The vintage winder, finally digitized without touching it

The problem

The three numbers that decide the drop test vanish at every batch change.

The machine that winds kraft paper into the drum body is typically 10 to 20 years old, with a proprietary panel showing paper tension, winding speed and layer count. These are exactly the three parameters that decide whether the drum will survive its rating's drop test. Replacing the winder costs several hundred thousand euros and weeks of downtime — yet today that critical data dies on screen every time the batch changes, leaving no trail to correlate against test failures.

• The tube winder that turns kraft paper into the drum body is typically 10 to 20 years old, with a proprietary panel and no data port anyone dares to open. The manufacturer may not even exist any more, or supports it only by phone.
• That panel shows paper tension, winding speed and layer count: exactly the three parameters that decide whether the drum survives its rating's drop test. Get one wrong and the drum body looks fine but gives way when it hits the ground.
• Replacing the winder costs several hundred thousand euros and weeks of downtime, so it keeps running — and its data dies on screen every time the batch changes.
• When a batch fails the drop test, there is no trail to correlate the failure with how that body was wound: too few layers, tension drifting, a speed pushed to catch up.

How it fits the IRIS system

Connect reading the panel from outside — the winder is observed, never touched.

An external camera points at the panel; the machine itself is never touched, only observed. iLEAN Connect reads the panel with OCR and an anchored LLM, structures the reading per batch, and cross-references the ERP to know which batch was being wound at each moment.

Nothing is installed on the winder, no wiring is opened and the manufacturer's warranty stays intact. The camera reads what the operator already sees, and the value appears the first time a drop test fails and someone can actually answer why. Without the winding data, that question ends in a guess; with it, it ends in a parameter and a batch.

See the full IRIS architecture →

Before and after

Today's winder panel versus the winder panel being read

Aspect	Today	With iLEAN Connect
Paper tension per batch	Seen on screen, then lost	Recorded and tied to the batch
Layer count of the drum body	Assumed from the recipe	Read from the panel
Drop test failure analysis	No winding data to look at	Root cause per batch in minutes
Which batch was on the mandrel	Reconstructed from memory	Crossed with the ERP automatically
Intervention on the winder	—	None: an external camera
Winder replacement	On the capex list	Deferred with data to back it

Before: zero correlation between winding data and test failures. After: a root-cause analysis per batch in minutes, with no machine replacement.

Impact estimate

Impact estimate — to be validated with your 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 5-10 months.
• Plus the avoided capex of replacing a winder that still winds well, which is the larger figure in most plants.
• From zero correlation between winding data and test failures to a root-cause analysis per batch in minutes.
• And drift in paper tension shows up as a trend before it turns into a batch that fails the drop test.

Estimated payback 5-10 months, plus avoided winder-replacement CAPEX. Estimate to be validated.

And the fair question from the production manager

“Can a camera really read an old panel with gauges and small digits?” — yes, because the task is anchored: the panel never moves, each value sits in a known position and has a physically possible range, and there the best models drop below 1.5% error [1]. A reading outside the possible range is discarded rather than stored, so a reflection never becomes a data point. The image behind every stored value is kept, so any reading can be checked by a person.

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