Re: [Discuss-gnuradio] FFT of FFTs

[Frank Brickle]

Eric Blossom wrote:

> Once you've got your individual streams of signals, for each one I
> would compute an estimate of whether it is occupied.

Good suggestions.

Elaborating on this a little, you might consider taking 
differences of successive magnitude or power spectra. If 
you're actually looking for periodicities in the onsets and 
releases, the resulting positive and negative pulses are 
very easy to track. This is a standard technique for 
first-order activity detection in spectral measurements.

Another productive technique is to take the wavelet 
transform of the magnitude, power, or difference spectrum, 
especially when the signals of interest are smeared over 
several bins. In this case you are looking for large wavelet 
coefficient magnitudes in the dilation region of the wavelet 
transform that corresponds to the spreading bandwidth in the 
 original spectrum.

There is yet another arsenal of methods to throw at this 
problem if you are in a position to keep a full tableau of 
past spectra in three-dimensional form. For example, a 
conventional histogram equalization on such a tableau of 
past spectra is quite effective at bringing "activity 
stripes" out of the noise, and there is a robust likelihood 
score comparing before- and after-equalization densities. 
But that's really getting ahead of the subject...

Frank