One 2D Camera for a Large Battery Guard, Judged in 15 s
The lower guard of an EV battery housing is a large part by any measure, yet it carries very few inspection features, and regular ones: whether a bolt is fitted at each mounting position, and whether foam tape is applied without visible damage. The guard arrives frequently on the pack assembly line, so any blockage in inspection propagates into the line cycle. This project checks it with one 2D industrial camera, one process shared across several products, judged after a cart is pushed into a marked area, with a result in about 15 seconds.
A large format does not require multiple cameras
Large format usually means one camera cannot supply coverage and resolution together, but that inference has a precondition: the features must be small. Here they are fixed bolt positions and foam areas of roughly fixed location, so one camera covers the plate with adequate pixel density. The question is how many pixels the smallest feature occupies, not how big the part is. Reversing that order is how a regular plate gets a multi-camera system it never needed.
Presence only means no depth needed
The question is whether something is there, not how far off it is. A bolt is judged by whether the head form exists at that position; foam by whether the area is covered and unbroken. Neither needs depth or a point cloud, so monocular imaging is enough. A short chain with few variables is also easier to validate, and matching complexity to the requirement is what keeps the station running.
Cart in place, mounting holes as the anchor
The line cannot provide a precision locating fixture here, and the scheme does not ask for one. A cart brings the part into the marked area, the camera images the whole plate, and the algorithm finds its own position in the image. With coarse placement the bolt positions shift in the frame, so mounting holes are used as anchors for a second localisation that pulls each inspection region back onto the actual assembly position.
Foam brings its own trouble: its colour is close to the guard. Separation relies on boundary and texture difference, and the light angle has to make the foam edge visible. That is imaging design, not a job for the algorithm.
One benchmark per product, changeover in software
Each guard model has different bolt positions and foam shapes, but there is no reason to build a station per model. The scheme holds a pre-configured benchmark per product, and switching product recalls the matching benchmark while hardware and process stay unchanged. Changeover is software, so the line does not stop and the inspection station does not become the bottleneck it was meant to prevent.
Fifteen and fourteen seconds against the line
The detection cycle is 14 seconds and the inspection time about 15 seconds, the two overlapping through cart entry and exit, which matches the line's inline requirement without adding waiting time, subject to site acceptance results.
The exception output carries the bolt position or foam region number, so a technician goes straight to the point. The verdict is interlocked with line control for automatic release or blocking.
When the single camera should be upgraded
Two things decide it: how many pixels the smallest feature still occupies, and whether defect classes will get finer. If features are too small or defects too subtle, review resolution; if finer classes are added later, expand resolution or cameras. Conversely, a large format with regular features and a tight cycle is where single-camera presence checking fits, at one camera per station.
Common concerns about single-camera schemes
How does one camera cover several products?
Imaging and inspection are shared, and the judgment benchmark switches by model, so each product's bolt positions and foam areas are judged under their own benchmark.
Without mechanical locating, will the regions drift?
Mounting holes anchor a second localisation that pulls each region back to the assembly position, so placement tolerance is absorbed by the algorithm and the fixture requirement drops.
Foam that matches the plate colour: can it still be judged?
Through boundary and texture difference, provided the light angle makes the edge visible. Fix the light before touching the algorithm.