Predictive maintenance
the engagement
client data — not shown
heard
"When the press goes down it costs us eleven thousand euro an hour and we always find out from the operator."
real problem
Maintenance ran on a calendar. Failure precursors were present in vibration and temperature data for hours beforehand, but nothing was reading it.
system built
Streaming ingestion from the PLC historian, a remaining-useful-life model per asset class, and alerts routed into the existing maintenance planner with a lead time the planners agreed was actionable.
what broke
Alerting on the model's raw output produced a false alarm every other shift. Trust recovered only after we tuned to a precision target the planners chose themselves.
the open rebuild
same architecture · public data
dataset
NASA C-MAPSS turbofan degradation
what the rebuild covers
The RUL model and the alert-threshold calibration are complete. No streaming layer in the rebuild — batch scoring stands in.
artifacts
NotebookCodeDatasetWrite-upDemo