Accelerometer Specs for Vibration Monitoring: What Matters

The new wireless sensor goes on the motor, the dashboard fills with tidy trend lines, and three months later a bearing fails without the trend moving. The sensor worked. It just couldn’t see the frequencies where an early bearing defect shows up, and it was held on with a magnet that cut off the high end anyway. Accelerometer specs for vibration monitoring only make sense when you read them against the fault you’re trying to catch. This article goes through the specifications that matter (bandwidth, noise density, g-range, resonance, axes and mounting), shows how to size each one with a little arithmetic, and explains why the datasheet’s best number often isn’t the one you get on the machine.

Key Takeaways

• Start from the fault. Imbalance and misalignment show up near running speed; bearing, gear-mesh and cavitation faults are most detectable at several kilohertz.
• Acceleration grows with frequency. The same vibration velocity is 100 times more acceleration at 3 kHz than at 30 Hz, which affects both g-range and noise.
• Noise density only means something with a bandwidth: total noise grows with the square root of the bandwidth you keep.
• The mounting is part of the sensor. Endevco’s installation guide shows that stud mounting keeps the high frequencies a magnet loses.
• MEMS and piezoelectric sensors each win somewhere. MEMS measures down to DC and has self-test; piezoelectric sensors usually have lower noise at high frequencies.

“The sensor didn’t see it coming”

Most disappointing vibration deployments fail in one of a few predictable ways.

The bandwidth stops below the fault. A sensor with a 1 kHz bandwidth tracks imbalance on a 1,800 rpm motor perfectly and is blind to the high-frequency impacts that an early bearing defect excites.

The noise floor sits above the healthy signal. A smooth, well-balanced machine vibrates very little. If the sensor’s own noise is larger than that, the trend is a flat line of sensor noise until the fault is already severe.

The range clips. A sensor sized for gentle running vibration saturates during start-up, impacts or a loose-foot event, and a clipped signal produces offsets and false harmonics.

The mounting filters the signal. A magnet, a thick adhesive layer or a bracket adds a mechanical resonance and a roll-off that the datasheet doesn’t know about.

Each failure points at a specific specification, and each specification can be sized before you buy.

How the key accelerometer specs work

Bandwidth: set by the fault, not the machine speed

Different faults live at different frequencies. Imbalance appears at running speed, misalignment at running speed and its low harmonics. Rolling-element bearing faults are different. In their widely cited tutorial in Mechanical Systems and Signal Processing, Randall and Antoni describe how each impact of a local bearing defect excites high-frequency resonances of the bearing and its housing, and how envelope analysis recovers the repetition rate (the defect frequency) from that high-frequency band. The impacts repeat at a few hundred hertz, but the energy you detect is often in the kilohertz range.

The defect frequencies themselves come from geometry. The outer-race frequency is BPFO = (n/2) × f_r × (1 − (d/D) cos φ), where n is the number of rolling elements, f_r the shaft frequency, d the element diameter, D the pitch diameter and φ the contact angle. An illustrative example: a 4-pole motor on 60 Hz mains running at 1,770 rpm (f_r = 29.5 Hz), with a bearing of 9 balls, d/D = 0.2 and φ = 0, has BPFO = 4.5 × 29.5 × 0.8 ≈ 106 Hz. Randall and Antoni note that the real frequencies typically differ from the calculated ones by a percent or two because of slip, which is one reason envelope spectra are read with a tolerance.

Analog Devices’ Chris Murphy, writing about condition-monitoring sensors, puts the practical consequence plainly: bearing, gear-mesh and pump-cavitation faults are most detectable above about 5 kHz, sometimes up to 20 kHz or beyond. A 1 kHz sensor is a good tool for imbalance and looseness; it is not an early bearing-fault detector.

Acceleration scales with frequency squared

Accelerometers measure acceleration, but vibration severity is usually judged in velocity, and displacement matters for clearances. The three are linked by frequency: for a sinusoid, acceleration = 2πf × velocity = (2πf)² × displacement.

Two worked numbers make the point. Benjamin Reiss of Analog Devices notes that a 250 nm displacement is about 1 g at 1 kHz but about 100 g at 10 kHz. And holding velocity fixed at 4.5 mm/s RMS, the acceleration is about 0.09 g at 30 Hz, 0.86 g at 300 Hz and 8.6 g at 3 kHz. That has two consequences. High-frequency faults can produce surprisingly large accelerations, so the g-range must cover them. And low-frequency vibration produces very small accelerations, so the noise floor must be low enough to see them.

