Bearing cutaway
SKF 6205–2RS JEM Inner ring / 29.95 Hz Cage / 11.93 HzBall spin / 70.59 Hz
Explanatory visualization — motion visually amplified; not reconstructed from sensor data.
01 / THE MACHINE SPEAKS IN SAMPLES
Listen to a
bearing fail.
A knock excites the housing. Its rhythm reveals the damaged part.
Make the damage. Follow the evidence.
Start quietly. iPhone silent mode mutes web audio.
Raw scope / the witness
AWAITING AUDIO— Acceleration┆ Predicted impacts
Each gold mark comes from the integrated bearing kinematics. Each tick comes from those same samples.
Listening: selected bearing
Seeded slow load wander (±10% by default) and smaller speed wander (±1%) let the reference capture normal operating variation.
The reference is still learning: whatever it hears now becomes its normal.
The Rust engine learns its reference this way too.
Simulation: every sound and number on this page is computed in your browser from a physics model; it is not sensor data.
Demonstration and educational software; not certified for safety-critical industrial use.
Healthy: small shaft components over a Gaussian noise floor. A defect adds knocks that make the housing ring.
↓ Follow the samples through the instrument02 / ONE RULE, A FIXED REFERENCE
The monitor
Demo deviation score
Score unavailable — learning demo baseline
AWAITING SAMPLESThreshold event: unavailable
A heuristic indicator derived from statistical deviation; not physical health, fault probability, or remaining useful life.
Reference mean —Population σ —Upper threshold —Reference records 0 / 20
RMS = √(Σx²/n). One RMS record per 0.25 s window. Records 1–N are training records without a score. On record N+1, the first N records give the fixed mean and population standard deviation; N+1 is the first compared record. Rust defaults to 500 records. Changing k moves only the threshold.
Overall RMS Silent-Ear computes this—
Crest factor Peak / RMS; not computed by Silent-Ear today—
Kurtosis Gaussian noise ≈ 3; teaching limit 4.5. Not computed by Silent-Ear today—
ENVELOPE-SPECTRUM TEACHING VERDICT
Awaiting samplesFault naming belongs to this teaching layer; not computed by Silent-Ear today.
RMS measures total vibration energy. Sparse early knocks raise kurtosis while total energy can remain small. As impacts overlap and the noise floor rises, kurtosis falls while RMS keeps rising.
03 / ZOOM INTO THE EVIDENCE
From seconds to single samples
Acceleration / 200 ms
4 seconds1 millisecond
Wheel or pinch to zoom. Drag to pan. Keyboard: + / − to zoom; arrow keys to pan. The slider does the same.
Samples→3 kHz housing resonance→Band-pass→Envelope→Knock rhythm
The housing resonance is the ringing after one strike. Zoom far enough and every dot is one accelerometer sample.
04 / SEPARATE THE KNOCK FROM ITS RHYTHM
Find the fault lines
Acceleration spectrum
OWN FFT / HANN WINDOWAwaiting samples.
FFT (fast Fourier transform) separates the samples into frequencies. A Hann window softens the block edges. A spectrum shows energy at each frequency. The metallic ringing gathers around the housing resonance.
Envelope spectrum
KNOCK RHYTHMAwaiting samples.
Listen to the rhythm of the knocks, not the knock itself. Band-pass isolates the ringing; rectification and low-pass recover its changing amplitude.
Band-pass → envelope
SAME ACCELEROMETER BUFFERAwaiting samples.
Cyan: the isolated resonance. Gold: its amplitude envelope. Sidebands are extra lines around the fault line, spaced by shaft or cage rotation.
Scrolling spectrogram
SHORT-TIME FOURIER TRANSFORMAwaiting samples.
Time moves right; frequency rises upward. Brighter pixels mean more energy. A new stripe records each 0.25 s window.
Advanced / bearing geometry and resonance
SKF 6205-2RS JEM: the drive-end bearing in the Case Western Reserve University bearing data. Contact angle uses degrees.
fᵣ = rpm / 60. Impact spacing includes ±2% slip jitter. Inner race load varies once per shaft turn; ball load varies with the cage. Each impulse excites e−ζωₙt sin(ωdt). Model follows the impact and demodulation approach of McFadden and Smith, and Randall and Antoni.
05 / BREAK IT ON PURPOSE
Run to failure
Manual laboratory
HealthyEarly defectOverlapping impactsSeizure
Kurtosis > 4.5 not observedRMS rule fired not observedScore CRITICAL not observedSimulated seizure 240 min / fixed
Run the scenario to measure the warning intervals from the computed samples.
Minutes from this page's scripted degradation, not field data. Real lead time depends on the machine, the fault and the sensor; Silent-Ear makes no prediction.
The script holds a healthy baseline, grows an outer race defect, then raises the broadband floor. The seizure is a fixed scenario endpoint, not a prediction. Scrubbing recomputes the samples and reference from the seed.
06 / WHERE THE DEMONSTRATION ENDS
From files to the factory
The actual Rust engine
Silent-Ear reads accelerometer files, computes one RMS value per channel per file, learns a fixed reference, and flags values strictly above mean + k·σ.
It does not compute crest factor, kurtosis or envelope spectra today. It does not diagnose a race or ball, predict a seizure, or estimate remaining useful life.
Results leave through REST /api/status and WebSocket /ws. Optional MQTT publishes silent-ear/status and silent-ear/health; silent-ear/alerts publishes when status is CRITICAL. Prometheus reads /metrics.
Measured performance, within its limits
40.96 M values in 1.29–1.44 s, Apple M3 Pro, release build, synthetic IMS-format files, RMS stage only; not measured on real IMS files or a Raspberry Pi.
Current record / output illustration
illustration; this page is not connected to a broker
silent-ear/health
Score unavailable — learning demo baseline
/metrics
Awaiting reference.
No illustrated CRITICAL alert.