Every digital vibration monitoring system converts the continuous analog signal from an accelerometer into discrete samples at a fixed rate. That rate — the sampling frequency — determines the maximum vibration frequency the system can measure. Choose too low a sampling rate and the system is physically incapable of detecting the high-frequency signatures that early-stage bearing defects produce. Choose an unnecessarily high rate and you generate massive data volumes that stress storage, bandwidth, and battery life without adding diagnostic value. This article explains the fundamental relationship between sampling rate and measurable frequency content, applies it to bearing monitoring for different machine speeds, and discusses the practical trade-offs that drive sampling rate selection in IoT sensor architectures.
The Nyquist-Shannon Sampling Theorem
The Nyquist-Shannon sampling theorem states that a continuous signal can be perfectly reconstructed from its samples if the sampling rate is at least twice the highest frequency component present in the signal. This minimum rate is called the Nyquist rate. The maximum frequency that can be represented at a given sampling rate is the Nyquist frequency, equal to half the sampling rate:
f_Nyquist = f_sample / 2For example, a system sampling at 20,000 samples per second (20 kS/s) can represent frequencies up to 10,000 Hz. A system sampling at 1,000 S/s can represent frequencies up to only 500 Hz.
Aliasing: What Happens Below the Nyquist Rate
When a signal contains frequency content above the Nyquist frequency, those high-frequency components do not simply disappear from the digital data. They are aliased — folded back into the measurable frequency range and appear as false spectral components at incorrect frequencies. A 7,000 Hz vibration component sampled at 10 kS/s (Nyquist = 5,000 Hz) appears in the digital spectrum at 3,000 Hz (10,000 – 7,000 = 3,000 Hz). This aliased peak looks identical to a genuine 3,000 Hz signal, and there is no way to distinguish it from real content after sampling.
Aliasing is not a minor nuisance — it produces false information. An aliased bearing defect frequency could masquerade as a gear mesh harmonic, or vice versa. An aliased structural resonance could create a phantom spectral peak that triggers false alarms. To prevent aliasing, every properly designed data acquisition system includes an anti-aliasing filter — an analog low-pass filter that attenuates signal content above the Nyquist frequency before the analog-to-digital converter samples the signal.
In practice, anti-aliasing filters are not perfectly sharp. They require a transition band to roll off from passband to stopband. A practical rule is that the usable analysis bandwidth is approximately 40% of the sampling rate (80% of Nyquist). A system sampling at 20 kS/s has a Nyquist frequency of 10 kHz but a usable analysis bandwidth of approximately 8 kHz after accounting for the anti-aliasing filter roll-off.
Frequency Requirements for Bearing Monitoring
What frequencies must a bearing monitoring system capture? The answer depends on the diagnostic technique being used.
Overall Vibration Level (Velocity RMS)
The ISO 10816 / ISO 20816 standard for evaluating machine vibration severity uses velocity RMS in the 10–1,000 Hz band. A sampling rate of 2,560 S/s (usable bandwidth ~1,000 Hz) is sufficient for this metric. Most low-cost IoT vibration sensors use this approach. It detects gross mechanical problems — severe imbalance, looseness, late-stage bearing failure — but cannot detect early-stage bearing defects.
Direct Spectral Analysis of Defect Frequencies
For a bearing with defect frequencies below 500 Hz (covering most bearings on machines running below approximately 3,000 RPM), the fundamental defect frequencies fall within the 10–1,000 Hz band. A 2,560 S/s rate captures these. However, the harmonics of defect frequencies — 2× BPFO, 3× BPFO, and higher — extend higher. For a BPFO of 107 Hz, 10 harmonics extend to 1,070 Hz. Higher harmonics at 15× or 20× BPFO (1,600–2,140 Hz) become important as defect severity grows. A sampling rate of 5,120 S/s (bandwidth ~2,000 Hz) captures these higher harmonics.
Envelope Analysis
Envelope analysis (amplitude demodulation) — the most powerful tool for early bearing defect detection — operates on the high-frequency structural resonance band, typically 2,000–10,000 Hz. To capture this band, the system must sample at 25,600 S/s or higher (usable bandwidth 8,000–10,000 Hz). This is a factor of 10 higher than the rate needed for simple overall vibration monitoring.
This sampling rate requirement is the central trade-off in IoT bearing monitoring sensor design. A sensor sampling at 25.6 kS/s for 1 second generates 25,600 samples — perhaps 50 kB of data at 16-bit resolution. A sensor sampling at 2.56 kS/s generates only 2,560 samples (5 kB). The 10× difference in data volume directly affects wireless transmission time, energy consumption (and therefore battery life), cloud storage costs, and processing load.
Ultrasonic / Stress Wave Monitoring
Some advanced monitoring techniques use ultrasonic frequencies (25–100 kHz and above) to detect lubrication breakdown and very early metal-to-metal contact. These require sampling rates of 100 kS/s or more and are typically limited to wired, permanently powered systems due to the extreme data rates involved.
Sampling Rate by Machine Speed
Bearing defect frequencies are proportional to shaft speed. Higher shaft speeds produce higher defect frequencies, requiring higher sampling rates. Lower shaft speeds produce lower defect frequencies that are easier to capture but require longer waveform records (more time) to accumulate enough cycles for reliable spectral analysis.
Low-Speed Machines (60–300 RPM)
Shaft frequency: 1–5 Hz. Typical BPFO: 5–25 Hz. Envelope analysis band: 500–5,000 Hz. Minimum sampling rate for envelope analysis: 12,800 S/s. The challenge at low speed is not the sampling rate but the record length: at 5 Hz BPFO, a 1-second record contains only 5 defect cycles. Reliable spectral analysis requires 10–20 seconds of data, generating 128,000–256,000 samples per acquisition at 12.8 kS/s.
