A supervised classifier learns each defect from examples, which works beautifully when defects are common and well sampled. The trouble on a mature production line is that the defects that matter most — contamination, a coating void, a safety-critical flaw — are rare, sometimes a fraction of a percent of output, and are precisely the ones that must never pass. There are simply too few examples to train a reliable classifier for them, and a fault the line has never produced before has no examples at all. Optimising a classifier for the frequent, benign defects while missing the rare, dangerous one is the classic failure.
Anomaly detection inverts the problem. Instead of learning what each defect looks like, it learns what a good part looks like — and good parts are abundant. Anything that deviates from that learned normal is flagged, whether or not the system has ever seen that particular defect. This is what makes it possible to catch the rare, the novel and the safety-critical: the model does not need an example of the fault, only a thorough understanding of normal. On a line where an escape is expensive, that is exactly the property you want.
The discipline moves to modelling normal well. A real part has legitimate variation — sheen, texture, print, tolerance — and a model that has learned too narrow a slice of normal flags that variation as a fault, driving false alarms. So the work is training on enough good-part variation that ordinary differences are understood as normal, and setting the sensitivity so genuine deviations stand out without burying the line in false calls. In practice the two framings are layered: classify the common defects where examples exist, and let anomaly detection catch everything else — including the fault no one has seen yet.