Vibration sensors in practice: types, sensitivity, mounting

You put a magnet on a fan's bearing housing and read 0.4 mm/s, on a machine you can hear through the wall. Or the opposite: 180 mm/s on a pump that's been running quietly for three years. In both cases, it isn't the rotor lying to you - it's the measurement chain: the sensor, the sensitivity you entered, the mount, the cable. Let's take that chain apart piece by piece, from the sensing element to the connector, and see exactly where trust in the number gets lost.

Updated 27 August 2026 · by AXILINE · Vila Nova de Gaia

Three working sensor types, and one special case

A piezoelectric accelerometer covers almost everything you'll do in the field. Inside it is a piezo element loaded by a seismic mass: under vibration, the mass presses on the crystal, and the crystal releases a charge proportional to the vibration acceleration. This type of sensor's bandwidth starts at a few hertz and runs into the tens of kilohertz, its housing tolerates a hot bearing, and its mass is small. Balanset-1A's standard sensors measure no more than 25×25×20 mm and weigh no more than 40 g, and that's not incidental: a heavy sensor loads a thin housing wall all by itself.

The price of that wide bandwidth is electronics. The charge from the piezo element has to be converted into a voltage, so the sensor needs either an external charge amplifier or a built-in preamplifier powered over the signal line. Capacitive (MEMS) accelerometers are several times cheaper and don't need a charge amplifier, but they have their own trade-off: after integration, their signal carries low-frequency noise around 0.5-3 Hz. For a machine at 1500 RPM that doesn't matter; for a slow-turning rotor at 120 RPM, that noise sits right in the working band.

A velocity sensor outputs a voltage proportional to velocity directly, with no integration at all. The classic inductive type is a coil and a magnetic core on springs; its conversion factor is high, typically tens of mV per mm/s, and on models with a mechanical amplifier it can run into the thousands. The trade-off here is different: mass, fragility, a narrow bandwidth, and cost. There's also a hybrid: a piezoelectric accelerometer with a built-in charge amplifier and integrator - from the outside it behaves like a velocity sensor, inside it's still the same piezo element.

Non-contact displacement sensors - proximity probes - solve a problem an accelerometer can't touch at all. An eddy-current, inductive, or capacitive sensor looks at the shaft across a gap and measures the shaft's relative displacement within the bearing, starting from 0 Hz. A sensitivity of around 1000 mV/mm gives good resolution with no amplification needed. The limitations are just as strict: the working gap is tied to the coil's diameter - a typical sensor has a gap of around 4 mm and a range of about ±2.5 mm - it needs a rigid mounting stand, and you have to account for mechanical and electrical runout on the shaft surface. This is the tool for machines on plain (sleeve) bearings and for slow-turning shafts, where the bearing housing itself barely shows anything.

The special case is the force sensor. On a below-resonance balancing machine with very rigid supports, a large imbalance produces almost no noticeable vibration: the rigid support barely deforms, even though the load on the bearings is real. There, what gets measured isn't vibration, but the force in the support. That gives a simple rule: flexible supports and above-resonance operation (running speed above the supports' natural frequency) call for accelerometers; rigid supports and below-resonance operation (running speed below the supports' natural frequency) call for force sensors. We cover this kind of support in more detail in the article on building a balancing stand.

Sensor and measurement-point selection for diagnostics is covered in ISO 13373. Check the applicable part and current edition for your machine type: the requirements for sensor frequency range and for compensating shaft runout differ there between rolling-element and plain bearings.

Sources: ISO 13373-3:2015

Why the instrument displays mm/s when the sensor measures acceleration

The Balanset-1A's signal chain looks like this: the accelerometer outputs a voltage proportional to vibration acceleration, the two-channel USB module amplifies it with a preamplifier, runs it through an integrator, and digitizes it at the ADC, and the software on the computer calculates amplitude, phase, spectrum, and weights. Integration happens in hardware, before the signal is converted to digital. So what you see on screen is mm/s, even though the sensing element has been measuring acceleration the whole time.

What integration does to the signal is visible from the formula. Vibration velocity equals acceleration divided by angular frequency: v = a/ω, where ω = 2π·f. At 10 Hz you're dividing by 63; at 1000 Hz, by 6283. In other words, the integrator suppresses high frequencies by two orders of magnitude, and boosts low frequencies by the same factor.

That gives two practical consequences. First: high-frequency content nearly vanishes from vibration velocity, so looking for a rolling-element bearing defect in mm/s is pointless. For that, you use vibration acceleration and envelope analysis - a method that pulls bearing impacts out of the high-frequency part of the signal. Second: any low-frequency noise and drift gets amplified by the integrator. That same MEMS-sensor noise around 0.5-3 Hz grows, after integration, into a tilted-up 'skirt' at the left edge of the spectrum - one that also eats into the channel's dynamic range.