Medium-Speed Machines (300–3,600 RPM)
Shaft frequency: 5–60 Hz. Typical BPFO: 25–320 Hz. Envelope analysis band: 2,000–10,000 Hz. Minimum sampling rate for envelope analysis: 25,600 S/s. This covers the majority of industrial machinery — motors, pumps, fans, compressors. A 1-second record at 25.6 kS/s provides adequate frequency resolution (1 Hz) and sufficient defect cycles for reliable detection.
High-Speed Machines (3,600–60,000 RPM)
Shaft frequency: 60–1,000 Hz. Typical BPFO: 320–5,000 Hz. Envelope analysis band: 5,000–20,000 Hz or higher. Minimum sampling rate for envelope analysis: 51,200 S/s or higher. High-speed spindles, turbomolecular pumps, and dental handpieces push defect frequencies into the kilohertz range, requiring correspondingly high sampling rates. These applications often use specialized high-bandwidth systems rather than general-purpose IoT sensors.
The IoT Sensor Trade-Off
Battery-powered wireless vibration sensors must balance diagnostic capability against energy budget. The energy cost of a vibration measurement is dominated by two factors: the data acquisition itself (powering the sensor and ADC) and the wireless transmission of the data.
Consider a concrete example: a sensor measuring a 1-second waveform every 15 minutes, 24 hours a day.
- At 2,560 S/s: 5 kB per acquisition × 96 acquisitions/day = 480 kB/day
- At 25,600 S/s: 50 kB per acquisition × 96 acquisitions/day = 4,800 kB/day (4.8 MB/day)
The 10× data volume difference translates to approximately 10× longer transmission time and 5–8× more energy per measurement cycle (ADC power also increases with sampling rate). For a sensor running on a lithium battery with a 5-year target life, this difference can mean choosing between a AA-sized battery and a D-sized battery — or between quarterly battery changes and annual changes.
Some IoT sensor architectures address this trade-off by using adaptive sampling: the sensor normally operates at a low sampling rate (2.56 kS/s) for overall vibration monitoring and periodically switches to a high rate (25.6 kS/s) for detailed spectral and envelope analysis. This reduces the average energy consumption while preserving the ability to perform full diagnostics on a scheduled or triggered basis.
Other platforms, like Fault Ledger, prioritize high-fidelity capture for every measurement, using sampling rates of 25.6 kS/s or higher as the standard operating mode. This approach ensures that every data record supports full envelope analysis and waveform-level forensic examination, at the cost of higher per-measurement energy consumption — a trade-off justified for critical bearing applications where early detection and failure evidence quality are paramount.
Spectral Resolution and Record Length
Sampling rate determines the frequency range, but record length (the duration of the captured waveform) determines frequency resolution:
Δf = 1 / Twhere T is the record length in seconds and Δf is the frequency resolution in Hz. A 1-second record provides 1 Hz resolution. A 0.1-second record provides only 10 Hz resolution — insufficient to separate a BPFO of 107 Hz from a shaft harmonic at 120 Hz.
For slow-speed bearings with closely spaced defect and shaft frequencies, high frequency resolution requires long records. A machine running at 60 RPM (1 Hz shaft frequency) with a BPFO of 3.57 Hz needs at least 0.5 Hz resolution to separate BPFO from harmonics of shaft speed (3 Hz and 4 Hz). This requires a 2-second record minimum. For reliable detection with clear spectral separation, 5–10 seconds is typical.
Total data per acquisition = sampling rate × record length × bytes per sample. At 25,600 S/s × 10 seconds × 2 bytes = 512 kB — a significant volume for a battery-powered wireless sensor.
Practical Guidelines for Sampling Rate Selection
- If you only need overall vibration severity (ISO 20816 compliance): 2,560 S/s is sufficient. Usable bandwidth ~1,000 Hz. Lowest power, smallest data volumes.
- If you need direct spectral analysis of defect frequencies and their harmonics: 5,120–10,240 S/s. Usable bandwidth 2,000–4,000 Hz. Moderate power and data requirements.
- If you need envelope analysis for early defect detection: 25,600 S/s or higher. Usable bandwidth 8,000–10,000 Hz. This is the minimum for serious bearing diagnostics. Higher power and data requirements, but enables 2–6 months additional warning time before failure.
- If you need ultrasonic monitoring for lubrication analysis: 102,400 S/s or higher. Typically limited to wired, permanently powered installations.
- Match record length to machine speed: Ensure the record contains at least 10–15 full cycles of the lowest defect frequency of interest. For a BPFO of 10 Hz, this means at least 1–1.5 seconds.
Conclusion
Sampling rate is not a specification to gloss over when selecting a bearing vibration monitoring system. It determines the frequency ceiling of everything the system can see. Below the Nyquist frequency, the system measures faithfully. Above it, information is either lost (filtered out) or corrupted (aliased into false peaks). For bearing diagnostics, where early defect detection depends on capturing high-frequency impulse energy for envelope analysis, the required sampling rate is typically 25.6 kS/s — ten times higher than what simple overall vibration monitoring requires. Platforms built for high-frequency capture, such as Fault Ledger, make this sampling rate the default precisely because the diagnostic techniques that matter most for bearing health — envelope analysis and waveform-level forensics — depend on it. Understanding this trade-off — and choosing a sampling rate that matches your diagnostic objectives — is a fundamental step in designing an effective bearing monitoring program.