A failing bearing announces itself long before it seizes — if you know how to listen. Predictive maintenance turns a machine's vibration into early warning, and the hard part is not the sensor but pulling the small signature of incipient failure out of the noise of a machine that is simply doing its job.
| Method | What it reveals | When it is used |
|---|---|---|
| Overall level (RMS) | Gross change in vibration energy | A crude first alarm; misses early faults |
| Spectrum (FFT) | Energy by frequency | Locating imbalance, misalignment, defects |
| Envelope / demodulation | Repetitive impacts | Early bearing and gear faults |
| Band energies | Trend in a chosen range | Watching a known fault frequency over time |
| Order tracking | Features vs shaft speed | Machines that change speed |
| Baseline comparison | Deviation from normal | Separating degradation from operation |
A healthy machine is not quiet. Under changing load and speed it produces a rich, variable vibration signal, and the early signature of a failing bearing or a growing imbalance is a small feature buried inside it, not an obvious spike. That is why a single overall vibration level is a poor early alarm: by the time the total energy has risen enough to trip a threshold, the fault is well advanced. The information about what is failing, and how early, lives in the frequency content, not the amplitude.
The physics is what makes diagnosis possible. Each fault mechanism excites vibration at frequencies tied to the machine's geometry — a bearing defect rings at frequencies set by its dimensions and speed, an imbalance shows at once-per-revolution, a gear defect throws sidebands around the tooth-mesh frequency. Spectral and envelope analysis expose those signatures, so a well-processed signal does not just say the machine is worse — it says which part is failing. But the fault frequencies shift with speed, so a machine that changes speed needs order tracking to keep the features from smearing.
None of this survives a poor sensing chain. Where the sensor is mounted, its bandwidth and its noise floor decide whether the small fault feature is even present in the data before any algorithm runs; a badly mounted sensor or too low a sampling rate loses the signature at the source. And because the baseline of normal shifts with load and speed, a fixed threshold either misses faults or floods false alarms — so the honest system models normal per operating point and trends features over time, catching a developing fault rather than reacting to one noisy reading.
Because it is a late alarm. Overall level rises only once a fault has grown enough to lift the total vibration energy, by which point the damage is well advanced. The early signature of a fault is a small feature at a specific frequency, hidden inside a signal that is already rich and variable, so catching it early means analysing the frequency content — the spectrum and the envelope — not the amplitude alone.
Because each fault excites vibration at frequencies tied to the machine's geometry and speed. A bearing defect produces energy at frequencies set by its dimensions, an imbalance appears at once-per-revolution, and a gear defect throws sidebands around the tooth-mesh frequency. Spectral and envelope analysis expose those signatures, so a well-processed signal identifies the failing component rather than just reporting that the machine is worse.
Because if the fault feature is not present in the captured data, no algorithm can recover it. Where the sensor is mounted, its usable bandwidth and its noise floor decide whether the small, early signature survives to be analysed. A poorly mounted sensor or too low a sampling rate loses the signature at the source, so faithful sensing is a precondition for the processing to work at all.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.