correct FFT settings Diadem

[Else]

Hi,
I am a new user with Diadem (version 10), and use the program to view and analyze EEG data. In order to calculate a powerspectrum from parts (ca. 5 seconds long) of my EEG data (sampled with 5000 Hz) I tested the analysis FFT function in Diadem for one-time signals. Being unexperienced with the FFT details, I have trouble chosing the correct settings:
 -type of window (Hanning ?)
-amplitude damping correction (none/periodic/random?)
 -FFT functions: should be amplitude I guess. Averaging arithmetic? (I also tried exponential, which did not look too bad, but I guess this puts more stress on the latest calculated parts..). Then calculating the average from amplitude ?(other option is autospectrum)
-time intervals: this of course is linked to my sample size (e.g. 5 sec) and sampling rate (5000 Hz). With my 5 sec of 5000 Hz I believe I can get up to large FFT sizes like 1024 or 2048. But what about the overlap, what are the rules for this? I guess too much overlap will distort the frequency spectrum, but what should be chosen? Reject no values sounds good, but if this is the choice, why?

 In the tests with the FFT I did, the 0 Hz peak is dominant. Can one choose the settings in a way that gets around this, or should one put a high-pass filter on the data to get rid of it?

Obviously ignorant, perhaps some expert can give me advice how to make the best powerspectra before I create non-sense. I checked the forum and examples in diadem on this, but I do not seem to find a clear explanation on the options that diadem provides.

Regards Else

[Ulrich_Bierwisch]

Hi Else,

there is no correct setting for FFT calculations. You can get a lot of different results and most of them are making sense. On the first view, this might look very complicated, but most of the time, the settings are only important if you try to compare your results with the results of other FFT Analysers or software. If you just need to find out more about your data, you can try out different settings and look for the results you need.

Here are some answers to your questions.

Your signal is cut of at the begin and the end of the measurement which results in an error. Most of the window functions are reducing this error. Hanning is the most common window. You should compare Hanning and Rectangle (no window) to see which one is giving you the better result. Forget all the other windows if you don't know them.

Hanning is taking a part of your signal away. In the first place, all FFT-values are smaller then the results without a window. It is possible to multiply the results with a correction. If you are looking for a sinus wave within your signal and you are interested in the amplitude of this sinus, you should use the periodic correction. In all other cases you should use the random correction. But the correction is just a factor. If you are not interested in the absolute value of the amplitudes at all, it doesn't matter.

FFT functions:
It is always the same result, but sometimes squared or multiplied with a factor.
Amplitude is OK, unless you compare your values with other results.
RMS = Amplitude/square root(2)
Autospectrum = Amplitude^2
Powerspectrum = RMS^2


Averaging is only important if you do more than one FFT calculation. In the first place, I would do one FFT as large as possible. If your signal is a continuous signal, you get better results if you take as much values as possible (4096, 8192, ...). Larger FFTs are giving you a much better frequency resolution. Most of the averaging and overlapping is not useful anymore. We have so much memory and speed in our computers, that we can do calculations with millions of values. Today averaging is still very important for 2-channel calculations like transfer functions on numbers of short measurements.

DIAdem can handle datasets with gaps where the equipment was not able to deliver any value. "Reject no values" is not important for you, because you don’t have "no values" in your measurement.

If 0 Hz is the dominant frequency, using a high pass filter is a good way to get better results.

This are just some short answers to your questions. Signal processing and FFT is a very wide area with lots of things you might need to know. If you need more specific help on analysing your data you can post a set of these data or send it by mail. It also would help to know what you are trying to find out.

regards,
Ulrich Bierwisch
ulrich.bierwisch@ni.com