Dismiss Notice
Join Physics Forums Today!
The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

Wiener deconvolution

  1. Sep 21, 2010 #1
    I have a system that acts as a low pass. I know the transient response quite well, and I am trying to do a Wiener deconvolution on the measurement data. I expect a clean signal with a few jumps and bursts. Basically piecewise continuous, with the occasional peak. For the Wiener deconvolution I need the spectral power density of the expected signal, and the spectral power density of the noise.
    The noise looks white falling a bit at the high frequencies. Is there a somewhat generic term to use for the spectral power density of a generic signal? Maybe 1/abs(f+epsilon). I could simulate a bit, but maybe someone here has used deconvolution and knows what kind of power densities tend to work well, and how sensitive the outcome is to this.
    Also: How well do these deconvolutions work, and what are the biggest problems?
  2. jcsd
  3. Sep 22, 2010 #2


    User Avatar
    Science Advisor
    Gold Member

    Wait. Are you saying you have an unknown signal x that you are trying to estimate from a filtered noisy version y = x*h + n? And you know the impulse response h (or frequency response H)?

    If so, you don't need Wiener's formulism, which attempts to estimate h and its inverse. Just apply h inverse to y (or H^(-1) to Y).
  4. Sep 22, 2010 #3
    Well, it's not quite as easy. The system lowpasses, and doing a deconvolution puts more emphasis on the high frequencies. But as signals tend to be dominantly in the low frequencies, most of the stuff that gets amplified is noise. So basically the H^-1 applied to N term is too large.
    The Wiener deconvolution is damping this effect by weighting the filter with the expected signal to noise ratio. I did a try with a 1/(1 + f^2) power spectrum for the signal but it does not look too good, I think I need more high frequencies.
    I am not sure, but it seems to me that by Wiener's formalism you mean blind deconvolution, where the impulse response is iteratively guessed.
  5. Sep 23, 2010 #4


    User Avatar
    Science Advisor
    Gold Member

    No, I was referring to Wiener deconvolution, aka Wiener filter.

    It sounds like you have a firm idea of the spectrum of your signal. You say
    1) it is predominantly low frequencies, below your filter cutoff
    2) a single pole rolling off at 1 Hz removes too much high frequency

    BTW, you have answered your question about how sensitive the results are to choice of signal PSD. It matters significantly.

    Try a response that corresponds to your understanding of the signal spectrum. Or measure the filtered spectrum directly--if you are right, you'll see signal predominantly at low frequencies, then noise, then the filter rolls everything off. Fit a spectrum to it and use that in the Wiener deconvolution.
Share this great discussion with others via Reddit, Google+, Twitter, or Facebook

Similar Threads for Wiener deconvolution Date
Non-causal wiener filters Aug 20, 2008