MATLAB Answers

Power Spectra estimation after FFT

조회 수: 9(최근 30일)
Luca Merolla
Luca Merolla 2020년 7월 12일
편집: LO 2020년 7월 14일
Hello guys!
I am working on a frequency analysis for skin conductance. The frequencies expected on those kinds of signals are very low (between 0.045 and 0.25 Hz), so I would like to get the PSD in specific low-frequency ranges, i.e. (0 - 0.045), (0.045 - 0.15), (0.15 - 0.25), and the total power of the signal. To do so, I computed the Fast Fourier Transform, but now the problem is computing the power spectrum after the fourier.
The tutorials from Matlab have not been of much help, and also there are so many functions that can be used that I'm feeling a little lost at the moment! Can anyone help me work out this issue? Among the many options, which one do you think works best, in terms of both efficacy and computational time? I'm not looking for anything too fancy really, just an efficacious and standard protocol.
The processing pipeline I am following is:
  1. 8th order Chebyshev Type I lowpass filter, with 0.8 cutoff
  2. Downsampling from 4 to 2 Hz (factor = 2)
  3. 8th order Butterworth highpass filter with 0.01 cutoff (do you think is appropriate, given that I'm mostly interested in low frequencies?)
  4. Welch's periodogram (50% overlap)
  5. Blackman window (128 points) applied to each segment
  6. FFT computer on each windowed segment
I also attached the code because it would be amazing to have examples or suggestions about how to proceed based on real examples! Any advice is always appreciated!
Thank you very much to any of you who is willing to help!
Luca
  댓글 수: 4
Luca Merolla
Luca Merolla 2020년 7월 13일
No worries, it's no secret!
So this is a skin conductance signal acquired with the Empatica E4 device (sampling frequency is 4 Hz), the tag lines are the beginning and end of a baseline acquisition, and the sharp spikes are most probably artefacts related to electrode drifts. All the filtering is computed offline, as it is a wearable device. The energy content I am expecting to find is very low, in the range of 0.045-0.25 Hz, related to sympathetic activation (like in the low frequency HRV analysis). Indeed, the aim of this analysis is to find features in the signal related to arousal and sympathetic nervous system activity: if that particular frequency band (0.045 - 0.25 Hz) holds the most of the power of the signal, that we can be pretty sure there is a sympathetic activity going on in the subject. This is what I read in the lietrature. Given the features of the signal, a time-domain analysis is not suitable because there isn't much peak activity, so I chose to focus on a frequency analysis.
I hope this can help to understand a little more, if you have more questions I am happy to answer! Thank you very much for your time!

댓글을 달려면 로그인하십시오.

채택된 답변

LO
LO 2020년 7월 13일
try this example:
sampling_freq = 5000
window = 8192,
noverlap = 4096,
nfft = 8192,
[p,f] = pwelch(your_signal, 8192,4096,8192,5000);
max_power=10*log10(max(p)); % this would be the max power
plot(f,10*log10(p)); % this would be the power plot
  댓글 수: 4
LO
LO 2020년 7월 14일
Well if my answer was working accept it 😊 For the rest it really depends on your data, the length of the segment you analyze. The frequency of the signal is irrelevant. However if it is a signal with frequency modulations (changes) then you want to use a nfft higher than 2^10 in order to better resolve frequency than time. Conversely if you want a better time resolution go lower. Regarding the other params it depends on the "data points" your signal is composed by. I suggest to get an idea, search for A Wikipedia page or try on your phone a spectrogram app or simply audacity on your computer and see how the different settings affect your spectrogram/fft analysis

댓글을 달려면 로그인하십시오.

추가 답변(0개)

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by