Steel galvanizing with AI — zinc is expensive; what leaves through excessive drainage never comes back.

The problem

The defect is born in the bath; you see it when it is already scrap.

On a hot-dip galvanizing line, defects rarely appear out of nowhere. The kettle chemistry drifts — a little more iron, dross forming slowly, a temperature off by a degree — and the parts coming out start looking different: irregular spangle, uneven thickness, a black area that did not coat properly, drainage that carries away too much zinc. All of that gets measured. The problem is not the measurement; it is that every data point lives on its own island:

1. The bath — temperature, lab composition, kilos of zinc added. Sometimes on a SCADA panel only the kettle operator sees, sometimes in a LIMS Excel file, sometimes in the kettle lead's head.
2. The part — you inspect it visually at the end, with a flashlight, and sign off the batch. Thickness is measured by sampling, not in line.
3. The run history — immersion speed, immersion time, pretreatment conditions, previous batch. In the MES, if you have one, or in the shift report.

When the customer complaint arrives (thickness out of spec, damage on site from insufficient protection), you have to reconstruct three weeks of production by hand to find the root cause. The zinc that left never comes back. Neither does the rejected load.

How it fits the IRIS system

iLEAN does not add a fourth system — it seals the cracks between the three you already have.

The galvanizing problem is not a lack of data; it is information living on islands that never reaches the decision-maker in time at the critical moment (the exit from the bath). iLEAN acts as the putty that fills those gaps, without asking you to change the kettle, the furnace SCADA or the lab's LIMS.

Edge sees the part as it leaves the bath. Connect reads the kettle wherever the data lives — an old panel, the LIMS, the kettle lead's Excel file. The agent cross-references it with the history and, if something does not add up, holds the load or alerts the air knife. The person signs.

The three iLEAN pieces applied to steel galvanizing control:

• Edge — a terminal with computer vision (CNN) over the part as it leaves the kettle. It detects uneven thickness, uncoated areas, adhered dross, ash inclusions, anomalous spangle and excessive drainage in milliseconds, and triggers a local actuator (stack light, alarm, per-load record). It works without a network. As long as it has power, it keeps seeing and recording.
• Connect — captures the bath parameters whether they come from a modern SCADA, from the kettle panel with an isolated computer, or from the LIMS Excel file the lab updates twice a day. And it also captures what comes from outside (the zinc supplier's analysis, the steel batch datasheet) at second zero.
• Agent — cross-references the Edge image, the bath parameters, the run history and the customer specification. If a series starts draining above the range, it does not wait for the end of the shift: it alerts the kettle lead with a proposed adjustment (air knife, exit speed, jig angle). The person validates and acts.

See the full IRIS architecture →

Before and after

Classic galvanizing vs. galvanizing cross-referenced with iLEAN

Aspect	Galvanizing with sampling + shift reports	With iLEAN Edge + Connect + Agent
Part inspection	Visual sampling at the end of the batch	100% in line, part by part, before cooling
Coating thickness	Measured by destructive / magnetic sampling	Continuous estimate by vision + bath cross-reference
Dross / ash detection	Operator's eye when pulling the jig	Edge vision on every exit, automatic record
Defect root cause	Hand-rebuilt from past shifts	Per-load dossier with cross-referenced bath data
Zinc over-consumption	Spotted at month-end on the dashboard	Alarm before the series turns into kilos
Operation without network	n/a	Edge keeps operating on the panel's power

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 hot-dip galvanizing line with one main kettle, a mix of coil and long products, coating thickness to standard (EN ISO 1461 / EN 10346 as applicable).
• Edge pilot over the bath exit + Connect reading of the SCADA or kettle panel. First value expected within a few weeks on dross/ash detection and continuous appearance tracking.
• Indicative payback between 4 and 9 months, depending on the current cost of zinc over-consumption and the share of loads regalvanized for cosmetic defects / out-of-spec thickness.
• A reasonable reduction to expect: ≥30% in parts reworked for cosmetic defects, once the model is calibrated to your plant's product mix.

The kettle lead's reasonable doubt

“What if the AI fires false positives when natural spangle varies?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI recontextualizes an image against a known reference (the good part from your own history), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the load and the person signs. The three safety rings exist precisely for this.

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