PU sole defect detection in safety footwear with AI — the bubble shows up at demolding, not at the customer.

The problem

Three families of defects the carousel never warns you about.

Injecting a PU sole onto an already lasted upper is a beautiful process to watch and a hard one to govern: the polyol/isocyanate system changes with humidity, the pigment arrives from a poorly homogenized drum, the insole activates or not depending on mold temperature, and the carousel carries 18 stations of which only one is emptied at a time. The typical defects:

1. Bubbles and porosity — moisture in the system, incomplete degassing. They come out as small craters on the sidewall or around the gate.
2. Insole debonding / poor upper-to-sole adhesion — mold temperature drifts, the adhesive does not activate, the upper separates from the outsole at the customer's first wash.
3. Off-shade color — the first pairs after a color change come out with pigment carryover, or the whole batch comes out darker than it should because the drum was not stirred.

The quality lead knows it, but the carousel does not stop for cosmetics and the final check is saturated. The classic system (inspector's eye + sampling + a complaint later) works 99% of the time — and that 1% is the complaint from the professional retailer with the batch returned, reverse logistics, a mark against your quality and, if it piles up, removal from the supplier list.

How it fits the IRIS system

iLEAN does not stop your Desma — it puts eyes on every demolding station.

The problem with PU sole defects is not a lack of technology in the injection machine: it is information living in islands (the machine setpoint, system humidity, the pigment batch, the inspector's eye) that at the critical moment (the first output after a color change, the last hour of the night shift) does not reach the person deciding in time. iLEAN acts as the putty that fills those gaps.

The camera looks at every pair at demolding. The agent cross-references it with mold temperature and pigment batch. The suspect piece is set aside before stitching or packing. The person signs.

The three iLEAN pieces applied to PU sole defect detection:

• Edge (Vision) — a terminal with a CNN camera over the carousel's demolding station. It detects bubbles, porosity, insole debonding, badly trimmed flash and off-shade color. It classifies by severity and, if the quality plan allows it, fires an actuator (stack light, ejector to the rework lane). It works with no network.
• Connect — captures the injection machine setpoint (mold temperature, pressure, closing time), the polyol/isocyanate batch and the pigment batch. It also captures the professional retailer's quality plan (what is acceptable and what is not) at second zero, with nobody having to forward it.
• Agents — cross-reference the visual inspection with the mold temperature of that specific station, the raw material batch and the injection machine history. If the combination points to process drift (not just a one-off piece), they propose a setpoint adjustment to the shift lead and assemble the per-batch dossier with visual samples. The person decides the correction; the system does not touch the injection machine.

See the full IRIS architecture →

Before and after

Final check vs. per-station AI inspection with iLEAN Vision

Aspect	Final check + sampling + complaint	With iLEAN Vision + Connect + Agents
Coverage	A sample of the batch, inspector's eye	100% of the pairs, every station
Bubbles and porosity	Detected at the end, sometimes at the customer	Flagged at demolding, before stitching
Color changeover	First pairs with pigment carryover	High sensitivity during the change, normal afterwards
Insole debonding	Detected at the customer's first wash	Detected in line through the visible interface
Operation with no network	n/a	Vision keeps inspecting on cabinet power
Per-batch dossier for the professional retailer	Rebuilt by hand	Per batch, automatic, with a photo of every rejection

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 diagnostic.

• Safety footwear plant with 1-3 PU injection carousels (typically 18-24 stations), multi-color and multi-model production, professional retailer as main customer.
• iLEAN Vision pilot on one carousel: CNN camera at the demolding station + integration with the injection machine control and the ERP. First value expected in a few weeks (bubble and off-shade color detection above 90% on known pairs).
• Indicative payback between 4 and 9 months, depending on the history of cosmetic complaints, the average cost of a batch return and how often pairs are discarded “on suspicion” today with no data behind it.
• Hard levers: ≥ 30% reduction in complaints over cosmetic defects, recovery of pairs discarded without cause, and per-batch/per-station traceability that closes the professional retailer's audit with no argument.

And the shift lead's reasonable doubt

“What if the AI mistakes a normal mold mark for a bubble?” — hallucination is a problem of free generation, not of anchored tasks. Classifying an image against a pattern trained on your own carousel is an anchored task. In that family of tasks, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN classifies and sets aside, and the inspector signs before the piece goes back into the flow. The three safety rings exist precisely for this.

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