That's exactly why the measurement band is deliberately limited. The Balanset-1A calculates the RMS (root-mean-square) vibration velocity over a 5-200 Hz range, and measures within 0.02-80 mm/s. That's more than enough for the running-speed component - vibration at the rotor's rotation frequency - and the first few harmonics on industrial machines. If a rotor turns at 120 RPM, its running frequency of 2 Hz already falls below that band, and you need a different approach: double integration through to vibration displacement, or non-contact sensors, which see displacement from zero hertz up.

The difference between RMS, peak, and peak-to-peak, and how to choose the band for acceptance testing, are covered separately in the article on measuring vibration correctly. What matters here is one thing: the 5-200 Hz band is not the 10-1000 Hz band that contractual tolerances are usually written against. Agree on the instrument and the band before the visit, not on the day of sign-off.

Sources: Balanset-1A operation manual

Sensitivity: the right picture with the wrong numbers

A sensor's sensitivity is the coefficient that converts physics into millivolts. For general-purpose piezo accelerometers it typically falls between 10-30 mV/(m/s²), and up to 100 or more for precision types. For velocity sensors, it's counted in mV per mm/s, usually 40-80. The figure is specific to the individual unit - it's on that particular sensor's datasheet, not in the model's general description.

In the Balanset-1A, integration happens inside the measurement module, so the coefficient you enter into the software is already in mV/(mm/s), nominally 20-30. It lives in the settings window under the F4 key, entered separately for each of the two channels, with the decimal fraction separated by a comma. There's exactly one occasion to touch it: when you've replaced the sensor. The rest of the time, that field is best left alone.

Now, the important part. A sensitivity error is a pure linear distortion of scale. The spectrum's shape stays correct, the ratio between the running-speed component and the overall vibration (the total level across the whole frequency band) stays correct, the phase stays correct, the polar plot stays correct. Balancing itself usually still converges: the instrument calculates influence coefficients - the machine's response to the trial weight - from those same distorted numbers, and the scale factor cancels itself out. What breaks is something else, and it happens to be exactly what you hand to the customer. The mm/s in the report is wrong, the ISO 20816 zone evaluation is wrong, the residual unbalance in g·mm is wrong, the setpoint in your monitoring system is wrong, and any trend you file that measurement into is corrupted for good.

It's worse when the two channels drift apart from each other. You fit a new sensor on channel one, enter its 27 mV/(mm/s), and channel two is left at the old 20. The amplitude ratio between the two supports is now off by almost a third, and in two-plane balancing, it's exactly that ratio that determines how mass gets split between the planes. The software will still get you to tolerance eventually - just by way of extra runs.

• Sensitivity is taken from the specific sensor's datasheet, not from the model's general description.
• Both channels are entered separately, and the values are checked against the labels on the sensors.
• The decimal fraction is entered the way the software expects, with a comma.
• You updated the coefficient the moment the sensor was replaced, not 'later, whenever I remember.'
• The entered values are recorded in the report, along with the sensors' serial numbers.
• Sensors are labelled by channel, so channel one and channel two can't get physically swapped.

You can check sensitivity in the field, roughly but usefully: mount both sensors side by side on the same bearing housing, same direction, and take a reading. A few percent difference is normal; a difference of several-fold means an error in the coefficient or a dead channel. Only a calibrator gives you a proper check, and for acceptance measurements, you need one.

Sources: Balanset-1A operation manual · ISO 20816-1:2016

Mounting sets the upper limit of the band

The sensor, the mount, and the surface together form a mass-spring system. The joint's stiffness acts as the spring, the sensor's mass plus part of the mount is the mass, and that system has a natural frequency: the stiffer the joint and the lighter the sensor, the higher it is. Below that frequency, the sensor faithfully tracks the housing's motion. Close to it, the mount starts amplifying the signal on its own; above it, the mount just rings and mixes its own peak into the spectrum. The working rule is simple: trust roughly the bottom third of the mount's natural frequency.

A stud into a prepared hole gives the maximum joint stiffness, and a bandwidth of around 10 kHz. A glued mounting pad is a bit softer. A magnet holds the sensor across a thin air gap and a layer of oxide, and its natural frequency drops to one and a half to two kilohertz. A hand-held probe provides almost no stiffness at all.

It's worth being direct about the hand-held probe, because people try to make it do things it can't. Its upper limit of 300-500 Hz means that at 1450 RPM you'll see the 24 Hz running-speed component and the first few harmonics, and that's it. Early-stage bearing defects don't live in that range. A chip on a raceway produces a short impact, the impact excites the natural frequencies of the ring and the housing in the low kilohertz range, and the entire envelope analysis is built on exactly that high-frequency energy. The probe cuts it off. What you get isn't 'the bearing is fine' - it's 'the instrument wasn't looking there.' Add poor repeatability on top: hand pressure, angle, and contact point change every time, spread easily exceeds 30%, and a month's worth of real growth in the level drowns in the method's own noise.

