Predictive maintenance of HTST pasteurizers — the unplanned stop costs entire shifts.

The problem

When the HTST stops mid-afternoon, the cost is not the part — it is four hours of shift and the milk in the flush.

An HTST pasteurizer that stops without warning is one of the worst scenarios in a dairy plant: the line downstream runs dry, packing stops, the cleaning crew has to come in early, and the milk in process ends up in the flush or in rework. And the pattern repeats:

1. The plate heat exchanger fouls with burnt proteins, differential pressure climbs, thermal performance drops — and one day it stops pasteurizing at the target temperature.
2. The centrifugal milk pump starts vibrating abnormally a few days beforehand — the operator “feels” it, but nobody is squaring that vibration with the history, so the warning arrives when the impeller has already broken.
3. The SCADA has the data, but it has it raw, locked inside the machine, with nobody cross-referencing it with the maintenance plan or the history of previous failures.

The veteran maintenance lead sees the pattern. But they cannot be in three places at once, nor review four signals per machine every hour — and the day they retire, that eye leaves with them. Having the data joined up without an agent is a library without a librarian.

How it fits the IRIS system

iLEAN does not replace the SCADA — it seals the crack between your signals and your maintenance plan.

The problem is not a lack of sensors: it is that the signals you already have are never cross-referenced with the history or the intervention plan, and nobody has time to do it by hand. iLEAN acts as the putty that fills that gap, without asking you to change the PLC, the SCADA or the CMMS.

Edge reads vibration and pressure where they live. The agent cross-references them with the failure history and predicts. The person decides when to intervene — the system only proposes, it never acts alone on the work order.

The iLEAN pieces applied to predictive maintenance of the HTST:

• Edge — a terminal on the floor next to the pasteurizer; it reads from the PLC over OPC-UA/Modbus, or adds its own accelerometer and pressure sensor where the machine does not deliver the signal. It processes locally: if the plant loses Internet, capture does not stop. What is critical cannot depend on WiFi.
• Maintenance agent — lives in Central, cross-references vibration + differential pressure + ΔT + flow with the maintenance reports of recent years and with this very machine's curve of previous failures. It identifies early failure signatures days ahead and prepares the reasoned work order for the CMMS. The person validates and schedules. The agent proposes — the maintenance manager decides.
• Connect — captures the maintenance reports that today live in the shift lead's email, the gasket-change Excels, the photos of the burnt thermocouple. Everything comes in and becomes the plant's permanent knowledge — it does not leave with whoever retires.

See the full IRIS architecture →

Before and after

Reactive + calendar-based preventive maintenance vs. predictive with iLEAN

Aspect	Reactive + calendar-based preventive	With iLEAN Edge + Agent
Exchanger fouling warning	When the outlet ΔT drops — already late	When differential pressure starts climbing, days ahead
Abnormal pump vibration	The operator “feels” it; nobody squares it	Cross-referenced with history, failure signature identified with margin
The veteran lead's knowledge	In their head — leaves with them	Captured in the agent, a permanent capability
Work order to the CMMS	By hand, after the failure	Automatic, reasoned, with the suggested window
Scheduling the intervention	During the unplanned stop	At shift change or over the weekend
Operation without plant network	n/a	Edge keeps capturing, powered from the cabinet

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 pasteurizer. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic, reading your report history and your SCADA.

• Dairy plant with 1-2 HTST pasteurizers of 5-15,000 l/h, a CIP cleaning cycle every 6-8 h, and unplanned stops recorded over the past 24 months.
• Edge + agent pilot on one pasteurizer (PLC reading + accelerometer on the pump + differential-pressure sensor). First value expected within a few weeks (signature of a known failure detected with margin).
• Indicative payback between 4 and 9 months, depending on the number and cost of historical unplanned stops and the cost per hour of stopped shift in your downstream packing.
• Expected reduction of unplanned stops ≥ 30% after the agent's first learning cycle — and the hard lever is a single avoided unplanned stop that pays for the pilot.

And the maintenance manager's reasonable doubt

“What if the agent hallucinates and makes me replace a pump that was fine?” — hallucination is a problem of free generation, not of anchored tasks. When the AI limits itself to cross-referencing real SCADA signals with the failure history of the same machine, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes and the maintenance manager signs. The three safety rings exist precisely for this.

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