The lab samples; the camera leaves no gaps

The problem

Chip defects are episodes, and the laboratory samples instants.

Laboratory control is exact, and the point is not to replace it. But it is discrete: it covers one instant every few hours. And chip defects are not continuous, they are episodes tied to process events: a start-up after a change, a temperature swing at the cutter, degraded material breaking loose from a dead zone in the melt circuit. A twenty-minute episode between two samples is invisible to the laboratory and perfectly visible to a camera. And there is a point of no return: once the affected chip is blended into the silo with the rest, the options are downgrading the whole lot to a lower-value use or reprocessing it. Both are expensive and both arrive late.

• PET chip sells on uniformity: the bottler is buying the guarantee that every pellet will behave like the one before it. Black specks, out-of-spec fines, agglomerated chip and irregular cut break that guarantee.
• Laboratory control is exact, and the point is not to replace it. But it is discrete: one instant every few hours.
• Defects are not continuous, they are episodes tied to process events: a start-up after a change, a temperature swing at the cutter, degraded material breaking loose from a dead zone in the melt circuit. A twenty-minute episode between two samples is invisible to the lab.
• Once the affected chip is blended into the silo, the options are downgrading the whole lot or reprocessing it. Both are expensive and both arrive late.

How it fits the IRIS system

Edge — local inference in milliseconds, before the silo.

Edge.

The episode is cross-referenced against the process event that caused it, so the root cause arrives ready-made instead of being reconstructed a week later from the lab log and the historian. Intercepting before the silo is the money; knowing why is what stops it happening again. And the laboratory keeps its role untouched: the camera adds coverage between samples, it does not replace a single analysis.

• An industrial camera over the stream with controlled lighting, at the pelletizer outlet and on conveying to silo or bagging.
• A neural network trained on the plant's actual chip: good, with specks, with fines, agglomerated, irregularly cut.
• Local inference in milliseconds, with no dependence on network or cloud.
• The affected section is flagged, the operator alerted and, depending on configuration, the section diverted before it contaminates the silo.
• The episode is cross-referenced against the process event that caused it: root cause arrives ready-made instead of being reconstructed.

See the full IRIS architecture →

Before and after

Sampling today versus continuous inspection

Aspect	Laboratory sampling	With iLEAN Edge
Coverage	One instant every few hours	Continuous
A twenty-minute episode	Invisible	Flagged and diverted
Where the defect is found	In the customer's blow molder	Before the silo
Affected volume	The whole silo	The section
Root cause	Reconstructed afterwards	Correlated with the process event
The laboratory	—	Unchanged: it remains the reference

from discrete sampling to continuous inspection. From finding the defect in the customer's blow molder to intercepting it before the silo. From reconstructed root cause to root cause correlated with the process event.

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 scrap reduction of at least 30% for the failure mode covered.
• Estimated payback 6-12 months, depending on volume and the cost of lot downgrading.
• From finding the defect in the customer's blow molder to intercepting it before the silo.
• And protection of the uniformity reputation, which is the product's commercial argument.

estimated payback 6-12 months depending on volume and the cost of lot downgrading, with a scrap reduction of ≥30% for the failure mode covered. Add the protection of the uniformity reputation, which is the product's commercial argument. *Estimate to validate*.

And the fair question from the production manager

«What if it diverts good chip?» — the false positive is the real risk of any vision system, which is why the network is trained on this plant's actual chip — good, with specks, with fines, agglomerated, irregularly cut — under that station's lighting, not on a generic model. Classifying a chip stream against a short catalog of defects learned from this plant's own samples is an anchored vision task, where the best models drop below 1.5% error [1]. And diverting is configurable: it can start as an alert to the operator, borderline sections are escalated to a person whose decision feeds back into training, and what is diverted is a section, not a lot — a false positive costs a few kilos of reprocessing, not a silo.

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