The situation
Turbine components were failing between inspections. The operator had vibration and temperature history going back years and wanted to know whether it contained enough warning to act on.
The constraint
A false alarm sends a crew to a turbine for nothing, and a missed failure costs a gearbox. The model had to be wrong in the cheaper direction, and it had to give warning far enough ahead to fit a maintenance window rather than merely confirming a failure in progress.
What I decided
I fixed the useful lead time first and built backwards from it, rather than optimising accuracy and discovering afterwards that the warnings arrived too late to use. Most of the work was in the sensor pipeline: which readings were trustworthy, what a gap in the record meant, and which degradation patterns were physically plausible instead of merely correlated.
What shipped
A prediction system over turbine sensor data, issuing early warnings against components with a history of expensive failure, with its confidence and its blind spots stated explicitly.
What it produced
Around a fifth less unplanned downtime, and maintenance scheduling that could be planned against the forecast rather than interrupted by the fault.