Preventive, predictive and condition-based maintenance: differences and examples

The three terms get used as if they were interchangeable in sales brochures, when in fact they describe different policies that do not call for the same data and do not cost the same. Confusing the three leads you to buy sensors for a problem a calendar would have solved, or the other way round.

The three policies, stripped of the sales vocabulary

Time-based preventive maintenance

You intervene at a fixed interval, without looking at the actual condition. An oil change every six months, a filter replaced every 2,000 hours, an annual visit.

This is the simplest policy to organise and the easiest to schedule. Its weakness is well known: it swaps out parts that were still perfectly good and misses failures that occur between two due dates. Its strength is less often acknowledged: it needs no data, no sensors, no history. On a low-criticality asset whose failure is repaired in an hour, it is often still the most economical answer. Replacing it with instrumented monitoring would cost more than the failures you would avoid.

Condition-based maintenance

You measure a quantity (wall thickness, vibration level, temperature, the particle content of an oil) and you intervene when it crosses a threshold.

This policy assumes three things: a quantity that genuinely reflects the state of degradation, a means of measuring it, and a threshold defined by someone who is putting their technical judgement on the line. The third point is the one that gets forgotten. A sensor that reports a value with no associated threshold does not produce condition-based maintenance: it produces a chart that nobody knows how to read.

Predictive maintenance

You do not simply note that a threshold has been crossed: you follow how the quantity is evolving in order to estimate when the threshold will be reached, and you plan the intervention beforehand, in a slot of your choosing.

This is an extension of condition-based maintenance, not a different technology. What it adds is the use of history, and therefore the need to have several comparable readings over time, taken at the same location. That is precisely where most installations come unstuck, not for lack of sensors, but because the readings that already exist have never been brought alongside one another. See tracking equipment degradation. On how AI makes this approach workable across an entire asset base, see predictive AI to extend equipment life.

The table that lets you decide

Table scrolls horizontally on small screens.

The last row is worth pausing on. Predictive maintenance applied to a history of two campaigns produces a date, and that date is reassuring. It has no value all the same: two points always define a straight line. Displaying a due date under those conditions is more dangerous than displaying nothing, because the decision is then taken on the strength of an illusion of knowledge.

What really determines the choice

Three questions are enough, in this order.

Is the failure progressive? If the asset degrades slowly and in a measurable way, condition-based and predictive maintenance make sense. If the failure is sudden and random, no measurement will announce it: the right answer is redundancy or repairability, not monitoring.

Is the intervention easy to plan? If it is, knowing the date in advance adds little. If the intervention calls for a shutdown, the booking of a contractor or a long-lead spare part, then anticipating it is worth a great deal, and predictive maintenance earns its place.

Do you have the history, or can you obtain it? This is the question that settles things in practice. Many sites already hold several measurement campaigns, locked inside PDF reports, never brought together. Before investing in new sensors, it is often more profitable to make use of what already exists. See from PDF reports to operating data.

The mix-ups that cost dearly

• Calling a plain threshold "predictive".A sensor with an alarm does condition-based maintenance: that is already very good, but it is not the same promise, and billing it as predictive sets up a disappointment.
• Instrumenting before you use what you already have.Adding sensors to an installation whose current readings go unused amounts to stacking one unread data source on top of another.
• Believing predictive maintenance replaces preventive maintenance.Regulatory obligations and manufacturer recommendations remain a floor; predictive maintenance organises what falls to your own decision, not what is imposed on you.
• Choosing a policy asset by asset without looking at criticality.The question is not "which is the best policy" but "what level of effort does this asset deserve". That is the subject of risk-based maintenance.
• Ignoring who sets the thresholds.A threshold with no identified responsible engineer will never be applied, because nobody will want to stop the line on its say-so alone.

Where to start if everything is time-based preventive

The move is not made in one go. The sequence that works is the following.

Begin by identifying the assets for which an anticipated due date would carry real value: those whose intervention requires a planned shutdown or a long-lead spare part. It is these, and not the most conspicuous ones, that justify the effort.

Then look at what readings already exist on those assets. Inspection reports, thickness surveys, contractor write-ups: in most cases the raw material is there. The point is not to measure more, but to bring together what has already been measured.

Finally, have the thresholds set by the person who will stand behind them. Without that step, you will end up with fine curves and no decision.

To place these policies within an overall strategy, see our Maintenance Intelligence page and the article prioritising maintenance by real risk. If your teams need to build up their judgement on these trade-offs, the AI training for maintenance managers works through them on your own cases.