Adaptive filtering / QUESTION ?

[likehim]

Hi all,

There is a little detail I 'm facing up within
signal processing application.
OK, the problem is to filter a signal, BUT with an adaptive algorithm.
The signal is not stationnary. I have already tested a lot of stuff :
standard numeric filter, time-frequency filtering (i.e wavelett based
denoising) etc BUT none of them gave satisfactory results.
I WANT TO TRY NOW ADAPTIVE FILTER.
Theory is nice, simple to understand, BUT I have a big problem
what shall I use as "desired signal", since I have only one the one
I want to filter ??
Thx in advance.
PS = The noise cannot be considered as Gaussian.

 

[brmadhukar]

Normally the desired signal is a pilot signal that is know before hand and as such is used to calibrate the coefficients. Check if there is possibility of finding the response to known inputs and use this. This depends on your application.
B R M

 

[fxiao]

Hi,
there are several situations in adaptive signal procesing. Your situation is source and channel are both unkonwn and you need to estimate the channel from the output signal. You definitely need to rely on the statistics characteristic of the source. If it is not stationary, is it cyclostationary? Hope this can give you some clue

 

[dsp4us]

First of all you have to be more specific. None stationary signal with none Gaussian noise is to broad description. What do you want to get after filter? If it is just signal without noise then you have to build model for that none Gaussian noise. Is it speech with impulse noise (clicks)? Is it two persons talking and you are trying to filter on out? I am really don’t think adaptive filter will help you much in those cases.
Cheers.

 

[nicleo]

Try Modified Probabilistic Neural Network by Anthony Zaknich. The MPNN was used as adaptive filter in digital signal processing. You can download his PhD Thesis at **broken link removed**

 

[zorro]

Hi likehim,

I agree with dsp4us. You have to specify why do you need to filter the signal, i.e., which characteristics have to be modified. Do you want to eliminate some disturbance (noise)? In this case, which are the known characteristics of the good signal an those of the noise? Without a clear specification of the problem it is not possible to solve it.
Regards

Z

 

[the_jackal]

This is indeed an ambiguous question. You say you have a desired signal and you need to apply adaptive filtering. Whenever there is a requirement to process signals that results from operations in an environment of unknown statistics or one that is inherently non stationary an adaptive filter is used because it offers a better solution to the problem. For instance, in communication systems we have ISI (intersymbol interference and noise) which can be eliminated/mitigated using adaptive filtering algorithms.

 

[hemanthraz]

why dont you look up something on blind source separation? i had a similar problem way back, [i wanted to separate 2 signals] but the project quickly grew out of hand.
I later did implement an nLMS filter on the TMS 5416 though, worked somewhat ok.
The main problem i had was with the overflow issues.

good luck

 

[ishtiaq]

Dear likehim,

Have u tried Wigner Filtering. It is a time-frequency filtering used for non-stationary signals. You can get too many informations on Wigner-Ville distribution based filtering. And a lot of information can be found on the following web address:

http://case.caltech.edu/tfr/

BR,
Ishtiaq