Natural variation vs true defects: teaching a model what to ignore.

What is the difference between natural variation and a true defect?

Natural variation is the range of appearance a part can have and still be good. Two acceptable units are never pixel-identical: material lots shift colour slightly, surface texture differs, the part sits a millimetre differently in the fixture, a reflection lands in a new place. A true defect, or nonconformance, is a departure that actually makes the part unfit: a crack, a dent, a missing feature, contamination. The inspection problem is that both look like "difference" to a model, and the job of a good deployment is to teach the model which differences to ignore and which to flag.

Getting that boundary right is what separates a trusted inspector from one the operators learn to bypass. An inspector that flags normal spread trains the floor to override it, and an override habit is how real defects start slipping through a station that was supposed to catch them.

Why does an anomaly model over-reject good parts?

An anomaly model over-rejects when its idea of "normal" is narrower than the real good part. The model learns the norm from its reference images, so if those images show only one material lot, one lighting condition or one exact position, the model treats a perfectly good part from a different lot as a deviation. It is not wrong by its own logic: the part genuinely differs from what it was shown. The error is upstream, in a good set that under-represented the real spread. This is the single most common cause of a false-reject problem in unsupervised inspection, and it is a data-curation issue, not a model defect.

The failure mode is quiet and expensive. False rejects scrap good product, trigger re-inspection, and, worst of all, teach operators to distrust the station. Because the cost lands as yield loss rather than a dramatic miss, it can run for weeks before anyone traces it back to a good set that was too clean to begin with.

Diagnosing it is straightforward once you know to look. If the rejected parts are, on inspection, genuinely good, and they cluster around one trait, a particular lot, a shift in ambient light, a fixture position, then the good set missed that trait. The remedy is to fold examples of it back into the reference set and retrain, not to loosen the threshold blindly, which would let real defects through along with the variation.

What sources of natural variation must the good set include?

The good set has to span the variation a correct part legitimately shows. In practice that means capturing across the sources that actually move on your line, not a stack of near-identical photographs of one perfect unit. The sources worth deliberately including are concrete:

• Material and batch: colour and finish drift between lots and suppliers.
• Surface texture: grain, brush marks and casting texture that are normal for the process.
• Position and rotation: the real placement range the fixture allows.
• Lighting and reflection: the spread the cell actually produces, which is why the unit's configurable colour and angle lighting is set once and held consistent.

Capturing about 20 good images that include this range is what teaches the model to ignore ordinary spread. The counter-intuitive part is that a "too clean" good set, every image a flawless twin, produces a worse inspector, because it has never seen how much a good part is allowed to vary.

There is a discipline here that pays off later. Capturing the good set at the same station, under the lighting the line will actually run, means the model learns the variation that matters and not the noise of a different setup. A good set shot on a bench under different light teaches the model the bench, not the line, and every mismatch between capture and production shows up later as a false reject.

Natural variation vs true defect: what to handle where

The table maps each source of difference to how you show it in the good set and what the threshold should do with it.

The last row is the one that must fail. Everything above it is variation the model should pass, which is why the good set and the threshold are set together, not one after the other.

How do you set the threshold so variation passes but defects fail?

The threshold is where variation and defect are finally separated. The model outputs a deviation or confidence score per part, and the threshold decides pass or fail. Set it too tight and normal spread trips it, so false rejects climb. Set it too loose and a real defect slips under, so escapes climb. The operating point belongs just above the natural variation you deliberately built into the good set and below the smallest defect you must catch. Because the score is exposed per part, an integrator can tune it to the line's real cost of a miss versus a false reject, and route borderline parts to a safe default instead of forcing a guess. The same per-part scoring drives variant recognition in the field, for example bumper variants recognised above 0.9 confidence.

In practice the operating point is not guessed once and left. It starts from the spread the good set shows, then moves as real production data accumulates, tightening where the line proves stable and loosening where a legitimate new source of variation appears. Treating the threshold as a living setting, reviewed as lots and tooling change, is what keeps the false-reject rate low across a whole run rather than only on the first day.

Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators. Separating acceptable variation from a real defect is a tuning step its per-part scoring is built to support.

For the metric behind escapes and false rejects, see the sibling post on the false-negative rate, and for the full method behind few-shot and anomaly training, see the pillar guide. To see where variation tuning sits among the tasks the unit runs, browse the real applications.