The chemical tank has a good pressure gauge — and its reading lives in a shift notebook.

The problem

The gauge works perfectly — the problem is that only the operator reads it, when he walks past.

A mid-sized chemical plant has dozens of tanks fitted with analog Bourdon gauges. And there are three reasons they are still there:

1. They are intrinsically safe. They need no power supply, they never freeze, and certifying them for ATEX costs a fraction of what an equivalent electronic transmitter costs.
2. Replacing them opens an instrumentation project. Shut down the tank, empty it, certify the new equipment, revalidate the safety management system. For a plant carrying tanks that are 25 or 30 years old, multiply that by 40 tanks.
3. OT is isolated — and properly isolated. In chemicals, opening a new port into the tank's OT network is not done “because it would be nice to have the data”. It is done only when it is unavoidable.

The result: the gauge reading is written down by the operator in a notebook every four hours. If, between one reading and the next, pressure creeps up because of a valve that closes badly, nobody sees it until the alarm dial comes around. And for Seveso III compliance, BREF or the internal safety system, that handwritten record is the worst of both worlds: too infrequent to anticipate anything, too frequent to sustain without human error.

How it fits the IRIS system

iLEAN does not instrument the tank — it puts a camera outside and reads the needle the way an operator would.

The problem is not a lack of data: the data exists (the needle moves in real time) and nobody reads it. iLEAN Connect acts as the filler that seals that crack between the gauge glass and the plant historian, without asking you to touch the panel, the tank or the OT network.

The camera lives outside OT. The CNN reads the needle and returns the value every minute. The validation chamber (ring 2) verifies that the reading adds up. The ring 1 historian imports the signed value. The operator keeps his notebook — and at the end of the shift he also has a continuous record.

The iLEAN pieces applied to reading the analog panel of a chemical tank:

• Connect — the fixed camera + vision. This is the graduated manual capture from the book: when the machine has no interface, a camera photographs the panel. The CNN locates the dial, segments the needle and returns the numeric value with its unit. It does not force you to replace the gauge. ATEX enclosure where the zone requires it, outside the tank's OT network.
• The three rings. The camera and the inference live in ring 3, where AI has autonomy and models can be updated without touching OT. The reading is packaged as signed JSON and passes through the validation chamber (ring 2): is the reading within physically possible ranges for this tank? does the delta against the previous one make sense? is it the right dial? Only then does ring 1 import the value into the historian. OT never accepts inbound connections — it reads from a signed mailbox.
• Agents — the intelligence on top. Once the data arrives every minute, agents detect slow trends that an operator with a notebook cannot see: a pressure drift rising 0.2 bar/hour over a shift and a half, a gentle drop indicating a leak at the base, an anomalous pulsation consistent with a valve that closes badly. The agent notifies the shift supervisor; the person decides.

See the full IRIS architecture →

Before and after

Shift notebook vs. continuous reading without touching the panel

Aspect	Manual entry every 4h	With iLEAN Connect + camera
Reading frequency	2-6 per shift, depending on rounds	Every minute (or every 10 s if useful)
Intervention on the tank	None (but no traceability either)	None — external camera, OT untouched
Instrumentation investment	€0, but hidden cost in operator-hours and risk	Camera + ATEX enclosure per tank, no civil works
Slow drift detection	Only if the operator compares shifts	The agent detects it and raises a notice
Record for Seveso III / BREF	Sheet signed per shift	Continuous signed time series
Human error risk	Biased reading, transcription, omission	Inferred confidence; if it drops, it warns

Impact estimate

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

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

• Chemical plant with 6-15 storage or reagent tanks fitted with analog Bourdon gauges, ATEX zone 1 or 2, manual reading rounds every 4 hours. Pilot on 2-3 critical tanks.
• IP camera with a certified ATEX enclosure per tank, installed without shutting down the tank (the camera goes outside). First expected value within a few weeks: the first continuous time series from the gauge entering the historian.
• Indicative payback between 4 and 9 months. Hard levers: operator-hours returned to the line, avoiding new electronic instrumentation with its project and revalidation, a continuous record for Seveso III/BREF inspection, early detection of slow drifts.
• Expected reduction in reading-round effort of ≥30% in the first quarter — without replacing the operator, giving him back the time spent on the notebook for higher-value tasks.

And the fair question from the OT security manager

“What if the camera or the vision system is an attack vector against my tank?” — it is not, because it does not touch OT. The camera lives in ring 3, inference happens outside, and the reading enters the ring 1 historian through a signed file drop with an asymmetric key per plant. OT never accepts inbound connections. It is the IEC 62443 pattern made formal: progressive trust from the outside in, transport over a signed folder, and the security manager understands “signed folder” far better than “new protocol”. And if one camera is compromised, you rotate its key — the other plants are unaffected.

As for reading reliability: hallucination is a problem of free generation, not of anchored tasks — reading a needle on a dial is the anchored task par excellence, and the best models brought the error below 1.5% ^[1]. If confidence drops below the threshold, the value is flagged as missing and the system raises a notice: never a false value, better a missing value.

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