0
$\begingroup$

I have a set of time series signals for which I have to develop a anomaly detection algorithm. I am considering using classifiers like SVM to do this.

However I am confused about how to properly process the features from the time series. I used FFT to get the frequency domain power. If I want to use these features to train the SVM model, is it a good practice to use logarithmic power values or the simple absolute power values ?

If I use the logarithmic values of the power features, is there a chance that the model performance might be negatively affected ?

$\endgroup$
2
  • $\begingroup$ Do you know that the frequency content of the different signals is actually different? $\endgroup$
    – Engineer
    Mar 25, 2020 at 11:56
  • $\begingroup$ Yes, I have observed them by visualising as a spectrum. They show different peaks. $\endgroup$
    – Kanmani
    Mar 26, 2020 at 0:57

1 Answer 1

0
$\begingroup$

Logarithmic scaling is more-or-less standard in most areas of audio machine learning. It reduces the range of values, which tends to be preferable for numerical optimization processes. Additionally, it is a bit more closely to human perception of sound.

It is rare that not scaling the values is better, but there is no guarantee. The best way to know for your specific data, task and evaluation metrics, is to do an empirical comparison of the two preprocessing options.

$\endgroup$

Your Answer

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

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