Predictive maintenance promises a warning before equipment fails, so a repair can be planned instead of rushed. The models behind it are well understood. What decides whether it works in your plant is everything around the model: which assets you choose, what data they produce, what your maintenance records say, and what your team does when an alert arrives.

Choose the right assets first

Not every asset is worth predicting. Rank candidates on three questions:

  • Does a failure stop the line? A failed filler drive matters more than a redundant pump.
  • Does it fail often enough to learn from? An asset that has failed once in ten years gives a model very little to work with.
  • Is it already instrumented? Assets with vibration, temperature, current or pressure signals can start sooner.

The best first assets score well on all three. Starting there gives the pilot a fair chance of showing a result in 90 days.

Check the signals

Most early warnings come from a small set of measurements: vibration, temperature, motor current and pressure. For each candidate asset, check that these signals reach a historian, how often they are sampled, how long they are kept, and whether there are gaps. A historian that keeps only a few weeks of data, or averages it heavily, may need its settings changed before a model can learn from it.

Read your failure history honestly

A model learns what a failure looks like from past failures. That history lives in your CMMS, and its quality matters as much as the sensor data. Look at the last six months or more of work orders for your candidate assets:

  • Are breakdowns recorded with a time close to when they actually happened?
  • Do failure codes describe the failure, or just say "repaired"?
  • Can a work order be tied to a specific asset, and that asset to its sensor tags?

If the history is thin, the first weeks of a pilot may need to focus on recording failures properly. That is useful work, but it should be planned rather than discovered.

Design the alert workflow before the first alert

An alert nobody acts on is worse than no alert, because it teaches people to ignore the next one. Before go-live, agree:

  • Who receives an alert, on which shift, and how: text, CMMS notification or both.
  • The response window. A warning 24-48 hours ahead allows the repair to move into a planned stop, but only if someone looks at it in time.
  • How to dismiss an alert, and how a dismissal is recorded, so false positives improve the model instead of eroding trust.
  • How an accepted alert becomes a work order, ideally drafted automatically and approved by a person.

Measure the outcome, not the activity

The number of alerts is not a result. The result is fewer unplanned stops and fewer hours lost to them. Agree the downtime baseline before the pilot, from the same systems you will use to measure the pilot. Useful supporting measures are the share of alerts that led to a real finding, and the ratio of planned to emergency work over time.

What predictive maintenance will not do

  • It does not replace preventive maintenance or OEM schedules. It adds an early warning on top of them.
  • It cannot reliably predict a failure mode that has never happened on your equipment.
  • It cannot fix records that were never kept. Better data capture sometimes has to come first.

A short readiness checklist

  • Three to five candidate assets, ranked by impact, failure frequency and instrumentation.
  • Sensor signals in a historian, with known retention and gaps.
  • Six months or more of CMMS history for those assets.
  • A named owner for alerts on every shift.
  • An agreed downtime baseline and success metric.

The NuruState assessment works through this list on site. It shows which assets are ready for a pilot now and what infrastructure the others need first.