Edge ultrasound on the refinery steam network — traps and joints, kg/h, immediate saving.

The problem

Hundreds of traps, thousands of joints, and prioritisation done by ear.

Saturated steam in a refinery is expensive energy: every kg/h lost is fuel burned in the boiler for nothing, plus the associated carbon footprint. A steam leak always follows the same script:

1. The refinery has a steam network that is practically a map of the plant — pipe-racks between units, traps at every low point, joints at every flange.
2. The trap survey comes round once a year (when it comes at all). It spots the ones failed open and notes them in a spreadsheet.
3. The repair plan is prioritised by the trap that hisses loudest or the one that is easiest to reach, not by the one losing the most kg/h.
4. Between one survey and the next, a joint can start leaking from thermal cycling and nobody finds out until the following year — months of steam venting to atmosphere.

The energy efficiency manager knows it. His people know it. What he does not have is continuous mapping that quantifies every point in kg/h and €/h and proposes a prioritised route the crew can execute between shutdowns. The veteran who could hear the traps like nobody else retires this year.

How it fits into the IRIS system

iLEAN Edge does not reinvent the trap survey — it makes it continuous and prioritises it by cost.

Ultrasonic trap surveying has existed for decades. What is missing is a layer that turns every inspected point into a quantified figure, cross-references it with the real steam generation cost of the refinery and proposes a repair route the manager can sign that same week. iLEAN is the filler that covers what the boiler SCADA, the maintenance CMMS and the inspector's spreadsheet left exposed.

Edge hears the leak and quantifies it. Connect captures the real cost of steam. The agent cross-references and proposes the route. The person signs — the steam network is never touched on its own.

The three iLEAN pieces applied to saturated steam leaks in a refinery:

• Edge — terminal with an ATEX (intrinsically safe) ultrasonic sensor, on the operator's own device or on a fixed station in safe areas. A CNN trained to tell a working trap from one failed open, estimate the steam flow lost in kg/h and translate it into €/h using the enthalpy of the steam and the generation cost. It works with no network: it records points with the light of the panel and syncs when connectivity returns.
• Connect — captures from inside the fuel consumption and steam production of the boiler house SCADA, the shutdown plans in the CMMS, the trap survey history, the hot work permits. And it captures from outside the email from the trap supplier (lead time of the spare model), the WhatsApp from the shift leader ("the trap on unit 7 is hissing louder").
• Agent — cross-references the Edge map with the real steam generation cost, how accessible each point is and the window of the next scheduled shutdown. It proposes a repair route prioritised by € lost per hour that fits the crew you already have. The efficiency manager signs the plan; every repaired trap is checked against the real saving in boiler fuel.

See the full IRIS architecture →

Before and after

Annual trap survey vs. continuous iLEAN mapping

Aspect	Classic annual survey	With iLEAN Edge ultrasound
Frequency	Annual, external	Continuous, with the operations team
Quantification	Yes/no, sometimes estimated	kg/h and €/h per point, with real enthalpy
Prioritisation	The loudest one or the easiest to reach	The one that costs most every hour
Detection between surveys	n/a — found at the next one	Continuous, a joint that starts leaking is detected
Crew route	Spreadsheet from the external report	Weekly plan prioritised by € lost
Saving verification	Estimated, no cross-check	Checked against boiler fuel from the SCADA

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 refinery. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized refinery with a saturated steam network from 6 to 16 bar, several hundred traps, a boiler house with 3-5 units, a dedicated energy efficiency team.
• Edge pilot on one process unit + integration with the boiler SCADA and the CMMS. First value expected within a few weeks: the first kg/h map and the first repair plan prioritised by €/h.
• Indicative payback between 4 and 9 months — steam is one of the most expensive utilities in the refinery and the accumulated leak is usually enormous.
• Expected reduction in the fuel consumption associated with leaks ≥ 30% against baseline in the first half-year.
• Hard lever: less fuel burned per kg/h of steam produced, verifiable month by month against the boiler house SCADA. The annual external survey becomes a complement, no longer the only snapshot of the year.

And the reasonable doubt of the CAIO

"What if the AI gets the flow wrong and I end up repairing the wrong traps?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where AI is limited to reading an ultrasonic sensor, classifying the trap by its acoustic signature and estimating flow with an anchored physical model, the best models brought the error below 1.5% ^[1]. And even so, nothing critical is decided alone: iLEAN proposes the route, the person signs, and the saving is checked against boiler fuel. If the estimate errs on the high side, the SCADA proves it wrong within a month.

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