I'm trying to classify some time series data and my goal is to convert it to a wavelet image and just use novel image classification techniques. However, my problem is the fact that my data doesn't have the same time scale (meaning some of them are about 20 seconds and some of them about 200, etc).

My first solution is that I could probably use some methods to cut the main part of the each data to a fixed size time, but I'm wondering if there's a better solution to this problem.

Can anyone point me in the right direction?

  • $\begingroup$ Which approach you use may depend on how the data differs e.g. is 200 seconds equivalent to 10 batches of 20 seconds, or one stretched version of the 20 second? Dynamic Time Warping (which was developed to help with voice recognition) can help to match similarly shaped sequences of different lengths. Other techniques such as using the Matrix Profile may help to identify common 'motifs'. Are you able to elaborate on how the series/sequences differ? $\endgroup$
    – Tom Bush
    Feb 7, 2022 at 10:31
  • $\begingroup$ @TomBush Well, to be more precise, my problem is about classification of earthquake records in which I have about 200 records of 2 dimensional data (of acceleration) over time. Although the time of each record may vary between 20 to 200 seconds, these records, have a main event that takes about 40 seconds, and the oscillations happen at that timeframe. There are some standard methods to extract the data of that main event so that each record have the same time, but I'm looking for a method to input all of my data without any "cutting". $\endgroup$ Feb 7, 2022 at 11:26
  • $\begingroup$ @TomBush (Sorry for breaking my reply in two parts). I guessed that voice recognition techniques might help in this problem, and I might even go to this direction. But my main idea is to transform my data to wavelet images and feed them to a simple CNN. So, the question is whether it is possible to do the said so using image classification. $\endgroup$ Feb 7, 2022 at 11:30
  • 1
    $\begingroup$ Ah, with you. I have limited experience of working with wavelets or CNNs, so it'd be wrong of me to comment further! Good luck, I hope someone picks this up as it's piqued my interest too :) I would add that in the past I have used some fairly easy automated techniques for cutting time series data such that the 'shape' begins in the same place (i.e. where the amplitude changes significantly) but understand if you want to avoid pre-processing of that sort. $\endgroup$
    – Tom Bush
    Feb 8, 2022 at 14:28


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.