EV traction inverter control with AI — the weld the final electrical test cannot see is the one that gets whole populations blocked in the field.

The problem

The final electrical test approves inverters with welds that will fail in the field.

The quality manager on an inverter line knows it even if they do not always say it out loud: the EOL test measures continuity, insulation, dielectric strength, switching — and a marginal busbar weld passes every one of them. A bead with borderline penetration or with spatter conducts perfectly on test day; it fails months later, with thermal cycling and vibration, in a high voltage vehicle out on the road.

Three things happen at once on a real inverter line and hardly any plant has them solved together:

1. Weld quality is not 100% verifiable at final test — the EOL confirms that the inverter works today, not that every busbar bead and every power module (SiC/IGBT) joint will survive the life of the vehicle. Sample-based visual inspection lets the rest through unlooked at.
2. Module↔inverter traceability lives in spreadsheets — the OEM requires knowing which power module serial goes inside each inverter, and that genealogy is maintained with loose scans and manual typing. Every keystroke is a transposed serial waiting for the audit.
3. A high voltage field failure forces you to scope with a blunt instrument — with no fine genealogy, the only defensible answer for the customer is to block everything built between two dates. Whole populations held, reverse logistics, weeks of engineering rebuilding paperwork.

The result is always the same: a defect born in one specific bead at one specific station ends up costing a massive scoping exercise, and the knowledge of “which units are genuinely suspect” does not exist in any system — it has to be rebuilt by hand, late and with a margin of error.

How it fits the IRIS system

iLEAN does not replace the MES or the EOL test — it stitches what today lives on islands.

The inverter's problem is not a lack of data: the welding station has its parameters, the EOL test has its measurements, the modules carry their serial. The problem is that none of it is stitched together unit by unit, and the missing piece — 100% visual verification of the bead — does not exist. iLEAN acts as the putty that joins welding, serials and test under each inverter's serial, without asking you to change the MES or the test bench.

Edge inspects every weld right after the process. Connect stitches serials and EOL test with no typing. JIDOKA AI holds on drift. Agents maintain the genealogy that scopes an incident in hours.

The four iLEAN pieces applied to the traction inverter:

• Edge — machine vision at the station, inspecting every busbar weld on 100% of units right after the process: bead geometry, continuity, spatter, pores. It works locally: if the plant loses its network, Edge keeps inspecting, recording and holding. What is critical does not depend on WiFi.
• Connect — stitches every power module serial to the inverter serial automatically, with the scan built into the station flow (not an extra step, no typing), and captures the EOL test results, hanging them off the same serial. The traceability spreadsheet disappears because it is no longer needed.
• JIDOKA AI — when bead geometry starts to drift (even though each unit is still within individual tolerance), it holds the affected units at the station instead of letting them move on to final test. It proposes the likely cause; the person decides whether to rework, adjust or release.
• Agents — keep the genealogy alive: module → inverter → welding evidence → EOL result → date and shift. When a field incident hits, they generate the exact list of affected units and the customer dossier in hours, not weeks.

See the full IRIS architecture →

Before and after

Classic inverter control vs. control with iLEAN

Aspect	EOL test + traceability spreadsheets	With iLEAN Edge + Connect + Agents
Busbar weld verification	Visual sampling, the rest goes through unlooked at	100% vision after the process, evidence per serial
Module ↔ inverter traceability	Loose scan + typing into a spreadsheet	Stitched automatically in the flow, no typing
EOL test results	On the bench, disconnected from the serial	Captured and hung off the genealogy
Welding process drift	Discovered at test or in the field	JIDOKA AI holds at the station
Scoping a field incident	Weeks, massive blocked population	Hours, exact list of affected serials
Dossier for OEM / IATF audit	Rebuilding spreadsheets and paperwork by hand	Cross-referenced genealogy, 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.

• EV traction inverter assembly line with laser busbar welding, SiC/IGBT power modules carrying their own serial, an EOL test at end of line and module↔inverter traceability kept today in spreadsheets.
• Edge vision pilot at the welding station + Connect stitching serials and EOL test. First value expected within a few weeks: the automatic genealogy and the 100% inspection work from day one; the process drift pattern sharpens later, with history behind it.
• Scoping a field incident in hours instead of weeks, with an exact list of serials — the blocked population goes from “everything built between two dates” to the units genuinely affected.
• Busbar welding verified on 100% of units, with visual evidence linked to every serial, defensible in front of the OEM both in audit and in an incident.
• Indicative payback between 5 and 12 months, estimate to be validated. The hard lever: a single field scoping exercise done properly — a minimal population instead of a massive one — can pay for the whole project.
• A recurring benefit that does not enter the ROI but carries weight: the fine genealogy stays as a permanent capability of the plant, not as a spreadsheet that depends on whoever maintains it.

And the quality manager's reasonable doubt

“What if vision flags a good bead as bad and holds units that were fine?” — the iLEAN system does not decide the release on its own. JIDOKA AI holds and proposes; the line manager reviews the evidence and signs. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI compares bead images against a pattern and cross-references serials with test results, the best models brought error below 1.5%^[1]. And even so, what is critical goes to the safety rings — the AI lives in the outer ring, proposes inward, and the release of a held unit is signed by a person. Never the other way round.

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