I need to classify an EEG signal (most likely using an artificila neural network) but for the inputs I would like to know the differences between using the result of the FFT or using the signal with different band-filteres for example butterworth.
The advantages of the FFT is that is much smaller (a 1 second window it will have ~60 inputs) and that the signal is still there. On the other hand using band-filters I would have as input all the signal (a 1 second window @ 256 Hz it will have 256 times the number of bands that I want to obtain). Also the signal gets distorted.
So my final question is, for classification purpouse, does the FFT contains the same information that the signal filtered?
Dear Jacob,
Hope you are well,
I may came late to this question but i found it is related to the basic principles of signal processing.
The filters would output discrete frequency bands according to the frequency ranges of the filters used. Filter would take much computation time than the FFT.
In fact FFT is a sampling process of the signal in frequency domain. So, i would propose that you use the method which gives you the sufficient signal features that you increase your identification. The filter bank also is frequency analysis method. But the FFT is a systematic method to get the frequency components of the signal. I believe the FFT is an elegant filter bank.
However you also run an experiment to assess these method experimentally.
Best wishes
FFT in the end provides a "set of coefficients", that correspond to the amplitude and phase of the DFT basis functions. Filter-banks however provide you with the "output signal" corresponding to each of the filter. Actually, filter-bank must be able to get your signal back without distortion if you construct them properly. Search the term, "perfect reconstruction filter banks". So, in the end both techniques do analysis. Obviously, both techniques should retain all the information of the signal, but for the EEG signal, you must know:
1. what are the important frequency bands (e.g., [0 - f_0 Hz], [f_0 Hz - f_1 Hz])
2. Are these bands of interest equal (same bandwidth)
And then use any technique. In order to reduce number of inputs to the filter-bank approach you might consider the variance of a block of output signal.
Not sure if this helps but the FFT or DFT is based on the assumption that, the signal to be characterized, can be represented by a basis set of periodic, time-stationary sinusoidal signals. I don't know if EEG signals meet this criterion. If you are wanting to develop 'signature functions' for an EEG, you can also try the Fourier Transform (FFT) of the autocorrelation function of the EEG signal. It is, in some respects, equivalent to the direct FFT of the EEG signal but it depends on the stationarity of the EEG signal. Good luck.
If you want to characterize a low frequency signal you have to take into account just to remove "unwanted" parts of your signal. This filter has to be designed carefully and often "good" parts of your signals will be lost.
I would prefer using FFT for characterization, because all frequency components still remain in frequency domain and it`s up to you to "ignore" unwanted parts of your frequencies.
Dear Jacob,
Hope you are well,
I may came late to this question but i found it is related to the basic principles of signal processing.
The filters would output discrete frequency bands according to the frequency ranges of the filters used. Filter would take much computation time than the FFT.
In fact FFT is a sampling process of the signal in frequency domain. So, i would propose that you use the method which gives you the sufficient signal features that you increase your identification. The filter bank also is frequency analysis method. But the FFT is a systematic method to get the frequency components of the signal. I believe the FFT is an elegant filter bank.
However you also run an experiment to assess these method experimentally.
Best wishes
- Badges
- Science topic
- Similar topics
- Biological Science
- Physiology
- Biosignals