Condition monitoring
A bearing fault hides underneath a 5 V offset — which makes this a resolution problem
The damaged race announces itself as a train of small, high-frequency impulses riding on top of a large low-frequency signal from the shaft. Recovering that impulse is a dynamic-range problem first, a sample-rate problem a distant second, and an analysis-method problem throughout.
IEPE excitation and compliance voltage, why 24-bit resolution decides what you can see, how to compute bearing defect frequencies, and what envelope analysis actually does.
Condition monitoring has an unusual property: the interesting signal is often a thousand times smaller than the signal it sits on. A shaft turning at 1500 rpm puts a large once-per-revolution component into every accelerometer on the machine, and a spalled bearing outer race adds a train of tiny impulses on top of it. If the acquisition chain cannot resolve the small thing in the presence of the large one, no amount of analysis afterwards will bring it back.
IEPE is a two-wire conversation, and the module has to hold up its end
An IEPE accelerometer has a small amplifier built into it, and that amplifier is powered over the same two wires that carry the signal back. The module supplies a constant current — 2.1 mA per channel on the USB-4431 and USB-4432 — and the sensor develops a DC bias of several volts across its own output stage. The mechanical signal rides on that bias, which is why the input has to be AC coupled before conversion.
The current loop needs headroom, and the requirement is easy to state and easy to violate: the excitation supply voltage on the card must exceed the sensor's bias voltage plus the largest signal the sensor will produce. Put numbers on it — a 100 mV/g accelerometer measuring ±10 g produces ±1 V. If its bias sits at 12 V, the card needs a supply above 13 V before the sensor can swing to its positive peak.
When it does not, the failure is asymmetric and distinctive: the positive peaks flatten while the negative half stays clean. Engineers often read this as a mechanical fault, because a clipped positive peak looks like an impact. It is not — it is the supply running out of room.
Why 24-bit is not a marketing number here
Start with what one bit is worth. On a ±10 V range, a 16-bit converter divides 20 V into 65,536 steps; a 24-bit converter divides the same 20 V into 16,777,216. That is the difference between 305 µV and 1.19 µV per step — a factor of 256, and it is the difference between a bearing impulse being a visible feature and being indistinguishable from quantisation.
| Front end | One LSB on ±10 V | What that means in practice |
|---|---|---|
| 16-bit | 305 µV | A 100 mV/g sensor sees roughly 0.003 g per step — coarse for envelope analysis, adequate for overall velocity trending |
| 24-bit | 1.19 µV | About 12 µg per step, which puts the quantisation floor well below the sensor noise |
Swipe the table sideways to see all columns
Compute the defect frequencies before you look at the spectrum
This is the step that separates a diagnosis from a guess. Bearing geometry fixes the rates at which a defect on each surface strikes the other surfaces, and those rates are all multiples of shaft speed determined by the number of rolling elements and their diameter relative to the pitch diameter. For a bearing with n rolling elements, element diameter d, pitch diameter D and contact angle φ:
| Defect | Frequency | Approximation |
|---|---|---|
| Outer race (BPFO) | (n·fᵣ/2) · [1 − (d/D)·cos φ] | ≈ 0.4 · n · fᵣ |
| Inner race (BPFI) | (n·fᵣ/2) · [1 + (d/D)·cos φ] | ≈ 0.6 · n · fᵣ |
| Rolling element (BSF) | (D/2d) · fᵣ · [1 − ((d/D)·cos φ)²] | — |
| Cage (FTF) | (fᵣ/2) · [1 − (d/D)·cos φ] | ≈ 0.4 · fᵣ |
Swipe the table sideways to see all columns
Take a bearing with eight rolling elements and d/D of 0.35 on a shaft turning at 25 Hz. The outer race defect frequency is 2.6 times shaft speed, or 65 Hz; the inner race frequency is 5.4 times, or 135 Hz; the cage frequency is 8.1 Hz. Note that BPFO is roughly n times FTF, and BPFI is roughly n times (fᵣ − FTF) — a useful cross-check when a peak does not land where you expected.
Now the part that catches people out. Those frequencies are low — tens to hundreds of hertz — and a spectrum taken over the raw signal will often show a healthy-looking peak there even on a good bearing, drowned among shaft harmonics and structural resonances. The defect is not detectable as a *tone*. It is detectable because each impact rings the bearing housing and the sensor mounting at their own natural frequencies, typically somewhere between 500 Hz and 20 kHz, and those rings are what actually appear in the spectrum.
