How to create indices for specific parts of signal to keep

Hi

So i have this signal as shown in the plot, this is currently a 50000x1 vector. I would like to extract only the parts where the amplitude increases, these correspond to the extraction and retraction of a piston rod with a 1s dwell time. I know i can do it manually from looking at the plot and find the indices, but i have 40 different signals and would thus like a automatic process for this.

I would like to find the point where each extraction and retraction cycle starts and ends so i can remove the other parts.

thanks for all help.

[dpb]

You'll need to use some thresholding level with some noise detection probably. Altho the plot is so small can't see much detail, looks as though is pretty clean signal so shouldn't be too difficult to find a discriminating level. Of course, not all datasets may be as clean as this one, too.

I suggest attaching a .mat file containing at least one of these time series if want somebody here to do anything more than simply pontificate similar to what I have done above... :)

[Jørgen Fone Pedersen]

Thank you for answer.

Yes i have tried that using the envelope function on this signal, and been able to zero out all other instances that i dont want.

However i still dont know how to kind of single out just one of the piston retraction/extension from the overall signal.

.mat file of the plotted signal attached here.

[Jan]

What is the nature of the noise? Does it have a specific band width? Then apply a filter at first. Or do you know the frequency of the wanted signal? Then FILTFILT will help also to remove the noise. Afterwards remove the mean and search for the change point. See: doc findchangepts

[Jørgen Fone Pedersen]

thank you.

I used your suggestions to find a very good estimation of where the extension and retraction begins using the spectrogram and the findchangepts functions.

However, i am still unsure of how to use this to find the correct indices of the signal. Any ideas on how to do that?

[s,f,t,pxx] = spectrogram(Short_mean,100,50,100,1e3,'yaxis');
figure(1)
plot(ts,Short_mean)
figure(2)
findchangepts(pow2db(pxx),'MaxNumChanges',20);

[Jan]

Isn't the output of findchangepts the wanted index already?

[Jørgen Fone Pedersen]

well yes. But I need to convert it somehow to fit the timeseries data which is a 1x47000 vector and the index findchangepts gives me is in the spectrogram data which is a 1x939 vector. I guess i dont understand the how the spectrogram works good enough. Any ideas on this?

[dpb]

Not a simple correlation to where in the time history the breaks are from the spectrogram that converts to the frequency domain...I tried just a straight discrimination in the time domain as

 ipts=findchangepts(Short_mean,'MaxNumChanges',20,'Statistic','std','MinDistance',1000);

It doesn't seem you get the plot if ask for output; the above options w/o assignment generated following figure:

which looks pretty good.

Had to set the 'MinDistance' parameter to get rid of some very short artifacts at transitions; these could probably be screened after the fact as well.

Might want to see what would do if did filter first as Jan suggested; above is on raw time series alone.

[Jørgen Fone Pedersen]

Wow, this looks great.

Thank you so much, im going for this method i think.

Regarding the filter I dont have much experience as to what filter to use or how to filter set the parameters. This is data gathered using acoustic emmision, and i dont have enough experience yet as to what is just noise and what is the signal. The signal i attached is actually a shortened version where i took the mean of each 1000 datapoints to sort of clean up the signal. The real signal is a 47e6x1 vector.

[dpb]

" The signal i attached is actually a shortened version where i took the mean of each 1000 datapoints..."

That's a moving-average filter right there. From the looks of the data, that's probably pretty good first cut. You might want to consider using median instead of mean to preserve the absolute magnitude since the mean will have somewhat of a constricting effect on the range of the observations as well as smoothing.

[Jørgen Fone Pedersen]

Thanks for your answers and tips. I had no idea it was a moving average filter it was called actually.

It seems to be working fine now, I will try with the median also to see how much of a difference it makes.