Playbook ยท Practical AI education

Predictive maintenance pilot playbook

A step-by-step approach to selecting assets, choosing sensing and modeling methods, and proving that earlier warnings lead to less unplanned downtime.

Key takeaways

  • Choose assets where an early warning leads to a cheaper action and where failures develop over time.
  • Run a short failure modes review to decide which signals matter and how much lead time is useful.
  • Match the modeling approach to your failure history; anomaly detection is often the realistic starting point.
  • Record the outcome of every alert and integrate confirmed findings with the CMMS.

Pick assets where an early warning changes something

Predictive maintenance only creates value when an earlier warning leads to a cheaper action. Before choosing technology, list candidate assets and ask, for each one, what happens today when it fails and what would happen if you had a week of notice.

Good pilot candidates usually share these traits:

  • Failures cause meaningful downtime, scrap, safety exposure or secondary damage.
  • The asset has a known failure mode that develops over time, such as bearing wear, misalignment, seal leakage, fouling or insulation breakdown, rather than sudden random failures.
  • There are several similar assets, so the approach can be tested on more than one and scaled afterward.
  • Maintenance has a way to act on a warning, such as a planned window, spare parts and a technician who can inspect.

Rotating equipment such as motors, pumps, fans, gearboxes and compressors is a common starting point because the failure physics is well understood and vibration and temperature monitoring are mature.

Understand the failure modes first

Sit down with the reliability engineer and the technicians who work on the asset. Run a short failure modes and effects review covering how the asset fails, what symptoms show up before failure, how long that warning period typically lasts, and what is measured today. This tells you which signals matter and how far ahead a prediction must be to be useful. A warning that arrives an hour before failure may be fine for a fan you can swap quickly, but useless for a gearbox that needs a week to source parts.

Choose sensing and data sources

Match the data to the failure mode.

SignalTypical failure modes it revealsNotes
VibrationBearing wear, imbalance, misalignment, looseness, gear damageWaveform or spectral data is far more informative than a single overall value
TemperatureLubrication problems, overload, electrical faultsOften slow to respond; best combined with other signals
Motor current and powerLoad changes, rotor bar issues, process blockagesCan often be read from drives without new sensors
Pressure and flowPump wear, cavitation, filter clogging, leaksUsually already in the historian
Oil analysisWear particles, contamination, lubricant breakdownPeriodic rather than continuous
Ultrasound and acousticEarly bearing faults, leaks, electrical dischargeUseful for slow-speed equipment

Start with data you already collect, then add sensors where a clear gap exists. Wireless vibration and temperature sensors have made retrofits cheaper, but check battery life, sampling rate, range in your plant environment and how data reaches your network.

Choose a modeling approach that fits your history

There are three broad approaches, and the right one depends mainly on how much labeled failure history you have.

  • Thresholds and condition rules. Alarm when vibration in a known frequency band exceeds a limit set by standards or experience. Simple, explainable, and a sensible baseline to beat.
  • Anomaly detection. Learn what normal operation looks like for each asset under each operating mode, then flag departures. Works with little or no failure history, which is the common case. The tradeoff is that it tells you something has changed, not necessarily what.
  • Failure prediction and remaining useful life. Supervised models trained on past failures to estimate the probability or time to failure. Needs enough recorded failures with good labels, which many plants do not have for any single asset type.

Account for operating context. A pump running at a different flow rate or a machine running a different product will look different. Models that ignore load, speed and product will raise false alarms every time the schedule changes.

Design the alert workflow

An alert is the start of a maintenance process, not the end of an analytics process. Define who receives each alert, how quickly they must review it, what inspection or check they perform, how they record the result, and how a confirmed finding becomes a work order in the CMMS.

Ask technicians to record the outcome of every alert: confirmed fault, no fault found, or inconclusive. Those outcomes are your most valuable data. They measure the false alarm rate and they become labels that improve the model. Keep the number of alerts low at first. A short list of credible warnings builds trust; a flood of them trains people to ignore the system.

Measure what matters

Track technical measures and business measures separately.

  • Technical: faults detected before failure, warning lead time, false alarm rate, missed failures.
  • Business: unplanned downtime hours on pilot assets, emergency work orders, overtime, secondary damage, and spare parts expediting, each compared with the baseline period.

Example: suppose a hypothetical plant has twelve similar pumps and recorded a known number of unplanned pump-related stops in the prior year. During a six-month pilot, each alert is logged with its outcome. At the end, the team counts how many developing faults were caught early, how many were missed, how many alerts found nothing, and how the stop hours compare with the same months the year before, adjusted for any change in run hours. That comparison, not the model's accuracy score alone, is what goes into the go or no-go decision.

Be careful with attribution. If the plant also improved its lubrication routes or replaced a problem pump model during the same period, some of the change belongs to those efforts. Note any other reliability work on the pilot assets, and where possible compare against similar assets that were not monitored. A modest, well-supported result is more useful for the scale-up decision than a large number nobody believes.

Common pitfalls

  • Choosing assets that rarely fail, so the pilot never sees an event.
  • Relying on one overall vibration value when the failure signature lives in specific frequencies.
  • Not recording alert outcomes, which leaves no way to measure accuracy or improve the model.
  • Ignoring operating modes, which drives false alarms.
  • Failing to connect alerts to the CMMS, so warnings live in a separate dashboard nobody checks.
  • Expecting the model to replace the reliability program. It works best inside an existing practice of inspections, lubrication and root cause analysis.

Related use cases

Maintenance & reliability

Predictive maintenance

Use sensor and maintenance history to spot developing equipment failures early and schedule repairs before they stop the line.

Machine learningIndustrial IoT & edge
Established7 software companies

Operations & production

Process anomaly detection

Models learn normal process behavior across many sensors and flag unusual patterns before they become quality, safety or equipment problems.

Machine learningIndustrial IoT & edge
Proven, scaling7 software companies

Operations & production

Machine monitoring & OEE analytics

Automatically capture machine states, downtime reasons and output to measure OEE accurately and target the biggest losses.

Industrial IoT & edgeMachine learning
Established8 software companies

Put AI to work

Find the AI opportunities hiding in your plant.

Genvo, our implementation partner, works with manufacturers to pick the right first project, connect the systems you already run and put AI agents and tools into daily use, with people approving what goes out.

Assessments and consultations are provided by Genvo. How our partnership works.