Predictive chiller maintenance with AI — when the chiller goes down, the plant goes down behind it.

The problem

The chiller gives no warning until it's too late — and it drags the whole plant down with it.

An industrial chiller is a treacherous asset: it runs fine for months, loses efficiency little by little without anyone noticing, and then one bad day it goes down with everything on it. When it falls, it is not just the equipment: it is the pharma cleanroom losing its particle class, the food fermentation chamber climbing a degree, the data center server rack heating up. The downtime costs a fortune and, above all, it arrives with no warning.

The BMS knows almost everything, but never together: it has temperatures, pressures and consumption on separate screens, never cross-referenced. Compressor vibration, if it is measured at all, lives in a separate system that only gets reviewed when someone gets suspicious. And nobody has the real COP curve of each chiller against the expected one in their day-to-day. The fleet degrades silently: each unit loses a few points of efficiency that show up on no dashboard, but do show up on the electricity bill.

The maintenance manager knows this, but cannot cover everything. The classic system (hours-based preventive + corrective when it trips) works 95 % of the time. That 5 % is what stops the plant.

How it fits the IRIS system

iLEAN doesn't add another dashboard — it seals the cracks between the ones you already have.

The problem with chiller prediction is not a lack of sensors: it is that the data lives on islands (vibration over here, current over there, BMS somewhere else, CMMS somewhere else again) and, when the first weak failure signal appears, no single system sees it. iLEAN acts as the filler that closes those gaps — without asking you to change BMS, SCADA or CMMS.

Edge reads vibration and current at the chiller. Connect captures the BMS and the service manual. The agent cross-references them with the expected COP curve and, when something drifts, opens a CMMS work order with the evidence right there.

The three iLEAN pieces applied to predictive chiller maintenance:

• Edge — a terminal with a CNN trained to recognize vibration signatures (bearings, imbalance, mechanical play) and to sample current and power factor at high resolution on the compressor motor. It works without a network. If the plant loses connectivity, the Edge keeps sampling and reacting locally.
• Connect — captures the BMS through whichever route applies (Modbus, BACnet, OPC-UA or reading an old panel), digitizes the manufacturer's service manual so the expected behavior is known, and absorbs whatever arrives from outside (an alert from the refrigerant supplier about a new batch, a manufacturer notice about a fault in a production series).
• Agent — cross-references the vibration signature, the current drift, the BMS pressures and the modeled COP. When the real curve moves away from the expected one, it opens a work order in the CMMS with a substantiated diagnosis and an estimated urgency. The person validates — the agent does not write to the chiller PLC; the safety rings exist for exactly this.

See the full IRIS architecture →

Before and after

Hours-based preventive vs. predictive with iLEAN

Aspect	Classic preventive + corrective	With iLEAN Edge + Connect + Agent
Diagnostic sources	BMS, manual vibration checks, manufacturer manual — never cross-referenced	Vibration + current + modeled COP, all cross-referenced on one board
Failure detection	When an alarm trips or an obvious symptom shows up	When the real curve drifts from the expected one — days of margin
Multi-brand fleet	Each chiller in its own software, each with its own dashboard	Every brand and generation — all on the same board
Silent degradation	Only visible on the bill, months later	8-15 % COP loss detected as soon as it appears
CMMS work order	«Check chiller 3» — the diagnosis happens during the intervention	Order already carries the vibration curve, the drift and the diagnosis
Multi-chiller in cascade	Static BMS rules, never re-optimized	Dynamic sequencing based on each unit's real curve

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 from your plant. We lay it out so the steering committee has an order of magnitude; we refine it during the diagnostic.

• A fleet of 3-6 industrial chillers (pharma, food or data center), multi-brand, mixed-age, in cascade or redundancy.
• Edge pilot on one critical chiller (vibration sensors + current reading + BMS and CMMS integration). First value expected within a few weeks — the first anomalous vibration signature usually shows up early if the fleet has some age on it.
• Indicative payback between 4 and 9 months, depending on the cost of the last unanticipated shutdown and the fleet's current electricity consumption.
• Hard levers: a single unplanned shutdown avoided (in pharma or a data center that pays for the pilot on its own) + a consumption reduction of the order of 8-15 % from running each chiller close to its optimal COP point.

And the maintenance manager's fair objection

«What if the AI gets it wrong and makes me open a work order on a false alarm?» — hallucination is a problem of free generation, not of anchored tasks. Comparing a real vibration signature against the one in the service manual, or a measured COP curve against the one expected for those conditions, are anchored tasks. On that kind of task, the best models brought the error rate below 1.5 % ^[1]. And even so, the order does not execute itself: the agent proposes it with the evidence in plain view; the person validates and signs. The three safety rings exist precisely for this — the agent does not write to the chiller PLC; anything critical belongs to the person.

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