Manufacturing & Maintenance 9 min read

Bearing fault detection: what the machine is telling you, months before it stops.

A bearing almost never fails without warning. It fails without anyone listening — and the difference between those two sentences is worth a great deal of production.

By Frank Guo · Technology & Product Leadership, addanode

TL;DR — Rolling-element bearings fail in roughly four stages, and each stage shows up differently: very high-frequency energy first, then bearing defect frequencies in the high-frequency band, then those frequencies with harmonics and sidebands in the normal vibration spectrum, and finally broadband noise with rising temperature. The diagnostic tell is that defect frequencies (BPFO, BPFI, BSF, FTF) are non-integer multiples of shaft speed — which is what separates a bearing fault from imbalance or misalignment. Envelope analysis exists because bearing impacts are low-energy and buried in noise. Practically: continuous trending tells you something changed; spectral analysis tells you what it is. Most plants need the first before they can use the second, and temperature alone is a late indicator — by the time a bearing is running hot you have lost most of your warning.

Why bearings are the asset worth watching first

Bearings sit under nearly everything that rotates: motors, pumps, fans, gearboxes, conveyor idlers and pulleys, mill and crusher assemblies. They are cheap relative to what they carry, and their failure is rarely contained — a seized bearing takes a shaft, a coupling, sometimes a gearbox with it, and always takes the production behind it.

They are also the asset that gives the most warning. A bearing does not decide to fail on Tuesday; it degrades through a sequence that is visible in vibration long before it is visible to a person walking past.

The four stages, and what each one looks like

Stage	What is physically happening	What you can detect
1 — Subsurface	Microscopic fatigue below the raceway surface. Nothing visible, nothing audible.	Only ultrasonic / very high-frequency techniques. Not visible in ordinary vibration readings, and not something most plants instrument for.
2 — Defect forms	A spall or pit breaks the surface. Each pass over it produces a small impact.	Bearing defect frequencies appear in the high-frequency band; envelope analysis shows them clearly. Overall vibration may still look normal — which is exactly why trending matters more than thresholds.
3 — Defect grows	The damaged area spreads; impacts get harder and more frequent.	Defect frequencies with harmonics and sidebands now appear in the normal velocity spectrum. Overall levels rise measurably. Often audible to an experienced ear. Temperature may begin to lift.
4 — Terminal	Clearance opens up, elements and cage degrade, lubrication breaks down.	Discrete frequencies get swamped by broadband noise as the noise floor rises; temperature climbs clearly. This stage is short — plan for hours to days, not weeks.

The practical consequence of that table: a plant that only watches temperature is watching stage 4. Temperature is a real signal and worth having — it is cheap, robust and unambiguous — but it arrives late. If your maintenance conversations regularly start with "the bearing was running hot", you are being told about failures you could have scheduled.

The four stages in the table above, read as time rather than as symptoms. What monitoring buys is the window, and the window is shortest exactly where the symptoms are most obvious — which is why the useful alarm is a change, not a level.

Why defect frequencies are the giveaway

Every rotating fault has a signature frequency, and the useful thing about bearings is that theirs are not whole multiples of shaft speed. Imbalance shows at 1× shaft speed. Misalignment typically at 1× and 2×. Looseness produces a run of harmonics. Bearings, by contrast, produce four characteristic frequencies determined by the geometry — number of elements, element diameter, pitch diameter, contact angle:

BPFO — ball pass frequency, outer race: an element passing a defect on the stationary outer race.
BPFI — ball pass frequency, inner race: a defect on the rotating inner race, usually with sidebands at shaft speed because the fault moves in and out of the load zone.
BSF — ball spin frequency: a defect on a rolling element itself, often appearing at twice BSF because the flaw strikes both races.
FTF — fundamental train frequency: the cage, which rotates at well below shaft speed. A rare but serious finding.

Because these fall at non-integer multiples of shaft speed, a spectrum that shows raised energy at, say, 4.37× shaft speed is saying something quite specific — and it is saying it in a language that imbalance and misalignment cannot speak. Bearing manufacturers publish these multipliers per part number, which is why knowing which bearing is actually installed is not administrative detail; it is the difference between a diagnosis and a guess.

What envelope analysis is for

Bearing impacts are sharp, small and fast — high frequency, low energy. In an ordinary vibration spectrum they sit underneath the far larger low-frequency energy of rotation, mesh and process noise. Envelope analysis (also called demodulation) filters to the high-frequency band where those impacts excite structural resonances, then extracts the envelope of that signal — the repetition rate of the knocking rather than the ringing itself. The result moves a fault that was invisible at stage 2 into plain view.