There's also an opposite effect people forget about. The sensor's own mass loads the structure. A forty-gram sensor on a massive cast bearing housing means nothing; the same sensor on a thin cover or a sheet of cladding noticeably lowers that section's natural frequency and changes the picture. The lighter the sensor and the more massive the mounting location, the less you influence the thing you're measuring.

The figures in the table are a guide, not a spec. Take the actual bandwidth for your sensor and your mounting method from its documentation. And a separate rule for trending: don't change the mounting method at a point that already has history behind it. Spectra from a magnet and from a stud differ, and you'll read that difference as a change in the machine.

The mounting pad: exactly where to put the sensor

The mounting pad matters more than the sensor's brand. A magnet on a layer of paint gets a springy gasket in the joint, and the mount's natural frequency drops to around a kilohertz - meaning you lose half the rated bandwidth without noticing. A magnet on a curved surface bears on two points instead of a flat plane and rocks on them. A magnet on metal chips or oily grime holds worse, and can come loose while running.

Choose the location by following the path of the force. Centrifugal force from imbalance travels from the rotor, through the bearing, into the housing, from the housing into the frame, and on into the foundation. So you need to measure on the bearing housing, on a massive part of the casting, ideally in the loaded zone - along the line the housing actually carries load on. On a belt-driven unit that's the direction of belt tension; on a pump, it's the direction of the resultant hydraulic force. The shorter the metal path from the bearing to the sensor, the less the structure interferes with the signal.

A properly prepared pad works for years. At the next shutdown, clean the spot down to bare metal, machine a flat spot if needed or glue on a pad, and, if the rules allow, tap a hole for an M4 stud. From then on you'll be mounting the sensor at the same point with the same stiffness every time, and the scatter between rounds will drop on its own.

• The point is on the bearing housing or its massive casting - not on a cover, not on a guard, not on a terminal-box lid.
• The spot is cleaned down to bare metal: paint, filler, rust, and grime removed.
• The surface is flat. A flat magnet needs a flat surface; for a cylindrical one, use a two-pole magnet or machine a flat spot.
• The location is as close to the bearing as possible and in the loaded zone, not on a thin rib.
• The point is marked with a punch mark, paint, or a label, and the bearing's designation stays the same from one visit to the next.
• Roll the magnet on from the edge, don't drop it flat: an impact overloads the input and produces 'ringing' in the first seconds of the recording.
• No rotating parts are reachable from that point, and the housing doesn't burn your hand.

Never reach a sensor, or your hand, through a protective guard. Skip an inaccessible point and record the reason. A measurement taken at the risk of injury isn't repeatable anyway: the second time, you physically won't place the sensor the same way.

Directions and consistency: across rounds, and within a single balancing job

The axis of maximum sensitivity is marked on the sensor's housing, and it needs to line up with the measurement direction. A misalignment costs you quietly: a 30° tilt gives a projection of 0.87 of the true value - a 13% understatement. You'll conclude the machine has calmed down, when it's only the sensor that has.

For a radial measurement, the horizontal direction is usually chosen. The foundation's stiffness horizontally is typically lower, the housing deforms more in that direction, and sensitivity to imbalance comes out higher. There's no fundamental difference between horizontal and vertical - what matters more is keeping whichever direction you chose consistent. The axial direction answers a different question: that's where shaft misalignment lives, and balancing won't remove it. How to read the picture across all three directions is covered in the articles on tracking down the cause of vibration, and on comparing imbalance and shaft misalignment.

A separate point about consistency within a single balancing job, because here a mistake wrecks the entire calculation. The instrument calculates the influence coefficient as a ratio of vectors: the numerator is the change in the vibration vector at the same point, in the same direction, with the same mounting. Move the sensor ten centimetres between the initial and trial runs, turn the magnet, set it on a different pad, and the phase will shift by tens of degrees all on its own. The instrument will faithfully calculate from that data and put the weight in the wrong sector. The same goes for the reflective marker: it sets the phase reference point, and it must not be repositioned between runs.

• One sensor, one point, one direction, one mounting method for the whole balancing job, from the initial run to the verification run.
• The phase marker goes on once and stays untouched until you're done.
• Cables are routed so the sensor never has to be repositioned partway through the job.
• For route-based rounds, record the bearing's designation, direction, and mounting method in the log, or the trend ends up built from measurements that can't be compared.
• If a point genuinely has to be changed, start a new baseline and note what changed and why, without editing the old records.

