About Services Tools Blog Contact us

Machine Audio Fault Monitor

A proposed sensor that spots machine faults from noise or vibration on the device, with a plant screen showing each fault and who has acknowledged it. Not built yet.

  • Audio Analysis
  • Manufacturing
  • Dashboard feature
  • Bespoke
  • Hybrid
  • 2025

The problem

Bearing and imbalance faults are heard late, because someone has to stand by the machine with a meter. Even a good reading is no use if maintenance cannot see which machine it came from. They also need to know how sure the call is and whether anyone has acknowledged it.

What it does

The design puts a microphone or vibration sensor on a small wireless board. The board summarises a short sample and labels it normal, bearing fault, imbalance, misalignment or looseness. A plant screen would store those readings, rank the alerts and keep a 24-hour view. Nothing has been built yet, so this is a design and not a running system.

Outcomes

  • A fault label and a confidence figure arrive without raw audio leaving the site.
  • One list of machines shows unacknowledged alerts ranked by severity.
  • A sound chart and a device-health panel show for the machine that was opened.
  • A data export lets later labels improve the model on the device.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Rookery press hall

1 unacknowledged

Design sample. No board is posting. Press 2 plays a synthesised test-bench clip, not a machine.

North bay press normal
Confidence 97% Battery 81% 36 C no alert
South bay press Press 2 unacknowledged
bearing fault Confidence 86% Battery 74% 41 C 09:12
Pack line imbalance
Confidence 71% Battery 63% 44 C acknowledged 9 Oct

South bay press

unacknowledged
Test-bench clip synthesised

Spoken label, then a normal hum, then the same hum with a fault tone. Not captured from a machine, and not a reading from the chart below.

Spectrum drawn sample, not measured from the clip
0 Hzpeak 187 Hz2 kHz
Confidence
0
Battery
0
Temperature
0
Feature vector
peak 187 Hz centroid 420 Hz rolloff 2.1 kHz zero-cross 0.08 cepstral peak 3
Sensor bearing fault

Wireless board, microphone on the frame. Class labelled on the device.

Battery 74% · last post 09:12 · piezo channel idle

Shift log
Acknowledge does not clear the fault class

Training export

Last 24 hours class counts
Labelled rows
TimeMachineClassConf.Ack
09:12South bay pressBearing fault86%unacknowledged
08:40Pack lineImbalance71%acknowledged
07:15North bay pressNormal97%none
06:50North bay pressNormal96%none
The file keeps the class, not the waveform

How it works

  1. Sample machine noise or vibration on the board.
  2. Summarise the pitch and tone pattern of the sample.
  3. Label the sample on the device and send a short reading with battery and temperature.
  4. File the reading against a machine and open an alert when the result leaves normal.
  5. Let an operator acknowledge the alert, and export labelled rows to improve the model later.

Pipeline

flowchart LR
  Sensor[Mic or piezo] --> Sample[On device sample]
  Sample --> Feat[Feature vector]
  Feat --> Class[Fault class]
  Class --> Post[Reading posted]
  Post --> Dash[Plant list]
  Dash --> Ack[Acknowledge]

Sample output

Synthesised test-bench recording: a spoken label, a normal hum, then a hum with a fault tone. Not captured from a machine.