AI manufacturing control of the on-board charger (OBC) — out-of-torque tightening does not fail at EOL; it fails months later, in the field.

The problem

The assembly defect the EOL test cannot see — and that shows up months later, in the vehicle.

The OBC is one of the most thankless modules to build in electrification: it mixes power (busbars with critical tightening torques), high-voltage electronics and thermal dissipation (thermal paste or gap filler whose dispensing stops being visible the moment the cover closes). The end-of-line test checks that the unit charges and insulates — today. It does not check that in eight months, after thousands of thermal cycles, it still will.

Three things happen at once on a real OBC line, and almost no plant has all three solved together:

1. Out-of-torque tightening does not fail at EOL — a busbar with low torque makes good contact during the test and works loose with thermal cycling in the field: contact resistance, heating, failure. The nutrunner records the curve, but nobody ties it to the unit's serial number.
2. Thermal dispensing cannot be seen once the unit is closed — a discontinuous or displaced thermal paste bead leaves a power module poorly cooled. The unit passes EOL cold and degrades in the field. Destructive sampling arrives late and covers few units.
3. Component traceability is done by hand — the serial numbers of the power modules and the transformer are contractually required, and on many lines someone transcribes or scans them into a spreadsheet. When the field incident arrives, reconstructing which units carry which batch takes weeks.

The result is always the same: the unit leaves the plant with a green EOL, the failure shows up months later in the vehicle, and only then does the plant discover it cannot prove what torque that screw got or scope how many more units share the problem. The cost is not the returned module — it is the campaign that cannot be scoped.

How it fits the IRIS system

iLEAN does not replace your nutrunners or your EOL test — it ties them to the serial number.

The OBC problem is not a lack of equipment: the electronic nutrunners already record torque and angle, the EOL test already gives results, the components already carry serial numbers. The problem is that every piece of data lives on its own island and nobody brings them together per unit. iLEAN acts as the putty that binds nutrunners, vision, EOL test and component serial numbers into a single file per OBC, without asking you to change a single piece of equipment on the line.

Connect integrates what already exists. Edge verifies at 100% what stops being visible once the unit closes. JIDOKA AI holds the unit if something does not add up. Agents keep the genealogy so you can answer in hours. The person decides.

The iLEAN pieces applied to OBC manufacturing control:

• Connect — integrates the electronic nutrunners (torque and angle of every critical screw, tied to the unit's serial number), the existing EOL test and the serial readers for power modules and transformer. The traceability that is transcribed by hand today starts building itself, unit by unit, without changing the equipment.
• Edge — a local terminal with vision that verifies, before closing, the dispensing of thermal paste or gap filler (continuity, width, position) and the presence and orientation of the components, on 100% of units. It runs locally: if the plant loses the network, Edge keeps inspecting, recording and holding. What is critical does not depend on WiFi.
• JIDOKA AI — if a torque value falls outside the band, vision flags irregular dispensing or a serial number does not match, the unit is held before moving on to closing. The alert reaches the supervisor with the evidence; the person decides whether to rework, deviate or release. The system never releases a held unit on its own.
• Agents — keep the complete genealogy of every serial number (components, tightening curves, image of the dispensing, EOL result, shift, line) and answer the critical question when a field incident hits: which other units share the same condition. The answer goes from weeks to hours.

See the full IRIS architecture →

Before and after

Classic OBC line vs. control with iLEAN

Aspect	Line with EOL + scattered records	With iLEAN Connect + Edge + Agents
Torque and angle per screw	In the nutrunner, never cross-referenced with the unit	Tied to every OBC's serial number
Thermal paste dispensing	Invisible once closed, destructive sampling	100% vision before closing, image per serial number
Component position before closing	Operator's eye, no evidence	Verified and recorded on every unit
Traceability of power modules and transformer	Manual transcription, error-prone	Serial number read and integrated automatically
Unit with something outside the band	Moves down the line, caught (or not) by the EOL test	Held by JIDOKA AI before closing
Response to a field incident	Weeks rebuilding records	Complete genealogy in hours

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.

• OBC assembly line with electronic nutrunners that record torque and angle, thermal paste or gap filler dispensing, an end-of-line test and component serial traceability that is contractually required but kept by hand.
• Connect pilot over your existing nutrunners and EOL test + Edge with vision at the dispensing station. First value expected within a few weeks: the file per serial number and the first JIDOKA AI holds appear before the history is complete.
• Expected reduction of ≥30% in field failures attributable to assembly causes (out-of-torque tightening, irregular dispensing, mispositioned component), an estimate to be validated against your current return rate.
• Response to a field incident with the complete genealogy of the serial number in hours, not weeks — and a campaign scoped to the units that really share the condition.
• Indicative payback between 5 and 12 months, an estimate to be validated. The hard lever: a field failure on a high-voltage component costs orders of magnitude more than the module — reverse logistics, analysis, campaign, customer penalties.

And the quality manager's reasonable doubt

“What if the AI holds good units and stalls my line?” — iLEAN's JIDOKA AI never decides the fate of a unit on its own. It holds with evidence (tightening curve, image of the dispensing, serial number); the supervisor decides whether to rework, deviate or release. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI compares a measured torque against a band or an image against a dispensing reference, the best models brought error below 1.5%^[1]. And even so, what is critical goes through 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.