nothing sent to the cloud.

The network is trained on good and bad examples of that specific closure under that line's lighting, not on a generic catalog. It decides in milliseconds at the line, ejects the defective piece and keeps the reject record by type, so the next mold maintenance knows which cavity to look at. Sampling does not disappear: it stops being the only barrier between the press and the bottler. The lab keeps its tests; the line gets a guard that never blinks between samples. And the toolroom gets a defect map per cavity it never had before.

See the full IRIS architecture →

Before and after

Manual sampling versus Edge on every closure

Aspect	Manual sampling	With iLEAN Edge
Share of closures checked	1-5% sample	100% on that line
Damaged thread or broken tamper band	Found by the bottler	Ejected before bagging
Short shot and flash	Seen only if sampled	Classified by type
First shots after a color change	The weakest moment	Full control from the first shot
Decision time per closure	—	Milliseconds, at the line
Closure images leaving the plant	—	None

1-5% manual sampling → 100% control of parts on that line.

Impact estimate

Impact estimate — to be validated against your current reject ratio.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on your current reject ratio and claim history. The more claims a closure has caused, the faster the return.
• From 1-5% sampling to 100% control of the closures on that line. Every closure is seen, not a statistical handful.
• Defective pieces are ejected before the outbound bag or octabin, so they never reach a bottler's capper. A claim that never happens is the saving that weighs most.
• Rejects are recorded by defect type, which points maintenance at the cavity or hot-runner tip that causes them instead of a general mold overhaul.

Payback 5-12 months · from 1-5% sampling to 100% control. From 1-5% sampling to 100% control on that line. Estimated payback of 5 to 12 months depending on the current reject ratio. Estimate to validate.

And the fair question from the production manager

“What if it starts ejecting good closures?” — the false positive is the real risk of any vision system, which is why the network is trained on good and bad examples of your own closures under that line's lighting. Classifying a known part against known defect types is an anchored task, where the best models drop below 1.5% error [1]. Borderline cases are escalated for a person to decide, and each decision feeds back into training. The false positive rate is measured and shown, not assumed.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.