Asked by Danupon Subanapong
on 28 Jul 2019

Hello!!!

I would like to ask for some help.

I am working with inputing a large number of time-series signals. Among these signals, there are some signals which the mean value has changed. I would ignore reading the signals which there is a change in mean value. I will give three examples of signal: 1) normal (without changing mean), 2) abrupt changing in mean value, and 3) linear changing in mean value. My goal is to detect signals which there is a change in mean and skip reading these signals.

Ps., I would like to set criteria that if the magnitude of change in mean is larger than 10,000, it will treat as there is a change of mean value. Lastly, the files of these three examples of signal are also attached.

1) Normal (✅ would like to input to matlab)

2) Abupt change (❌ do not want to input to matlab)

3) Linear change (❌ do not want to input to matlab)

Answer by Akira Agata
on 31 Jul 2019

Accepted Answer

It may need to apply "smoothing" before detecting changes larger than 10,000. Looking at your data, ~2000 points movmean will clarify whether data contains changing point or not, as shown in the followign plot.

But, please note that ~2000 points moving mean will clean-up changes in short time (such as spilke noise). So, if some of your data contains such a short-time changes, you should consider some additional detection methods.

Sign in to comment.

Answer by Jan
on 28 Jul 2019

Edited by Jan
on 28 Jul 2019

It is impossible to detect a change, if it occurs in the first or last few frames.

Ist there only 1 change of the mean? Then fit a line to the data and set a limit for the slope. If there can be more changes, fit a parabola or a polynomial of higher order to the data and check the factors.

Alternatively use findchangepts.

Is the sine wave with a frequency of about 4/100 the search data. Then you can filter it out at first to reduce the noise for the detection of a change in the mean.

Danupon Subanapong
on 29 Jul 2019

Sign in to comment.

Answer by dpb
on 28 Jul 2019

Edited by dpb
on 28 Jul 2019

The linear or similar trend is probably detectable as Jan says by fitting linear trend line and testing for nonzero slope..._IF_ the trend is something like your example and doesn't also include returning to the same or near same baseline...in that case you could have essentially zero overall slope with two trend lines buried inside.

I'd guess more robust would be to use windowing to compute means over the duration of the signal and compare those for stationarity--perhaps some preliminary smoothing might help, then again might not.

Which is the kind of thing as Jan recommends findchangepts does...I had not been aware it had been introduced to Signal Processing TB -- it's a fairly recent addition.

Danupon Subanapong
on 29 Jul 2019

"I'd guess more robust would be to use windowing to compute means over the duration of the signal and compare those for stationarity--perhaps some preliminary smoothing might help, then again might not."

I also think about this idea. I think this also might work. I will try both of your ideas. Thank you so much for your help.

Steven Lord
on 31 Jul 2019

Sign in to comment.

Opportunities for recent engineering grads.

Apply Today
## 0 Comments

Sign in to comment.