Remove noise from strain gauge measurements

Hi, I have been working on filtering strain gauge data but having problems with filtering the Strain gauge data. My data looks as shown below.

After using low-pass 5th order butter worth filter with a cut off frequency of 85Hzt, i am able to cut down the noise but i still could not able to get a smooth curve that i am expecting.Here is how data looks after filtering

As you can see in the image, the data still shows some noise. Can anyone help in filtering the data for a smoothed curve?

i tried using other cut off frequencies but nothing is working.

Any help will be a lot appreciated.

Thank you.

Answer 1

[Star Strider]

You probably need to look closely at your signal with the fft (link), specifically the code between the first (top) two plot figures, and use that information to determine your cutoff frequencies.

With respect to designing the filter, the Signal Processing Toolbox designfilt function is likely easiest to use. My IIR filter design procedure is here: How to design a lowpass filter for ocean wave data in Matlab?

[Vishwa Teja Kasoju]

Thank you for the answer.

I worked on the steps you provided but i couldnot able to decide on wp,Ws value in buttord function. How can we decide the values of wp,ws? I am collecting my data at a sampling rate of 100 kHzt.

[Star Strider]

My pleasure.

I would have to have some or all of your data to help you design your filter. I need to plot the Fourier transform and experiment with the filter design.

Another option is the Savitzky-Golay filter. See the documentation for sgolayfilt for details. It is essentially a low-pass filter but with nonlinear characteristics.

[Vishwa Teja Kasoju]

Iam attaching sample data for a single cycle below. I will also check the other option (Savitzky-Golay filter). Thanks for the help.

[Star Strider]

My pleasure.

You need to use a very low passband. See if this does what you want:

[d,s,r] = xlsread('Vishwa Teja Kasoju Sample data_Single cycle.xlsx');
Fs = 100E+3;                                                            % Sampling Frequency (Hz)
Fn = Fs/2;                                                              % Nyquist Frequency (Hz)
L  = length(d);                                                         % Signal Length
t  = linspace(0, 1, L)/Fs;                                              % Time Vector

% dmc = d-mean(d);                                                        % Mean-Corrected Data
FTd = fft(d)/L;                                                         % Fourier Transform
Fv = linspace(0, 1, fix(L/2)+1)*Fn;                                     % Frequency Vector (One-Sided Fourier Transform)
Iv = 1:length(Fv);                                                      % Index Vector

figure(1)
semilogy(Fv, abs(FTd(Iv))*2)
grid
set(gca, 'XLim',[0  5E+3])

Wp =  50/Fn;
Ws =  75/Fn;
Rp = 10;
Rs = 20;
[n,Wn] = buttord(Wp,Ws,Rp,Rs);
[b,a] = butter(n,Wn);
[sos,g] = tf2sos(b,a);

figure(2)
freqz(sos, 4096, Fs)

filt_d = filtfilt(sos,g,d);

figure(3)
subplot(2,1,1)
plot(t, d)
grid

subplot(2,1,2)
plot(t, filt_d)
grid

If necessary, continue reducing the passband and stopband frequencies (the numerators of the ‘Wp’ and ‘Ws’ assignments) until you do. I would keep the stopband frequency about 15-25 Hz higher than the passband, since this produces a stable filter at these very low frequencies. (You might see if 20 and 30 Hz do what you want. Those work, but I have no idea if they produce the result you want.)

[Vishwa Teja Kasoju]

I think (Wp,Ws) = (20,45) is working good for me.Thanks:)

[Star Strider]

My pleasure!

[PIYUSH MOHANTY]

How did you decide on Wp and Ws?

Answer 2

[KSSV]

Try this:

x = 0:.1:20;
y1 = 5*sin(x) + 2*x - x.^2 +.3*x.^3 - .2*(x-15).^4 - 10*x.^2.*cos(x./3+12).^3 + .5*(x-12).^4;
r = randi(1000,1,201) - 500;
y1 = y1+r;   % add noise 
g = gausswin(20); % <-- this value determines the width of the smoothing window
g = g/sum(g);
y3 = conv(y2, g, 'same') ;
figure;
hold on; 
plot(y1, 'b'); 
plot(y3, 'g', 'linewidth', 3);

[Vishwa Teja Kasoju]

Thank you for the answer.

I tried running the code you provided with my data set and it looks like this.

we can see that the peaks are lost. My intention is to filter the high frequency noise without loosing the peak information.

Is there any way to work that out?

Thanks again for the reply.

[Image Analyst]

What did you use as the width of the Gaussian Window in gausswin()? It looks like you used something far, far too big. With 90,000 data points, looking at your plot, it looks like you should use something around 3000 or so. But it looks like you used a number like 50,000 or something way too big.

[Vishwa Teja Kasoju]

Yes, i used 86,000 for Gaussian Window.But after using 3000 in Gausswin(), the graph looked pretty smooth.Can you please tell me how to decide on this value?