Define the failure you want to detect
Separate a failure mode from a general anomaly. Agree which action an engineer could take, how much warning is useful and the cost of a false alarm.
Build a usable baseline
- Signals
Align timestamps for vibration, temperature, current, load and operating mode.
- Events
Connect work orders, inspection notes, component changes and failure labels.
- Context
Compare similar operating conditions and account for sensor maintenance.
- Evaluation
Review held-out time periods and compare alerts with a simple threshold baseline.
Keep the maintenance decision with the engineer
An alert should show the source signals, time window, uncertainty and relevant history. Record whether the engineer accepted, dismissed or investigated it. Count useful warnings and operating cost over an agreed period.




