How it works
The Nyquist theorem says you must sample at more than twice the highest frequency in the signal to represent it. Anything above half the sampling rate folds back as a false lower frequency (aliasing), so an anti-aliasing filter should remove it before the ADC.
In practice you sample faster than the theoretical minimum. Vibration and FFT analysers commonly use 2.56× the maximum frequency, and capturing waveform shape in the time domain often needs 5–10×.
f_min = 2 × f_max
f_s = k × f_max (k ≈ 2.56 for FFT analysis)
Data rate = f_s × bits × channelsWorked example
To analyse vibration up to 5 kHz, the theoretical minimum is 10 kHz. At the usual 2.56× that's 12.8 kHz. Three 16-bit axes then produce 12,800 × 16 × 3 = 614.4 kbit/s, about 4.6 GB per day if streamed raw, which is why edge devices compute features instead.