How to select a proper method for detecting onset of signals?

[yijun1994]

I have a need for detecting change point of seismic waves, the onset seem quite obvious by visual inspection. But costAR and costL2 seem failed to detect the change point. Could you give me some suggestions about how to select a proper cost function among presented ones?

[oboulant]

Hi @yijun1994 ,

Thx for your interest in ruptures.

Would you have a visual graph of those data ?

In the mean time, you can find in the documentation User Guide the list of all the cost function available in ruptures : https://centre-borelli.github.io/ruptures-docs/user-guide/

[yijun1994]

The visual graph of some signal samples is displayed above. Actually, they are earthquake signals.

[deepcharles]

To detect changes in amplitude, you can use CostNormal. You can also transform you signal to only keep its enveloppe (see here) and use CostL2 on the transformed signal.

Hope this helps

[yijun1994]

Thanks! I also employed kernel-based (Gaussian) method to identify the change point. I found it useful but not so precise. I am wondering if the parameters I set is not right.

[deepcharles]

The detected changes might not be at the exact location you want, because there is a small amplitude phase between the flat phase and the full amplitude phase. Some ah-hoc post-processing might be needed in any case.

[yijun1994]

is 'ah-hoc post-processing' a specific method in signal processing?

[deepcharles]

No, it is whatever method you can find to add more precision if the detected changes are a bit "off".

[deepcharles]

Closing for now. Feel free to reopen.