Using Sound to Detect Impending Defects

I am intrigued by the use of sound to detect things that are about to break down. Examples where a “listening device” might be put to use:

• If the sound of a car’s engine starts to change over time it might be an indication that a gear, belt, tube, or bearing starts to develop an issue due to wear, scraping, deformation, lack of lubrication, et cetera.
• Inside a laptop a change to its internal sound signature could be an indication of dust accumulation or a bulging battery causing external sound to enter the case.

It is not a novel idea, for instance there is a patent that describes this idea to detect bearing defects in trains:

An apparatus for the detection of impact frequencies in moving railway train bearings which are characteristic of bearing defects. The apparatus transduces acoustic vibrations of the bearings into an electrical signal and extracts an envelope modulated carrier frequency component from the electric signal. The envelope is extracted from the carrier frequency component and is analyzed by a bandpass filter. The bandpass filter includes a switch capacitor filter controlled by a master clock which, in turn, is controlled by train speed and/or direction sensors positioned along the track on which the train travels.

Related to this, SpaceX engineers were are able to triangulate a failed strut based on audio recordings inside the upper stage:

“We’ve got microphones, technically accelerometers, at various points on the upper stage, and by looking at the exact timing of high-frequency events on the stage, we can, by acoustic triangulation, identify the location where the snap occurred or the breakage occurred via sound.”

The use of this idea might see an increase in the future when it becomes easier to fit devices with tiny listening devices, creating a massive baseline of sound signatures, making it easier to detect anomalies.

Obviously there are privacy issues related to listening devices close to conversations. I don’t know enough about audio frequencies to determine if filtering out the voice frequency used by human speech (between ~85 to 255 Hz) still yields meaningful data to detect impending defects.