Fourier amplitude of a frequency spectrum depends on the value of Nfft. If we take Nfft as 1024, the Fourier amplitude of fundamental frequency is 22.7 and if we take Nfft as 100000, it is 50.9. What is the optimum value of Nfft ?
Can you please explain what Nfft is in your case, is it the length of your timeseries? Also do you perform a sliding window analysis and average the spectrum after?
By the 3 images you have supplied it seems that you are using a sliding window FFT which gives you an average spectrum of a timeseries and with Nfft you vary the length of this sliding window.
Nfft works like the width of the classes in a histogram. You choose. If you choose Nfft to be too big/wide, then the classes are too narrow and you have a 'noisy' frequency representation, if you choose Nfft too small/narrow, then the width of the classes is too wide and you get an over-smooth frequency representation. Think it like this: You want to represent the estimated age distribution of students in a class of 200 first year students at University. A width of a week is too narrow, A width of a year is too wide.
Your amplitude will increase if you are using an unscaled FFT. Moroever, as you vary NFFT (and keep the sample rate of your signals the same) the frequency resolution of each bin changes and your fundamental may sit slightly differently in each bin. The variation of power in your fundamental using an unweighted FFT is ~3dB (look up scalloping loss).
- Badges
- Science topic