Predictive maintenance means servicing a machine when measurements show it is wearing out, rather than on a fixed calendar, and machine learning is one way to turn sensor readings into early warnings. The best-known U.S. federal estimate, in a Department of Energy guide dated August 2010, puts the saving at eight percent to 12 percent against a preventive-only program. That figure comes from studies that predate modern machine learning, so it describes the maintenance strategy, not the AI.
Key takeaways
- The Department of Energy's Operations and Maintenance Best Practices Guide (Release 3.0, August 2010) estimates eight percent to 12 percent savings for predictive over preventive maintenance and 12 percent to 18 percent for preventive over reactive.
- The guide's larger figures, such as 70 percent to 75 percent fewer breakdowns, are attributed to unnamed independent surveys and are not specific to AI.
- Machine learning enters mainly through estimates of remaining useful life, which need run-to-failure data that is costly to collect.
- NASA publishes a simulated turbofan engine dataset that researchers use to practice these predictions, but simulated engines are not real fleets.
- A savings claim is only as good as its baseline, so the comparison group matters more than the headline percentage.
Three ways to schedule maintenance
The Department of Energy guide, prepared by Pacific Northwest National Laboratory for the Federal Energy Management Program, separates maintenance into three types. Reactive maintenance repairs equipment after it fails. Preventive maintenance follows a schedule based on calendar time or machine run time, much like changing a car's oil every 3,000 to 5,000 miles. Predictive maintenance bases the work on measured condition, so the oil would be analyzed and changed only when its properties call for it.
| Approach | Trigger | Estimated saving in the guide | Stated drawback |
|---|---|---|---|
| Reactive | Failure | Baseline | Unplanned breakdowns and emergency labor |
| Preventive | Calendar or run time | 12 percent to 18 percent over reactive | Some unneeded work; catastrophic failures still occur |
| Predictive | Measured condition | Eight percent to 12 percent over preventive | Spending on diagnostic equipment and staff training |
The guide adds that a facility relying heavily on reactive maintenance could see savings above 30 percent to 40 percent. It describes all of these as estimates from past studies and does not tie them to any one technology.
Where machine learning fits
Predictive maintenance existed long before current AI. Vibration, temperature, and oil analysis were compared with limits by technicians. What machine learning adds is a way to learn the link between a stream of sensor readings and later failure, then report either a health score or a forecast. A National Institute of Standards and Technology report on prognostics and health management in manufacturing (NIST IR 8012, June 2014) describes two outputs: a health index from zero to one for diagnostics, and remaining useful life for prognostics. The report says a prognostic approach should give a predicted time of failure along with a confidence interval, not a single guess.
A public practice dataset, and what it cannot show
To train and compare prediction methods, researchers need histories that run from healthy to failed. NASA's Prognostics Center of Excellence hosts one of the most used, the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) turbofan degradation data. Its authors, Abhinav Saxena and Kai Goebel, created it in 2008. The data are simulated, not recorded from real engines.
| Subset | Training engines | Test engines | Operating conditions | Fault modes |
|---|---|---|---|---|
| FD001 | 100 | 100 | One | One |
| FD002 | 260 | 259 | Six | One |
| FD003 | 100 | 100 | One | Two |
| FD004 | 248 | 249 | Six | Two |
Each engine contributes one row per operating cycle, with three operational settings and a set of sensor measurements, plus noise. Faults grow until failure in the training data, while test series stop before failure, and the goal is to predict the cycles left. NASA's data portal lists the licensing as not specified. A method that scores well here has shown it can learn a pattern in clean, simulated data. It has not shown that it works on a plant's mixed fleet, missing readings, and unrecorded repairs.
Why measured results are hard to compare
NIST IR 8012 states that standards for system development, data collection and analysis, data management, and software interoperability appeared to be partly lacking in 2014. It also notes that no consistent method existed for assessing both the technical and economic benefits of these systems. Data from seeded fault testing, accelerated testing, or field monitoring is expensive to collect, and the report says such data typically makes up less than 10 percent of what thorough verification needs. The report is more than a decade old, and later standards work has not been reviewed for this article, but the data problem it describes is the same one that keeps real-world savings figures scarce.
The Department of Energy guide is equally candid about costs. Starting a predictive program is not inexpensive, it says, with much of the equipment costing more than USD 50,000, and it lists staff training and savings that management cannot easily see as disadvantages. Its list of average results from "independent surveys" (10 times return on investment, 25 percent to 30 percent lower maintenance costs, 70 percent to 75 percent fewer breakdowns, 35 percent to 45 percent less downtime, and 20 percent to 25 percent more production) gives no study names in the passage, so those numbers cannot be traced from the guide alone.
Questions for any predictive maintenance claim
- What was the baseline: reactive, preventive on a fixed schedule, or an earlier monitoring program?
- How many machines and how many months did the comparison cover?
- Were real failures counted, or only alerts raised by the model?
- Did the savings include sensors, integration, training, and staff time to act on alerts?
- Was the model tested on machines it had not seen during training?
- Does the number come from a peer-reviewed study, a government report, or a seller's brochure?
FAQ
Does AI predictive maintenance save eight percent to 12 percent?
Not as a measured AI result. The range is the Department of Energy guide's estimate for predictive maintenance in general, drawn from earlier studies, compared with a preventive-only program.
Is predictive maintenance the same as AI?
No. Condition-based maintenance can use simple threshold alarms on vibration or temperature. Machine learning is one method for reading the same data and forecasting failure.
What data does a machine learning maintenance model need?
Sensor histories linked to known failures or repairs. NIST notes that run-to-failure style data is expensive to obtain, which is one reason public simulated sets such as C-MAPSS are popular for research.
Is the NASA turbofan dataset real engine data?
No. It was produced with a simulation, and each engine starts with unknown wear and manufacturing variation that the dataset treats as normal.
Related reading
- How AI helps companies cut costs
- AI for safety and efficiency at sea
- AI across U.S. industries
- Ten AI technologies used in industry
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Operations and Maintenance Best Practices: A Guide to Achieving Operational Efficiency, Release 3.0 (PNNL-19634), Pacific Northwest National Laboratory for the U.S. Department of Energy, August 2010
- Vogl GW, Weiss BA, Donmez MA, Standards Related to Prognostics and Health Management (PHM) for Manufacturing (NIST IR 8012), National Institute of Standards and Technology, June 2014
- Turbofan Engine Degradation Simulation (C-MAPSS), NASA Prognostics Center of Excellence Data Repository, Saxena A and Goebel K, 2008
- CMAPSS Jet Engine Simulated Data, NASA open data portal, updated May 29, 2025
- Saxena A, Goebel K, Simon D, Eklund N, Damage Propagation Modeling for Aircraft Engine Run-to-Failure Simulation, International Conference on Prognostics and Health Management, 2008