Powder coating defects with AI — a crater found late is scrap; found on the line, it is nothing.

The problem

Coating quality only shows at the end — and at the end it can no longer be fixed.

Classic quality control on a tier-1 powder coating line works like this: the operator at the end looks at the part under the inspection booth light, marks what they see, and it gets scrapped. The problem is that by then:

1. The coating is already cured — before the oven the powder could be reclaimed, the part blown off and repainted. After the oven, neither the part nor the powder can be recovered: real scrap, in kilos and in euros.
2. The defect may have been repeating for a while — dozens of parts can go by between the moment it starts and the moment it is seen. The root cause (a change in powder viscosity, dirt on the overhead conveyor, a drifting applicator parameter) is still active.
3. The OEM demands 25 PPM — the automotive quality standard is in the order of 25 parts per million ^[1]. At that level of demand, every false positive costs and every false negative gets seen.
4. The knowledge walks out with the veteran — the people who can read a part as it goes past are always one or two experienced operators. The day they retire, part of the control goes with them.

The quality manager knows it, the plant manager lives with it, the OEM's buyer bills for it. The classic setup — operator + controlled light + a trained eye — works 99% of the time. That 1% is what reaches the OEM, and at this level of demand that 1% is enough.

How it fits the IRIS system

iLEAN Edge does not replace the veteran operator — it captures their eye as a permanent capability.

The problem with powder coating control is not information: it is the capacity to look at every part, all the time, with the same attention as in the first minute. And no person, however good, can sustain that. iLEAN acts as the putty that fills the gap between the veteran's expertise and full coverage of the line, without asking you to change your applicator, your oven or your PLC.

Edge sees the part before the oven and at the exit. Connect reads the old applicator without having to replace it. The agent cross-references root cause and work order. The person decides the corrective action. The warm bed: the shift walks in and finds everything already sorted.

The three iLEAN pieces applied to automotive powder coating:

• iLEAN Edge (Vision) — physical terminals with a CNN trained on your parts, installed before the oven and at the exit. They detect craters, sags, orange peel, seeds and coverage gaps. They fire an actuator in 45 ms (stack light, ejector, dry contact to the PLC). They work with no network: if the plant loses its internet connection, the critical detect-and-intervene cycle carries on for as long as the device has power.
• iLEAN Connect — reads the old applicator with no need to replace it (graded capture: manual, intermediate, integrated, depending on the age of the machine). And it captures what comes in from outside (a powder batch change from the supplier, a new technical data sheet, a WhatsApp from the maintenance tech) at second zero.
• iLEAN Agent — cross-references the detected defects with the applicator, the raw material, the overhead conveyor and the oven. When a pattern repeats (seeds every time powder viscosity drifts), it does not send an email at 10 p.m.: it prepares the work order for the maintenance tech and the adjustment proposal for the coating supervisor. The person signs — Agents have no hands on critical OT.

See the full IRIS architecture →

Before and after

Manual end-of-line control vs. digital jidoka with iLEAN Edge

Aspect	Manual control + final booth	With iLEAN Edge + Vision + Agent
When it is detected	End of line, coating already cured	Before the oven + at the exit, inside the cycle
Recovering part and powder	Impossible — part goes to scrap	Blow off and repaint before the oven
Coverage	100% of the booth, real % lower through fatigue	100% of every part, no fatigue
Root cause detection	Later analysis, days	Pattern cross-referenced with the applicator in real time
Operation with no network	n/a	Edge keeps detecting on panel power
Defect that reaches the OEM	Whatever slips through in the 1% — expensive	Drops to a defensible level — > 98% caught on the pilot line

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 line. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Powder coating line for automotive parts (subframes, cross members, exhaust pipes, leaf springs), with defect scrap at 2-5% and an OEM requirement in the order of 25 PPM.
• Edge pilot before the oven + cameras at the exit + integration with the ejection actuator and the line PLC. First value expected within a few weeks: detection of the most frequent defects above 90%.
• Indicative payback between 3 and 6 months, depending on your current scrap cost, the cost of wasted powder and your exposure to an OEM complaint. The real pilot line closed at roughly 3 months; the wide range reflects plant-to-plant variability.
• A reduction in defects reaching the end of the line of ≥ 50% is achievable within 60 days; with root cause discipline from the coating team, it can reach > 95% as in the textbook case.

And the plant manager's reasonable doubt

“What if the AI gets it wrong and lets a serious defect through to the OEM?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI recontextualizes a piece of data (looking at a part and saying whether it has a crater), the best models brought error below 1.5% ^[2]. And even so, the operator is still there: Edge acts as a safety net, not as a replacement. The operator's warm bed is knowing that a hundred invisible agents secured the shift; theirs is the decision that matters.

^[1] Symestic — automotive quality standard in the order of 25 PPM (parts per million).
^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.