0
$\begingroup$

I have an acceleration measurement for a full day with sampling rate 10.24 Hz. I have divided the signal into 30 minute intervals because I want to compare the amplitude spectra for each 30 minute interval to other data I have for 30 minute intervals of the day.

I have not used any windowing, overlap or derating of the time signal before applying DFT by the fft functions from numpy (Python).

As a quality check, I want to calculate RMS for the time domain and RMS for the frequency domain, because my understanding is that they should be approximately equal. So questions are:

  1. Should time signal RMS and amplitude spectra RMS be approximately equal?

  2. If so, will amplitude spectra RMS always be smaller than time signal RMS?

I have calculated RMS for both, and found that there are larger variation in time (RMS decreasing/increasing) for the amplitude spectra RMS compared to the time signal RMS.

  1. Assuming 1 is correct, should the two RMS calculations have similar trends (increase/decrease), or will frequency domain RMS vary more depending on the what my DFT/FFT can detect with my sampling rate of 10.24 Hz?

Thanks for reading.

$\endgroup$
5
  • 1
    $\begingroup$ From Pareseval's theorem, the energies in the time and frequency domains will be equal. $\endgroup$
    – jojeck
    Oct 27, 2022 at 8:02
  • $\begingroup$ @birki Are you using complex or real FFT? If real: is your input even/odd length, and are you doing anything special for bin 0 and Nyquist when you calculate your RMS? $\endgroup$
    – cloudfeet
    Oct 27, 2022 at 8:24
  • $\begingroup$ @jojek strictly speaking, and also very practical, because not all FFTs are unitary, proportional, not identical but yes. $\endgroup$ Oct 27, 2022 at 8:36
  • $\begingroup$ @cloudfeet real FFT, one-sided. Even, length 18432. Not doing anything for the 0 bin and nyquist. The amplitude spectra RMS is smaller, so I assume something is wrong. The 0 bin amplitude and nyquist amplitude should be halfed in the RMS calculation, or does that depend on if the length of the input is even or odd? $\endgroup$
    – birki
    Oct 28, 2022 at 8:57
  • $\begingroup$ @jojek Is this the case for random vibration signals as well? Or signals with noise? WIll they be exactly equal? $\endgroup$
    – birki
    Oct 28, 2022 at 9:20

1 Answer 1

1
$\begingroup$

If you use proper DFT scaling, then Perceval's Theorem holds and RMS values in time and frequency will be identical, i.e.

$$\sum |x[n]|^2 = \sum |X[k]|^2 $$

This equality answers your other questions I think. The scaling you need to deploy is $\frac{1}{\sqrt{N}}$ for both the forward and inverse transforms

$\endgroup$
2
  • $\begingroup$ Is this the case for random vibration signals as well? Or signals with noise? Will they be exactly equal? $\endgroup$
    – birki
    Nov 2, 2022 at 14:01
  • $\begingroup$ There's no such notation as signal, noise and randomness in this theorem. You have an set of numbers as a time series, no matter there it came from. Computationally there is an exact RMS number, and an exact FFT. This theorem states that the FFT will have the same RMS as the time series. $\endgroup$
    – Michael
    Aug 30, 2023 at 5:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.