Explainer · 5 min read
Predictive maintenance starts with the maintenance log
Most refineries and chemical plants have recorded vibration, temperature and pressure for years. The historian is full. So it is natural to assume predictive maintenance is a matter of pointing a model at that data. In practice, the sensor data is only half of what is needed.
A model learns to recognise a failure by seeing what the signals looked like before past failures. That means knowing when each failure happened, on which machine, and what kind of failure it was. That information lives in the maintenance system, and it is often the weaker half.
Why the work orders matter
Work orders are written for the people doing the repair, not for analysis. The same failure is described in a dozen ways, the date recorded is when the order was closed, and planned and emergency work look alike. Before any model is trained, those records have to be read, cleaned and matched to the equipment and the sensor tags.
This is unglamorous work, and it decides whether the project succeeds. A model trained on muddled failure history will raise warnings nobody trusts.
Start with a few machines
The best first project is narrow: a handful of critical pumps or compressors, with known failure modes and enough history. Prove that the warnings arrive early enough to plan the work, then extend to more equipment.
Keep the model honest
Every warning should be recorded along with what was found when the machine was inspected. That record shows the reliability team how often the model is right, and gives the model better examples over time. A warning system that is never checked tends to drift into one that is ignored.
This is the kind of work our advanced analytics & predictive ai line does.