The leak test bench starts talking (without replacing it)

The problem

The plant's gatekeeper leaves no trace of why it rejects.

The bench decides pass or fail, and the operator writes the reject count in the shift log. The information that really matters — the curve: initial pressure, decay slope, margin against the threshold — never leaves the screen. The consequence is that no root-cause analysis can answer the question every 8D asks: are this week's rejects concentrated in one cavity, one shift, one material reel or one crimping head? The plant knows how many parts failed. It does not know why. The apparent alternative is to replace the bench: a six-figure capex plus a cell shutdown, plus revalidating the test method with the customer. It gets deferred year after year, with good reason, and the information gap stays.

• The leak bench is the only thing between a defective fuel line and a field leak. Many are ten to twenty years old and run on a machine kept off the network by the manufacturer's decision.
• The bench decides pass or fail, and the operator notes the reject count on the shift report.
• The information that actually matters — initial pressure, decay slope, margin to threshold — never leaves the screen.
• So no analysis can answer the question always asked in an 8D: do this week's rejects concentrate in one cavity, one shift, one material reel or one crimping head?

How it fits the IRIS system

Connect in photo mode on the panel — solved from outside, not inside.

Connect in panel-photo mode. IRIS solves the interface-less machine from the outside, not from the inside: no integration with the bench software, no demand for a data output it does not have, and nothing revalidated. Step 1 — Fixed camera. An industrial camera pointed at the panel, triggered in sync with the bench cycle. Step 2 — Grounded reading. Optical recognition interpreted by a grounded model that understands that specific screen: initial pressure, decay, threshold and verdict. Step 3 — MES cross-reference. The time series is cross-referenced with the MES to know which part number, which cavity and which operator were active in that cycle. Step 4 — Correlation. With curve and context in the same memory, the correlation between process parameter and leak result stops being a hypothesis and becomes a query.

It does not integrate with the bench's software, does not touch its validation and opens no ports. On the machine that is the quality gatekeeper, that is the difference between an approvable case and a project quality never authorizes.

See the full IRIS architecture →

Before and after

Today's bench versus the bench being read

Aspect	Today	With iLEAN Connect
The leak curve	Invisible and volatile	Digitized part by part
What survives a reject	A count on the shift report	The curve, the cavity and the lot
Root cause by cavity, shift or material	Impossible	Minutes
Closing an 8D	Weeks	Days
The bench's software	—	Untouched and not revalidated
Replacement capex	In the plan	Deferrable

Impact estimate

Estimated impact — to validate with your own numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-10 months.
• Scrap reduction from the curve-to-parameter correlation that is impossible to make today.
• Avoided replacement capex for the bench.
• And closing an 8D goes from weeks to days, which is where the customer actually measures you.

Estimated payback of 5 to 10 months, counting the scrap reduction the curve-to-parameter correlation delivers and the replacement capex avoided. *Estimate to validate* against the plant's reject history. The saving that is hardest to quantify — and the one a quality director values most — is the shorter response time when a customer demands a demonstrated root cause.

And the fair question from the production manager

«Does a camera reading a panel hold up in an 8D?» — here the task is anchored: fixed screen, fields in known positions and expected physical ranges, where the best models drop below 1.5% error [1]. And the system discards any frame whose value falls outside the possible range instead of accepting it, which is what would invalidate the series in front of a customer.

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