In predictive maintenance, data analytics combines machine readings such as vibration, temperature, current and operating hours with maintenance records to decide when maintenance is actually needed. Its job is not to know about every failure in advance; it is to catch early signs of degradation and move maintenance from an unplanned breakdown to a time you can plan. The methods range from simple threshold rules to anomaly detection and remaining useful life estimates.
We first published this article in April 2026 and expanded it in October 2026. It now covers maintenance strategies, which data matters, the main methods and what to watch out for when you start.
Maintenance strategies: after failure, by the calendar or by condition
| Strategy | When maintenance happens | Strength | Weakness |
|---|---|---|---|
| Reactive (corrective) | After something breaks | Cheap at first, few staff | Unplanned downtime, risk of secondary damage |
| Preventive (scheduled) | By calendar or run hours | Easy to plan | Unneeded work, failures can still slip through |
| Condition-based | When a measured value crosses a limit | Decisions based on actual condition | Limits have to be set well |
| Predictive | When the trend in readings points to degradation | Moves maintenance to the right moment | Needs investment in data, tools and training |
The operations and maintenance guidance published by PNNL, a US Department of Energy laboratory, describes reactive maintenance as “run it until it breaks”. Preventive maintenance follows a schedule based on time or machine run hours, while predictive maintenance schedules tasks based on the measured condition of the equipment. The same source estimates that a well-functioning predictive programme can save 8% to 12% compared with preventive maintenance alone, and notes that it requires extra investment in diagnostic equipment and staff training.
The line between condition-based and predictive maintenance is thin. The first asks “is the value above the limit right now?”, the second “at this rate, when will it get there?” Most plants don’t pick a single strategy: they use predictive maintenance on critical machines and run cheap, redundant equipment to failure. Reliability-centred maintenance (RCM) makes this mix systematic by ranking equipment by the consequences of its failure (PNNL).
The choice matters. In a NIST study of US manufacturers (2016 data), the 25% of establishments that relied most on reactive maintenance had 3.3 times more downtime than the 25% that relied on it least. That is a correlation rather than proof of cause, but it shows how closely the maintenance approach is tied to production.
Which data matters?
- Vibration: The most widely used indicator for rotating equipment such as motors, pumps, fans and bearings. Imbalance, misalignment and bearing wear all change the vibration pattern. PNNL lists vibration analysis alongside thermography, oil analysis, motor analysis and ultrasound as core condition monitoring techniques.
- Temperature: Bearing, winding and cabinet temperatures reveal friction, overload or cooling problems. Read without the ambient temperature, they can mislead.
- Current and power: The current a motor draws reflects load changes and electrical problems, and it can often be measured from the electrical cabinet without touching the machine.
- Operating hours and cycle counts: Wear depends on use more than on the calendar, so a plan based on run hours usually makes more sense than a fixed schedule.
- Maintenance and failure records: Which part was replaced, when, and with what symptoms? These records are one of the few places a model can learn what a failure looks like. NIST researchers point out that text documents account for a significant share of the data collected across the asset management life cycle, and that this source is underused.
However good the sensor data, record its context too: which product was the machine running, and at what speed? The same vibration level can be a warning sign at low load and perfectly normal at high load. We explain how to collect this data from the field in our post on IoT remote monitoring.
From simple thresholds to remaining useful life models
You can think of the methods as four steps of increasing complexity. Each builds on the previous one, and for most plants the first two already deliver real value.
- Fixed thresholds. “Alert if bearing temperature exceeds the set value.” Easy to set up and easy to explain. Limits come from manufacturer data, standards or the team’s experience.
- Trend monitoring. Look at the rate of change rather than the value itself. If vibration has been creeping up for weeks, you can plan maintenance before it crosses any limit.
- Anomaly detection. Learn the machine’s normal behaviour from data and flag deviations from it. The key advantage is that it needs no past failure examples, only healthy operating data covering different conditions. It can also catch variables that move together, such as current rising while speed stays constant.
- Remaining useful life (RUL) models. These try to answer “how many hours does this bearing have left?” They are the most informative and the most data-hungry. RUL models are usually trained on histories that run from a healthy state to failure, which is how the datasets in NASA’s prognostics data repository are structured. Real plants rarely have many such examples. A study by Krokotsch and colleagues notes that RUL data can only be labelled after a machine has failed and that data from machines near the end of their life is rare.
Where you stop depends on how critical the machine is and how much data you have. Trend monitoring may be enough for a cheap, redundant pump, while anomaly detection can be worth the investment for a compressor that stops the whole line when it fails. We cover other uses of AI in production in our post on Industry 4.0 and AI.
Where to start
- Pick the critical machines. You don’t need sensors on everything. Start with the few machines whose failure hurts production, safety or customers the most. PNNL’s RCM approach also weights equipment by the consequences of failure.
- Talk through failure modes. List with your maintenance team how these machines have failed in the past. That list tells you which data to collect: vibration for bearing problems, current and temperature for winding problems.
- Collect a baseline. Gather a few weeks of data while the machine is healthy, across different products, speeds and shifts, so you know what normal looks like. Anomaly detection doesn’t work without this reference.
- Design alerts technicians trust. Every alert needs an owner and a clear action. Start with a few meaningful alerts and tune the limits with technicians’ feedback.
- Close the loop. Record what was checked after each alert and what was found: a real problem or a false alarm? Over time, these records become the model’s most valuable training data.
Common pitfalls
Waiting for labelled failure data. “Let’s collect enough failures first, then build a model” can take years. Trend monitoring and anomaly detection deliver value without failure examples, and you can move to more advanced models as labelled data builds up.
Alert fatigue. A system that produces lots of unnecessary alerts is soon ignored. The ISA-18 series of standards on alarm management in the process industries calls for minimising nuisance alarms and alarm floods and for rationalising and prioritising alarms. The same principle applies to predictive maintenance alerts.
Ignoring maintenance records. Sensor data tells you what changed; maintenance records tell you what it meant. Even free-text records with missing codes are valuable once cleaned and matched with sensor data. If records still live on paper, moving them into a digital maintenance system may be the first step.
Promising certainty. No model catches every failure in advance. The goal is fewer unplanned stoppages and better-timed maintenance, and setting that expectation early protects trust in the project.
Frequently asked questions
Do we need AI for predictive maintenance?
No. Much of the benefit can come from well-chosen thresholds and trend monitoring. Machine learning earns its place when variables interact and simple rules are not enough.
How much data do we need?
For anomaly detection, healthy data covering the machine’s different operating conditions is enough. Remaining useful life estimates only become reliable once several degradation examples have been recorded.
Does it work on older machines?
Yes. You can add external vibration, temperature or current sensors to machines whose control systems provide no data, usually without modifying the machine itself.
Conclusion
Data analytics for predictive maintenance starts with the right data: vibration, temperature, current, operating hours and maintenance records. Methods range from simple thresholds to anomaly detection and remaining useful life models. For most plants, though, the real gain comes from collecting baseline data on critical machines and setting up a small number of meaningful alerts that technicians trust.
We build software for asset tracking, maintenance management and sensor data analysis. If you are wondering where to start with your machines, take a look at our asset management and AI solutions pages, or get in touch and we can look at it together.
