I'm trying to wrap my head around the proper use of a Wiener or error-prediction filter for filtering data. It seems to me that it is only a whitening filter, so how is it used when the data you want to recover isn't an AWGN signal?

For instance, I have a signal that has several distint interfering signals - I can see them on a PSD, but I don't know that they are a) stationary and b) what properties they have. I can use a method like Yule-Walker equations to recover the AR model for the whole signal, but in this case I only want to recover the model of the interfering signals, not the portion I want to recover.

I tried implementing an adaptive LMS notch filter, with the reference signal being a single sinewave, but this turned out to me much too narrow and didn't track frequency changes in the signal very well.

I guess basically my question is this, if I'm using an error prediction filter to filter real data, then how do I separate the data portion from the noise portion? In other words, I don't want to whiten the whole signal, only the noise portion. What am I missing?

  • $\begingroup$ +1 Good question. Can you give some more details about your application and signal you are dealing with? $\endgroup$ – Dipan Mehta May 18 '12 at 11:27

i am not sure i understand correctly the question (feel free to update me if not so).

There is the MUSIC algorithm, which extracts signals embedded in a background noise, as a sum of sinusoidals signals

There is also the option to use SVD (or Karhunen-Loeve transform) and reduce the dimensionality of the input data while retaning maximum information (this will discard most of the background noise components).

If this is online or real-time, this could be done adaptively.

Hope this helps

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.