Noise density: always multiply by the bandwidth

Accelerometer noise is usually specified as a density in µg/√Hz. To compare it with a vibration level you need the RMS noise over the bandwidth you keep, which is the density times the square root of the noise bandwidth. For a single-pole filter the noise bandwidth is π/2 times the −3 dB frequency.

Murphy’s example: 80 µg/√Hz through a 1 kHz single-pole low-pass filter gives about 3.2 mg RMS. The same arithmetic shows why wide bandwidth is not free: 25 µg/√Hz over a 5 kHz single-pole bandwidth is about 2.2 mg RMS, and 100 µg/√Hz over the same band is about 8.9 mg. Murphy relates these figures to the vibration levels of known-good machines in the ISO machine classes and concludes that a sensor for a small, smooth asset needs a noise floor of a few milli-g to see the healthy baseline at all.

g-range: cover the worst event, including gravity

Reiss’s advice on range is simple and often ignored: choose a range that covers every acceleration in the system, and remember that gravity alone puts 1 g on a vertical axis. A signal that exceeds the range clips, and clipping produces asymmetric offsets and harmonics that look like faults. Murphy notes that condition-monitoring requirements start around ±16 g for smaller assets and reach ±500 g for some industrial equipment.

There’s a tension here. In many sensors a larger range comes with more noise, so the range should be large enough for the worst real event and no larger.

Resonance: the frequency the sensor exaggerates

Every accelerometer has a mechanical resonance where its response rises sharply. Reiss points out that accelerations at the resonant frequency are amplified and distort the measurement unless they are damped or filtered. The usable bandwidth sits well below the resonance, and the resonance of the sensor as mounted is what counts.

Mounting: the spec the datasheet doesn’t control

Jim Matthews’ installation guide for Endevco (TP319) states that stud mounting gives higher transmissibility than any other method, that no adhesive is as stiff as a normal mounting stud, and that magnetic adapters are generally useful only for low-frequency measurements below a few hundred hertz. It also specifies the surface preparation that stud mounting assumes: a flat, smooth spot face, a perpendicular tapped hole and the manufacturer’s torque. A sensor with a 10 kHz datasheet bandwidth, held by a magnet on a curved, painted housing, is a sub-kilohertz sensor.

MEMS or piezoelectric

Albarbar and colleagues tested three capacitive MEMS accelerometers back to back with a calibrated piezoelectric reference on a shaker and on a CNC machine, published in Sensors in 2008. Two of the three tracked the reference well, including running speed and harmonics on the machine; the third showed unstable sensitivity, noise and phase shift. Their conclusion was that MEMS sensors can be suitable for condition monitoring if they are chosen by specification and validated, not assumed to be equivalent.

Murphy summarizes the trade-off from a vendor’s perspective (and the article promotes Analog Devices’ own parts): piezoelectric sensors typically reach lower noise at high frequency, while MEMS sensors respond down to DC, include self-test and are easier to integrate with wireless nodes.

Standards set the yardstick, not the sensor

ISO 20816-1 sets out general conditions and procedures for measuring and evaluating machine vibration, with criteria expressed in terms of both vibration magnitude and change of vibration; machine-specific limits are in the other parts of the series. Whatever sensor you choose has to resolve the quantity and frequency range that the relevant part asks for.

Accelerometer specs compared by fault type

Fault	Where it shows up	Bandwidth needed	Noise priority	Range priority	Mounting
Imbalance	1× running speed	Up to ~10× running speed	High on smooth machines (small accelerations)	Low	Magnet often acceptable
Misalignment, looseness	1×, 2×, 3× and low harmonics	Up to ~1 kHz	Moderate	Moderate (impacts on looseness)	Magnet or adhesive
Early bearing defects	Kilohertz resonances, envelope at defect frequencies	Several kHz, ideally >5 kHz	Moderate	High (impacts)	Stud or stiff adhesive
Gear mesh	Teeth × shaft speed and sidebands	Above mesh frequency and harmonics	Moderate	High	Stud
Pump cavitation	Broadband high frequency	>5 kHz	Moderate	High	Stud

The bandwidth and mounting columns follow the sources above; the “up to” figures for imbalance and misalignment are rules of thumb for illustration, not standard limits.

What to actually do

Write down the fault before the part number

List the faults that matter on this machine and estimate their frequencies from running speed, gear teeth and bearing geometry. Those frequencies set the minimum bandwidth.

