PET, PE and PVC extrusion with AI — the drift that reaches the customer when there are already 8 hours of bad reel.

The problem

By the time the gel shows up, eight hours of reel are already gone.

PET, PE or PVC extrusion looks stable until it isn't. Every production manager knows the story: the SCADA is green, the laser sizer says thickness is within tolerance — and two hours later the customer's quality department returns a reel for gel, for a longitudinal band or for a variation in clarity. The pattern repeats because the three real causes live in three different places:

1. The moisture of the material entering the screw, which depends on the dryer, on the recipe and on ambient temperature. It lives on the dryer panel, sometimes integrated with nothing at all.
2. The masterbatch or resin batch, with its variability declared on a certificate of analysis that arrives by email from the supplier and almost never crosses the factory to reach the extruder.
3. The response of the screw and the calender to those two variables — torque, pressure, barrel temperature, haul-off speed — logged in the SCADA, but with nobody correlating it with the other two in real time.

The shift lead has three screens and the SCADA charts. The veteran operator knows that “that resin, at this moisture, gives you a band after 40 minutes”, but that knowledge walks out the door the day he retires. The classic system works 99% of the time. That 1% is eight hours of reel discovered at the customer — and every time it is hard cost plus reputational damage.

How it fits the IRIS system

iLEAN does not replace the extruder's SCADA — it captures the islands the SCADA could not see.

The extrusion problem is not a lack of sensors: it is information living on separate islands (dryer, batch certificate, the veteran's eye, laser sizer) that, at the critical moment (the batch change, the recipe change), does not reach the decision-maker cross-referenced. iLEAN acts as the putty that seals those cracks, without asking you to change extruder, SCADA or ERP.

Edge sees gel and band in line, before the reel ships. Connect captures the batch certificate and the dryer panel. The agent cross-references the three and proposes the adjustment. The person signs — the line never reconfigures itself.

The three iLEAN pieces applied to PET, PE and PVC extrusion:

• Edge — a computer-vision terminal (CNN) over the calender or the sizer. It detects gel, longitudinal band, fish-eyes, patches of poorly dispersed pigment and clarity drift in milliseconds. It cross-references with the laser sizer and the IR camera. It triggers the actuator (stack light, ejector, controlled stop) when the pattern is clear. It works with no network: if the plant loses WiFi, Edge keeps reading and holding.
• Connect — captures the supplier's certificate of analysis whether it arrives by EDI, email or PDF, the dryer readings (a photo of the panel if it is standalone, an integration if it has an industrial PC), the ambient humidity and the batch recipe. Graduated capture means it works just as well with modern lines as with extruders built twenty years ago.
• Agent — cross-references actual moisture, the masterbatch/resin batch, the screw's response and the defect pattern seen by Edge. It recognizes the characteristic pattern (“with this batch and this moisture you get a band after 40 minutes”) and proposes a recipe adjustment before the defect reaches the film. It generates the per-reel dossier with full traceability for the customer and for the auditor.

See the full IRIS architecture →

Before and after

Extrusion run by eye vs. extrusion cross-referenced with iLEAN

Aspect	SCADA + dryer + the operator's eye	With iLEAN Edge + Connect + Agent
Gel and band detection	Post-reel quality check or customer complaint	In-line vision, milliseconds, before winding
Resin batch change	Certificate by email, never reaches the operator	Cross-referenced at second zero with the screw's response
Material moisture	Dryer panel, standalone, read by hand	Continuous capture, correlated with historical defects
The veteran's knowledge	Lives in his head, leaves with him	Captured as a persistent pattern in the agent
Operation without a network	n/a	Edge keeps running with its own light
Per-reel dossier for the customer	Rebuilt by hand when a complaint arrives	Automatic traceability, with a photo of the defect if there was one

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the real data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• A plant extruding film, sheet or pipe in PET/PE/PVC, multi-format and multi-batch, with recurring scrap attributed to gel, band or thickness variation.
• Edge pilot on one line (camera at the calender/sizer + Connect for the dryer and the batch certificate + an agent that cross-references). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on hours/year of scrap from undetected drift, the average cost of a rejected reel and your current level of customer complaints.
• The hard lever is threefold: reel scrap reduction ≥30%, fewer unplanned stoppages caused by gel, and a per-reel dossier that cuts customer complaints.

And the quality manager's reasonable doubt

“What if the AI gets it wrong and lets a reel with gel through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares what the camera sees against the recipe and the history (recontextualizing a piece of data), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds and the person signs. The three safety rings are there for exactly this.

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