Laser marking verification on circuit breakers — one wrongly engraved digit is a different product, and another claim.

The problem

The laser marks whatever it is told to — and nobody checks that the marking matches the batch.

On a modular MCB line, the laser puts several pieces of data on every part that define the product electrically:

1. Rated current In and tripping curve (B, C, D, K, Z) — the combination that determines when the breaker trips.
2. Breaking capacity Icu and rated voltage — the limits above which the breaker no longer guarantees protection.
3. Commercial part number, batch, CE marking and, in many markets, local approval symbols.

When a whole shift runs a single SKU, human control works. The problem appears at the SKU changeover: the laser operator loads the new job, the line operator changes the packaging, and between the last part with the old marking coming out and the first part with the new marking entering the new packaging, there are minutes in which any slip can get through. And if on top of that the marking contrast is poor (a laser out of adjustment, a dirty lens, a misaligned part), the part passes the outgoing inspection because “you can read it”, reaches the customer half blurred, and a claim follows.

Classic control works 99% of the time. That 1% is the claim, the returned batch, the rework cost and, at worst, an installed MCB that does not protect the way its label says it does.

How it fits the IRIS system

AI vision looks at every part the way we would — but without getting tired at part 7,000.

Correct marking is the intersection of three pieces of data that live on islands: the job loaded in the laser, the batch SKU in the MES/ERP and the physical part coming out of marking. iLEAN acts as the putty between the three, without asking you to change the laser or rewrite the MES.

Vision reads the marking on every MCB. The agent cross-references it with the expected SKU. If it does not match, it rejects the part before packing — the line keeps running.

The iLEAN pieces applied to laser marking verification on MCBs:

• Vision (Edge specialized in AI vision) — a CNN-equipped terminal between the laser marking and the packing station. It reads every marked field on each part (In, Icu, curve, part number, batch, symbols) and cross-references them with the expected SKU. Works with no network: if the plant loses WiFi, Vision keeps reading, rejecting and storing the trace locally.
• Connect — captures the expected batch SKU from wherever it lives: a vertical MES, the ERP, the shift lead's planning sheet or, on older lines, a photo of the laser's panel. And also whatever arrives from outside (a customer order change that comes in by email at 8:00 and modifies the 8:01 batch).
• Agent — on top of the Vision+Connect flow, it validates the completeness and legibility criteria of the marking, builds the per-batch dossier with a photo of every part, and raises alerts when the reject rate for poor contrast climbs (a sign that the laser needs servicing).

See the full IRIS architecture →

Before and after

Human inspection vs. inspection with iLEAN Vision

Aspect	Inspection today	With iLEAN Vision + Agent
SKU changeover	Laser operator + line operator	Vision verifies every part against the expected SKU
Wrong marking (curve, In)	Found at the customer	Rejected before packing
Poor contrast / blurred marking	“You can read it, it passes”	Minimum legibility criterion; holds if it drops below
Laser servicing	Fixed periodic plan	Alert when the hold rate for contrast climbs
Defense against a claim	“We had inspected it”	Signed photo of the marking on every part
Operation without network	n/a	Edge keeps running on the light in the panel

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.

• A modular MCB line, multi-curve (B/C/D), multi-In, multi-Icu, running several thousand parts per shift with a handful of claims for wrong marking or poor contrast over the past few years.
• Vision pilot at end of line between laser and packing (camera + lighting + reject actuator + terminal + integration with MES/ERP). First value expected within a few weeks: detection above the human baseline at SKU changeover from day one.
• Indicative payback between 4 and 9 months, depending on the number of SKU changeovers and the average cost of a claim / return.
• Expected reduction in marking-related claims of ≥ 30% against the baseline — conservative.
• A sector with a demanding quality standard (on the order of 25 PPM in automotive ^[2]); in MCBs for industry, where voltages are critical, the cost of a mismarked part in the wrong place weighs far more than the part itself.

And the quality manager's reasonable doubt

“What if the AI rejects parts that were fine?” — hallucination is a problem of free generation, not of anchored tasks. Reading marked characters and comparing them against an expected SKU is exactly an anchored task: the best models brought error below 1.5% ^[1]. And even so, Vision does not sign: it rejects the part, drops it in the review bin, and the operator decides at the end of the shift. A false positive costs minutes; a false negative costs a claim.

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