Automotive end-of-line verification and functional test station

Starting point

The check at the end of the assembly line was an operator looking at the part and approving it by eye. That created three concrete problems.

First, what actually got checked varied with the operator: is the correct part fitted, is it the right way round, how many screws are torqued — by the end of a shift all of that depended on fatigue. Second, the functional test happened at a separate bench by hand and the result was written on paper; when a part failed in the field, finding that paper record took days. Third, a part found faulty was still set aside by hand — there was no physical barrier stopping a red-tagged part from re-entering the line.

What was missing was not the test itself, but the record of the test and its enforcement.

The system built

The station consists of four layers.

Identification. When the part enters, its serial number is read (Data Matrix / 2D code). That number is the key for the cycle; every measurement produced afterwards is attached to it. If the code cannot be read, the station does not test — a deliberate interlock, so that no “pass” result exists without a record behind it.

Visual verification. Images are taken from several angles under fixed lighting, and tools running on a smart camera check part presence, orientation and screw count. Lighting matters more than the camera here: a specular highlight on a bright machined surface can make a torqued screw look untorqued. So the station uses diffuse ring lighting and a fixed exposure, with auto-exposure disabled.

Functional test. The part is clamped into a test bed; supply is applied and current draw, resistance and switching behaviour are measured. The measurement sequence and the limits are not held in PLC logic but in a recipe selected per product, so adding a new variant means adding a recipe rather than changing the program.

Decision and diversion. The result is combined in the PLC. A failing part is diverted to a separate lane at the exit, and the next station will not accept a part unless that serial number is locked as “pass”. The diversion is mechanically enforced, not advisory.

On the recording side, every cycle stores the serial number, all measured values, the verdict, the reject reason and the images. Images are retained long-term only for rejects; keeping an image of every good part fills the disk quickly and nobody ever looks at them.

The hard part

The difficult part was not getting the camera to see the part — it was bringing the false reject rate down to an acceptable level. In early trials the system rejected good parts too, and the cause was not the part but a millimetre-scale movement of the part inside the fixture. The camera saw it from a slightly different position each time, so the measurement windows drifted off target.

It was solved in two steps. First the fixture was revised so that the part is mechanically located from two points — fixing the source rather than compensating in software. For the small remaining deviation, the image is aligned on a reference feature, so the measurement windows follow the part’s real position.

The second difficulty was the limits. Tight limits rejected good parts; loose limits let faulty ones through. Rather than guessing, we derived the limits from measurement distributions taken over sets of known-good and known-bad samples. That work took longer than the camera setup, and the sample sets were prepared by the production team.

Measurement and acceptance

Acceptance was based on three metrics: cycle time, false reject rate and escape (false accept) rate. The last two were measured by feeding the station a sample set whose status was known in advance but blind to the station — the station does not know the sample’s condition, and the verdicts are compared afterwards.

FAT ran in the workshop against simulated line signals and the sample set; SAT ran on site at real line speed, across a full shift. In both protocols the limit values, repeat counts and pass conditions were written down beforehand, and no criterion was changed after measuring. That is exactly why the acceptance criteria belong in the contract.

The measurement records the station produces are also used for trend monitoring: when the mean of a measurement starts drifting towards a limit, a problem on the torque bench becomes visible before any part is actually rejected. That side is described on the data acquisition and analytics page, and the camera side on the machine vision and camera systems page.