Poka-yoke in engine assembly with AI vision — the right part at every station, without a sensor per variant.

The problem

The classic poka-yoke was brilliant — until the 30 variants arrived.

The classic poka-yoke is one of the most elegant ideas in lean: a mechanical detail that makes it impossible to fit the wrong part. It works perfectly when the line does one thing. A modern engine line does not do one thing — it does 30: different blocks, different brackets, different injectors, different gaskets, different sensors depending on the market. And then the classic poka-yoke breaks in three places:

1. New variant, new sensor — every NPI adds another piece of hardware to the station. Two years later the station is full of sensors that break, get wet, shift with vibration, and in the end people bypass them.
2. Right part, wrong revision — the mechanical sensor cannot tell rev. F4 from rev. F5 of the same bracket. The part fits; the engine should not be carrying it.
3. Variant change on the line — the recipe change does not always reach the station. The person keeps building with the last set-up because the notice never got through, and it surfaces 12 stations downstream.

By the time the mixed-up part reaches the finished engine, it is no longer a few minutes of rework: it is a partial teardown, a batch containment, and sometimes a call from the OEM if the engine has already shipped. The classic system works almost every time. The few times it fails, it fails expensively.

How it fits the IRIS system

iLEAN Vision does not replace the poka-yoke — it extends it to high-mix.

The idea behind the poka-yoke (make the error impossible) is still the right one. What changed is the line: variability went from 3 to 30, and the mechanical solution no longer scales. iLEAN does not ask you to rip out the sensors that work — it adds a vision + agent layer on top, as the putty that seals the gap the classic poka-yoke left exposed when high-mix arrived.

Edge sees what the operator is holding. The agent cross-references the MES recipe and says which engine is running now. If the part is not the one that belongs, the station locks before the bolt goes in. The person signs — never the other way round.

The iLEAN pieces applied to poka-yoke in engine assembly:

• Edge + Vision — a terminal with machine vision (CNN) and controlled lighting at the station. It sees the part in the person's hand (or in the feeder, or on the shelf) before the bolt goes in. It tells fine variants apart: same silhouette, different gasket color, different etching, different dimension. It returns OK/NOK to the station PLC over a dry contact or a bus. It works with no network: the model recipe travels with the part, Edge reads it, and if the plant loses WiFi it keeps verifying.
• Connect — captures the recipe of the model in progress from the MES or from the head-of-line PLC, and syncs to each station which variant is due. It also captures the active part revisions (which revision is valid right now?) from the engineering change list, which very often lives in a PDM folder the station never read.
• Agent — governs the rule for each station: for every engine in progress, which parts it should expect to see, in what order and with what tolerance. If a part does not match, it does not send an email at 10 p.m.: it locks the station, opens the incident with a photo and the reference of the expected part, and alerts the shift lead on whichever channel they use. No bolt ever goes in without verification.

See the full IRIS architecture →

Before and after

Classic mechanical poka-yoke vs. poka-yoke with iLEAN AI vision

Aspect	Classic poka-yoke (mechanical sensor)	With iLEAN Vision + Edge
Cost per new variant	New sensor, installation, wiring, calibration	A capture session in the plant; software, not hardware
Telling rev. F4 from rev. F5 of the same bracket	No (the part fits either way)	Yes (the CNN sees the etching, the agent cross-checks PDM)
Variant-change notice reaching the station	Paper / email — arrives late	MES recipe propagated to the station in seconds
Broken / bypassed sensors	Frequent with vibration and oil	Rugged camera + Edge self-diagnostics
Traceability of which part was fitted	“We assume the right one”	An automatic photo per engine at the moment of assembly
Setup at SKU changeover	Manual, blind window	Automatic reconfiguration from the recipe

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 immersion.

• Engine assembly line with 20-40 active variants and 8-12 critical stations (the ones that concentrate the most mixed-up parts according to the MES Pareto).
• An Edge pilot on 1-2 stations from the Pareto. First value expected within a few weeks: detection above 90% on variants already in series production.
• Indicative payback between 4 and 9 months, depending on the historical frequency of mixed-up parts caught in final assembly and the average rework cost on a finished engine.
• A reasonable reduction of ≥ 30% in escaped mixed-up parts during the pilot; the rollout to the remaining critical stations done by your own trained people, in 2-3 months after validation.

And the quality manager's reasonable doubt

“What if the camera confuses two similar variants?” — visual inspection with the recipe as the anchor is an anchored task, not free generation. On that kind of task, the best models brought error below 1.5% ^[1]. And even so, what is critical is not decided alone: iLEAN locks the station and the person validates. The three safety rings exist precisely for this.

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