Earning a plant’s trust in predictive maintenance
Program lead across four organizations, plus data preparation and validation design.
Over 95% accuracy · under 3% false alarms
How I ran the proof of conceptSimplified view of the workflow
Context
At a U.S. vehicle plant, engineers checked the paint-shop ventilation fan motors by hand, and the rule-based vibration alerts depended on those checks. A motor failure can stop the line, but the alerts kept flagging motors as abnormal even after repair, and the maintenance team had stopped trusting them.
Decision & trade-off
The plant had little time for an outside project. I chose the use case that asked the least of the plant instead of the one with the largest potential impact. I also defined success as fewer false alarms and a score that returns to normal after a repair, not accuracy alone.
My actions
I started by asking the leads of five shops about their problems before proposing anything. I then brought together the innovation center, an AI partner, a sensor supplier, and the plant, and scheduled data collection around production so the line never stopped. I prepared and labelled the vibration data myself, compared normal and abnormal periods, and designed the test that checked the partner’s model against the plant’s existing alerts.
Outcome
Over 12 weeks on live production equipment, the model reached more than 95% accuracy with a false-alarm rate under 3%. I delivered a roadmap for extending the approach, and smart factory projects of this kind later expanded across the company. The AI partner built the model; my part was the program, the data preparation, and the validation.