Predictive maintenance

There are three ways to maintain a turbine. Run it to failure, which is cheapest until the failure happens and most expensive after. Service it on a calendar, which is predictable and replaces plenty of healthy components. Or service it on evidence, which is what predictive maintenance means: watching for the signatures that precede a failure and intervening inside that window.

The evidence is usually already being recorded. Bearing degradation, gearbox wear, and generator faults develop over weeks and show up as drift in vibration, temperature, and power signatures long before an alarm threshold is crossed. Detecting them is a question of knowing what normal looks like for that specific machine, which is the same per-turbine baseline that reveals yaw misalignment and gaps against a learned power curve.

The economics are decided by the cost of being wrong, which is why precision matters more here than in most machine learning applications. A missed failure is an unplanned outage and possibly a crane. A false positive sends a crew up a tower for nothing - and offshore, that is a vessel. A model that flags everything is not useful, because a fleet cannot act on it.