CASE 11 · CASE ATLAS · LEVEL B WORKSHOP SCENARIO

Gearbox condition-monitoring alarms

SITUATION

A wind farm's condition-monitoring system generates hundreds of vibration and oil-debris alarms a week across its gearbox fleet, most of them benign. The reliability engineer needs the genuinely urgent alarms triaged before the next scheduled service visit, not a fresh unranked list every morning.

EVIDENCE

Time-series vibration spectra and oil-debris sensor readings per gearbox, the fleet's historical alarm-to-failure record, and prior technician dispositions (false alarm, monitor, intervene) for a sample of past alarms.

AI APPROACH

A statistical anomaly-scoring method — not a language model — ranks open alarms by similarity to the sensor signatures that preceded past confirmed failures, presented as a prioritized triage queue rather than a failure prediction.

KEY CONTROLS

The ranking is explicitly a triage aid; no alarm is auto-closed or auto-escalated to a work order without a reliability engineer's review, and the model's training set (which failures it has and has not seen) is stated alongside the ranking.

VALIDATION MEASURES

Back-test the ranking against the fleet's own alarm-to-failure history on a held-out period, tracking both missed high-risk alarms and false-urgent flags, since a missed bearing failure and a wasted site visit carry very different costs.

HUMAN DECISION

The reliability engineer decides which alarms trigger an inspection or intervention and owns the resulting work order; the ranking only orders the queue, it does not close it.

DISCUSSION QUESTIONS

  1. What failure mode has this fleet had that the model has never seen an example of, and what does that mean for trusting its ranking on a novel signature?
  2. How would you notice if the queue quietly started deprioritising a real recurring failure pattern?
  3. Who re-baselines the ranking after a gearbox design or oil-sampling interval changes?

TRY IT LIVE

Run this scenario's approach through Correlated Error Simulator — How often can a clean sample hide a failing class?

Open Correlated Error Simulator →