1
$\begingroup$

The following is a plot of the FFT of a noisy electronic component. It's not white noise, but is stochastic in nature. I used Python's matplotlib plt.magnitude_spectrum(...) function. It was a 1MHz sample rate.

fft

What is the actual magnitude? Is it the top or the bottom of the purple? Or am I supposed to interpolate a mean through the middle? Or something else?

$\endgroup$
  • $\begingroup$ "FFT of a noisy component": um, I'd presume that's the FFT of a sampled voltage or current over or through that component, right? Can you define what you mean with "actual magnitude"? We don't really know what you're supposed to do, because we neither really know what we're looking at, nor what you really want, nor what you want to do with the result afterwards! Application context might really massively help here :) $\endgroup$ – Marcus Müller Nov 17 '19 at 18:12
  • $\begingroup$ @MarcusMüller Oh sorry. It's nothing complicated. I'm just wondering which part of an FFT you read. I'm unfamiliar with the Python implementation. Do you just fit a line through the middle of the purple? I'd like to be able to summarise it as eg. rolls off at 20dB/decade , or something... $\endgroup$ – Paul Uszak Nov 17 '19 at 21:25
  • $\begingroup$ The problem is that what you're showing us is a "crunched together" graph, and very far from the output of an FFT. So, you do neither, because as said, what you do is dictated by what you need, and you still haven't told us! $\endgroup$ – Marcus Müller Nov 17 '19 at 21:39
  • $\begingroup$ It's a log frequency plot so not as intuitive to look at. But if you take the direct FFT output and multiply it by the complex conjugate of itself you can get a power estimate per bin. You can then do various moving averages to access the average power across bins. Do not average the log magnitude output as that has other errors depending on the noise distribution from a true rms power. $\endgroup$ – Dan Boschen Nov 17 '19 at 22:08
  • 2
    $\begingroup$ If it makes you feel better call it "volt-squared spectral density" in your head, or "spectral density of the autocorrelation expectation value". But it's still called "power spectral density" in the literature, even though it's used in a lot of places where the word "power" has no physical meaning. The method suggested by @DanBoschen is one quick way of getting an estimate of PSD. $\endgroup$ – TimWescott Nov 18 '19 at 23:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.