It is worth understanding the limitation as well: envelope analysis is sensitive to the band you choose and to how the sensor is mounted. A magnetically mounted sensor on a painted surface will not deliver the high-frequency fidelity that a properly stud-mounted one will — which is one of the reasons apparently identical monitoring programmes produce very different results, a theme we cover in why most vibration monitoring projects fail.

Where we draw our own line, plainly: addanode's condition monitoring provides continuous vibration, temperature and motor-current trending — it is built to tell you, reliably and without a site visit, that an asset's behaviour has changed and when it started. Naming a specific defect from a spectrum is specialist analysis work, and on a critical asset it is worth commissioning. The two are complements, not competitors: continuous trending decides which machine deserves an analyst's afternoon, and an analyst turns "this changed three weeks ago" into "outer race, plan the swap". A plant that buys the second without the first pays a specialist to walk past healthy machines.

How much warning do you actually get?

Honestly: it varies more than vendors like to admit. Detection at stage 2 typically buys weeks to months on a moderately loaded asset at steady speed — enough to order the part, plan the outage and change it on your terms. But that window compresses under load, contamination, poor lubrication or high speed, and it can collapse quickly once stage 3 sets in.

Which argues for the discipline rather than the promise: trend continuously, act on change rather than on absolute thresholds, and treat the alarm as the start of a planning conversation rather than an emergency. The value is not that the machine gets fixed — it would have been fixed either way. The value is that it gets fixed during a planned window, with the right part on site, instead of at 02:00 on a Sunday with a production line behind it. We put numbers to that in the cost of unplanned downtime.

Where to start on a real plant

Rank by consequence, not by count. Instrument the bearings whose failure stops production, floods a level or parks a fleet — usually a much shorter list than the asset register suggests.
Record what is installed. Bearing part numbers on the monitored assets, so defect frequencies can be calculated when it matters.
Get the mounting right. Sensor location and mounting method determine whether stage 2 is visible at all. This is where most of the technical quality of a programme is decided.
Establish a baseline before drawing alarm lines. Machines differ; a level that is alarming on one pump is normal on its twin. Change from that machine's own baseline is the reliable signal.
Write down what happens on an alarm. Who looks, within how long, and what authority they have to schedule work. A programme with no agreed response is a data collection exercise, not maintenance.

FAQ

Bearing fault detection — common questions

How does vibration analysis detect bearing faults?

A surface defect produces a small impact each time a rolling element passes over it. That repetition happens at a frequency set by the bearing's geometry — BPFO, BPFI, BSF or FTF — which falls at non-integer multiples of shaft speed. Detecting raised energy at those specific frequencies, usually via envelope analysis in the early stages, identifies both that there is a fault and which part of the bearing it is on.

What are the stages of bearing failure?

Four. Subsurface fatigue (detectable only ultrasonically); defect formation (bearing frequencies in the high-frequency band, visible with envelope analysis while overall levels still look normal); defect growth (harmonics and sidebands in the velocity spectrum, rising overall levels, often audible); and terminal degradation (broadband noise, clearly rising temperature, hours-to-days remaining).

Can temperature alone detect a failing bearing?

Only late. Temperature rises meaningfully at stage 3 to 4, when most of your warning has already elapsed. It remains worth monitoring — it is cheap, robust and unambiguous, and it catches lubrication problems well — but a programme built on temperature alone will convert catastrophic failures into slightly-less-catastrophic ones rather than into planned work.

What is envelope analysis and when do I need it?

It filters the signal to the high-frequency band where bearing impacts excite resonance, then extracts the repetition rate of those impacts rather than the ringing. It is what makes a stage-2 fault visible while the overall vibration level still looks healthy. You need it when you want the earliest possible warning on critical assets; for general trending, continuous overall vibration, temperature and current will tell you that something has changed.

How much warning does bearing monitoring give?

Detected at stage 2 on a moderately loaded, steady-speed asset, typically weeks to months — enough to order parts and plan an outage. That window shortens under heavy load, contamination, poor lubrication or high speed, and compresses fast once stage 3 begins. Treat any specific number quoted without knowing your duty and speed as marketing.

Do I need machine learning for bearing fault detection?

No — and it is worth saying clearly, because searches on this topic are full of research datasets and model repositories. The physics is well understood and deterministic: defect frequencies are calculable from bearing geometry, and the diagnostic method has been stable for decades. Machine learning has real uses in classifying large fleets and cutting false alarms, but a plant that has no baseline, no consistent sensor mounting and no agreed alarm response will not be rescued by a model.

Primary sources

US DOE / PNNL — Operations & Maintenance Best Practices Guide, Release 3.0 (2010), PNNL-19634

Find out which of your bearings changed last month.

Continuous vibration, temperature and current trending on the assets whose failure actually stops you — installed on what you already own.

Start a Conversation Condition monitoring →