Bearing Failure Prediction: Weeks of Warning From Vibration Data

In one line: A rolling-element bearing doesn't fail suddenly: it degrades through four predictable stages, each with a distinct vibration signature. The earliest stages are invisible to fixed thresholds but clearly detectable to models that learn a bearing's normal behavior and normalize for speed and load. Catch the degradation in stage two or three and you get two to four weeks of warning, enough to plan the swap instead of surviving the failure.

Why bearings are where you should look first

If you monitor one failure mode across your rotating equipment, make it bearings. Bearing degradation is behind a large share of motor and pump failures: the landmark IEEE and EPRI motor-reliability studies both rank bearings as the leading cause of electric-motor failure (the exact percentage gets misquoted, but the ranking is consistent across both surveys). They're also the failure mode that responds best to condition monitoring, because a degrading bearing broadcasts its condition through vibration long before it seizes.

The catch is that "long before" only helps if you're listening in the right way. Most plants aren't: they set a single vibration alarm and wait for it to trip. By the time a fixed threshold fires, the bearing is usually deep into its final stage. The warning was there for weeks. The monitoring just couldn't read it.

The four stages of bearing failure

A rolling-element bearing degrades along a well-understood path, often described using the P-F curve (from Potential failure to Functional failure). Four stages, each with a characteristic signature:

Stage 1: Subsurface (the earliest warning). Microscopic fatigue begins below the raceway surface. There's no visible defect yet, but ultrasonic and very-high-frequency vibration energy starts to rise. This stage can appear months before failure. It's real, but it's subtle and easy to dismiss as noise.

Stage 2: Defect frequencies emerge. A tiny spall or crack forms on a raceway or rolling element. Now every time a ball rolls over that defect, it produces a small, repeating impact, at a frequency set by the bearing's geometry and shaft speed. These bearing defect frequencies show up in the vibration spectrum (more on them below). This stage typically gives weeks to a couple of months of lead time and is the sweet spot for planned intervention.

Stage 3: Harmonics and sidebands. The defect widens. Its defect frequency grows harmonics and sidebands, broadband energy rises, and the bearing may become audibly rough and start running hot. Lead time here is days to weeks. Still actionable, but the window is closing.

Stage 4: Imminent failure. The vibration signature actually gets noisy and erratic as the bearing tears itself apart, and overall levels can even drop briefly before catastrophic seizure. Lead time: hours to days. This is the stage where a fixed threshold finally trips, far too late to plan anything.

The lesson is uncomfortable for threshold-based monitoring: the alarm you set almost always fires in stage 4. The value is all in stages 2 and 3, and reading those requires more than a single number.

The signatures: bearing defect frequencies

Here's what makes bearings so predictable. When a defect exists on a specific part of the bearing, it generates impacts at a specific, calculable frequency derived from the bearing's geometry (number of rolling elements, ball and pitch diameters, contact angle) and the shaft speed. The four canonical ones:

BPFO (Ball Pass Frequency, Outer race): a defect on the outer raceway. The most common bearing fault.
BPFI (Ball Pass Frequency, Inner race): a defect on the inner raceway.
BSF (Ball Spin Frequency): a defect on a rolling element itself.
FTF (Fundamental Train Frequency): a cage or lubrication problem.

Identify a rising peak at one of these frequencies and you don't just know the bearing is degrading: you know where the defect is. That's diagnostic gold: an outer-race defect and a cage problem call for different responses and different urgency.

The problem is that these signatures are small and buried. In a raw vibration spectrum, an early-stage defect frequency is often drowned out by the much larger peaks from shaft rotation, imbalance, and misalignment. This is why envelope analysis (also called demodulation) exists: it isolates the high-frequency impact energy and demodulates it to make the defect frequencies visible. It's a genuinely useful technique (our guide to getting started with vibration analysis walks through the defect-frequency math), and it's also why bearing analysis has historically needed a trained vibration analyst: someone has to run the demodulation, know the bearing's defect frequencies, and interpret the result.

That expertise doesn't scale. You can't put a certified analyst on every one of 200 assets, checking every week. Which is where machine learning earns its place.

Why fixed thresholds miss the early stages

A single overall-vibration threshold (say 4.5 mm/s RMS from ISO 10816 / ISO 20816) has three blind spots that map directly onto early bearing failure:

It ignores frequency. Overall RMS is one number. A stage-2 defect adds a small amount of energy at a specific frequency while the overall level barely moves. The threshold sees nothing.
It ignores speed and load. Vibration amplitude scales roughly with speed squared. On any VFD-driven asset, the "normal" level moves every few minutes as the drive changes speed, so a fixed limit is either constantly false-alarming at high speed or blind at low speed.
It ignores the trend. A bearing creeping from 1.8 to 2.4 mm/s over six weeks is screaming "I'm degrading." A fixed limit at 4.5 says "still fine" the entire time.

Envelope analysis fixes the frequency blind spot but still needs a human to run and read it. Fixed thresholds on envelope metrics reintroduce the speed and trend problems. The gap is a system that watches the full picture continuously and knows what normal looks like for this bearing under these conditions.

What ML-based detection adds

Machine-learning condition monitoring doesn't discard vibration analysis: it automates and scales the reading of it, across every asset, continuously.

