Bearing Fault Detection by Sound

Rolling-element bearings are in almost every rotating machine: motors, fans, pumps, conveyors, gearboxes. They also fail in a familiar way. A small defect on a race or a ball grows, the bearing gets noisier and hotter, and if nobody acts, it eventually seizes or breaks apart. Somewhere along that path it becomes audible.

That makes sound a reasonable first check for bearing faults, and a poor last one. This post covers what bearing faults sound like in general terms, why a recording at 16 kHz cannot catch the earliest stages, what the published results for Saina's mechanical specialist show on bearings, and when vibration analysis or ultrasound is the better tool. For the broader picture of acoustic screening, see the pillar on machine fault detection from sound.

What a failing bearing sounds like

Descriptions vary between machines and between the people describing them, so treat these as rough guides:

• A healthy bearing gives a steady, even sound that blends into the rest of the machine.
• A damaged race or rolling element tends to add a repeating tick, click or knock as the balls roll over the defect, or a rough rumble once damage spreads. The repetition rate follows the shaft speed.
• Lubrication problems are often described as a squeal, a whine or a dry, harsh sound.
• Late-stage damage gets louder and coarser: grinding, rattling, and often heat you can feel on the housing.

A change from the machine's usual sound matters more than any one description. That is why a recording from the same spot each time, compared with earlier ones, is worth more than a single clip.

Why 16 kHz limits early detection

Saina Helm 2 Omni, and the mechanical specialist that runs on it, read each clip as a 16 kHz log-mel spectrogram. At that sample rate, sound above 8 kHz is not in the input at all.

For bearings this matters. The specialist's card notes that some early bearing defects first show above 8 kHz. The tools built to catch bearings early work in exactly that region or at frequencies far above it: ultrasound instruments listen well above the range of human hearing, and vibration analysts use techniques such as envelope analysis to pick out the repeating impacts of a small defect before they are audible.

The practical consequence: a phone recording scored by this model is more likely to catch a bearing fault once it is audible to a person standing next to the machine than in its first stages. That is still useful on machines nobody is watching. It is not early warning in the sense vibration and ultrasound programmes mean it.

What the mechanical specialist says about bearings

The specialist can answer two questions about a bearing recording: whether the machine sounds normal, and which condition it most likely has, from inner race, outer race, ball or cage faults, plus healthy.

Two of its seven training sources are bearing sets. Its card reports these held-out results for them:

Across all seven sources, each counted once, the card reports 88% for the specialist against 30% for the base model, on 1,165 held-out clips.

Read the bearing figures with their limits:

• Few bearings, lab conditions. One bearing in the stethoscope set and 20 units in UORED, recorded in a lab. A bearing in your gearbox, on a different mount, in a louder room, will sound different. Expect lower accuracy and check it on your own recordings first.
• Naming the fault is harder than spotting one. On the 20-unit UORED set, the specialist says normal or faulty correctly 91% of the time but names the condition correctly 64% of the time. Treat the condition answer as a hint for the technician, not a diagnosis.
• Bearings inside motors are a different case. For electric motors the specialist only says whether the motor sounds healthy. A bearing fault was one of seven motor faults in its training data, and they sound almost the same in its 16 kHz input, so it does not name them.
• Fans and pumps. On MIMII recordings of fans, pumps, valves and slide rails, machines outside its training data, the specialist detects faults worse than the base model (50% against 70%). If your bearings sit in those machines, test both.
• Nothing above 8 kHz, as above.
• A screening aid. It is not a substitute for inspection, vibration analysis or a qualified technician, and not for safety-critical shutdown decisions.

Asking about a bearing

The card's recording advice: phone held against or next to the housing, 1 to 10 seconds, machine running steadily. In the state, name the machine and how it runs, including the speed if you know it. Offer a healthy option and only the faults possible for that machine.

Illustrative; the shape follows the bearing example in the API docs, and the base64 string is cut short.

{
  "model": "saina-helm-2-omni-0.8b",
  "specialist": "mechanical",
  "mode": "decision",
  "threshold": 0.8,
  "state": "The clip is a recording of a rolling-element bearing on a fan shaft at 1,450 rpm, recorded with a phone held against the bearing housing.",
  "audios": ["data:audio/wav;base64,UklGRi..."],
  "questions": {
    "normal": {"type": "yes_no",
               "question": "Is the machine operating normally (no fault or anomaly audible)?"},
    "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",
                              "cage": "Cage fault"}}
  }
}

POST it to https://api.saina.run/v1/ask with your API key. In decision mode, an answer the model is not sure of comes back with no selection and a reason (below_threshold, below_margin or tie). Route a confident "not normal" to a technician, an unsure answer to a re-recording and a person listening, and log a confident "normal". Acoustic predictive maintenance turns that into a weekly routine across a list of machines.

When to use vibration or ultrasound instead

• You want warning weeks ahead, not days. Vibration analysis and ultrasound see bearing defects in stages a 16 kHz recording cannot. If the point is to plan a bearing change well before the noise starts, use them.
• The machine is critical. A seized bearing that stops a line, damages a shaft or puts people at risk justifies permanently mounted vibration sensors, or a condition-monitoring vendor's service.
• You need to know which component and why. A trained vibration analyst can separate bearing, misalignment, imbalance and looseness, and estimate how far a defect has progressed. The specialist names a likely bearing condition at best.
• Lubrication programmes. Ultrasound is widely used to guide greasing, which a phone recording does not help with.
• You have many identical machines. Recording their normal sound and training your own anomaly detector, or buying a vendor's, may outperform a general model on that machine.

Sound fits as the cheap first layer: machines that have no monitoring at all, recorded on an existing round, so that the noisy ones get a vibration reading or a technician sooner.

Try it

The home page's Machine fault example compares the base model and the specialist on a compressor recording, with no account (5 free runs). In the console, the Compressor fault preset takes your own bearing clip and options. The specialist's card, with the full evaluation and training data, is at run-saina/saina-helm-2-omni-0.8b-mechanical. Introducing Saina specialists explains how specialists work.

Related posts

• Machine fault detection from sound (pillar)
• Acoustic predictive maintenance: a practical routine
• Introducing Saina specialists