Cables, interference, and signs of a broken signal chain

The cable is part of the sensor, not just a wire running 'to the instrument.' A cable hanging free vibrates along with the machine, its shield flexes, and flexing generates triboelectric noise - a parasitic charge from layers rubbing against each other. For charge-output sensors this is the single most common source of spectrum clutter; for sensors with a built-in preamplifier the effect is weaker, but it doesn't disappear. Secure the cable to the housing so a short strain-relief loop is left at the connector, with the rest of the cable lying still - neither pulled taut nor free to swing.

Interference arrives both through the air and through the ground. Don't run the measurement cable alongside power lines, and especially not alongside a VFD's output: a VFD generates broadband interference, and it'll land in your spectrum as lines at 'impossible' frequencies that no amount of filtering afterward will remove. If the site's power is poor, run from the laptop's battery - that's the instrument's standard recommendation. And check the connectors: a connector that isn't fully seated, or a dirty contact, produces exactly the same picture as a broken shield.

An implausible number almost always means a problem in the signal chain, not in the machine. Before raising any alarm, remove the sensor and set it down on the floor. The reading should drop to almost zero. If it doesn't, you weren't measuring the machine.

Checking the signal chain in five minutes

These five minutes before the first run save you a repeat site visit. The sequence is the same whether you're balancing a fan or taking a round-based reading.

1. 1

   Inspection and connection

   Channel 1 and channel 2 sensors into connectors X1 and X2, the laser phase sensor into X3, the module over USB to the laptop. Connectors latched all the way, cables free of kinks or crush marks.

2. 2

   Zero level on a stopped machine

   Switch to vibrometer mode before starting up. The level should be small but alive. Exactly zero means a dead channel; 3 mm/s on a stationary machine means interference or a fault in the chain.

3. 3

   Tap the housing

   Give the housing a light tap with your hand, near the sensor. The response has to show up on exactly that channel. In ten seconds, you've confirmed both channels are alive and not swapped.

4. 4

   RPM and the marker

   The tachometer should give a stable reading close to what's expected. Jumps to multiples of that value mean a stray reflection or a second mark. Mount the tachometer stand on a stationary, rigid surface, not on a cover.

5. 5

   Stability at operating speed

   The running-speed component's amplitude and phase shouldn't change by more than 10-15% over the course of the measurement. More than that means either operation near resonance or a problem in the chain, and you shouldn't balance from data like that.

6. 6

   Cross-checking the channels

   If in doubt, mount both sensors side by side on the same housing, same direction, and compare. A few percent difference is normal; several-fold means an error in sensitivity or a faulty channel.

Write the mounting method, the entered sensitivities, and the sensors' serial numbers into the report right away. Six months from now, that's the only way to work out why the old numbers don't match the new ones.

Sources: Balanset-1A operation manual

Two channels for two bearings

The two-channel setup isn't there for convenience. Two-plane balancing needs the software to have four influence coefficients: how a weight in plane one changes vibration at bearing one and at bearing two, and the same again for a weight in plane two. The cross coefficients - how a plane affects the far bearing - are usually smaller than the direct ones, and it's exactly those that carry the information about moment (couple) unbalance, where the heavy points at the two ends of the rotor face in opposite directions. The cross coefficients are also the easiest ones to corrupt.

With one channel, you physically can't measure both bearings at the same time, so you take them on separate runs. Between runs the machine changes: it warms up, RPM drifts slightly, the load or the damper position shifts. The vectors the instrument later subtracts come from two different states, and part of the difference between them is explained not by the weight, but by drift in operating conditions. The small cross coefficients suffer first. Two channels capture both bearings at the same instant, from the same revolution, referenced to the same timing mark, and that problem disappears.

The second argument is the number of runs. Two-plane balancing on two channels needs three runs: an initial one and two trial runs, one per plane. On a single channel, the same job stretches to twice that. For an induced-draft fan or a mill that can't be started and stopped at will, the difference between three runs and six is the difference between one shift and two. We break down exactly what a site visit takes, and where the time goes, in a separate article on the number of runs.

We're engineers who design and manufacture the Balanset instruments and use them to balance rotors on site ourselves, so the measurement chain that ships with the kit is built for shop-floor work, not for a lab bench. Two single-axis accelerometers, no larger than 25×25×20 mm and no heavier than 40 g, mounted on a magnet or an M4 stud. A laser phase sensor reading off a reflective marker, on a magnetic stand, with a range of 100-100,000 RPM and phase accuracy of ±1°. A two-channel USB module with preamplifiers, integrators, and an ADC, simultaneous acquisition on both channels, overall vibration and the running-speed component, phase, spectrum, time waveform, an archive, and reports. There's a Balanset-1A OEM version without the case, for building into machine tools and test stands. Consulting support on measurement methodology and on the instrument itself is included: if you're unsure where to mount the sensor on your particular machine, ask before the visit, not after.

Sources: Balanset-1A manufacturer specification

Frequently asked questions

Related content