It learns a per-bearing baseline. An LSTM autoencoder is trained to reconstruct a specific bearing's normal vibration behavior. When the bearing is healthy, reconstruction is near-perfect. When a defect frequency starts rising (even a small one, even while overall RMS looks fine), reconstruction error climbs and an anomaly is flagged. The model effectively learns "normal for Pump 7A at this speed and load," not a fixed number.

It normalizes for operating conditions. Because the model learns behavior as a function of speed and load, it can flag "2.1 mm/s at 900 RPM is abnormal for this bearing" while correctly treating "3.0 mm/s at 1,800 RPM" as fine. That's the speed-dependent problem fixed thresholds can't touch.

It combines sensors. An outer-race defect rarely shows up in vibration alone: bearing temperature drifts up, motor current gets subtly erratic. A model watching all of them together catches the pattern earlier and with fewer false alarms than any single channel.

It estimates remaining useful life, with honest uncertainty. Detection tells you the bearing is degrading. The next question is always "how long do I have?" A remaining-useful-life (RUL) model estimates that, and a well-built one reports it as a range, not a false-precision single number. Prevly's RUL models report conformal prediction intervals, validated on real NASA C-MAPSS data, so an estimate comes with a defensible confidence band an engineer can actually plan around, not a black-box "14 days" with no error bar.

Fourteen days of warning, with the evidence

Here's a representative example of what actionable bearing prediction looks like (illustrative of the pattern, not a specific customer result). On a centrifugal pump, an ML-based system flags an anomaly 14 days before the bearing would have seized. The alert doesn't just say "anomaly." It carries feature attribution, the specific signals that drove the decision:

vibration_x_rms: +0.34 (elevated radial vibration, the dominant contributor)
temperature_delta: +0.21 (bearing temperature rising faster than housing temperature)
current_kurtosis: +0.12 (subtle spikes in motor current, indicating intermittent mechanical resistance)

A reliability engineer reads that and immediately forms a hypothesis: elevated radial vibration + thermal rise + current spikes is a classic outer-race defect pattern. They confirm with a targeted envelope measurement, schedule the swap during the next planned downtime, and replace the bearing in a two-hour window, instead of a catastrophic seizure and eighteen hours of unplanned downtime.

The fixed threshold, for the record, was still green. It would have stayed green for another twelve days.

What this looks like with Prevly

Prevly is built to make bearing prediction routine across a whole plant, not a specialized project on a handful of assets:

Read-only ingestion. Prevly Edge subscribes to your existing vibration, temperature, and current sensors over read-only OPC-UA. No new hardware, no writes to your PLCs.
Detection that scales the analyst's judgment. LSTM-autoencoder anomaly detection learns each asset's normal baseline and flags the early-stage degradation that overall-vibration thresholds miss, on every monitored asset, continuously.
Explainable by design. Every alert reports which sensors drove it (via Integrated Gradients for the deep models, SHAP for the gradient-boosted RUL model), so an engineer verifies the finding instead of trusting a score.
A prediction becomes a plan. Detection plus RUL plus attribution becomes a drafted work order (asset, likely fault, recommended action, and the sensor evidence attached), so you order the bearing before you need it and the technician arrives knowing what to open.

Bearings tell you they're failing weeks in advance. The only question is whether your monitoring is built to hear it.

Frequently asked questions

How early can you detect a bearing failure? It depends on the stage and the method. Subsurface fatigue (stage 1) can appear months ahead in ultrasonic and high-frequency energy but is subtle. The reliable, actionable window is stage 2, when defect frequencies emerge: typically weeks to a couple of months of lead time. Fixed thresholds usually only catch stage 4, hours to days before seizure.

What causes rolling-element bearing failure? The most common causes are inadequate or contaminated lubrication, misalignment and excessive load, contamination and moisture ingress, electrical fluting (especially on VFD-driven motors), and normal fatigue at end of life. The failure mode that results (outer race, inner race, rolling element, or cage) shows up at a distinct defect frequency.

What are BPFO, BPFI, BSF, and FTF? They're the four bearing defect frequencies: Ball Pass Frequency Outer race, Ball Pass Frequency Inner race, Ball Spin Frequency, and Fundamental Train Frequency. Each corresponds to a defect on a specific part of the bearing and is calculated from the bearing's geometry and shaft speed. A rising peak at one of them tells you where the defect is.

Can you predict how long a bearing will last? You can estimate remaining useful life once degradation is detected, but honestly it's a range, not a fixed date. Good RUL models report a confidence interval (Prevly's RUL models use conformal prediction intervals, validated on real NASA C-MAPSS data) so you can plan around a defensible window rather than a false-precision number.

Do I need a vibration analyst to do this? Traditionally yes: envelope analysis and defect-frequency interpretation are specialist skills. ML-based condition monitoring automates the continuous reading across every asset, so a certified analyst's time is spent confirming and acting on flagged findings rather than manually screening hundreds of machines.

See it on your own bearings

Prevly brings AI-based bearing-failure detection, RUL estimation, and explainable attribution to your existing sensors: read-only, on-premise, no analyst required to screen every asset. Within days of connecting your data, you'll see what your bearings have been signaling.

Request a Prevly demo and find the degradation your thresholds are still calling green.