Vibration signatures: what the accelerometer hears that the microphone can't

Every phone carries a sensor that most people only meet through screen rotation: the accelerometer. It samples the phone's own motion hundreds of times a second, in three axes, and it does so whether the phone is being waved around or lying perfectly still on top of a washing machine that is trying, in its quiet mechanical way, to tell you about its bearings.

The Vibration tool turns that sensor into an instrument: set the phone on a machine or a surface, record a stretch of low-frequency motion, and get a picture of how the thing under it shakes. It is the touch-based sibling of the sound spectrum — same idea, different physical channel — and the channel matters more than it first appears.

What an accelerometer sees that a microphone doesn't

A microphone hears sound that made it into the air. That's a real limitation, twice over. First, air is a poor reporter of low frequencies: the slow, heavy shaking of an unbalanced drum or a settling floor radiates into the air weakly or not at all — much of the mechanical story sits at frequencies you feel through your feet rather than hear through your ears. Second, the air mixes everything: the compressor you care about, the traffic outside, the conversation in the next room.

An accelerometer in contact with a machine reads structure-borne vibration — motion that traveled through the machine's own body into the phone resting on it. The slow frequencies arrive at full strength, and the sensor is nearly deaf to airborne noise, so the recording is the machine's, not the room's. Where the microphone excels at hums and whines, the accelerometer owns the territory below them: the shakes, wobbles, and rumbles.

In the frequency domain, rotating machinery is as legible here as it is in sound. A drum spinning at a steady rate puts a peak at the rotation frequency; an unbalanced load makes that peak grow; bearing wear tends to add energy at higher multiples and a general roughening of the signature. You don't need to know which pattern means which fault to benefit — you need the signature from when the machine was healthy, and the willingness to compare.

Where to put the phone

The use cases are as domestic as measurement gets:

• The washing machine. The classic. Set the phone on top mid-cycle: an unbalanced load and worn bearings both make it shake, but they shake differently, and both drift over weeks in a way memory can't track.
• The fridge. A compressor has a signature; a compressor working too hard, or one whose mounts have aged, has a different one.
• The 3D printer. Frames loosen, belts wear, and print quality quietly degrades. A vibration signature during a standard move pattern is a health check that takes a minute.
• The car at idle. Phone on the dashboard, engine idling — a roughening idle shows up as a change in the signature long before it becomes a noise.
• The wobbly table, the trembling floor. Sometimes the question is about the structure itself: how much does this floor actually move when a truck passes? Now there's a number instead of an argument.

Placement beats phone model

Here's the practical rule that decides whether your measurements are comparable: where and how you place the phone matters more than which phone it is.

The reason comes from how the features are built. Different phones have different accelerometers with different sensitivities, so the absolute magnitude of a reading is partly a property of the sensor — which is exactly why the analysis doesn't lean on absolute levels. What it leans on is the structure of the vibration: where the peaks sit, how they relate to each other, how the signature is shaped. Those are properties of the machine, and they survive the trip across phone models. This is the same design rule that runs through every tool on the platform — relative, excitation-robust features, never absolute levels — and it's what makes readings meaningful without per-device calibration.

Placement, on the other hand, changes the physics of the measurement itself. A phone on the machine's corner rides a different motion than one at its center; a phone on a folded towel is measuring the towel; a phone pressed flat against a rigid panel is coupled, while one rattling loosely on top is adding its own percussion. So the discipline is simple: same spot, same orientation, firm flat contact, machine in the same operating state — mid-spin compares with mid-spin, not with fill.

Signatures over time: the catalog does the remembering

This is where the platform's structure carries the load. Each capture files under an object in the catalog — this washing machine, that printer — and repeated measurements stack up under it in order. The question "is it worse than it was?" stops depending on anyone's memory of how it felt three months ago; the answer is two signatures, side by side, on the same screen.

And when the story resolves — the repair tech opens the machine and finds the worn bearing, or the terrifying wobble turns out to be a lost sock — the measurement can take a ground-truth label after the fact. A vibration signature plus what it actually turned out to be is a labelled sample from a real machine on a real phone, and the collection of those, across many machines and many users, is what any future automatic diagnosis has to be trained on. The measurements you take to answer today's question are the dataset for tomorrow's.

Everything runs on-device: the analysis happens on the phone, offline, and nothing about the measurement requires a network. The accelerometer also has an active sibling — instead of waiting for a machine to shake, the phone can do the shaking itself with its vibration motor and read how an object answers; that's the haptic response tool. For the full range of what phone sensors can measure, start with the tools overview.