I have a time series data of accelerometer in X,Y,Z axis. Data is not sampled at a constant sampling rate(but is close to 100 Hz). In the paper I am referring, it mentions that for feature selection I have to use 'windowing technique', with window of interval 4 seconds.

And the 'Features' that are extracted can be either time or frequency based features for the 'window'. So below features are extracted per window.

Time Based features : Mean, Maximum, Minimum, SD, Median
Frequency Based features : Mean, Maximum, Minimum, SD, Kurtosis

I understand the time based features but what does frequency based features mean ? Viz, How frequency based mean is different from time based mean.


Your model needs to have standardized input features, so that the features can be compared to each other and to other inputs in general. With frequency data, the metric that encompasses this is period.

The goal here is to determine the period of the perturbation over a given time domain, which is determined by finding its frequency over said domain.

It is not necessary to take the FFT in order to convert from time to frequency domain. Check out this post for more information.

  • $\begingroup$ do i need to take FFT of the window to get the 'Frequency Based features'? $\endgroup$
    – OSK
    Jun 2 '17 at 21:09
  • $\begingroup$ updated my answer, see above $\endgroup$
    – redress
    Jun 2 '17 at 21:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.