What predictive maintenance is
Maintenance has traditionally been reactive, fixing a machine after it fails, or scheduled, servicing it at fixed intervals whether or not it needs it. Predictive maintenance adds a third option: sensors on the machine measure signals such as vibration and temperature, and a model learns what normal behaviour looks like so that drift away from it can be flagged early.
For a plant manager the useful definition is practical. It is a way to plan a repair before an unplanned stop, using evidence from the machine itself.
What the evidence says
The anchor study is a peer-reviewed review, Artificial intelligence and robotics in predictive maintenance: a comprehensive review, published in Frontiers in Mechanical Engineering on January 7, 2026. The authors selected 85 full-text studies from 1,864 database records.
- Support vector machines and neural networks reach roughly 85% to 94% accuracy for fault classification.
- Reported downtime reductions are 15% to 40% across manufacturing, aerospace and energy.
- The main challenges are data quality, computing cost, interpretability, generalisation between machines and system integration.
These figures are collected from many studies with different machines, data and definitions. They are ranges to plan around, not a guarantee for any single plant.
Why false alarms decide whether it works
A predictive system that raises too many alerts is quickly ignored. A 2021 case study in Sensors worked on two real machines: a coal crusher at a coal-fired power plant and a 500-tonne gantry in a steelworks converter. The authors added operator feedback and explanations on top of an outlier-detection model and reduced false alarms by 90.25% on average, compared with the stand-alone method.
That result is against the authors’ own baseline on two machines, not against other published systems. The lesson is about design: alerts must be specific enough that the maintenance team acts on them, and operators must be able to see why an alert was raised.
Why small plants lag, and what that means
A 2016 pilot survey affiliated with NIST, The present status and future growth of maintenance in US manufacturing, found large manufacturers ahead of small and medium ones in adopting advanced maintenance strategies. Cost, technical support and workforce skills were the main barriers, and reactive maintenance was still common among smaller firms.
Most of the strongest published results come from heavy industry and large manufacturers. A small plant should read them as evidence that the method can work, and test it on its own equipment before committing.
A one-machine pilot
Our own suggestion, not a finding from the studies:
- Pick one critical machine where an unplanned stop is costly.
- Start a failure log on day one. Record every failure and repair, with date, cause and downtime.
- Add a small number of sensors, such as vibration and temperature, to that machine only.
- Judge success by action. Are alerts specific enough that the maintenance team acts on them, and are false alarms low enough that they trust them?
- Expand only when the result holds for several weeks.
The failure log needs somewhere to live. Manufacturing software that already tracks production, inventory and quality, such as our Gaamma manufacturing ERP, is one place it can sit next to the rest of the plant’s data.
Sources
- Azeta et al., Artificial intelligence and robotics in predictive maintenance: a comprehensive review, Frontiers in Mechanical Engineering, January 7, 2026
- Hermansa et al., Sensor-Based Predictive Maintenance with Reduction of False Alarms: A Case Study in Heavy Industry, Sensors, December 29, 2021
- Jin et al., The present status and future growth of maintenance in US manufacturing, Manufacturing Review 3, article 10, 2016
Frequently asked questions
What is predictive maintenance?
It uses sensors on machines, such as vibration and temperature, and a model that learns what normal looks like, so signs of failure can be flagged before a breakdown. It sits between fixing things when they break and servicing them on a fixed schedule.
How accurate is AI at predicting machine failures?
A January 2026 review of 85 studies in Frontiers in Mechanical Engineering reports roughly 85% to 94% accuracy for fault classification and downtime reductions of 15% to 40%. These are averages across many studies and sites, not a promise for any one plant.
Can a small factory use predictive maintenance?
Possibly, but the evidence is thinner. A 2016 NIST pilot survey found small and medium manufacturers behind large ones on advanced maintenance. A sensible start is one critical machine, a few sensors and a written failure log.
What is the biggest practical problem?
Data quality and false alarms. The same review names data quality, computing cost, interpretability, generalisation and integration as the main challenges, and a 2021 heavy-industry study was built specifically to cut false alarms.
What you get:
- What predictive maintenance is and where it sits between reactive and scheduled upkeep
- What an 85-study 2026 review reports on accuracy and downtime
- Why false alarms decide success, with a heavy-industry case
- A one-machine pilot plan with a failure log and clear success test
- Where small manufacturers stand, per a NIST-affiliated survey