The closed induction furnace that finally talks

The problem

The most expensive asset in the plant leaves no record of what it does.

Replacing or retrofitting the panel means seven-figure CAPEX and weeks without melting, which in a foundry is the same as stopping the entire plant. So the data stays inside: every time the screen refreshes, the information is gone. The result is that nobody correlates furnace behavior with metallurgical outcome or with consumption. In an electricity-intensive process, where energy per metric ton is one of the big lines in the P&L, that correlation is direct money. The relationship between superheat, holding time, refractory wear and defect is written down nowhere: it lives in the experience of two or three people.

• The induction furnace is ten to twenty-five years into service and its builder shipped it closed: power, bath temperature, holding time and alarms live on a screen that refreshes and forgets.
• Retrofitting the panel means seven-figure CAPEX and weeks without melting, which in a foundry is the same as stopping the whole plant. So the retrofit stays in the plan, year after year.
• So nobody correlates superheat and holding time with nodularity, with defects or with energy per metric ton, in a process where electricity is one of the big lines of the P&L.
• The relationship between refractory wear, tapping temperature and a shrinkage run is written nowhere: it lives in the experience of two or three melters, and it leaves the plant when they do.

How it fits the IRIS system

Connect in panel-photo mode — solved from outside the furnace, not inside.

Connect in panel-photo mode — the same mode as the previous case, applied to a machine instead of a document:

It touches nothing: no ports opened, no PLC modified, no warranty at risk. The camera reads the screen the way the melter does, only every few seconds and for every heat, and the result is the first complete furnace curve the plant has ever had. The same bracket-and-camera pattern then goes on the heat treatment furnace recorder, the sand plant or the spectrometer. And because the series is crossed with the heat sheet and the lab result, the plant can finally ask the question it has been guessing at for years: which superheat and which holding time give the nodularity the customer pays for, at the lowest kilowatt-hours per metric ton.

• A fixed industrial camera points at the panel, capturing periodically.
• A grounded model reads each frame and turns it into a structured time series.
• The series is cross-referenced with the heat sheet (which heat was in that furnace at that moment) and with the lab result.
• The plant gets the full curve per heat for the first time, without having touched the furnace electronics.

The same pattern works for any other silent asset in the plant: the heat treatment furnace recorder, the sand plant or the spectrometer.

See the full IRIS architecture →

Before and after

Today's furnace versus the furnace being read

Aspect	Today	With iLEAN Connect
Furnace curve per heat	None: the screen forgets	Complete time series
Superheat and holding time	In the melter's memory	Crossed with the heat sheet and the lab
Root cause of a defect run	Reconstructed from memory	The curve in front of you, in minutes
Energy per metric ton	A monthly invoice	A figure per heat and per furnace
Furnace electronics	—	Untouched, warranty intact
Panel retrofit CAPEX	Unavoidable	Deferrable, without losing the data

from no furnace history at all to a complete time series per heat · from root-cause analysis from memory to analysis with the curve in front of you, in minutes · from unavoidable retrofit CAPEX to deferrable CAPEX without losing the data.

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.
• Panel retrofit CAPEX avoided or deferred, in the hundreds of thousands, with no weeks without melting.
• A 3-8% reduction in energy per metric ton once superheat and holding times can be tuned against the actual outcome.
• And root-cause analysis of a defect run goes from memory to the curve, in minutes.

estimated payback 5 to 10 months. Two levers: avoiding the panel retrofit CAPEX (in the hundreds of thousands) and a 3-8% reduction in energy per metric ton once the correlation allows superheat and holding times to be tuned. *Estimate to validate*.

And the fair question from the production manager

“Does a camera reading a screen give data you can trust for a metallurgical decision?” — the task is anchored: fixed screen, fields in known positions and physical ranges the value cannot leave, where the best models drop below 1.5% error [1]. Any frame whose reading falls outside the possible range — a bath temperature that jumps 300 degrees between two captures — is discarded instead of stored, so the series is clean before anyone looks at it. And the curve is never used alone: it is read next to the heat sheet and the lab result.

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