3

I'm looking for a Python package or method to accurately estimate the arrival time of time signals.

A single time signal is stored inside a 1-dimensional numpy array, so every component of the array represents the value of the signal at a given timestep.

Here is an example of the signals shape:

enter image description

Setting a threeshold is not always efficient, it depends on the signal behaviour and amplitude.

I would like to find a gradient based method to identify the starting timestep. Do you have any relevant experience or suggestion?

MBaz
  • 15,314
  • 9
  • 30
  • 44
Matteo Scarponi
  • 133
  • 6
  • Does this have to be done in real-time? If not, there are many more options available. – Peter K. Jan 29 '17 at 22:20
  • 1
    No they are data from a tsunami simulation database. Each Recording represents water elevation associated with a specific geographical point (x,y). To compare different mareograms (signals) collected in different locations, hence with different arrival times, I need to estimate their arrival time as accurately as possible. – Matteo Scarponi Jan 30 '17 at 12:33
  • 1
    Cool! Sounds like a fun project. I'll try to write an answer later (probably tomorrow), but look up CUSUM, particularly this answer. You may have to apply it to the high-pass signal (gradient signal) or you may want to look at an "energy" measure. – Peter K. Jan 30 '17 at 12:39
  • I'll look at it! Seems pretty useful – Matteo Scarponi Jan 30 '17 at 17:32

3 Answers3

3

A gradient-based method is essentially a high-pass filter, if you think about it. ("gradient" being "derivative" for 1D signals, and the derivative being small for non-changing signals, and very high for high-frequency signals). And that's exactly how I implement it:

  • Signal in
  • high-pass filter
  • abs()
  • find first above threshold
  • subtract delay of filter¹

You can do so by designing a high-pass filter using scipy.signal.fir_filter_design and apply it using scipy.signal.filter.

Another popular method, which is mainly interesting if you want to do this on a live stream of samples, is doing the same within GNU Radio, where you can take the filter taps from the filter design, and put them into a readily available FIR filter. I kind of covered that in a blog post once, and you might want to read the first chapters of the GNU Radio Guided Tutorials, too.

¹: this requires your filter to have constant group delay, i.e. be linear-phase.
Marcus Müller
  • 30,525
  • 4
  • 34
  • 58
3

As mentioned before, a high pass filter can add delays to your detector. If your filter is not linear-phase, the delay depends on the frequency of your signal. Another approach is the called Onset Detection, or a detection of a sudden burst of energy. A great paper about this is here.

luciano kruk
  • 316
  • 1
  • 4
3

I think Marcus' technique is a good approach. However, if you want to avoid the phase-delay introduced by the high-pass filter, you could filter the signal twice. See Zero-phase Filtering. Granted you are filtering a given signal twice, but there is no added phase delay and, again, this is off-line data.

Community
  • 1
Michael_RW
  • 434
  • 2
  • 5
  • You can always save your data to memory and filter it twice. So this works even in online setups, as long as your filtering operations can be done under schedule. – The Dude Oct 26 '17 at 18:22