Skip to Content
DocsUse CasesDetecting Bearing Failure

Detecting Bearing Failure

Every rotating machine depends on its bearings. When one starts to fail, vibration patterns shift — subtle at first, unmistakable by the end. Catching that shift early is the difference between a planned replacement and a catastrophic breakdown.

This is a real bearing failure from the NASA IMS Bearing Dataset  — one of the most cited run-to-failure experiments in predictive maintenance. A bearing degrades from healthy to catastrophic over 7 days while accelerometers sample every vibration at 20 kHz.

Each 10-minute snapshot is condensed to window statistics — RMS, kurtosis, and more — using the same winkComposer building blocks that would run at the edge in production. A winkComposer flow then scores the bearing’s health from these statistics in real time.

Drag the slider and watch it detect the failure before it happens.

Loading bearing health demo...

What You’re Seeing

The chart shows vibration RMS (cyan) — the overall energy in the signal — tracked by an exponentially smoothed mean (lavender dashed). The filled envelope around the mean is the peak-tracking vibration envelope — tight when the bearing is healthy, widening as degradation introduces new vibration modes.

For about the first day the system shows Learning — it is observing the equipment’s baseline vibration automatically, with no manual threshold tuning required.

The bearing holds steady for about 3.75 days. RMS and kurtosis stay normal, and the assessment reads Healthy. Around day 4, multiple evidence sources fire at once — energy spikes, a rising trend, and falling signal clarity — and the assessment jumps straight to Critical. The change-point markers (vertical dashed lines) visually confirm the shift.

The key drivers table shows each evidence source’s intensity and persistence — distinguishing sustained degradation (Continuous) from transient spikes (Intermittent). This is the first question maintenance engineers ask: is this real or a false alarm?

What you’re watching is a P-F curve — the trajectory from the point where a fault becomes detectable (P) to functional failure (F). The entire discipline of predictive maintenance exists to maximise the interval between P and F.

How It Works

One flow, six building blocks — each contributes a different perspective on bearing health:

twStatsextracts features from raw 20 kHz vibration(offline here; live in production)esStatssmooths RMS and tracks the ±σ envelope→ the bearing’s volatility signaturermsTrenddetects rising vibration energy→ the degradation ratekurtTrenddetects rising impulsiveness→ often the first sign of a defectrmsCPDfires when RMS shifts to a new level→ visual marker on the chartkurtCPDfires when kurtosis shifts to a new level→ corroborates the RMS change-pointbearingHealthself-calibrates, then fuses 4 evidence sources→ health state + intensity × persistence

ES Stats smooths the raw RMS signal and tracks the vibration envelope — the lavender band on the chart. It also computes the z-score (how far the current reading deviates from normal) and signal-to-noise ratio (SNR), which drops as the bearing’s behaviour becomes erratic.

RMS Trend watches whether the smoothed RMS is climbing consistently — the degradation rate. Only a rising trend with sufficient confidence passes through to the assessment.

Kurtosis Trend does the same for impulsiveness — often the first sign of a defect.

RMS Change-Point Detection and Kurtosis Change-Point Detection run Page-Hinkley  tests that fire when the signal shifts to a new level — the vertical markers on the chart. When both fire at roughly the same time, two independent physical signals confirm the same shift.

The health assessment at the end of the pipeline fuses four evidence sources into a single confidence score:

  • Energy spikes — z-score from ES Stats (current deviation from the running average)
  • Impact events — raw kurtosis (spiky vibration patterns)
  • Energy trend — trend confidence, gated to the rising direction only (zero when stable, regardless of confidence)
  • Signal clarity — SNR from ES Stats (drops as behaviour becomes erratic)

The system learns the equipment’s baseline automatically during the first ~1.2 days — no manual threshold tuning per asset. After that, each source builds evidence strength independently. The combined confidence determines the health state (Healthy, Monitor, Degraded, Critical), the recommended action, and shows both the intensity and persistence of each driver.

References

  • Qiu, H., Lee, J., Lin, J. & Yu, G. (2006). Wavelet filter-based weak signature detection method and its application on rolling element bearing prognostics. Journal of Sound and Vibration, 289, 1066–1090. doi:10.1016/j.jsv.2005.03.007 
  • Page, E.S. (1954). Continuous inspection schemes. Biometrika, 41(1/2), 100–115. doi:10.2307/2333009 
  • Dataset: NASA IMS Bearing Data  — University of Cincinnati, 2004. Test 2, Bearing 1: inner race defect, ~164 hours run-to-failure. Four bearings under 6,000 lb radial load at 2,000 RPM. Accelerometers captured 20,480 samples at 20 kHz every 10 minutes (982 snapshots).

Next Steps

Last updated on
OverviewCatching Process Drift