Predictive maintenance of refinery pumps and compressors — one unplanned failure empties the whole unit.

The problem

The monitoring is there, but the alarm arrives when it is already too late.

Every serious refinery has vibration monitoring systems (Bently, GE, Emerson AMS) installed decades ago, collecting data from the critical pumps and the axial compressors 24/7. And even so, unplanned shutdowns keep happening. For three reasons that repeat themselves:

1. A generic threshold is blind to context — the alarm trips when vibration exceeds X mm/s RMS. But a bearing enters fatigue by slowly raising the spectral envelope long before the overall RMS moves, and a mechanical seal warns you through the gland temperature, not through vibration. The generic threshold finds out late, or does not find out at all.
2. The signals live in systems that never talk to each other — monitoring on its own console, the DCS with the process pressures on a separate one, the CMMS (SAP PM, Maximo) with the intervention history on a third, the route inspection reports in the supervisor's spreadsheet. Nobody cross-references the four.
3. Route inspection depends on the veteran's eye — and the veteran inspects each pump once a month. Whatever drifts in the second week gets discovered when the pump beeps on the operator's console — or when it runs too hot.

The result is unplanned failures that stop an entire process unit and force a restart worth several hundred thousand euros — all because of a five-kilo bearing.

How it fits the IRIS system

iLEAN does not replace your monitoring — it seals the cracks between monitoring, DCS, CMMS and the veteran's route.

The problem is not a lack of sensors: it is information living on islands plus a generic alarm that arrives late. iLEAN acts as the putty that fills the gaps between the systems you already have, and gives the maintenance planner an agent that cross-references the four sources and names the failure mode.

Edge sees and measures at the equipment. Connect reads monitoring, DCS, CMMS and the supervisor's spreadsheet. Brain recognizes the failure pattern. The agent opens the work order — the person schedules it.

The iLEAN pieces applied to predicting pump and compressor failure:

• Edge — a terminal with a CNN over a camera and thermography of the bearing and the mechanical seal, whenever the existing monitoring needs to be complemented. It reads the seal, sees the drip, measures the spot temperature of the bearing without entering the DCS. It works with no network, as a non-negotiable design rule: in a process unit, what is critical cannot depend on WiFi.
• Connect — captures the data wherever it comes from: OPC UA integration with the monitoring system (Bently, Emerson AMS, GE), the DCS over Modbus or a proprietary driver, the CMMS through an API or a scheduled export, the route supervisor's spreadsheet from a shared folder. Integration is not a magic button: every island is real work. What changed is that it is now affordable.
• Brain + Agent — on top of all that, Brain recognizes the pattern of each failure mode (bearing fatigue, seal drying out, cavitation, misalignment, unbalance) by cross-referencing vibration + thermography + motor current + differential head. The agent names the probable failure mode, estimates the intervention window (weeks, not hours) and opens the work order in SAP PM/Maximo with the evidence attached — spectral plot, thermographic image, comparison with sister pumps, spare part recommendation. The planner validates and schedules.

See the full IRIS architecture →

Before and after

Generic-threshold monitoring vs. multivariable prediction with a work order

Aspect	Classic monitoring + the veteran's route	With iLEAN Edge + Brain + Agent
Alarm trigger	Generic vibration RMS threshold	Multivariable pattern by failure mode
Warning window	Hours — sometimes minutes	Weeks — a work order you can schedule in a planned shutdown
Diagnosis	“High vibration” — with no context	“Radial bearing fatigue — replace within the next 4 weeks”
Cross-referenced sources	Monitoring, DCS, CMMS and spreadsheet live on islands	All four unified at second zero
Work order	Manual, in the weekly meeting	Opened in SAP PM/Maximo with the evidence attached
Does it touch the SIS or the DCS	n/a	Never — Connect only reads, the OT ring stays intact

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.

• Refinery with FCC + HDS + reformer, a fleet of ~150-300 API 610 pumps and 4-8 critical centrifugal compressors, existing Bently/Emerson AMS monitoring, SAP PM or IBM Maximo as the CMMS.
• Edge + Brain pilot on the Pareto family: the 20-30 pumps that historically account for 80% of unplanned failures. First value expected within a few weeks.
• Reduction of unplanned failures of ≥ 30% in the pilot family. The system starts by giving useful warning on the most frequent failure modes (bearing and seal) and learns the less frequent ones over time.
• Indicative payback between 4 and 9 months, dominated by the cost of a single unplanned unit shutdown avoided. In an FCC, the saving from one single shutdown usually covers several complete pilots.

And the maintenance manager's reasonable doubt

“What if the agent opens false work orders and swamps the planner?” — hallucination is a problem of free generation, not of anchored tasks. When the AI limits itself to comparing the current pattern with a history of real failures and naming the most probable mode, the best models brought error below 1.5% ^[1]. And even so, the agent does not touch the SIS or the DCS: it opens a proposed work order in your CMMS with the evidence, and the planner decides. The three rings exist precisely for this — the OT ring accepts only what has been validated and signed.

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