Void detection in XLPE insulation with AI — a dielectric failure is not a warning, it is a breakdown.

The problem

A void in XLPE is invisible to the eye and catastrophic in service.

The XLPE insulation on an MV cable is like good cheese: either it comes out homogeneous, or it is worth nothing. A void, a bubble or an inclusion is a point where the electric field concentrates — that is where the partial discharge starts that eats the insulation from the inside until it fails on site, months or years later. And the defect slips through for three reasons that almost always come together:

1. Different raw material with every batch — XLPE pellets vary in residual moisture, melt flow index and cross-linking agent content. What came out perfect yesterday can gas up today with exactly the same recipe.
2. Line speed against quality — sales pushes for more meters per hour; the quality manager pays for it when a run goes to the final test and bounces.
3. Inspection too late — the dielectric test and the partial discharge test happen at the end, when you have already made kilometers of cable. If the run does not pass, all of that is scrap with copper inside.

The plant manager knows this, but between the extruder and the final test the insulation is already cured and opaque — the defect becomes visible once it can no longer be touched. The classic system works 99% of the time. That 1% is what reaches the end customer as a switchboard failure — and costs orders of magnitude more than the whole run.

How it fits the IRIS system

iLEAN Vision does not replace your dielectric test — it seals the gap between the extruder and the lab.

The problem with a void in XLPE is not a lack of measurement technology — the partial discharge test does its job perfectly. The problem is that the information arrives too late and, by the time it does, the copper is already inside the defective insulation. iLEAN acts as the putty that fills that gap between the extruder and the lab, without asking you to change your CV line, your MES or your test protocol.

Vision watches the XLPE as it comes out. Connect reads the extruder parameters wherever they live. The agent cross-references pellet moisture, pressure and speed with the pattern learned in your own plant, and holds the run when something does not add up. The person signs — never the other way round.

The iLEAN pieces applied to void detection in XLPE insulation:

• iLEAN Vision (Edge with CNN) — a box with high-speed cameras and calibrated lighting, installed right at the exit of the insulation crosshead, before the curing tube. It watches the insulation while it is still hot and semi-translucent, detects voids, bubbles and discontinuities, and sends a signal to the PLC to mark or hold the run. It works with no network. If the plant loses WiFi, Edge keeps watching.
• Connect — captures the line parameters whether they come from a modern PLC, an old HMI or a control sheet the shift lead updates every hour. And it also captures what arrives from outside: the technical data sheet from the XLPE pellet supplier, the residual moisture of the latest batch, alarms from the curing tube, all at second zero.
• Agent — cross-references the image of the insulation, the extrusion parameters and the defect pattern learned in your plant. If there is a deviation, it does not send an email at 10 p.m.: it alerts the quality manager through whichever channel they use and, if the configured level of autonomy allows it, holds the run before it enters curing. The person validates and signs.

See the full IRIS architecture →

Before and after

Manual XLPE insulation inspection vs. inline AI vision with iLEAN.

Aspect	Classic inspection	With inline iLEAN Vision
Moment of detection	Final electrical test, off line	At the crosshead exit, before curing
Material at stake when it is detected	Kilometers of cable with copper inside	Meters, while the line can still be adjusted
Pellets with out-of-range moisture	Discovered when the run gases up	Cross-referenced with the batch data sheet at second zero
Change of recipe or XLPE supplier	A silent risk until the final test	Pattern learned per plant, immediate alert
Operation with no network	n/a	Edge keeps watching with its own lighting
Dossier for auditor / end customer	Rebuilt by hand, weeks	Photo and metrics per reel, automatic

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.

• Medium-voltage cable plant, one CV (continuous vulcanization) line with an XLPE insulation extruder and a dry or nitrogen curing tube.
• iLEAN Vision pilot on one line: a box with cameras + lighting + hold actuator + integration with the extruder's PLC. First value expected within a few weeks, in assist mode (it alerts the shift lead), then moving up to automatic stopping once the quality manager gives the go-ahead.
• Indicative payback between 4 and 9 months, depending on the partial discharge rejections documented over recent years and the average cost of an MV cable drum rewound or destroyed (copper included).
• Expected reduction in defective insulation scrap of ≥ 30%, with a conservative floor: the hard lever is avoiding a single long run rejected at the final test.

And the plant manager's reasonable doubt

“What if the AI gets it wrong and lets a bad cable through or, worse, stops a good one?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares an image of the insulation against the pattern learned in your own plant, the best models brought error below 1.5%^[1]. And even so, what is critical is not decided alone: iLEAN marks and holds, the person signs. The three safety rings exist precisely for this.

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