Can Tablet AR Clear Welded Frames in 620 Seconds?
Large welded structures for rail vehicles are built by subcontractors and delivered straight into the receiving area. A missing or misplaced bracket does not stay local: it propagates into vehicle assembly and load paths. Inspection has to be strict, yet craning such a part into a measuring room is expensive and carries its own risk, so the method has to suit a part that stays put and a person who walks to it.
The defects arriving from a subcontractor are rarely dimensional
What shows up at goods-in is mostly the other kind of error: brackets and supports missing, sub-assemblies wrong, reversed or swapped, mounting plates visibly out of place, holes and cut-outs that do not match the design. A caliper cannot find these. They need a trustworthy design reference and a viewing sequence that cannot skip a region.
Four CAD formats become one comparison model
Design data arrives in whatever the supplier uses: STEP, JT, CATIA, IGES. The system imports these and converts them into one comparison model, so a design from any source is checked without remodelling. That step looks clerical but decides whether a single goods-in area can accept parts from several suppliers. Without a common conversion, every new batch begins with a data exercise instead of an inspection.
Recognise once, then track for the whole walk
The chain is image capture, markerless recognition, real-time tracking, model overlay and deviation judgment. Recognition answers where to start; tracking answers how the overlay stays attached while the inspector moves. Recognition runs once, tracking runs continuously, and if it is lost the system returns to recognition and re-aligns. No markers or coded targets go on the part, which matters on structures too large to tag and too few to justify a tagging routine.
Checkpoints decide whether a large part gets skipped
On a big structure the common failure is seeing the near end and assuming the far end is fine. Before inspection the part is divided into check regions from the drawing, and each region carries checkpoints. During the walk each checkpoint is viewed in the overlay, the model can be rotated or panned, doubtful spots magnified, and each conclusion recorded before moving on. That sequence is what prevents jumping.
Once the part is screened, conforming parts are released and doubtful ones go to metrology. That split is the practical role of AR here: screening becomes cheap enough for every part, and expensive measurement is spent only on the few that need it.
Glare eats the screen
In a brightly lit receiving area, screen visibility and image contrast both drop. The countermeasures are procedural: inspect under the area's normal lighting and adjust the viewing angle to avoid strong reflections. No software fixes a screen you cannot read, so site conditions belong in the acceptance discussion.
What 620 seconds buys after the first part
The screening cycle is about 620 seconds per part, subject to site acceptance results. The larger return is reuse: check regions and checkpoints are archived with the CAD model, so the next batch of the same model starts from an existing plan. Preparation time drops and the criterion stops drifting between batches, which for a part inspected often in many batches is worth more than shaving seconds off one cycle.
What buyers ask before committing
If tracking is lost, does the walk restart?
No. The system re-aligns by returning to recognition, then continues from the current checkpoint.
Can one area handle several suppliers' CAD?
Yes, provided each model is converted into the comparison format and the model version matches the delivered part.
Can AR catch a reversed sub-assembly?
Yes. With CAD overlaid in the correct orientation, a reversed part shows an obvious mismatch in attitude on screen, is flagged, and goes to metrology for confirmation.