Hear the failing bearing before it fails

The step

Condition monitoring usually means accelerometers, a historian, and someone to build and maintain a model per machine class. It is worth it on the critical assets and never happens on the other ninety percent.

• Threshold alarms on vibration or current catch the end of a failure, not the beginning, and they need the sensor installed in the first place.
• Per-machine anomaly models need weeks of “normal” data for every unit and retraining after every overhaul.
• Listening by ear works, when the person who can do it is there.

Sound carries the signature of most mechanical faults, and a phone or a cheap microphone is enough to capture it. The specialist turns the clip into a scored decision you can put on a schedule.

What the request looks like

{
  "model": "saina-helm-2-omni-0.8b",
  "specialist": "mechanical",
  "state": "Ten-second recording of the drive-end bearing on conveyor
motor M-14, 1,450 rpm, phone 30 cm from the housing, normal load.",
  "audios": ["data:audio/wav;base64,UklGRi..."],
  "mode": "decision",
  "threshold": 0.7,
  "questions": {
    "condition": {
      "type": "single_choice",
      "question": "Which condition does the machine in this clip have?",
      "options": { "healthy": null, "inner": "Inner race bearing fault", "outer": "Outer race bearing fault", "ball": "Ball (rolling element) fault" }
    },
    "normal": { "type": "yes_no", "question": "Is the machine operating normally (no fault or anomaly audible)?" }
  }
}

And what comes back:

{
  "model": "saina-helm-2-omni-0.8b",
  "specialist": "mechanical",
  "answers": {
    "condition": {
      "type": "single_choice",
      "selection": "outer",
      "reason": "accepted",
      "probabilities": {
        "outer": 0.78,
        "inner": 0.12,
        "ball": 0.07,
        "healthy": 0.03
      },
      "confidence": 0.71
    },
    "normal": {
      "type": "yes_no",
      "yes": 0.04,
      "no": 0.96,
      "confidence": 0.92,
      "selected": false,
      "reason": "accepted"
    }
  }
}

Illustrative answer for a faulty bearing. The compressor example on the home page is a recorded one: the specialist hears the fault (no: 1.00) where the base model calls the compressor normal (0.72).

The yes/no question is the alarm; the single-choice question is the diagnosis. Score them together from the same clip. When the clip is noisy or the sound sits between two faults, the diagnosis falls back (below_threshold) while the alarm may still fire, which is exactly the order a maintenance team wants: “something is wrong, and we are not sure what” beats silence.

On held-out machine-sound sets, the Mechanical specialist lifts condition accuracy from 30% (Helm 2 Omni alone) to 88%. It is trained on bearings, electric motors, air compressors and 3D printers; see the model card for what that covers.

Putting it in place

Record the same way every time

Ten seconds, same distance, same load. Say so in the state: machine, speed, where the microphone was. The model reads the text and the audio together, and the context sharpens the decision.

Name the faults you can act on

The options are yours. Start with healthy plus the two or three failure modes your team actually repairs; a short, concrete option list scores better than a long one. Use the fallback to catch the rest.

Wire it in

A clip is a base64 WAV, FLAC, MP3 or OGG data URL in audios, up to four per request and ten seconds each (longer clips are rejected, not cut). Add "specialist": "mechanical" and the specialist answers instead of the base model; the response names both. In n8n, a schedule trigger, a Read Binary File and an HTTP Request node are enough; the Saina Helm community node is text-only today, so audio requests go over HTTP. To run it on the plant network with no internet: pip install 'saina[local,server,media]' and load the specialist weights; it is 0.8B and runs on a CPU.

What it won’t do

• Ten seconds per clip. Send several clips from one machine as separate requests, or up to four in one.
• It scores the conditions you list; it does not estimate remaining life or severity in hours.
• Machines outside the training domain (gearboxes, fans, hydraulics) may work; measure on a few known-good and known-bad recordings before trusting it.

Frequently asked questions

Read next