Creating AI Anomaly Detector Models

The process of creating a model for our AI anomaly detector is mostly automatic - you only need to provide two things.

1) Any properties we should ignore, and 2) Example times where the asset has been behaving 'normally'.

There are some properties that shouldn't be included in anomaly detector models. These include…

Points with Radically Different Time-Scales

If you're detecting anomalies using properties that change every second, you should avoid mixing in properties that only change daily (such as daily performance metrics and KPIs) unless you're willing to include large amounts of sample data, as false-positives are common when mixing data across time-scales.

Points Already Powered By Anomaly Detectors

If you've already made an anomaly detector for this asset and are using it live, you should make sure you exclude that point from your AI. Otherwise you end up with a circular dependency - a situation where your AI needs data from itself to function.

By default, the system will take a snapshot of the last 2 hours of time as 'normal'. But you can define your own set of times you use as examples of normal operation for the asset.

Normal might also include situations where…

* The asset is switched off, * The asset is on, but unutilised, * The asset is operating in extreme environmental conditions

Read our walkthrough on creating an anomaly detector.