Hallo, I am quite new to this topic and I hope I can get some help here. suppose we have a system driven by white noise resulting in a Random Process. The measurement I get is the addition of the Random process with the measurement noise. my question is : Is it possible to identify the measurement noise from the recieved noisy signal? Do anyone has a paper or a book approaching such topic? what I am trying to do afterward is to estimate the AR parameters for the De-Noised measured signal.

# Measurement Noise of PSD estimation (AR)

Started by ●July 13, 2009

Reply by ●July 13, 20092009-07-13

On 13 Jul, 17:07, "m.alwaieh" <m.alaw...@gmail.com> wrote:> Hallo, > > �I am quite new to this topic and I hope I can get some help here. > suppose we have a system driven by white noise resulting in a Random > Process. The measurement I get is the addition of the Random process with > the measurement noise. > > my question is : > Is it possible to identify the measurement noise from the recieved noisy > signal? Do anyone has a paper or a book approaching such topic?The best you can do is to estimate the noise power, subject to the assumptions of the AR model. Once you have come up with a set of parameters, you also have an associated noise variance, which caracterizes the noise over the data segment. This noise will contain several components: 1) Model mis-match, in the case the data do not comply to the AR(p) model (including additive noise not complying to assumptions about noise autocovariance function) 2) Additive noise 3) Instrumentation noise, like hum from power supplies and quantiation noise You can't separate between these components unless you have done additional investigations. As for literature, "Random Data" by Bendat & Piersol is a classic. Rune

Reply by ●July 13, 20092009-07-13

m.alwaieh wrote:> Hallo, > > I am quite new to this topic and I hope I can get some help here. > suppose we have a system driven by white noise resulting in a Random > Process. The measurement I get is the addition of the Random process with > the measurement noise. > > my question is : > Is it possible to identify the measurement noise from the recieved noisy > signal? Do anyone has a paper or a book approaching such topic?Probably not, but there are more things on heaven and earth ...> what I am trying to do afterward is to estimate the AR parameters for the > De-Noised measured signal.How about determining the measurement noise and subtracting its power from your "noisy" measurement of the random process? That can be a bit tricky, but it's certainly possible. If the measurement noise is 20 dB below the process noise, it will have negligible effect. log(10)/log(10.1) =.9957 Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������

Reply by ●July 14, 20092009-07-14

On Jul 13, 8:07�am, "m.alwaieh" <m.alaw...@gmail.com> wrote:> Hallo, > > �I am quite new to this topic and I hope I can get some help here. > suppose we have a system driven by white noise resulting in a Random > Process. The measurement I get is the addition of the Random process with > the measurement noise. > > my question is : > Is it possible to identify the measurement noise from the recieved noisy > signal? Do anyone has a paper or a book approaching such topic? > > what I am trying to do afterward is to estimate the AR parameters for the > De-Noised measured signal.No you cannot. You may be able to find the ratio of the measurement noise variance to the innovations variance. Not straight forward. Hardy

Reply by ●July 14, 20092009-07-14

*Suppose that I have the measurement noise variance, How to carry on with power spectral density estimation without including the measurment noise? I mean how can i obtain the transfer function (H(z)) charcterizing the system from a noisy measurment given I know this noise? *Do Sub-Space Estimators (like MUSIC) as they decompose the observation into a signal subspace and a noise subspace?>As for literature, "Random Data" by Bendat & Piersol >is a classic.Thanks, I will check this one.

Reply by ●July 14, 20092009-07-14

On 14 Jul, 10:10, "m.alawieh" <m.alaw...@gmail.com> wrote:> *Suppose that I have the measurement noise variance, How to carry on with > power spectral density estimation without including the measurment noise? > > I mean how can i obtain the transfer function (H(z)) charcterizing the > system from a noisy measurment given I know this noise?One way is to pre-whiten the noise. Another is to subtract the noise covariance from the signal + noise covariance. Both methods have its risks.> *Do Sub-Space Estimators (like MUSIC) as they decompose the observation > into a signal subspace and a noise subspace? �Do what? Those kinds of things can also exploit knowledge about noise covariance. Rune

Reply by ●July 14, 20092009-07-14

>One way is to pre-whiten the noise. Another is to subtract the noisecovariance from the signal + noise>covariance. Both methods have its risks.I didn't deal with this problems since a while, I need to refresh my information first. Thanks for the highlights though.>Do what? Those kinds of things can also exploit >knowledge about noise covariance. >Sorry this wrong statement, I read that MUSIC algorithm is based on dividing the observation space to a signal space and noise space. The way I understand it is; this algorithm estimates the signal from the observed noisy measurement and the noise. This is done by obtaining the eigen-vector approximation of the auto-correlation matrix for each of the noise and signal spaces. Here I might be way too wrong, 1-we take the AC matrix obtained from AR estimation for example. 2-use MUSIC (or any subspace method) to get the eigen values for the corr matrix 3- Yeooo, I have signal and noise separated spaces.

Reply by ●July 14, 20092009-07-14

On 14 Jul, 14:03, "m.alawieh" <m.alaw...@gmail.com> wrote:> Here I might be way too wrong, > 1-we take the AC matrix obtained from AR estimation for example. > 2-use MUSIC (or any subspace method) to get the eigen values for the corr > matrix > 3- Yeooo, I have signal and noise separated spaces.Not quite. The idea behind MUSIC is that the noise covariance matrix is of full rank, whil ethe signal covariance matrix is singular. Which means that the signal+noise covariance matrix is of full rank. What happens with MUSIC is that one isolates the signal subspace, that also contains noise. So you don't *remove* the noise; you *reduce* it. And anyway, with MUSIC it's the noise-only subspace you use in the computations. But your general idea is correct. If you apply the same sorts of ideas to an AR estimator, the resulting coefficient estimate might be better. However, there are a couple of details to watch ut for: 1) These things require very expensive computational routines, like eigen value decompositions or SVDs. 2) Order estimation is a moot point, since order estimators as developed for the Levinson recursion do not *formally* match EVDs or SVDs. This last point doesn't mean the order estimators do not work (they do), but depending on the context you need to be very cautious. This is the kind of thing that might seed discussions if you try and publish such a method in one of the journals. Rune