denoise non stationary signal

how an non-stationary signal could be denoised ? In most material, is only written about denoising stationary signals.

Is there any specific filter?

Besides that I just need my signal in the frequency domain and if you suggest denoise with wavelet how can I transform the signal from wavelet to frequency domain directly?

Answer 1

[Star Strider]

I have nothing against using wavelets to denoise a signal, however I generally find the sgolayfilt function useful in eliminating broadband noise from a signal. If you have band-limited noise, then use an appropriate digital filter such as bandpass to eliminiate it (the filter type will be dependent on the nature of the noise).

[mehrab zamanian]

TNX, The method you suggested, is for denoising for the STATIONARY signal, not my signal that is non-stationary The frequency content of my signal varied with time I can't use only one bandpass filter!

[Star Strider]

See if sgolayfilt will do what you want.

The digital filters only require that the sampling times (sampling intervals) be consistent. They are appropriate for band-limited noise.

[mehrab zamanian]

Tnx a lot

Unfortunatly I cant find best order and framelength for this filter that suitable for my signal. In frequency domain I need my signal have certain peak and the frequency content between peaks become smooth.

please help me for this problem! this is a part of my thesis and I have been working on it for months.

my signal and sample rate are attached in .mat file. my code is written below. the main frequency is between [20-100]KHz

load("Bnew")
order = 9;
framelen =301;
sgf = sgolayfilt(Bnew,order,framelen);
fsgf=fft(sgf,2^20);

asgf=abs(fsgf);

Ze=asgf(1:end/2);
L=size(Ze,1);
f = fs*(0:floor(L)-1)/(2*L);
figure
plot(f,Ze) 
xlim([20000 100000])

[Star Strider]

I should have been a bit more specific in my recommendations. I usually choose a 3-degree polynomial. Higher orders than 5 will model the signal very closely and not give good noise elimination. I always have to experiment with the frame length.

Try this —

LD = load(websave('Bnew','https://www.mathworks.com/matlabcentral/answers/uploaded_files/1120910/Bnew.mat'))

LD = struct with fields:
    Bnew: [198657×1 double]
      fs: 10000000

Bnew = LD.Bnew;
fs = LD.fs

fs = 10000000

L = numel(Bnew);
t = linspace(0, L-1, L)/fs;

figure
plot(t, Bnew)
grid
xlim([0 0.005])

Bnew_filt = sgolayfilt(Bnew, 3, 301);

figure
plot(t, Bnew_filt)
grid
xlim([0 0.005])

Fn = fs/2;
NFFT = 2^nextpow2(L);
FTssf = fft([Bnew Bnew_filt], NFFT)/L;
Fv = linspace(0, 1, NFFT/2+1)*Fn;
Iv = 1:numel(Fv);

H = FTssf(:,2) ./ FTssf(:,1);

figure
plot(Fv, abs(FTssf(Iv,:))*2)
grid
xlabel('Frequency')
ylabel('Magnitude')
title('Fourier Transforms')
legend('Original','Filtered', 'Location','best')
xlim([0 2.5E+5])

figure
plot(Fv, mag2db(abs(H(Iv))))
% set(gca,'XScale','log')
grid
xlim([0 2.5E+5])
xlabel('Frequency')
ylabel('Magnitude')
title('Transfer Function of Savitzky-Golay Filter For This Signal')

The frequency characteristics of this signal are interesting, leading me to believe it has already been lowpass-filtered. The derived thansfer function of the Savitzky-Golay filter reveals it to have a ‘comb’ characteristic with significant filtering at three different frequencies.

Experiment with it (and this code) to get the result you want.

.