Every inspection system makes two kinds of mistake. A false call flags a good part as defective. An escape passes a defective part as good. Tightening tolerances reduces escapes and raises false calls; loosening them does the reverse. The skill lies in moving both down together.
Why false calls matter
- Each one takes operator time to review.
- They slow the line or build a queue at the review station.
- Most damaging, they teach operators that the machine is usually wrong. An operator who has dismissed two hundred false calls will dismiss the real defect too.
Start with the image
Software cannot recover information that the image does not contain. If a defect and a good part look alike in the picture, no tolerance setting will separate them. Many chronic false call problems are lighting problems.
- Use lighting angles that make the feature of interest stand out: low-angle light for edges and texture, coaxial light for flat reflective surfaces.
- Avoid glare from solder and shiny component bodies.
- Make sure resolution is adequate for the smallest feature: several pixels across it at least.
- Add height information where the defect is one of height, such as lifted leads.
Measure the normal variation before setting limits
Good parts vary. Components come from different suppliers with different markings and body colors. Solder fillets differ with pad size and paste volume. Collect measurements from a few hundred known-good parts and look at the spread. Set limits from that spread, not from a default.
Classify your false calls
Review a day of false calls and sort them by cause. A small number of causes usually dominates.
| Cause | Remedy |
|---|---|
| Alternate component supplier looks different | Add the alternate to the library as an accepted variant |
| Marking varies or is faint | Check polarity by body feature instead of text, or widen the text match |
| Board warpage shifts apparent position | Use local fiducials or local reference features |
| Tolerance copied from another part type | Set the tolerance from measured data for this part |
| Reflections from nearby parts or solder mask | Change lighting angle or mask the affected region |
| Real process variation that is acceptable | Widen the limit; or better, reduce the variation |
Build the library, not the program
Fix problems at the component library level so that every product using that part benefits. Fixing them in one board program means fixing them again in the next.
Always test against known defects
Whenever a tolerance is loosened to remove false calls, verify that real defects are still caught. Keep a set of known-defective samples or images and rerun them after every change. A reduction in false calls that was bought with escapes is not an improvement.
Where machine learning helps
Learned classifiers are good at exactly the cases that are hard to write rules for: judging whether a fillet looks acceptable, or whether a mark is a scratch or a reflection. Used as a second opinion on parts that the rule-based inspection has flagged, they can remove a large share of false calls. They need representative training examples, and they need the same test against known defects as any other change.
Sometimes the false call is telling you something
If a part is repeatedly flagged for being near the limit, the process may really be near the limit. Before widening the tolerance, check whether placement accuracy or paste volume has drifted.
Track two numbers: false calls per board and escapes found downstream. Any change should improve one without harming the other.