The o-ring nobody sees: Edge vision at the assembly station

The problem

The poka-yoke cannot tell what a camera can.

O-ring assembly is manual or semi-automatic and its verification is visual. Poka-yokes exist, but they do not cover partial seating or a twisted ring: geometries a mechanical fixture cannot distinguish. Rejects that reach the leak test cost one bench cycle each, and the bench is usually the cell's bottleneck. Worse: if the defect is marginal, the cold test can pass it — and then the part leaves the plant. The cost of a leak caught at the customer versus the same leak caught at the station is three orders of magnitude apart. And under thermal cycling and vibration, partial seating is exactly the failure mode that shows up late.

• O-ring missing, twisted, cut, doubled or badly seated in its groove: the part looks perfect, assembles perfectly, and fails at sealing.
• Poka-yokes exist, but they do not cover partial seating or a twisted ring: geometries a mechanical fixture cannot distinguish.
• At cadence, a human eye checking thousands of couplings per shift lets the borderline cases through — and the borderline case is precisely the one that fails.
• Rejects reaching the leak test cost one bench cycle each, and the bench is usually the cell's bottleneck.
• Worse: if the defect is marginal, the cold test can pass it, and then the part leaves the plant.

How it fits the IRIS system

Edge — part-by-part classification in milliseconds, before closing.

Edge: AI vision with local inference, the IRIS component that works at line rate and without depending on the network. Step 1 — Camera and lighting. An industrial camera with controlled illumination over the o-ring groove, mounted before the joint closes and before the leak test. Step 2 — Local inference. The neural network runs on the device itself, in milliseconds. The network can go down and the station keeps deciding. Step 3 — Trained on real parts. It is trained on conforming and non-conforming parts from the plant's real part numbers, including the rare defects quality has already collected, and classifies: present and seated, missing, twisted, doubled or damaged. Step 4 — Ejection with evidence. On a non-conforming part, automatic ejection and a record with the image that proves it. Root-cause analysis stops being an exchange of opinions.

Catching it at the previous station instead of at the bench frees cycles on the cell's bottleneck. That side effect is often worth as much as the scrap avoided, and it almost never makes it into the business case.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	Depends on fatigue	Objective and constant at cadence
Partial seating and twisted ring	The poka-yoke cannot see them	Classified part by part
Point of detection	The leak test	The previous station
Bench cycles spent on rejects	One per bad part	Freed
Marginal defect	The cold test can pass it	Caught earlier
Borderline cases	They slip through	Escalated to a person

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 scrap and rework reduction of at least 30 % in the covered family.
• Estimated payback 5-11 months.
• Leak bench cycles freed: if it is the cell's bottleneck, that is capacity.
• And the marginal defect the cold test could pass stops leaving the plant.

Estimated reduction in scrap and rework of at least 30% in the covered family, with estimated payback of 5 to 11 months. *Estimate to validate* against the plant's reject history and mix. What pays for the whole project, though, is the cut in field-leak risk.

And the fair question from the production manager

«What if it rejects good parts?» — the false positive is the real risk of any vision system, which is why the model is trained on good and bad parts of your specific references and with that station's lighting, not on a generic model. Borderline cases are not simply rejected: they are escalated for a person to decide, and every decision feeds back into training.

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