anomaly detection

Flagging readings that don't look like the ones a device has seen before, without examples of faults.

Anomaly detection identifies observations that deviate from a learned model of normal behaviour, typically trained on normal data only; on a device the simplest form is a z-score against a running baseline.

Examples

Each fan learns its own normal vibration and flags windows with z above 4, with no labelled examples of faults needed.
The lab's detector catches the imbalance trace at window 106 using only what healthy running taught it.