Normal behavior model

A normal behavior model, often abbreviated NBM, learns the relationship between a target channel and the conditions that drive it, using a period when the machine was known to be healthy. Predict gearbox oil temperature from power, ambient temperature, and rotor speed; predict power from wind speed, direction, air density, and turbulence intensity. Once trained, the model runs on live data and the residual - measured minus predicted - carries the information.

The point of this construction is that it removes the conditions from the comparison. A raw temperature rising in July tells you nothing; a temperature five degrees above what the model expects for that load and that ambient tells you a great deal. It is the general form of the argument that makes a learned power curve more useful than a datasheet one, applied to every channel a turbine records.

It also has real failure modes worth understanding. If the training period already contains the fault, the model learns the fault as normal. If a channel used as an input is itself degrading - a drifting nacelle anemometer is the classic case - the model compensates for the problem it is supposed to find. And after a genuine repair or a control change, the baseline has to be re-established rather than carried forward.

Handled carefully, this is the standard basis for condition monitoring on SCADA-rate data and for the detection of slow underperformance. Its great practical advantage is that it needs no failure examples to train on, which matters in a domain where the events being predicted are, thankfully, rare.