The camera that never gets tired on the line

The problem

At assembly cadence, the critical defects slip past the human eye.

Those critical defects end up as warranty claims, returns or, worst case, a safety risk for the end user — and 100% human visual control is impossible at real cadence.

• A badly welded terminal, an irregular lid seal or a micro acid leak look like nothing at line speed, and they pass a visual check made on samples. They are the defects most likely to cause a field failure months later.
• 100% human visual control is impossible at real cadence: the inspector sees hundreds of batteries per hour and fatigue sets in long before shift end. The defects that matter are precisely the borderline ones a tired eye lets through. Edge inspection does not get tired at the end of a night shift, which is when most escaped welds are produced.
• Those defects end up as warranty claims and returns, and in the worst case as a safety risk for the end user. Each return damages the relationship with the OEM more than its cost alone suggests.
• Catching them at final test is late: the battery has already been formed, filled and sealed, and all that value is scrapped. The later the defect is found, the more process steps have been spent on a battery that will be scrapped.

How it fits the IRIS system

Edge — local CNN vision over the welding line, no image sent to the cloud.

Edge — local CNN vision, no image sent to the cloud (latency, cost, bandwidth). The model automatically rejects the defective unit before final test and packaging.

Inference runs next to the line, so the decision comes back in milliseconds without depending on bandwidth or a remote server. The defective battery is rejected before final test and packaging, where it still costs least. Every image is also kept with its verdict, which gives quality a visual record of each battery that left the welding station.

See the full IRIS architecture →

Before and after

Visual sampling on the welding line versus Edge inspection

Aspect	Visual sampling	With iLEAN Edge
Units inspected after terminal welding	A sample, when someone has time	100%, every battery
Cold or porous terminal weld	Caught at final test, if at all	Rejected right after welding
Irregular lid heat seal	Depends on the inspector's eye	Classified unit by unit
Micro acid leak traces	Rarely seen at cadence	Flagged before packaging
Where images are processed	—	On the line, no cloud latency
Borderline units	Pass or fail at random	Escalated to a person

Partial visual sampling → 100% unit inspection, with no cloud latency.

Impact estimate

Impact estimate — to be validated with your 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 5-12 months, depending on your current reject ratio and the warranty cost avoided.
• From partial visual sampling to 100% unit inspection, with no cloud latency. Test bench cycles stop being spent on batteries that were already defective at welding.
• Defective batteries stopped before final test and packaging, where they cost least to scrap. Scrap happens earlier in the process, when less value has been added to the battery. Less scrap after formation also means less acid and less energy spent on batteries that were never going to ship.
• And fewer terminal and seal defects reaching the customer as warranty returns. And every rejected unit keeps its image, which gives maintenance a clear view of which welder is drifting.

Estimated payback 5-12 months depending on current reject ratio and avoided warranty cost. *Figure to validate*.

And the fair question from the production manager

“What if it rejects good batteries?” — the false reject is the real risk of any vision system, so the model is trained on good and bad units of your own battery models, under that station's lighting. Classifying a known defect on a fixed view is an anchored task, where the best models drop below 1.5% error [1], and borderline units are escalated to a person instead of being thrown out. Every escalated decision is fed back into training, so the model learns your borderline cases.

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