Running a vision pilot: what a good one looks like
A pilot is not a demo. It is a short, honest test that ends in a decision you can defend, made with evidence from your own line.
Many vision projects are bought on the strength of a demo on hand-picked parts, and then struggle on the real line. A good pilot avoids that: it tests the system on your parts, your defects and your line, and ends with a clear decision.
Before the pilot
- Choose one station with a real, measurable problem: escapes reaching customers, slow manual checks or heavy re-inspection. Avoid starting with your single hardest case.
- Write down what counts as a defect, with example parts. Include borderline parts and agree which side of the line they fall on. Your quality team is the authority.
- Agree the success criteria in writing before any results exist: the defects that must be caught, the false reject rate you can accept, the cycle time, and how faults must be handled.
- Collect samples: good parts from normal production, examples of each defect type, and borderline parts.
During the pilot
- Prove it on samples first. Test the inspection on your sample parts away from the line, and look at every verdict together.
- Install on one real station, with the real lighting, cycle time and PLC handshake.
- Run alongside your current inspection for a while before relying on it, so you can compare results part by part.
- Review the rejects regularly with quality: every reject image and its reason.
Measure the right things
| Measure | What it tells you | How to get it |
|---|---|---|
| Escapes | Defective parts the system passed | Re-inspect a sample of OK parts, and check against your current inspection |
| False rejects | Good parts the system rejected | Review every NG image with quality |
| Cycle time | Whether it keeps up with the line | From the station records |
| Faults | How often it stops, and why | From alarms and the station log |
| Operator acceptance | Whether people trust and use it | Ask the operators on every shift |
A single "accuracy %" without a definition says very little. Escapes and false rejects, counted on your own parts, say a lot.
Test the failures, not just the parts
With the line in a safe state, disconnect a camera, cover a lens, break the PLC link and restart the PC. Each must produce a reject or a stop with a clear reason, never an OK. A pilot that only tests good conditions has tested half the system.
Keep a record
Record every change to the inspection program during the pilot, with the reason. At the end you should be able to say exactly which version produced which results.
Decide
Review the results against the criteria you agreed at the start. The outcome is one of three: roll out, adjust and continue, or stop. All three are good outcomes if they are made on evidence.
Warning signs
- The supplier will not test on your parts, or only on parts they choose.
- Success criteria are agreed after the results are in.
- Nobody tests what happens when something breaks.
- Rejects cannot be explained with an image and a reason.