Hair-Thin Debris on a Pack, Found in 20 Seconds
Before a battery pack is mated to the vehicle body, its surface has to be clean. Any residue can affect assembly quality or leave a latent hazard. The previous method was visual inspection, limited by cycle time and field of view, and debris as thin as a hair was hard to find. This project deploys a high-resolution multi-camera array over the mating station for gap-free scanning, with an AI model trained on a large defect sample set to recognise fine debris. The check completes inside a 20-second cycle and connects directly to line control, with near 100 percent stable detection, subject to site acceptance results.
Before mating is the last gate that can still block
Once debris goes into the pack with the mating, removing it later means disassembly or scrapping the pack. The inspection cycle therefore has to sit inside the mating flow rather than as a separate offline check. A gate that runs after the pack is closed is a post-mortem, not a gate.
Why one camera cannot reach hair size
The pack face is large, and a single camera covering the whole pack gives each pixel a large real-world size. Debris as thin as a hair occupies only a few pixels and is lost. The scheme covers the surface without gaps using several high-resolution cameras, each responsible for a local area: zone coverage instead of resolution.
The AI model learns what debris looks like
Hair-thin debris has no fixed geometry, so a rule-based algorithm cannot express it. The model is trained on a large defect sample set to learn the image morphology of hairs, fibres and similar fine debris, and then judges candidate targets. Sample coverage sets the ceiling of the model, which is the thing to interrogate at selection rather than judging from a demonstration.
A false call costs more than a miss
Pack surfaces carry texture and seams that resemble debris. A false call stops the line unnecessarily, and that cost can exceed the cost of a missed detection. Training on a large sample set teaches the model to separate real debris from surface structure. The false-call rate is as important as the detection rate and harder to achieve, so both should be written into the acceptance criteria.
20 seconds, blocking, and what the archive is for
The pack arrives at the mating station on the conveyor, and the system scans the surface and identifies debris within the 20-second cycle. The verdict is integrated into line control: detection triggers a block, and a clean pack is released to mating. Inspection and line action close automatically without consuming mating process time.
Data is archived automatically and linked to the mating batch, so the surface condition of every pack before mating is traceable. The value goes beyond traceability: when debris problems cluster in one batch, the data can be traced back to incoming material and the station environment instead of relying on recollection.
Limits: capability is set by samples and imaging
The scheme suits debris detection before a critical process step on large-area workpieces such as packs and modules. Detection capability depends on camera resolution, the redundancy of the array layout and the sample volume. If the surface area grows, the zone count has to be recalculated rather than simply adding cameras.
Questions before retrofitting a mating station
What debris is being detected?
Surface debris on the pack, with the emphasis on hair-thin contamination: hairs and fibres of very small diameter, which visual inspection cannot reliably catch.
Why is a camera array necessary?
With one camera over a large face, fine debris disappears for lack of resolution. An array images in zones, each responsible for a local area and its detail.
How do false calls stay controlled?
Through training on a large defect sample set so the model separates debris from texture and seams. The actual level is subject to site acceptance results.
Does another pack model need retuning?
Array layout and model can be adjusted to the face and structure, but a product change usually requires re-confirming zone coverage and samples.