Are you doing too much preventive maintenance? What your data answers

Open the preventive maintenance plan of any industrial site. You will find dozens of tasks carried over from one year to the next, set to a calendar that no one questions any more. A grease point redone every month, a seal replaced every quarter, a valve stripped down twice a year.

Then ask a simple question: this frequency, where does it come from? Often the answer fits in one phrase, "we have always done it this way", or "it is the manufacturer's recommendation", frozen on the day of commissioning.

The problem is not that preventive work is useless. It is that a plan never reviewed drifts in two directions at once. Some tasks come round too often, they cost money and weaken the equipment. Others, too rare, let failures slip through. Your history already knows which is which.

Are you doing too much preventive work?

Reducing breakdowns comes to mind straight away. Reducing preventive work, far less. Yet a preventive intervention is never neutral. It takes the equipment out of service, ties up a technician, consumes parts. And above all, it opens a window for error.

Every strip-down is an opportunity for an imperfect reassembly: a seal put back slightly wrong, an approximate tightening torque, an abrupt restart on start-up. A share of early failures does not come from wear, but from the last human intervention on the machine.

Over-maintenance therefore carries a double cost:

• A direct cost, visible: labour, parts changed before the end of their life, production stoppage.
• An induced cost, invisible: the risk of a fault introduced by the intervention itself, on equipment that was working.

On stable, low-criticality equipment, multiplying interventions sometimes amounts to buying risk at the price of peace of mind. The right level of preventive work is not the maximum. It is the point where you prevent more failures than you cause.

How do you spot useless or mistimed tasks?

The history answers a precise question: for each task, does the maintenance frequency match the real failure frequency? The misalignment shows up in a few concrete signals.

• The task never finds anything. Months of checks, no fault recorded, no part out of tolerance. The check may be too frequent, or poorly targeted.
• The replaced part is still good. A part that was not degraded is changed on the calendar. Systematic replacement runs well ahead of wear.
• The breakdown happens despite the preventive work. The equipment fails between two interventions set too far apart. Here, you need to tighten up, not space out.
• The task is never recorded. A procedure exists on paper, but no report documents it. You do not know whether it is done, or what it turns up.
• Two tasks overlap. The same point is checked by two different procedures, inherited from two eras.

Cross-referencing these signals assumes a usable history. That is the whole point of making use of CMMS data before adding anything: without reliable reports, you recalibrate blind.

Frequency is not reasoned the same way either, depending on the criticality of the equipment. On a critical asset, the analysis starts from risk, a logic developed in risk-based maintenance.

From calendar to condition-based: which policy for which equipment?

Recalibrating is not only changing a frequency. Sometimes it means changing the logic. Systematic preventive work is only one rung of a ladder. The detailed definitions of each policy are laid out in preventive, predictive, condition-based; the point here is knowing which one to choose.

Table scrolls horizontally on small screens.

Predictive is not the summit to reach everywhere. It is the most demanding rung, reserved for equipment where equipment life prediction pays for its cost. For the majority of a fleet, optimised preventive work and a few condition-based points are enough.

Do you need sensors, or is your data enough?

A common reflex: to maintain better, instrument. Fit sensors, feed back measurements, monitor continuously. It is sometimes justified, often premature.

Before adding a sensor, a question of sequence: has the existing history already been used? In many cases, the CMMS reports are enough to spot mistimed tasks and recalibrate a procedure. The sensor does not create this analysis, it assumes it.

Instrumentation is justified when three conditions stack up:

• Degradation is fast, too fast for a periodic check to see it coming.
• It is invisible to standard inspections, it cannot be read by eye or by touch.
• Its consequence is costly, an unplanned stoppage, a safety incident, a chain failure.

When these conditions are not met, a sensor adds data without adding a decision. Better to first draw out everything the history contains, then instrument the few points where the human eye and the calendar are no longer enough. This ties in with the logic of reducing unplanned downtime: you target the effort where the breakdown really hurts.

• Spacing out a task on the cost argument alone.An expensive task that finds nothing can still protect against a serious, rare risk. Cost does not decide on its own, criticality does too.
• Adding sensors before making use of the history.You pay for instrumentation to rediscover what the reports already said.
• Confusing the absence of breakdowns with proof of usefulness.A machine that does not fail does not prove that the monthly task protects it. It may simply be robust, and telling the two apart calls for comparing populations, not observing a single case.

What does AI bring, and what does it not do?

Across a whole fleet, comparing each procedure to its history by hand is disheartening. This is where automated analysis really helps.

The tool cross-references the intervention history and the breakdown history, task by task. It spots the procedures that never precede a fault, those that arrive after the breakdown, the slow drifts in failure intervals. It can propose, for each task, a frequency recalculated from the facts.

Let us stay honest about the limit. AI creates no new truth: it re-reads, connects and highlights what your data already contains. If your reports are poor or false, its proposals will be too. And a proposed frequency remains a hypothesis: safety, criticality and knowledge of the process belong to the engineer, who validates or refuses.

It is in this division that the value lies. Building this reading on top of your data, without replacing your judgement, is the object of AI systems built around your data. The tool sheds light on the calendar, it does not sign off the procedure.

What should you take away?

A preventive plan is never fixed, it ages. Too much maintenance costs money and weakens equipment, too little lets breakdowns slip through, and your history already shows which way each task leans. You compare the real frequency of failures against the calendar, you space out or tighten up, you move to condition-based work where degradation is observable. Sensors come after the history, not before. And the last decision stays human.