Predicting drive failures on a cement plant’s mills and kiln

Two unplanned stoppages of the raw mill in a year had each cost several days of production. Existing vibration rounds were monthly and hand-held, so faults that developed between rounds were missed.

Process load varied strongly with feed and moisture, so simple vibration thresholds produced alarms the team had learned to ignore.

We defined the measurement strategy per asset: tri-axial accelerometers on gearbox bearings, motor current signature on the main drives, and synchronisation with the plant historian for load, speed and feed.

Order tracking and envelope analysis separated bearing and gear signatures from load-dependent components. Condition indicators were normalised against operating state before any model saw them.

A failure-likelihood model was trained on eighteen months of history and the first confirmed faults, then deployed to a monitoring service with a review workflow for the maintenance planner.

Four developing faults were identified during the pilot year, each confirmed on inspection and repaired during planned stops.

The planner now receives a weekly condition report with ranked assets, evidence spectra and the indicator history, rather than threshold alarms.