What envelope analysis actually does
Envelope analysis, sometimes called demodulation or high-frequency resonance technique, is a four-step recipe for moving the defect information from where it is hidden to where it is legible.
- Band-pass the raw signal around a structural resonance — the band where impacts ring, not where the defect tone sits.
- Rectify the band-passed signal, so that every ring contributes a positive-going burst.
- Low-pass the result to extract its envelope. The high-frequency ringing is now gone; what remains is the rate at which rings arrived.
- Take the spectrum of that envelope. The bearing defect frequencies now appear as clear peaks, no longer buried under shaft harmonics.
Two practical consequences follow. First, the band-pass centre frequency has to be chosen per machine — it is a property of the bearing housing and the mounting, not of the bearing. Second, because the resonance amplifies weak impulses, envelope analysis finds a defect long before the overall vibration level moves at all, which is exactly what makes it worth doing.
The anti-alias filter has to track the sample rate
Classic FFT analysers set their anti-alias filter corner at the sample rate divided by 2.56, and that constant is worth remembering: it is what makes an analyser's displayed span honest. The reason the constant is not simply 2 is that a real analogue filter rolls off gradually, so the corner has to be placed low enough that everything above Nyquist is already well attenuated.
On the USB-4431 and USB-4432 that filter tracks the sample rate automatically, which means changing the span cannot silently invalidate the spectrum. On a general-purpose module there is no such filter, and the responsibility moves to you: keep the sensor's usable bandwidth below Nyquist by choosing the sensor and any external conditioning accordingly, or accept that high-frequency content is being folded down into the band you are trying to interpret.
Phase coherence: the reason a multiplexed card cannot do modal work
On a multiplexed module the converter visits one channel at a time, so two accelerometers on the same machine are never sampled at the same instant. For overall level trending this does not matter. For anything that depends on the *relationship* between channels it is fatal: operational deflection shapes, modal analysis, order tracking across a speed ramp and cross-correlation all assume the channels were measured simultaneously.
The USB-4431 and USB-4432 put a converter on every channel, so all inputs are sampled at once at the full 102.4 kS/s. That is the property to buy when the answer depends on phase, and it is not a property any 62xx module has, at any sample rate.
Mounting decides your upper frequency limit
The most carefully specified sensor in the world is limited by how it is attached. Every mounting method has a resonance of its own, and above that resonance the reading is no longer a measurement of the machine.
| Mounting | Typical usable range | Where it is appropriate |
|---|---|---|
| Stud or bolt | Highest — limited by the sensor itself | Permanent monitoring points, envelope analysis, anything above a few kilohertz |
| Adhesive | High, but depends on the bond and the surface | Survey work where a stud is not practical |
| Magnet | Often around 7 kHz and below | Quick route-based checks at low to medium frequency; not for envelope analysis |
| Probe or tape | Low | Confirming that something is running, not measuring it |
Swipe the table sideways to see all columns
The other half of the mounting problem is mass loading. The accelerometer and its mount add mass to the surface they are attached to, and once the sensor approaches a tenth of the effective mass of the structure at that point, the structure's response changes because of the measurement. On a thin panel or a small bracket, that is easy to do by accident.
Three parts, three jobs
- USB-4431 — four 24-bit IEPE inputs at 102.4 kS/s each, sampled simultaneously, with one 24-bit analog output. The output is what lets the same instrument drive a shaker or an exciter while it measures the response, which is what turns a monitoring setup into a test setup.
- USB-4432 — five 24-bit inputs of which four carry IEPE conditioning, on a ±40 V range, with no analog output. The wider range and the fifth channel suit high-level measurements and pressure or force channels sitting alongside accelerometers.
- USB-6212 — sixteen 16-bit inputs at 400 kS/s with 32 digital lines, for slow multi-point surveys where an external IEPE conditioner supplies the excitation. It is multiplexed and has no IEPE conditioning, so it is not a substitute for the two above in phase-dependent work; it is the right part for scanning many measurement points at low frequency.
Mistakes worth checking for first
- Reading a clipped positive peak as a mechanical impact when the excitation supply has simply run out of headroom.
- Running envelope analysis off a magnet-mounted sensor, so the resonance band being demodulated is the magnet's, not the bearing housing's.
- Comparing channels on a multiplexed card and attributing the time offset to a real phase difference.
- Setting thresholds from a textbook instead of trending one machine over weeks.
- Treating a 24-bit specification as a noise-floor guarantee without checking the front end.
Every figure in this note is either arithmetic from published resolution and sample-rate specifications, or a value stated by the standard it cites. Where a number depends on a particular module, confirm it against that module datasheet before design freeze.