The alarm log is a lagging indicator

A turbine's alarm system is a protection system. Its purpose is to stop the machine before damage becomes dangerous or catastrophic, and it does that by comparing signals to fixed thresholds. Those thresholds are necessarily set where intervention is unavoidable, because a protection system that stopped the machine on early evidence would be stopping it constantly. An alarm is therefore, by design, the last moment rather than the first.

The interval between the first detectable evidence and the threshold crossing is where predictive maintenance lives. For gearbox and main bearing degradation, temperature and vibration signatures typically depart from their own normal weeks to months before a protection threshold is reached, and the departure is small enough that no fixed limit would catch it. The signal is there; what is missing is a reference for what normal looks like on that specific machine.

This is why per-turbine modeling matters more here than anywhere else. A gearbox temperature that is three degrees above where that gearbox usually sits under the same load and ambient conditions is meaningful. The same three degrees against a fleet-wide limit is nothing, because the limit has to accommodate the hottest legitimate machine in the population. A threshold that is safe for every turbine is uninformative for each one.

The commercial argument is straightforward, and it is stronger offshore than anywhere else. A planned intervention performed inside a scheduled campaign, with the part already on the vessel, and an unplanned response to a failed component are different by a large multiple in both cost and lost production. Notice measured in weeks is what converts one into the other, and there is no version of that conversion that starts from the alarm.

The counter-argument deserves a hearing. Early warning has value only if the organization can act on it, and a finding that arrives eight weeks before failure but three weeks after the last scheduled campaign has bought very little. This is a real limit, and it is an argument about planning horizons rather than about detection: the operators who get the most from predictive findings are the ones who changed how campaigns are scheduled, not the ones who added a feed to an existing process.

It is also worth stating what this does not do. A finding narrows where to look and establishes what looking is worth. It is not a diagnosis in the sense that a borescope inspection is a diagnosis, and it cannot see a failure mode that leaves no trace in the recorded channels. A platform that claims otherwise is describing an instrument set it does not have.

One thing to check on your own fleet: take the last three unplanned component failures and pull the twelve months of history preceding each one. The question is not whether the failure was predictable in hindsight - it usually is - but whether the departure from that machine's own normal was large enough that a model would have raised it before the alarm did. That exercise is the most honest test of predictive maintenance available, and it costs nothing but the data you already hold.