Predictive maintenance: what it is, how it works and real examples

Predictive maintenance means intervening when a measurement shows that a failure has begun, not when the calendar says so. Instead of replacing a bearing every 8,000 hours, you measure something that changes before it seizes, track that measurement, and replace the bearing when it starts to rise.

The more precise technical name is condition-based maintenance. This article explains how the cycle works and gives concrete examples by equipment type, with the symptom that gives early warning of each failure. For the comparison with the other regimes, see predictive, preventive and corrective maintenance.

How it works, step by step

The procedure is standardized by ISO 17359, and it always has the same shape:

1. Choose the asset and the failure mode. You do not monitor a machine in general: you monitor the bearing, the alignment, the insulation. Start with what has failed before.
2. Choose the parameter. Vibration, temperature, current, particles in the oil, ultrasound, differential pressure. Each failure mode has a parameter that responds first.
3. Measure the baseline. With the machine in a known good condition, at its normal operating point. Without it, no later number means anything.
4. Set alert and alarm levels. From the applicable standard, or from the machine’s own baseline, which is better when you have it.
5. Follow the trend. This is where predictive maintenance beats one-off inspection. A single reading tells you where the machine is; a series tells you where it is heading.
6. Diagnose and act. The alert tells you something changed; the spectrum, the thermal image or the oil report tells you what.
7. Record the outcome and redo the baseline. Without this, the program does not learn.

Notice what predictive maintenance really delivers: lead time. The gain is not predicting the future; it is turning an emergency shutdown into a scheduled one, with the part already purchased.

Example 1: electric motor

What fails most in a motor is the bearing, and it warns in stages: first at high frequencies, then at the natural frequencies of the components, then at the fault frequencies visible in the velocity spectrum, and only at the end as noise and heat.

• Early symptom: a rise in the overall vibration value at the bearing, together with the housing temperature at the same point.
• What confirms it: the spectrum, with peaks at non-integer multiples of running speed. Full detail in the four stages of bearing failure.
• Other modes the same measurement catches: unbalance (a peak at 1x running speed), misalignment (2x running speed and high axial vibration), base looseness (a series of harmonics) and electrical causes (a component at twice line frequency, 120 Hz on a 60 Hz supply, with no exact relationship to running speed).

Example 2: centrifugal pump

A pump mixes mechanical problems and process problems in the same signal.

• Cavitation: a rise in broadband noise at high frequency, with no defined peak, and the pump sounds as if it were pumping gravel. The cause is on the suction side, not in the part.
• Fouled or eroded impeller: unbalance, with a dominant peak at 1x running speed in the radial direction.
• Misalignment with the motor: 2x running speed and high axial vibration on both sides of the coupling. It is the most frequent cause and the cheapest to fix.
• Bearing and mechanical seal: temperature rising before the leak.

The details are in cavitation and misalignment in the vibration signal.

Example 3: gearbox and speed reducer

Gears are the case where oil analysis usually warns before anything else.

• Early symptom: metal particles in the oil, by ferrography or a particle sensor, and changes in viscosity and water content.
• In vibration: a peak at the gear mesh frequency (number of teeth times shaft speed) with sidebands spaced at running speed. Growing sidebands mean a tooth problem; the gear mesh frequency on its own is normal.
• In temperature: a housing running hotter at the same load means friction or insufficient lubrication.

Example 4: air compressor

• Leaks in the distribution network: detectable by ultrasound, and where much of a plant’s wasted energy goes. A leak does not stop the machine; it makes the compressor run more hours a day.
• Clogging filter: rising differential pressure. It is the simplest predictive measurement there is, and it needs no sophisticated sensor.
• Operating hours: a compressor that now runs more hours to deliver the same output is warning of leaks or lost efficiency before any alarm goes off.

Example 5: electrical panel

Here the parameter is heat, and the technique is thermography.

• Loose or oxidized connection: a hot spot concentrated at the terminal, with the temperature falling along the conductor. It is the most common defect.
• Unbalanced phase: one of the three phases hotter than the others under the same load.
• Circuit breaker with worn contacts: heat at the contacts or terminals.

Two rules change the result: the criterion is the temperature difference between comparable points, not the absolute value, and the inspection is only valid with the panel under load. Details in thermography on electrical panels.

What predictive maintenance does not catch

Being honest about this is what makes the program credible:

• Sudden failure. A bolt that works loose, an electronic component that fails open, a foreign object in the pump. There is no trend to follow.
• Defects outside the sensor’s range. An early-stage bearing defect requires high-frequency response and rigid mounting. A poorly mounted sensor does not see it.
• Ignored alerts. The most common failure mode of a predictive maintenance program is the inbox. Alerts that arrive all the time stop being read; that is why a notification should go out only when the level changes, not at every reading.

What to do now

1. Pull the work-order history for the last two years and list the five failures that hurt the most.
2. For each one, write down which parameter would have changed beforehand. If there is none, that item is not a candidate for predictive maintenance.
3. Measure the baseline of those machines while they are in good condition.
4. Set alert and alarm levels, and agree on who reads them and what they do when one arrives.
5. At the end of the first year, compare planned shutdowns with emergency ones. That is the number that justifies the program.

If you want to know what your plant needs, talk to one of our engineers.

---

Sources: ISO 17359:2018, Condition monitoring and diagnostics of machines — General guidelines (procedure, baseline, alert and diagnosis); EN 13306, Maintenance — Maintenance terminology (condition-based maintenance); ISO 20816-3:2022 (evaluation of vibration in rotating machines); ANSI/NETA MTS, Table 100.18 (thermographic survey criteria); F. S. Nowlan and H. F. Heap, Reliability-Centered Maintenance, 1978.