True vs False anomalies
Three reasons why time-series leave the normality domain.
1 State/parametes/context
In this first set of slides, we recall the problem of anomaly detection from blind normality characterization using healthy data.
The slides below explain that the time-series produced by an industrial equipment depend on three items, namely:
The so-called state vector of the equipment
The vector of parameters of the equipments and
The exgenous items representing the context of operation.
Ideally, we need to detect excursions, outside the normality domain, that are due to changes in the system’s parameters or internal relationships and not because of a new unseen context or because of the state visiting regions of operation that were not encountered in the training data.
2 The state/context induced false alarm
The next slides show a visual explanation of the above mentioned three reasons for which the time-series indicators leave the domain of normality as learned using the healthy training data.
