Predictive maintenance: how to get started

Predictive maintenance means intervening when a measurement indicates the equipment is degrading: neither on a calendar schedule nor after it has already failed. It’s the strategy that wastes the fewest hours of the three, and also the one with the most demanding requirements.

This article is about those requirements, because most predictive projects don’t fail because of the technology.

What gets measured

The usual variables are few and have been working for decades:

• Vibration, for bearings, imbalance and misalignment in rotating machinery.
• Temperature, via thermography or a probe, for electrical connections, bearings and insulation.
• Power draw, which reveals abnormal strain before noise does.
• Oil analysis, which detects internal wear without opening anything up.
• Differential pressure, which tells you when a filter is genuinely clogged rather than just due for a calendar-based change.
• Noise, the most subjective one, and the first thing the operator notices.

The choice isn’t arbitrary: it depends on the failure mode you’re trying to anticipate. If the equipment fails from clogging, measuring vibration won’t tell you anything.

Where the thresholds come from

This is the question that decides whether the project actually works, and the answer is uncomfortable: from your own history, not from a catalog.

A generic reference value can point you in the right direction at first, but the point at which that specific pump, in that specific installation, starts to degrade is something you only learn by measuring it for a while while it’s healthy. Without a baseline, any reading is just a number with no context.

Something counterintuitive follows from that: predictive maintenance starts by measuring equipment that’s working fine. If you only measure when you’re suspicious, you’ll never have anything to compare against.

You don’t need sensors to get started

This is the most useful idea in the whole article. An inspection round with logged values is already a time series, and a time series lets you see a trend.

In GMAO Cloud, checklists support fields with their type, their label and a minimum and maximum value. When a reading falls outside that range, it gets recorded as an anomaly right then, not as a note on the side that nobody’s going to reread. Checklist templates resolve in cascade — asset, model, subfamily, family — so a round covering twenty identical pumps gets defined once.

With that, one person with a thermometer and a portable vibration meter can do real predictive maintenance on critical equipment, at almost no cost. Continuous sensing speeds that up and automates it; it doesn’t invent it.

Where the CMMS fits

Worth stating precisely, because some vendors sell a CMMS as if it were a signal analytics platform, and it isn’t.

What a CMMS brings to predictive maintenance is the place where the history and the decision live:

• The asset record, with its model, serial number, installation date and its own fields with their units, in asset management.
• The readings, captured within the preventive order, with who took them and when.
• The anomalies, which turn into issues with their priority and owner.
• The resulting intervention, with its time and material, and the effect it had.
• The metrics to tell whether all of this is actually working.

Orders that trigger themselves when a threshold is crossed

This is the piece that turns everything above into something automatic, and it’s worth explaining in detail because it’s what truly sets a predictive system apart from just an inspection round.

In GMAO Cloud an asset can carry an associated counter — kilometers, operating hours, cycles, units produced — with a configured limit: how many units that component is expected to withstand. A warning percentage is also defined.

When a technician logs a reading, the system totals up the accumulated figure, calculates what percentage of the limit has been used, and if it crosses the configured percentage, it automatically generates the preventive work order. The order is created with its asset, its address, the checklist template that matches that equipment, and a description explaining why it was generated.

In other words: preventive maintenance doesn’t depend only on the calendar. It can depend on actual use, which is exactly what’s needed for a fleet, for a machine running irregular shifts, or for a component that wears out by cycles rather than by months.

It’s worth combining both. Time-based periodicity covers what degrades whether it’s in use or not — gaskets, oils, corrosion — while a counter-based threshold covers what degrades with use. A single piece of equipment can have both at once.

The accumulated-cost alert

There’s a second automation that often goes unnoticed and that answers the most expensive question in maintenance: when to stop repairing.

An asset can have its replacement cost recorded along with a warning percentage. The system periodically compares that machine’s accumulated repair cost against that figure and, once it crosses it, sends an alert.

It doesn’t generate an order: it notifies, because replacing equipment is a business decision, not a task. But it puts the figure in front of you right when it’s time to look at it, instead of buried in a report nobody opened.

The step that kills projects

Letting an alert lead to nothing.

A technician logs a vibration reading that’s gone up, and nothing happens. By the third time, they stop logging it, and rightly so. From there on you have sensors, charts, and zero predictive maintenance.

An anomaly has to end in one of two things: an issue with an owner and a date, or an explicit decision to do nothing. Both are valid; silence isn’t. And it’s worth making sure the notification system only alerts about things that matter: if it alerts about everything, people stop reading alerts, which is worse than not having them.

When predictive maintenance does NOT pay off

For balance, and because it saves money: there’s equipment where setting all of this up is just wasted effort.

When the failure has no consequence. If the equipment is redundant, cheap, and its failure doesn’t stop anything, corrective is the right strategy. It’s not neglect: it’s putting effort where it pays off.

When the failure gives no warning. There are sudden failure modes — a brittle fracture, a blown fuse — with no prior degradation to measure. There, predictive maintenance simply can’t work by definition, and what’s called for is preventive maintenance or redundancy.

When measuring costs more than the failure. If anticipating the failure means stopping the machine to measure, or an hour of technician time every week on equipment that gets replaced for two hundred euros, the math doesn’t work.

When there’s nobody to interpret the data. This is the most common case and the least recognized. A vibration chart with nobody who knows how to read it ends up being a file nobody opens.

Each asset’s criticality — a dedicated field in asset management — is exactly what lets you decide this without arguing it case by case.

How to know if it’s working

With reports, on the same equipment over time:

• MTBF of the monitored equipment: if it goes up, it’s failing less.
• Accumulated downtime: translates the above into lost production.
• Corrective ratio: should drop for equipment under a monitoring round and not for the rest. If it drops for everything, the change isn’t due to predictive maintenance.
• Accumulated cost per asset: the figure that decides whether it’s still worth maintaining.

If after a few months nothing has moved on the monitored equipment, either the wrong variable is being measured or the threshold is set wrong. Both are fixable; what isn’t fixable is not having measured at all.

Where to start

With three or four critical pieces of equipment, not the whole plant. For each one: which failure mode you want to anticipate, which variable reveals it, how often it’s measured, and who acts when the threshold is crossed. With the round set up as preventive maintenance and its checklist with values, you’ll have a baseline within a few months.

And before anything else, the boring part: that orders get closed with real data. Predictive maintenance on top of an empty history doesn’t exist.

If you’d like to set it up on your equipment, you can request a demo.