The CIP dosing pump is cavitating and nobody hears it — until the next batch comes out with detergent residue.

The problem

A €4,000 pump can ruin a €40,000 batch.

The CIP (Clean-In-Place) system of a cheese plant is the silent guardian of the batch. It cleans tanks, heat exchangers and pipework without dismantling anything — and the whole cycle rests on one small, cheap part: the dosing pump that injects caustic, nitric acid or peracetic into the water line. When that pump cavitates, three things happen at once, and none of them shows up on the panel:

1. The vibration signature changes — a broadband peak appears, rising above the vane-pass harmonic. Audible to a trained ear at 30 cm; inaudible at 2 m over plant noise.
2. The real concentration drifts — the pump draws intermittent air, actual flow drops, and the conductivity of the cleaning solution never reaches range.
3. The rinse comes out poor — the cycle finishes apparently fine (pressures OK, times OK), but detergent residue stays in the line. The first pack of the following day carries it out.

The quality manager finds out two days later through a complaint, or through an odd pH in a routine sample, or never finds out at all and the end customer perceives a soapy taste. The classic system — an operator who hears it, maintenance replacing the seal every X hours — works 99% of the time. That 1% is the one that reaches RASFF or the retailer's contact centre.

How it fits into the IRIS system

iLEAN Edge does not replace your maintenance team — it gives them the ear they cannot have.

The problem with incipient cavitation is not a lack of data: it is information living on an island (the pump housing) that, at the critical moment (the 5 minutes before the rinse turns poor), never reaches the person who decides. iLEAN acts as the filler that closes that gap, without asking you to change the pump, the CIP panel or the ERP.

Edge hears what the operator cannot hear. Connect takes the alert to the channel people already use. The agent crosses the signature with the CIP recipe and the maintenance calendar. The person signs off — never the other way round.

The three iLEAN pieces applied to the CIP dosing pump:

• Edge — a terminal with a MEMS or piezoelectric sensor bolted to the pump housing. It runs a CNN trained to recognise the spectral signature of incipient cavitation, seal blowout and air in the line. It fires a dry contact to the CIP PLC in milliseconds. It works without network: if the plant loses WiFi during the night cleaning cycle, Edge keeps reading and keeps alerting.
• Connect — captures the context the vibration signature alone cannot resolve: the detergent batch received from the supplier (what arrives by email or messaging app at hour zero), the CIP recipe from the MES, the date of the last mechanical seal change noted in a forgotten spreadsheet owned by the maintenance manager. And it takes the alert to the shift supervisor's earpiece or tablet through whichever channel they use.
• Agent — crosses the vibration signature, the CIP recipe, the maintenance calendar and the next scheduled packaging batch. If cavitation shows up 30 minutes before the end of the cycle, it does not send an email at 03:00: it holds the cycle, proposes an extra rinse and alerts the quality manager. The person validates and signs off; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Calendar-based maintenance vs. anticipation through the vibration signature

Aspect	Classic CIP maintenance	With iLEAN Edge on the pump
Cavitation detection	An operator who hears it walking past (if they walk past)	Spectral signature recognised by a CNN in seconds
Mechanical seal replacement	Every X calendar hours, failing or not	When the signature justifies it — no sooner, no later
Rinse validation	Final conductivity + faith in the cycle	Pump signature + conductivity + recipe cross-checked
Cross-contamination	Detected through a complaint or a sample	Anticipated before the first pack of the day
Operation without network	n/a	Edge keeps running on the panel's power
Traceability for the IFS/BRC auditor	Maintenance report rebuilt by hand	Automatic record per CIP batch, with the signature attached

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.

• Mid-sized plant making cured or fresh cheese, automatic CIP per shift, caustic and acid dosing pump on the tank and pasteuriser branch.
• Edge pilot on the dosing pump (sensor + trained CNN + dry contact to the CIP PLC). First expected value in a few weeks: healthy signature recorded, first useful alert of incipient cavitation.
• Expected reduction of poor-rinse events of ≥ 30% against the baseline of the last few years.
• Indicative payback between 4 and 9 months, depending on the historical frequency of rinse incidents and the average cost of reworking or destroying the first batch of the day.
• The hard lever is a single cross-contamination avoided: first batch of the day recovered, no retailer complaint, no food-safety authority inspection. One single event pays for the pilot.

And the maintenance manager's reasonable doubt

«What if the CNN gets it wrong and stops the CIP over some odd vibration that was not cavitation?» — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recognises a signature over the real sensor signal, the best models brought error below 1.5% ^[1]. And even so, the critical call is never made alone: Edge alerts, the person signs off. The three safety rings are there for precisely this.

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