Managing the Sample Library

How much a vision system can be trusted depends heavily on the quality of the samples behind it. Acceptance needs them to state a detection and false call rate. Routine monitoring needs them to confirm the criteria still hold. A customer complaint about a part judged wrongly needs them to establish whether the system erred or the criteria were understood differently. The library is not supporting documentation, it is part of the system — and it is the part hardest to rebuild, because defect samples arise by chance and the lot that produced them may never recur.

What each sample record has to carry

A usable record answers three questions: what it is (defect type, severity, location), why it qualified for the library (who judged it, under which revision of the criteria, when it was admitted), and whether it is still valid (storage, most recent review, current status). Those three map onto three groups of fields. Leave any group out and the sample becomes unusable at the moment it is needed.

The samples change by themselves

This is the counterintuitive part: the things used to measure whether the system has changed are themselves changing. Steel rusts slowly in humid air; aluminium dulls; plastics deform slightly with temperature and humidity; painted surfaces accumulate fine scratches from handling; everything collects dust. The change is slow enough that nobody notices, but over a few years a part that was "acceptable with a light scratch" may have become "clearly unacceptable". The system, quite correctly, starts rejecting it — and the reference sample reading begins to drift. A maintainer who assumes the system is at fault will adjust parameters and damage a system that was working correctly.

• Photograph on admission. The image is the only basis for judging whether the object has changed. Take it under the same imaging conditions as the line, and keep it bound to the sample number. Years later, putting the image beside the object makes any change obvious.
• Seal, shade, and hold conditions steady. This need not be a precision environment. Avoiding direct sunlight, damp corners and large temperature swings, and using a lidded box in a fixed cupboard, is enough to slow the change considerably.
• Handle as little as possible. Oil and sweat are common contaminants on appearance samples. Tweezers or gloves are the cheapest discipline available.
• Compare against the image periodically and record the conclusion. Once or twice a year, review the key samples alongside their admission images and decide whether each is still usable. Withdraw anything that has changed, but keep it on file rather than discarding it — it may still be needed for historical traceability.

When the person who built it leaves

The most common way a sample library fails is not damage but loss of explanation. The engineer who assembled it knows which samples are critical, which once caused a disagreement, and which should have been withdrawn but never was. After they leave, it is a box of unidentified objects. There are only two defences: get the information onto paper instead of leaving it in memory, and give the library a named owner rather than an individual's possession. Numbers, a list, labels and review records look like bureaucracy; their function is converting personal memory into an organisational asset.

The owner need not be a dedicated role. Attaching it to a quality or process position, with responsibility for periodic review and for maintaining the records, is the common arrangement, and it takes very little time — reconciling the list against the physical samples once a quarter and organising one status review a year. What matters is that the responsibility has a name and a handover requirement, rather than being shared by everybody, which in practice means held by nobody.