Vision inspection of MAG welds on HEB profiles — the defect that slips through the shop costs ten times more on site.

The problem

A defective bead in the shop is a touch-up. On site it is a crane standing idle.

Visual inspection of 100% of the MAG bead in EXC2 or EXC3 steelwork is a job of gauges, a flashlight and eyesight. When it goes well, it goes well. When there are three beads to check before the truck leaves at five, the part that always fails is the one that fails — inspector fatigue. The defect not caught in the shop reaches the site in one of three ways, all of them expensive:

1. The engineer of record spots it during erection — the piece is rejected with the crane already hooked up, back to the shop, the profile stuck on site.
2. The 10% NDT uncovers it at the end of the phase — forcing an expanded sample, a batch on hold, several pieces reworked.
3. It shows up years later — fatigue cracking starting at an undetected undercut. That one is the worst case of all.

The welder is good. The RWC is good. The classic system works 95% of the time — and the other 5% are the odd days: end of shift, a WPS change, a new wire spool, a piece inspected in a hurry because the truck is waiting. What costs money is not the defect: it is not having caught it in the shop.

How it fits the IRIS system

iLEAN does not replace the RWC — it gives him a flashlight that never tires and a notebook that fills itself in.

The problem with visual inspection is not one of judgment: the responsible welding coordinator knows perfectly well what counts as unacceptable undercut and what does not. The problem is one of constant coverage and archived evidence. Nobody can inspect 100% of the bead with the same patience on the first profile of the shift as on the last. And nobody can reconstruct, three years later and unaided, exactly which bead was inspected and with which flashlight. iLEAN acts as the putty that seals that crack between human visual inspection, the applicable WPS, the profile's traceability and the EN 1090 dossier — without asking you to change the table, the robot or the ERP.

Vision runs along the bead with the same patience at six as at eight. Connect captures the WPS, the wire batch and the profile's piece mark. The agent cross-references, proposes "accept" or "rework" and prepares the dossier. The person signs.

The three iLEAN pieces applied to MAG welding of HEB profiles:

• Vision (Edge) — a CNN-equipped camera over the inspection table, next to the welder or at the end of the robotic gantry. It segments the bead and marks undercut, clustered porosity, excess weld metal and spatter with coordinates, according to the declared quality level (B/C/D). It works with no network: if the plant loses WiFi, Vision keeps looking and flagging.
• Connect — captures the applicable WPS, the wire batch and spool, the gas, the welder's WPQR or the robot program, and the profile's erection mark (everything that normally lives somewhere between the inspection plan, the ERP and the shop manager's head).
• Agent — cross-references the image, the WPS and the profile's traceability, proposes accept or rework, assembles the EN 1090 visual inspection dossier with the marked image and the context, and hands it to the responsible welding coordinator for signature. The structure does not go out to site with a profile whose bead Vision has flagged without a human decision first.

See the full IRIS architecture →

Before and after

Visual inspection with a flashlight vs. cross-checked inspection with iLEAN Vision

Aspect	Flashlight + gauge + the RWC's eye	With iLEAN Vision + Connect + Agent
Coverage	100% mandatory for EXC2/EXC3, subject to fatigue	100% covered by Vision, same patience all day long
Defect types	The inspector's eye + EN ISO 5817 criteria	An eye trained on samples from your shop + EN ISO 5817
Traceability of WPS/wire batch/welder	Noted on a report if someone asked for it	Linked to the bead automatically
Marked image of the defect	If there is one, a loose photo on a phone	Coordinates, size and type, archived with the bead
EN 1090 dossier for the RWC	Rebuilt by hand	Automatic pack per mark/profile
Operation with no network	n/a	Vision keeps running on the panel's power alone

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

• An EXC2/EXC3 steel fabrication shop with manual MAG welding on HEB profiles and robotic gantries for repetitive connections.
• Vision pilot on one inspection table and one robotic gantry (camera + lighting + Edge + integration with the inspection plan and the piece-mark ERP). First value expected within a few weeks: consistent detection of the surface defect most frequent in your production.
• Indicative payback between 4 and 9 months, depending on the documented frequency of touch-ups on site, the RWC's hours per week spent on inspection and the average cost of re-welding a profile at height.
• The hard lever is a single defective piece kept off the site (cutting, re-preparation, re-welding with the crane standing idle), plus a reduction ≥ 30% in the RWC's hours spent on routine visual inspection, which shift to reviewing what Vision flags.

And the responsible welding coordinator's reasonable doubt

"What if the AI flags a bead as defective when it actually meets EN ISO 5817 level B?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply compares an image against samples trained on your own shop's criteria, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: Vision flags, the RWC signs. The three safety rings are there precisely for this.

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