Pressure and valve crimping in cosmetic aerosols with AI — the shelf leak is born on the line, not in the truck.

The problem

The crimping defect is silent — which is why it always reaches the customer.

Valve crimping is the aerosol's critical operation: a collet presses the valve flange against the can neck and seals the pressurized system. Done right, the unit works for months and dispenses product without losing gas. Done wrong by a small margin — an aging collet, a valve with a slightly different tolerance from a new supplier, a temperature variation during gassing — the unit looks perfect at the outfeed and loses gas over the following days.

The three realities that are almost never in the same place:

1. The seal geometry — flange height, collet deformation, crimp symmetry. The operator sees it in visual inspection, usually by sampling.
2. The internal pressure — set by the gasser. Read by a sensor or by the line's own gauge. Often logged in a separate system.
3. The defect on the shelf — the return, the retailer complaint, the call from the brand manager. It arrives weeks later, and reconstructing what happened in that run costs the plant manager half a morning.

The classic system (sampling + visual inspection + manual adjustment between batches) works 99% of the time. That 1% is what reaches the shelf, where the damage is multiplied by the cost of the return and the cost to the brand.

How it fits the IRIS system

iLEAN doesn't change your crimper — it closes the loop between the crimp, the pressure and the rejected unit.

The problem with cosmetic aerosols is not a lack of machines: it is information living on islands that, at the critical moment (the unit leaving with an invisible defect), does not reach the decision-maker in time. iLEAN acts as the putty that fills the gaps between the crimper, the gasser, the operator and the MES, without asking you to replace any of them.

Edge sees the crimping and reads the pressure. The actuator rejects the unit before packing. The agent prepares the per-batch dossier — and a person signs every rejection.

The iLEAN pieces applied to cosmetic aerosols:

• Edge — a terminal with computer vision (CNN) over the aerosol neck after crimping. It measures the seal geometry — flange height, symmetry, collet deformation — and triggers the actuator (rejector) if it is out of tolerance. In parallel it reads the internal pressure after gassing, via sensor or via vision on the line's own gauge. Reaction time in milliseconds. It works without a network — the non-negotiable rule: what is critical cannot depend on WiFi.
• Connect — captures the valve supplier's data (batch change, tolerance change notified by email), the shift's ambient temperature, the wear reported by maintenance. At second zero. Without asking anyone to forward anything.
• Agent — cross-references crimp geometry + pressure + shift conditions + defect history. When a drift starts to appear (the geometry runs toward the high limit in the fourth hour of the shift, just as the collet heats up), the agent warns before the drift becomes a defect. Maintenance acts on data, not hunches.
• Three safety rings — the critical crimping parameters live in the isolated ring 1. Whatever comes from outside (supplier notifications, parameters suggested by the agent) goes through ring 2 first. For IT/CAIO: it fits the AI Act and regulatory traceability.

See the full IRIS architecture →

Before and after

Sampling + manual adjustment vs. in-line control with iLEAN

Aspect	Sampling + visual inspection	With iLEAN Edge + Connect + Agent
Crimp geometry verification	Sampling every N units, operator's eye	100% in line, in milliseconds
Internal pressure reading	Sensor → separate system	Cross-referenced at second zero with the crimp
Drift detection from collet wear	Reactive, after a run of defects	Predictive, before the drift becomes a defect
Valve supplier change	Manual adjustment, trial and error	The model adapts with data from the new batch
Operation without a network	n/a	Edge keeps running on panel power alone
Dossier for the auditor / customer complaint	Rebuilt by hand, weeks	Per-batch dossier, automatic, with geometry + pressure

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.

• Cosmetic aerosol packing line, multi-format (hairspray + deodorant or deodorant + foam), modern filler and gasser, crimper with a history of manual micro-adjustments.
• Edge pilot on one line (camera over the crimp + pressure reading + actuator + MES integration). First value expected within a few weeks: the first shift with 100% geometry and pressure verification.
• Indicative payback between 4 and 9 months, depending on last year's volume of leak returns and the average cost of a returned run.
• Hard lever: a single avoided leak-return run pays for the pilot. Maintenance starts acting on data, not hunches, and defensive over-crimping drops.
• Expected reduction in crimping out-of-tolerance units ≥ 30% once the model has its first full shift and the agent starts anticipating wear.

And the operations director's reasonable doubt

“What if the AI rejects good units or, worse, lets a bad one through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely measures a geometry and compares it against a known tolerance, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: doubtful units go through a validation screen before the actuator. The three safety rings exist precisely for this.

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