Float glass with AI — catching the bath drift half an hour before the ribbon falls from prime to B.

The problem

The ribbon never stops — and by the time you see the defect, you already have hundreds of meters of it.

A float glass line is one of the most demanding continuous processes in the industry. Ribbon quality is the crossing of four realities that live in different systems:

1. What happens in the melting furnace — composition, zone temperatures, charging cycles. SCADA data that almost nobody cross-references in real time with the inspection camera.
2. What happens in the tin bath — thermocouples by zone, reducing atmosphere, ribbon speed. Small drifts the operator spots too late.
3. What the ribbon shows — optical defects (bubbles, inclusions, distortion, roller marks). Classic optical inspection detects them, but with the defect already fully formed.
4. How the lehr anneals — cooling gradient, residual stress. Anneal it badly and the sheet breaks spontaneously at the cutting line or on site.

The line manager knows all this, but cannot hold four systems in one head. The ribbon keeps coming out, and by the time it is confirmed that it is falling from prime to B you have lost 20-30 minutes at thousands of euros an hour of margin — not a minor incident. The classic system works 99% of the time. That 1% is where the month's margin goes.

How it fits the IRIS system

iLEAN does not replace the bath SCADA — it seals the cracks between bath, camera and lehr.

The problem in float is not a lack of information: it is information living on islands that, at the critical moment (the bath drift, the bubble forming in one specific band), does not reach the decision-maker cross-referenced. iLEAN acts as the putty that fills those gaps, without asking you to change the SCADA, the optical inspection system or the annealing lehr.

Edge watches the ribbon with deep vision at real speed. Connect reads the bath, the furnace and the lehr wherever they live. The agent cross-references the defect pattern with the process signature and flags the drift — 20-30 minutes early — to the shift lead, who decides.

The three iLEAN pieces applied to float glass:

• Edge — a machine-vision terminal (CNN) on the line. Cameras at the lehr entrance or at the bath exit, with coherent lighting to classify bubbles, inclusions, distortion and roller marks at ribbon speed. It tags the zone and sends the trace to the agent. It works without a network.
• Connect — captures bath parameters (thermocouples, atmosphere, speed), furnace parameters (curve and charging) and lehr parameters (gradient), whether they come from a modern SCADA, a local data logger or the lab spreadsheet. And it captures what arrives from outside: a change of soda ash from the supplier, a new specification from a glass processor.
• Agent — cross-references the defect pattern on camera with the thermal/chemical signature of the process and anticipates the drift. It does not send an email at 10 pm: it alerts the shift lead through whatever channel they use, with the probable root cause and the suggested actions. The person decides what to adjust; the line does not correct itself.

See the full IRIS architecture →

Before and after

Classic optical inspection vs. cross-referenced inspection with iLEAN

Aspect	Optical inspection + standalone SCADA	With iLEAN Edge + Connect + Agent
Defect detection	At the cutting line, defect already formed	On the ribbon at high speed, classified in line
Root cause	Investigated afterwards; the bath data is in another system	Automatic cross-reference with bath, furnace and lehr at second zero
Anticipating the drift	n/a — you see it once it is already scrap	20-30 minutes early, from defect pattern + thermal signature
Fall from prime to B	Discovered at the end of the shift	Alert to the shift lead with the probable root cause
Operation without a network	n/a	Edge keeps classifying on cabinet power
File for the glass processor	Rebuilt by hand when a claim comes in	Per-coil/batch dossier, with a photo of the zone

Impact estimate

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

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

• Continuous float line running several formats, multi-thickness, supplying glaziers and glass processors.
• Edge pilot at the lehr entrance (cameras + lighting + integration with the bath and furnace SCADA). First value expected within a few weeks.
• Reduction of ribbon-hours downgraded to B against the historical baseline ≥ 30% — a conservative estimate; that is the order of leverage of predictive jidoka.
• Indicative payback between 4 and 9 months, depending on the prime-to-B margin gap and the frequency of drifts documented over the last few months.
• The hard lever is every hour gained in prime grade and the root cause closed for the next shift.

And the line manager's reasonable doubt

“What if the AI flags false positives and makes me touch the bath when I shouldn't?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply cross-references a defect pattern with a process signature, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN alerts and the person decides what to adjust. The three safety rings exist precisely for this.

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