Size the noise floor from the healthy machine

Convert the smallest vibration you need to see (often the healthy baseline) into acceleration at its frequency, and make sure the sensor’s RMS noise over your bandwidth is well below it.

Size the range from the worst event

Include gravity, start-up transients and impacts. If you can’t estimate them, measure them with a high-range sensor first.

Specify the mounting with the sensor

Treat the mounting method, surface preparation and torque as part of the bill of materials. If the requirement includes bearing faults, plan for a stud or a stiff adhesive mount.

Validate against a reference

As Albarbar and colleagues did, mount the candidate next to a calibrated reference on the real machine and compare spectra before rolling out a fleet.

What not to do

Don’t compare noise densities from two datasheets without multiplying each by its bandwidth. And don’t read a flat trend from an under-specified sensor as proof of a healthy machine.

How to put it to work

If you have an hour

Pick one monitored machine and calculate its running speed, its bearing defect frequencies and the acceleration of its healthy vibration. Compare them with the sensor’s bandwidth, noise and range.

If you have a sprint

Mount your production sensor next to a reference accelerometer with a stud mount, capture both on the running machine and overlay the spectra. Repeat with the production mounting. The difference is what your installation is costing you.

If you want it to last the product’s lifetime

Record the sensor model, mounting, location and orientation for every installation, and repeat the side-by-side check when any of them change. A remount on a different surface is effectively a different sensor.

The question to ask yourself: “Which fault am I trying to catch, at what frequency, and have I proved that this sensor, mounted this way, can see it?”

Frequently asked questions

What accelerometer specifications matter for vibration monitoring?

Bandwidth, noise density, measurement range (g-range), resonant frequency and mounting matter most, followed by axes, cross-axis sensitivity and temperature range. Each should be sized from the faults you need to detect and from the machine’s healthy vibration level.

What bandwidth do I need for bearing fault detection?

Early bearing defects excite high-frequency resonances, so detection usually needs several kilohertz of bandwidth; Analog Devices cites faults most detectable above about 5 kHz, sometimes up to 20 kHz. The defect frequencies themselves are lower and are recovered with envelope analysis.

Are MEMS accelerometers good enough for condition monitoring?

Some are. A 2008 study in Sensors found two of three tested MEMS sensors tracked a piezoelectric reference well, while the third did not. MEMS sensors respond down to DC and include self-test; piezoelectric sensors usually have lower noise at high frequencies. Validate against a reference before deployment.

How does mounting affect accelerometer readings?

Mounting adds stiffness and mass between the machine and the sensor, which changes the high-frequency response. Stud mounting gives the best transmissibility; magnetic mounts are generally useful only below a few hundred hertz.

How do I convert accelerometer noise density to RMS noise?

Multiply the noise density by the square root of the noise bandwidth. For a single-pole low-pass filter, the noise bandwidth is π/2 times the −3 dB frequency. For example, 80 µg/√Hz with a 1 kHz single-pole filter gives about 3.2 mg RMS.

Related reading: sensor noise sources explained, choosing an anti-aliasing filter and vibration data for predictive maintenance.

Sources

1. Randall, R. B., & Antoni, J. (2011). “Rolling Element Bearing Diagnostics—A Tutorial.” Mechanical Systems and Signal Processing, 25(2), 485–520. https://doi.org/10.1016/j.ymssp.2010.07.017
2. Albarbar, A., Mekid, S., Starr, A., & Pietruszkiewicz, R. (2008). “Suitability of MEMS Accelerometers for Condition Monitoring: An Experimental Study.” Sensors, 8(2), 784–799. https://doi.org/10.3390/s8020784
3. International Organization for Standardization (2016). ISO 20816-1:2016, Mechanical vibration – Measurement and evaluation of machine vibration – Part 1: General guidelines. https://www.iso.org/standard/63180.html
4. Murphy, C. (2021). “Why MEMS Accelerometers Are Becoming the Designer’s Best Choice for CbM Applications.” Analog Devices. https://www.analog.com/en/technical-articles/why-memes-acceler-are-best-choice-for-cbm-apps.html
5. Reiss, B. (2023). “Maximize Predictive Maintenance with Careful Selection of MEMS Accelerometers.” Analog Dialogue, RAQ Issue 212. Analog Devices. https://www.analog.com/en/resources/analog-dialogue/raqs/raq-issue-212.html
6. Matthews, J. “Guide to Accelerometer Installation.” Technical Paper TP319, Endevco. https://www.endevco.com/contentStore/mktgContent/endevco/dlm_uploads/2019/02/TP319.pdf