An eye on every cavity, at the real speed of the line

The problem

The cost is not your own scrap: it is the customer's line stopping.

At real line speed, human control of a PET line is statistical by necessity. The eye tires, and sampling by definition does not see most of the parts. The characteristic defects of the sub-sector are well known: black specks or contamination in the preform, bubbles, crystallinity or pearlescence from a drifted oven profile, malformed neck finish or flash, scuffing from guide rail friction, ovality and wall thickness distribution outside the window. The cost is not your own scrap, which is minor: it is your customer's line stopping. A bottle with irregular wall thickness fails at high speed on the bottler's filler, and there the damage is somebody else's downtime plus the commercial relationship. And there is a trend that worsens the picture: the more rPET in the blend, the higher the input variability — color, viscosity, residual contamination — and the defect rate rises with it.

• At real cadence, human control of a PET line is statistical by necessity: one or two percent is sampled and sampling leaves gaps.
• The defects are known and repetitive: black speck or contamination, bubble, crystallinity or pearling from a deviated oven profile, badly formed neck or flash, scuffing, ovality and thickness distribution out of window.
• Your own scrap is the smaller part. The expensive part is a defective bottle stopping the bottler's filler.
• And without cavity-level attribution, the finding stays at "this lot has more scrap", which is not actionable.

How it fits the IRIS system

Edge — AI vision at the line, trained on the plant's real product.

Edge: AI vision at the edge of the line. Step by step:

What changes is not only coverage: it is attribution. Knowing the defect comes from cavity seven of that mold turns a quality problem into a specific maintenance work order.

• Cameras at the two key points. Preform injection outfeed and blow molder outfeed.
• Model trained on real product. A convolutional network trained on good and bad examples of the plant's specific formats, not on a generic catalog.
• Inference at the edge. Milliseconds per part on a local GPU. The image never leaves for the cloud: no latency, no traffic cost and no bandwidth problem.
• Rejection. The defective part is removed before it reaches the pallet.
• Attribution to mold and cavity. Every detection is tagged with its source mold and cavity. This is what changes the nature of the data: it stops being quality control and becomes a maintenance work order — "cavity thirty-seven has been producing black specks for two shifts".
• Closing the loop with labeling. The same camera can read the printed label and check it against the live order.

See the full IRIS architecture →

Before and after

Today's sampling versus Edge vision

Aspect	Sampling control	With iLEAN Edge
Inspection coverage	1-2 % of pieces	100 %
Point of detection	The customer's filler	The line itself
Nature of the finding	"This lot has more scrap"	Root cause by mold and cavity
Action it allows	Block the lot	Intervene on the specific cavity
Where the image is processed	—	On site: no latency, no traffic cost
With high recycled content	More variability, same sampling	More variability, 100 % control

Impact estimate

Estimated impact — to validate with your own numbers.

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 current scrap rate and the weight of customer complaints.
• The return grows with the percentage of recycled resin in the blend, because input variability is higher.
• Bottler line stoppages avoided, which is where the cost stops bearing any relation to the price of the bottle.
• From an aggregate finding per lot to a root cause attributed to mold and cavity.

Estimated payback 5-12 months, depending on the current scrap rate and the weight of customer complaints. The return grows with the recycled resin percentage in the blend, because input variability is higher. The justification changes a great deal between a 12,000 bottles/hour line and a 60,000 one: it is worth calibrating against real line speed. *Estimate to be validated.*

And the fair question from the production manager

«What if it rejects good pieces?» — the false positive is the real risk of any vision system, which is why the network is trained on good and bad examples of the plant's specific formats, not on a generic model. Borderline pieces are escalated for a person to decide, and every decision feeds back into training. Seeing one hundred percent also means drift shows as a trend before a lot is compromised.

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