Traditional anomaly detection AI models are outlier detection models.

They're similar to a check engine light in that they explain that there is an anomaly, but they do very little to help explain what the anomaly actually is, which results in users needing to explore their data to find the cause of the alert.

Value Predicting Models instead try to predict what the status of the system should be, given the data it has available.

For instance, if you're running this on a machine with vibration, power and temperature monitoring, it tries to estimate what each of these values should be based on each-other, and on the overall state of the system around it (ie. ambient conditions, current product, load on the conveyor etc.).

By comparing these predicted values against the actual values, we can then see when there are discrepancies. If these discrepancies are large enough, we not only detect an anomaly, but we can show which specific sensor values are incorrect - for example, the temeprature being too high for the current motor load and ambient conditions.

Unfortunately, they aren't able to explain why they are indicating an anomaly - they only let you know that the data coming from the asset(s) is unexpected.

Generally, these types of AI generate a value - you can adjust exactly how sensitive your anomaly detector is to changes by choosing the threshold the system uses to go from saying 'OK' to 'Anomalous'.

A high threshold will result in a less sensitive anomaly detector.

The AI Anomaly Detector addon uses Support Vector Machines to provide this style of anomaly detection. See the differences between model types for more information.