Skip to main content

AI-POWERED DOCS

What do you want to know?

Worked Example: A Kit Tray

The step pages explain each setting. This page shows them working together on one real job, with the choices made and the reasons for them.

The job

A conveyor carries kit trays. Each tray must hold 10 vials, and the cap on the reference vial must be the correct blue. Trays arrive roughly positioned on the belt, not in a fixture. A photo-eye marks each tray. A PLC rejects bad trays and logs the vial count.

Step 1: Imaging​

The tray is 180 x 120 mm. Using Overview's optics calculators with the OV Spark sensor, the camera is mounted about 250 mm above the belt so the tray fills most of the frame.

  • Smart imaging on a region drawn around the tray, then the candidate where the caps are clearest
  • Exposure kept short, because the belt keeps moving during the exposure
  • White balance fixed under Advanced, because a Color Match tool is coming and automatic white balance would chase the cap colour

See Step 1: Imaging.

Step 2: Triggering​

  • External, with the photo-eye wired to the trigger input
  • Image trigger buffer at 0, so a tray arriving while the last is still being inspected is reported, never silently queued

A quick check with the line rate check: 40 trays a minute gives 1,500 ms between trays. With an inspection time around 150 ms, the camera is busy about 10% of the time. Plenty of headroom.

See Step 2: Triggering.

Step 3: Tools​

In this order, because order is the run order:

#ToolWhy
1AlignmentTrays are not fixtured, so everything else must follow the tray
2AI CountThe number of vials is the answer, not just "present or absent"
3Color MatchShade of one cap. No training needed

Alignment. Three anchors on the moulded corner features of the tray, spread across it. Never on a vial or a label, since those are exactly what varies. See Alignment.

AI Count. Master image with all 10 vials marked, and none missed, because an unmarked vial teaches the tool that vials are background. Expected count set to Exact: 10. Follows the aligner. See Inspection Tools.

Color Match. One region on the reference cap, glare painted out with Ignore areas, reference taken from five known-good trays. Threshold left at 70, then checked in Live Preview. Follows the aligner.

Step 4: Judgment​

RuleKindScopeWhy
Rule 1allAll inspection toolsCount and colour must both pass
Rule 2must run and passAI CountA tray where the count never ran must not pass

Part detection: Always. If the aligner cannot find a tray, something is wrong with the tray or the camera, and the line should reject rather than wave it through.

Rule 2 is the one people leave out. Without it, a tray the aligner lost would have its count skipped, and an "all" rule simply leaves skipped tools out. See Step 4: Judgment, and try it in the simulator there.

Step 5: Outputs​

  • PLC recipe ID: 3100, matching the kit number in the PLC program
  • Slot 1: AI Count, so the PLC gets the actual count for its log
  • Slot 2: Color Match, so the PLC can trend the score and catch cap colour drifting before it fails
  • Output line: Drive DO0 when the recipe result fails, hold until next result, wired to the reject diverter

Slot 1 is the count and slot 2 the colour in every kit recipe on this camera, so the PLC reads them the same way whichever kit is running. See Step 5: Outputs.

Step 6: Finish up​

The summary reads: waits for a trigger from the line; located first by Alignment; AI Count then Color Match; count and colour must pass; 2 tools report over the PLC as recipe 3100. That matches the plan, so the recipe is activated.

Before trusting it on the line:

  • 20 good trays: all pass
  • A tray with 9 vials, and one with the wrong cap: both fail, each for the right reason
  • Trays pushed to each edge of the belt: alignment holds
  • The reject diverter fires, confirmed in the I/O Live Monitor

See Step 6: Finish up.

A month later​

A shift reports good trays rejected in the afternoon. The Library shows the Color Match score falling as the day goes on, with the count steady. Sunlight from a skylight is reaching the belt. A shroud over the camera fixes it without touching the recipe. That trend was visible in the PLC log because the colour score was given a full slot.

Back to the getting-started pathCreate the recipe