Predictive maintenance means monitoring the condition of equipment through measurements (vibration, temperature, oil, electrical current) to forecast when it will fail, and stepping in just before. The EN 13306 standard calls it predictive maintenance, a form of condition-based maintenance. Its purpose is to avoid unplanned production stops without replacing parts that are still good.
What is predictive maintenance?
The EN 13306 standard (2018) defines predictive maintenance as condition-based maintenance carried out following a forecast derived from repeated analysis or known characteristics and evaluation of the significant parameters of the degradation of the item. In other words: a sign of wear is measured regularly, its trend is followed over time, and the probable failure date is derived from it.
The principle rests on a simple observation: equipment almost always shows signs of degradation before it fails (rising vibration, a bearing heating up, oil carrying particles). Predictive maintenance aims to spot these signs early enough to plan the intervention.
Predictive, preventive or condition-based maintenance: what is the difference?
- Corrective maintenance: repair after the failure.
- Predetermined preventive maintenance: intervene on a fixed schedule (running hours, cycles, calendar) without checking the equipment's condition. Example: an oil change every 500 hours.
- Condition-based maintenance: intervene when a measurement crosses a threshold. The trigger is a condition measurement, not a date.
- Predictive maintenance: follow the trend of the measurement and project when the failure will occur. The output is not an alert but a horizon: so many days or cycles left before the probable failure, known as the remaining useful life.
The standard places predictive maintenance within condition-based maintenance, itself a form of preventive maintenance. In practice, many sites combine all four according to how critical each asset is.
Monitoring techniques
- Vibration analysis: detects mechanical faults in rotating machines (imbalance, misalignment, bearing wear).
- Infrared thermography: finds hot spots (overheating bearing, faulty electrical connection), without contact and while running.
- Oil analysis: measures internal wear through metal particles, viscosity and water contamination.
- Ultrasound: finds compressed air or gas leaks, electrical discharges and lubrication faults.
- Electrical current analysis: changes in the current drawn point to problems in the motor or in the machine it drives.
The choice depends on the failure mode to monitor: a vibration sensor is no use on an electrical cabinet.
How it works, from sensors to work orders
1. Select the assets whose failure costs the most in lost production or safety.
2. Select the measurement that reveals their failure mode, and fit the sensors or organise the readings.
3. Establish the baseline in normal operation, then the alert thresholds.
4. Collect and store the data, linked to each asset.
5. Analyse: with thresholds and trends, or with machine learning models that compare an asset with its own history and with similar assets.
6. Trigger the intervention in the CMMS, with the parts and the technician, at the least costly time for production.
The CMMS is the link that turns an alert into a planned intervention. See our CMMS guide.
Predictive maintenance and AI
Machine learning has long been used for predictive maintenance: it spots drifts that no fixed threshold catches, by combining several measurements. Generative AI brings a more recent use: finding, in a technical manual that can run to 2,000 pages, the procedure that addresses the detected problem. Vendors also show agents that, starting from a sensor alert, check the history, create the work order and assign it to an available technician.
Examples in industry
- SNCF: from 2013 to 2020, the French railway operator fitted its trains with sensors, first on air conditioning, then on doors, compressors, batteries and motors. In 2021, it reported monitoring data from 900 trains and having cut their failures by 20% (Journal du Net, 02/06/2021). It also stressed that sensor data alone is not enough: it has to be combined with other sources.
- Air France Industries KLM Engineering & Maintenance: its Prognos tool detects a failure 10 to 20 flights before it happens, about a week to replace the part. It took 2 to 3 months of data to build a reliable model for a new system (Journal de l'Aviation, 13/07/2017).
- TGV M: for its new high-speed train, SNCF is moving from scheduled inspections to maintenance triggered by actual condition, based on sensors reporting continuously (Le Monde Informatique, 27/05/2024).
What manufacturers said at Tech for Industry Show
At Tech for Industry Show 2026, at Paris Expo Porte de Versailles:
- Alstom presented Hub, its predictive maintenance platform that uses sensor data to anticipate failures and schedule replacements at night or at weekends. For the Riyadh metro, the customer required its data to stay in Saudi Arabia. Read the article
- AVEVA cited Air Liquide, which monitors more than 300 sites worldwide and more than 40,000 assets, and detects drifts months in advance to prepare maintenance and spare parts. Read the article
- Renault described a predictive maintenance project on its robots' gearboxes, stopped for lack of results, then relaunched with a PhD student: an example of the patience these projects require. Read the article
- Astriis, a young French company specialising in vibration analysis of rotating machines, won the show's Startup Award. Read the article
The talks are available as replays: Alstom and Exide Technologies: AI, simulation and data, Scaling sovereign AI.
Prerequisites and limits
- Reliable historical data: without an operating history, and ideally a failure history, a model cannot forecast anything. A new site starts with condition monitoring.
- Start small: one or two critical assets, or a pilot site, before scaling up. Many projects stay at pilot stage for lack of a business case and value tracking (Deloitte).
- Connect the tools: sensors, analytics platform, CMMS and ERP must share their data.
- Not every failure can be predicted: a sudden breakage or a human error leaves no measurable warning sign.
- Keep the expertise: interpreting a vibration spectrum or validating an alert remains a skilled job.
2027 exhibitors equipping predictive maintenance
To meet at Tech for Industry Show 2027 (in alphabetical order):
- CARL Software: maintenance management software and an IoT platform dedicated to predictive maintenance;
- Probayes: artificial intelligence and predictive analytics;
- SpinalCom: operational digital twin bringing together OT, IT and CMMS data.
Frequently asked questions
What is predictive maintenance, in one sentence?
Maintenance that forecasts an asset's failure from the trend of wear measurements, in order to intervene just before.
What is the difference between predictive and preventive maintenance?
Predetermined preventive maintenance intervenes on a fixed date, whatever the asset's condition; predictive maintenance intervenes when the measured wear announces the failure.
What is the difference between condition-based and predictive maintenance?
Condition-based maintenance reacts when a measurement crosses a threshold; predictive maintenance projects the trend of the measurement to estimate when the failure will occur.
Do you need AI for predictive maintenance?
No. Thresholds and trend curves are enough for many assets. Machine learning helps when several measurements need to be combined.
Where to start?
With the assets whose failure costs the most, a measurement suited to their failure mode, and a CMMS that turns alerts into interventions.
