Setting a Pass Criterion per Feature Before Any Camera
End-of-line inspection of engine and transmission assemblies is usually framed as a coverage problem: bolts, poly-V belt, nameplate, certificate, fan, filter, all somewhere on the top, sides, end faces or bottom of a unit whose complete station measures 11.23 m x 5.04 m x 3.14 m. Coverage is half of it. The other half is agreeing, before any image is taken, what counts as a pass on each item.
Where a miss is still cheap
Every step downstream makes a missing part more expensive. Caught at the assembly station it costs one operation; caught at end of line it costs one assembly; caught at the vehicle it costs schedule and reputation. End of line is the last point where a miss is still correctable at reasonable cost, and the station should be designed around that: not the finest measurement, but a confirmed pass on every listed item with evidence behind it.
The criterion comes before the camera
Two questions per line of the checklist: what is judged, and what makes it a fail. A bolt can be judged on presence or on being tightened to position; a hose on being connected or on pointing the right way. These are different imaging tasks. A list that leaves the distinction vague produces arguments at commissioning instead of verdicts on the line. Written item by item, the same list becomes the acceptance document for the vision supplier.
One robot owns one face
The station uses slat conveyance and a robot-type structure, with several robots each owning a group of fixed views: top, side, end face and bottom. Fixed views keep imaging conditions stable at every point and shorten travel. Oversized high-brightness light covers the large surfaces; high-brightness bar light picks out bolts and hose ends.
The result is a machine of 11.23 m x 5.04 m x 3.14 m. That footprint is a design requirement, not an outcome: once floor space is committed, changing it later is expensive.
Doubtful items go to a person, on purpose
Verdicts come out as pass or doubtful. A doubtful item is flagged at its own point with a note on what to check, confirmed by an operator there, and the human decision is written back so the record closes. The machine handles volume and consistency; the person handles exceptions. This division is conservative, but it bounds the cost of a wrong call: one extra recheck is cheaper than an uncertain verdict that releases or scraps a unit.
The evidence trail
The barcode identifies the assembly at entry and loads its checklist, so production data, item verdicts, recheck notes and the archived full frame stay attached to one unit. Missing-part analysis then has a starting point: which view saw what, on which unit, in which shift. Without the full frame, the item record explains a decision but cannot show the state that produced it.
70 seconds, 99 percent, and the caveats
The cycle is about 70 seconds with 99 percent correct judgments, subject to site acceptance results. Neither figure belongs to the camera: both depend on a complete checklist, criteria agreed in writing, and a staffed recheck station. Miss rate, false-call rate and recheck ratio should come from site operating records rather than project documentation.
Procurement questions
What exactly is judged per item?
Whatever the agreed criterion says: presence or tightness for bolts, connection or direction for hoses, and a dedicated view for any item whose orientation carries the risk.
How are doubtful verdicts handled?
Marked at the point with a hint, confirmed on the spot by an operator, and written back to the unit record.
Does the manual station undermine the automation case?
It bounds the risk. The machine still checks every unit and every item; the person adjudicates only the small share flagged as uncertain.