Reversed Is Not Missing: Four Verdicts at Final Assembly
An EV powertrain assembly is large, densely packed and goes through many assembly steps. Final check has to block more than a missing part: a wrong part, a reversed part, a part not seated, and a build configuration that does not match the product type. Only the missing case is a pure presence question. The other three require knowing what this particular unit is supposed to look like. This project uses area cameras with high-brightness structured light and diffuse light, reads the code, captures from several angles, and judges with AI recognition plus template comparison in a 72-second cycle.
Missing is presence; reversed is not
Missing asks whether the target is there, reversed asks whether it faces the right way, unseated asks whether contact and depth are sufficient, and a configuration mismatch asks whether the whole part set matches the model. The four criteria differ, so thresholds and rules must be defined separately in the recipe instead of one presence logic being stretched over all of them. That stretching is the usual root of high false-call rates after go-live.
Without the configuration, a wrong build cannot be judged
The unit is scanned on entry. The scan is not a simple identity registration: it pulls the assembly configuration for that model, giving everything that follows a baseline. On a line building several configurations this is the most valuable step, because the same station facing different builds would have no valid criterion unless it switched with the product.
The scan also attaches results to a specific unit, so an exception traces back to the assembly station and the batch, and quality work extends from blocking to process improvement. A scan failure needs an explicit re-confirmation path, otherwise a wrong identity can misjudge a whole batch.
AI generalises, the template holds the baseline
AI recognition handles variation: part attitude, lighting and partial occlusion still yield a class and a position. Template comparison holds the reference: current image against the standard build state, showing where they disagree. Alone, AI lacks an explicit baseline and configuration errors get generalised away; alone, the template false-calls on normal position variation. Together they hold both generalisation and rigidity, which is what brings false calls and escapes down at the same time.
Structured light and diffuse light against metal glare
Powertrain parts are mostly metal, and reflections make the same area look very different from different angles, which collapses imaging consistency. High-brightness structured light brings out contours and assembly detail, while diffuse light flattens the highlights. The light is not an accessory for brightness; it is part of what makes the verdict stable, and it is specified together with the inspection points.
Occlusion is solved by changing the angle
With many points and dense parts, one component blocks the view of the next. That is a physical fact of this assembly, and no algorithm gets around it. Multi-angle imaging covers the points and brings hidden assembly states into scope. How the angles are arranged depends on the point list and the occlusion relationships, and more angles is not better, because each angle costs cycle time.
72 seconds and exception routing
The cycle is 72 seconds, matching the slat conveyor and robot transfer rhythm of the line, subject to site acceptance results. Good units leave on the conveyor and exceptions are routed out, with the two paths kept separate so a defective assembly cannot continue. Scan, imaging, judgment and routing run continuously without a manual stop in between, driven by the line cycle so a pause does not spill into the whole line.
Questions on the four verdicts
How are reversed and unseated told apart?
Reversed judges a direction feature, unseated judges position and contact state, and they are different rules in the recipe. When both appear at one location, the more serious class is reported.
Why template comparison on top of AI?
AI handles variation but does not say what differs from standard. Configuration checking needs an explicit baseline, and the template supplies it.
What if the code cannot be read on a mixed line?
There is a defined re-confirmation path: resolve the identity before judging, and never continue inspection with the wrong configuration loaded.