The machine stays itself, but now it speaks

The problem

The parameters that explain the crack do exist. They last one cycle.

Flash butt welding is the most expensive structural failure point in the product. The panel displays exactly what would be needed to correlate parameter with defect and with test rejection, but that data is overwritten every cycle. When a lot fails radial fatigue testing, or a crack shows up at the customer, nobody can go back and check the upset factor that specific rim was welded with. Root cause investigation runs blind, and the prudent reaction — blocking a whole week of production — is the most expensive one. The electrode also wears progressively, and its drift is invisible until the weld no longer conforms. Maintenance is scheduled by calendar because no other signal is available. And the obvious way out is blocked: replacing the welder costs over half a million euros and weeks of downtime on a line with no inventory buffer. That is why it has been sitting in the capital plan for years without leaving it.

• The flash butt weld is the most expensive structural failure point in the product: this is where the trapped oxides and inclusions that end in rim cracking and fatigue failure come from.
• The panel shows flashing current, upset time and length, pressure and electrode displacement — exactly what would be needed. And it overwrites all of it every cycle.
• When a lot fails the radial fatigue test, nobody can go back and check what upset factor that particular rim was welded with. Root cause is investigated blind.
• The prudent reaction — blocking a whole week of production — is the most expensive one, and it is the only one available without data.
• The electrode wears progressively and its drift stays invisible until the bead no longer conforms. Maintenance runs on the calendar because there is no other signal.

How it fits the IRIS system

Connect in photo mode on a machine with no interface — and without touching its qualification.

Connect in photo mode over an interface-less machine panel.

The obvious alternative is blocked by its price: replacing the welder runs above half a million euros and weeks of downtime on a line with no stock buffer. An external camera costs three orders of magnitude less and leaves the machine's qualification status untouched.

• A fixed industrial camera pointed at the panel, with optics suited to its flicker and screen rotation.
• Optical recognition with a grounded model that interprets each frame and structures the time curve per rim and per lot.
• Cross-referenced with the manufacturing system — which part number and which heat was on the machine at each instant — and with electrode replacement history in the maintenance system.
• No new controller, no change to the original control system and no impact on the machine's compliance status.

See the full IRIS architecture →

Before and after

Replacing the welder versus reading its panel

Aspect	Replacing the machine	With iLEAN Connect
Investment	Above 500 K€	Industrial camera and commissioning
Line downtime	Weeks	None: installed while running
Machine qualification	Starts over	Untouched: original control not modified
Parameter-to-crack correlation	Yes, from zero history	Yes, from the first captured cycle
Electrode replacement	By calendar	By observed drift
Scope of a containment	One week of production	The suspect lot, bounded

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 payback 5-10 months, combining three effects that do not overlap.
• Deferred replacement capex, above 500 K€, taken out of the investment plan.
• Reduced rim scrap attributable to weld parameter drift.
• The ability to bound the suspect lot instead of blocking a full week of production.

estimated payback 5-10 months. It combines deferred replacement capex (above 500 K€), reduced rim scrap attributable to parameter drift, and the ability to bound lots instead of blocking them. *Estimate to validate.*

And the fair question from the production manager

«Isn't a camera reading a screen a fragile workaround?» — it would be if the model were asked to interpret an arbitrary image. Here the task is anchored: fixed screen, fields in known positions, expected ranges. In that regime the best models drop below 1.5% error [1], and on top of that the system discards any frame whose value falls outside the physically possible range instead of accepting it.

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