Vibration condition monitoring with edge processing

Vibration monitoring is often dismissed as too expensive or too complicated for a mid-size plant. Most of that cost is in transmitting and storing raw waveform data that nobody looks at. Processing at the edge changes the arithmetic substantially.

The constraints

• 01

  Raw vibration data at kHz sample rates cannot be streamed continuously

• 02

  A useful baseline must cover all normal operating states, not just one

• 03

  Fixed absolute thresholds produce either false alarms or missed faults

• 04

  Alerts that are not explainable will be ignored by maintenance teams

• 05

  The system must survive its own false alarms without losing credibility

The approach

1. 01

   Process on the device

   FFT and envelope analysis run on the gateway. Bearing defect frequencies, overall RMS, crest factor and band energies are extracted locally; the raw waveform is retained briefly and discarded unless an event captures it.

2. 02

   Baseline per operating state

   The machine's normal states are identified from load and speed, and a separate baseline is established for each — because comparing a loaded machine to an unloaded baseline is where most false alarms originate.

3. 03

   Trend before threshold

   Deviation from the machine's own baseline, and the rate of that deviation, are the primary signals. Absolute limits are a backstop, not the mechanism.

4. 04

   Explainable alerts

   An alert states which feature moved, by how much, relative to which baseline, and over what period — so a maintenance engineer can agree or disagree with it rather than simply trusting or ignoring it.

5. 05

   Record the outcome

   Every alert is closed out as true, false or premature, and that feedback adjusts the thresholds. Without this the system degrades into noise within a year.