Utilities stop being islands, without touching their PLC

The problem

The utilities are islands, and a degraded island stops the train.

Desalination, nitrogen generation, instrument air, chillers, gensets — each with its own local HMI. And those screens carry data that genuinely matters: desalinated water flow and conductivity, nitrogen purity and pressure, instrument air dew point, overall vibration levels. The orthodox answer would be to integrate each unit by the book: negotiate with the vendor, buy a gateway, pay for the engineering, review the warranty. For every single unit. Months and budget. Which is why it almost never gets done. And the consequence is always the same: the slow degradation of a utility only surfaces once it is already hurting the main process. In a single train, a degraded utility eventually becomes a complex shutdown.

• Desalination, nitrogen generation, instrument air, chillers, gensets: each vendor delivers a turnkey unit with its own PLC, its own screen and a warranty voided if anyone touches the configuration.
• Those screens carry data that matters: desalinated water flow and conductivity, nitrogen purity and pressure, instrument air dew point, overall vibration. It exists only while an operator walks past.
• Integrating each unit by the book — vendor negotiation, gateway, engineering, warranty review — costs months and budget per unit. Which is why it almost never gets done.
• So the slow degradation of a utility only surfaces once it is already hurting the main process. In a single train, a degraded utility eventually becomes a complex shutdown.

How it fits the IRIS system

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

Connect photo mode on panels.

Nothing is installed on the unit, no port is opened and the vendor's PLC is never touched: the unit stays theirs, the data becomes the plant's. On equipment under a vendor warranty, that is the difference between an approvable case and a project purchasing never signs. And it scales one panel at a time, at the pace the budget allows.

• A fixed industrial camera framing the screen, capturing every few seconds.
• Optical recognition plus a model anchored to that specific panel's template: which field sits in which zone, in what units, what range is valid.
• A time series per variable in central memory.
• Automatic cross-referencing against main process state, which is where the real value appears. Zero intervention on the vendor's PLC and zero risk to its warranty.

See the full IRIS architecture →

Before and after

The isolated panel versus the panel being read

Aspect	Today	With iLEAN Connect
Instrument air dew point	Visible on the local screen	A time series in central memory
Nitrogen purity drifting	Discovered through its consequence	Seen in its trend, days earlier
Cross-check against main process	Impossible	Automatic
Vendor PLC and warranty	Intact — and unreachable	Intact — and read
Integration per unit	Gateway, engineering, months	One camera, one template
Data when nobody is looking	Does not exist	Exists

from data that only exists while somebody looks at it, to a historical series you can cross against main process events. From discovering degradation through its consequence, to seeing it in its trend.

Impact estimate

Estimated impact — to validate with your own 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 4-10 months, depending on how many panels are covered.
• Integration CAPEX avoided: gateways, vendor engineering and warranty renegotiation, multiplied per unit.
• Early detection of degradation in the auxiliaries, before it reaches oxidation or polycondensation.
• From data that only exists while somebody looks at it, to a history you can cross against main process events.

estimated payback 4-10 months depending on how many panels are covered. Two return paths: the integration CAPEX avoided (gateways, vendor engineering, warranty renegotiation, multiplied per unit) and early detection of degradation in the auxiliaries. *Estimate to validate*.

And the fair question from the production manager

«Can a camera reading a screen be trusted for a utility that feeds the main train?» — the task is anchored: a fixed screen, fields in known positions, known units and expected physical ranges, where the best models drop below 1.5% error [1]. Any frame whose value falls outside the possible range of that variable — a dew point the dryer cannot produce, a purity above one hundred — is discarded rather than stored, and a glare or a hand in front of the screen leaves a gap in the series, not a false value. A reading is only ever stored as what it is: a screen reading, timestamped, never a control signal.

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