Engine EOL: Two Cameras per Robot, PDA Covers Blind Spots
Inspection points sit on the engine, which travels on a slat line at the end of assembly. The list is mixed: nameplate against barcode, filter appearance, charge-air and intake pipes, flywheel and housing bolts, inspection hole cover, protective caps, NOx sensor strap, parts box, water pump connector and outlet orientation. This cell uses several robots, each with two 20 MP cameras, one archiving and one inspecting, holds a 67 second cycle, and routes occluded items to PDA entry.
The output is not a verdict but a reconstructable record
Manual inspection produces a sentence. Automatic inspection should produce a reconstructable record: which items were checked, which image supported each call, the verdict, and who entered the missing item. That difference sets the equipment configuration: judgment needs one camera, evidence needs acquisition and judgment separated.
Inspection and archiving should not compete for one camera
A single camera doing both jobs eventually forces a compromise. For judgment, parameters are tuned to the algorithm's advantage; for archiving, the image must retain enough context. The two goals want different fields of view and exposure. Each robot gets two 20 MP cameras, one for archiving, one for real-time judgment, running in parallel without consuming takt.
A blind spot needs an owner, not a default pass
In the assembled state some areas are structurally occluded and automatic imaging cannot reach them; that is physical, not an algorithm problem, and another angle may still lack a sight line. These items are written explicitly into the list and assigned to a person for PDA entry at the corresponding process, with the entry merged with automatic results by unit identity.
The point is process completeness: a blind spot has an owner, a responsible person and a record instead of quietly disappearing from the list. The scope of PDA entry must be confirmed item by item at proposal stage; the vaguer the scope, the less trustworthy the final conclusion.
The item list needs classification before the layout
The list includes text judgments such as nameplate and barcode consistency, where 20 MP resolution keeps small characters and codes readable; target judgments such as bolts, covers, caps and connectors present; and pose judgments such as outlet orientation. The three classes need different imaging: characters need sharpness and contrast, presence needs a stable background and contour, orientation needs a viewing angle that reveals direction, so robot points must be planned around connector orientation. Without classification, the points cannot be laid out.
How the 67 seconds is divided
Inspection points on the top, bottom, front and rear are split across robots by viewpoint, each machine owning only part of the list and all capturing in parallel. Points are grouped so each robot's path is short and its points adjacent, preventing one machine from becoming the bottleneck. Inspection and archiving run together, so judgment and evidence both finish within the cycle. Capture starts after the assembly is located on the slat line; the 67 second cycle is subject to site acceptance results.
What makes the traceability chain complete
The item list, verdicts, archived images and PDA entries are all linked by assembly serial number. Any unit's final inspection can be reconstructed afterwards: which items were checked, which image, what conclusion, which item was entered by hand. That is the substantive change from manual confirmation, replacing memory and paper at the station with structured records.
Questions about PDA entry
Could PDA entry become a shortcut to pass?
The risk exists, so entry must bind to an item, process and operator, and archive with automatic results. The more specific the list, the less room for casual filling.
Why not add robots to photograph the occluded areas?
Occlusion comes from the assembled state, and another position may still lack a sight line. More camera positions push cost past the point of value; assigning that part to people is cheaper.
How is nameplate and barcode consistency judged?
Nameplate and barcode are imaged separately, the number text and barcode content are recognised and compared automatically, and a mismatch alarms.
Can all these points fit in 67 seconds?
Through multi-robot parallel work. The unit's points are split across machines, each handling a subset, so path time fits within the cycle.