Predictive maintenance: without data from legacy equipment, AI is running on empty.
At a thermal power plant operator we equip, a gas turbine trips on a fault. From an airport, a manager sees it in Mobapi and alerts the plant technician. The fault is cleared on site in under six minutes, and four hours of downtime are avoided. Nothing in this story was predicted: a measurement was coming in continuously, and someone saw it in time. Before you buy prediction, you need that continuous data across the whole fleet, including from the machines that have never transmitted anything.
AI is spreading, downtime is not coming down.
MaintainX’s 2026 report1, published on 5 May, surveyed 2,234 maintenance and operations managers in the United States and Canada. In 58% of cases, their team already uses AI. And yet 79% saw their unplanned downtime stay flat or rise over the year.
Two caveats. This is a vendor study, and MaintainX itself sells maintenance tools. The panel is not French. Its co-founder nonetheless draws a lesson from it that we share: reliability comes from execution maturity, not from system adoption alone.
What “predictive” requires.
The NF EN 13306 standard defines predictive maintenance as condition-based maintenance carried out following a forecast, derived from analysis of the parameters of the item’s degradation. To apply it, you need either models of the equipment’s behaviour or a large quantity of data from multiple sources, as Techniques de l’Ingénieur2 sums it up. Either way, it all starts with measurements.
On legacy equipment, the data is what’s missing.
This is where the real fleet gets stuck. A machine from 2005 was not designed to send its data anywhere but its own screen. A machine bought second-hand sometimes arrives with no documentation. And when the fleet mixes five makes, each manufacturer has its own software, its own format and its own history, when there is one.
What you find there is data in silos, readings on paper, maintenance work noted in a logbook. A model trained on three recent machines out of twenty says nothing about the other seventeen.
The turbine at the start of this article needed no model. It needed a measurement coming in, and someone to see it.
That is the foundation we lay with you. We read what your machines already give out, right down to a control system that only talks over a serial link or through its printer output, and we add 4-20 mA sensors to the ones that stay silent. Thresholds, alarms and maintenance work then sit in the same history. The day you want to try a predictive model, your tool will be able to retrieve the raw data through the API.
Where to start, with no promises.
Start with what costs most. First pick the equipment whose downtime costs you the most. There is no need to instrument everything on day one.
Read what is already there. Collect what PLCs, meters and analogue outputs already provide, then fit a sensor to the machines that say nothing.
Set thresholds. Monitoring a parameter and acting when it drifts is already condition-based maintenance as defined by NF EN 133063.
Record maintenance work. Log it in the same place as the measurements, so that a drift can be linked to what was done.
Consider predictive maintenance, later. Once you have the history in hand, not before.
One last question.
How much history before you can predict?
There is no universal duration. It all depends on how the equipment degrades and on how regular the measurements are. Put the question to your supplier, and be wary of an answer that doesn’t start with your data.
Sources.
-
MaintainX, press release of 5 May 2026 on the State of Industrial Maintenance report, 2,234 respondents in the United States and Canada.
-
Techniques de l’Ingénieur, article MT9280 on maintenance methods (definitions from the NF EN 13306 standard), public extract.
-
travail-industrie.com, page on choosing a type of maintenance (definition of condition-based maintenance according to the NF EN 13306 standard).
