Edge ultrasound on wine bottling inertisation — catch the leak before the batch oxidises.

The problem

The leak cannot be seen, smelled or heard — and the batch oxidises.

Wine bottling lives obsessed with oxygen: TPO (Total Package Oxygen) decides whether the batch ages well or shows oxidative notes within four months. Nitrogen inertisation is the barrier, and the leak is the silent enemy:

1. Nitrogen is distributed through hoses, quick connectors and valves at three points of the line (pre-tank sparging, jetting under the cork, headspace pressure).
2. A loose hose or an old gasket creates a leak that cannot be seen, cannot be smelled and cannot be heard by ear.
3. The operator notices high nitrogen consumption at the end of the shift —if they notice at all—, but by then 8,000 bottles of young white have already been bottled with dissolved O₂ off target.
4. The winemaker tastes it four months later. The batch has oxidative notes. The winery absorbs it as a loss or releases it to market and pays later in brand image.

The line manager and the winemaker know this. Their people know it. What is missing is a sensor that sees the leak the moment it happens and an agent that decides whether that specific wine can take it or not — because a premium young white is not the same as an aged red with high free sulphur. The veteran winemaker's knowledge about which wines are sensitive and which are not does not digitise itself.

How it fits the IRIS system

iLEAN Edge does not replace the winemaker — it gives them the data they could never have in real time.

The nitrogen problem is not a lack of investment in inertisation: premium wineries have been investing in jetting, vacuum and headspace pressure for years. What is missing is a layer that sees the leak at second zero and cross-references it with the sensitivity of the batch in progress. iLEAN is that filler that closes what the bottling line SCADA, the winery MES and the winemaker's notebook left exposed.

Edge hears the leak. Connect captures the recipe and the sensitivity of the batch. The agent cross-references and proposes. The winemaker signs. The batch does not oxidise.

The three iLEAN pieces applied to nitrogen leaks in wine bottling:

• Edge — terminal with ultrasonic sensor mounted fixed over the inertisation area (quick connectors, critical valves) or portable for start-of-batch inspection. A CNN trained to tell a real nitrogen leak signature from the normal noise of the bottling line, estimate the flow rate and map the point. Works with no network: it keeps looking and alerting through the panel light.
• Connect — captures from inside the wine recipe held in the winery MES (young/aged/reserve, free sulphur, pH), the historical nitrogen consumption per batch and the data from the inertisation head itself. And it captures from outside the winemaker's note on which whites are extra-sensitive this harvest, the importer's email with their maximum TPO threshold, the maintenance manager's WhatsApp about the hose replaced last week.
• Agent — cross-references the detected leak with the sensitivity of the batch in progress and proposes a decision: watch closely (small leak on a non-sensitive batch), stop and repair (moderate leak on a premium young white), switch to the redundant line (large leak with a high-value batch in progress). The line manager, usually together with the winemaker, signs. The line does not restart by itself on a critical batch.

See the full IRIS architecture →

Before and after

Inertisation by ear vs. inertisation with Edge leak mapping

Aspect	Classic bottling	With iLEAN Edge ultrasound
Leak detection	High consumption at the end of the shift, if spotted	The moment it happens, at the inertisation head
Decision per batch	Fixed rules — every batch treated the same	Cross-referenced with the sensitivity of the wine in progress
Expected TPO	Random within a wide range	Under control, with an alarm if the leak compromises it
Oxidation detection	Winemaker's tasting, weeks later	Anticipated at the moment of bottling
Winemaker's knowledge	In their notebook and their nose	Captured as an agent rule
Operation with no network	n/a	Edge keeps alerting through the panel light

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

• Mid-sized winery with 1-2 bottling lines, full nitrogen inertisation (sparging + jetting + headspace), a mixed portfolio (premium young whites + aged + reserve) with very different sensitivity to O₂.
• Edge pilot on the main line + integration with the winery MES and the nitrogen head SCADA. First value expected within a few weeks: the first leak map and the first agent decisions validated by the winemaker.
• Indicative payback between 4 and 9 months, assuming the cost of historical losses and returns due to oxidation.
• Expected reduction of off-target TPO incidents ≥ 30% against baseline over the first full campaign.
• Hard lever: a single batch of premium young white saved from oxidation pays for the pilot comfortably. And you avoid the damage to brand image, which never shows on an invoice but weighs more in the medium term.

And the winemaker's reasonable doubt

«What if the AI stops my line on a batch that was running fine?» — hallucination is a problem of free generation, not of anchored tasks. On tasks where the AI limits itself to reading an ultrasonic sensor, classifying the signature and cross-referencing it with the sensitivity of the batch, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN proposes, the winemaker or the line manager signs. On non-sensitive batches the agent can operate with self-approval; on premium young whites, human validation is mandatory. The sensitivity of the wine decides the sensitivity of the system.

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