1
$\begingroup$

I'm looking for a good way of extracting features from the frequency domain of vibration data for a one-class support vector machine.

The image below shows an example from the dataset found in this link. The dominant peak found at around 960 Hz corresponds to the "main frequency" of the system. E.g rotation*number of gears.

Im keeping this question fairly open ended, because i know there are several methods for doing this and I'm mainly looking for your suggestions and experience.

Would it be possible to analyze this so I get some specific information regarding these peaks caused by the faulty gear tooth?

Normal (Left) Faulty (Right) Time and Frequency

$\endgroup$
  • $\begingroup$ The question is unclear. I didn't get what you're after. $\endgroup$ – Royi Jul 22 '18 at 22:22
1
$\begingroup$

Assuming you have epochs/segments of data. For this kind of signals, it is a safe approach to extract features using wavelet representations.

Using FFT might work as well, but I dont know how problematic would be the stationarity assumption in this kind of applications. Besides, FFT estimates for this kind of signals is sometimes very noisy. If you insist on this kind of representation, I would suggest estimating the power spectral density using Welch or multitaper methods. These usually deliver a nicer spectral estimation.

In any case, using any spectral features as direct input to your SVM might not be the best idea. Irrespectively of what features you extract (FFT, Wavelets, etc..) I would include a data dimensionality reduction step (PCA would be the first thing I'd try)

$\endgroup$
  • $\begingroup$ Thanks for the feddback. Im not insisting on using FFT at all, it just seemed to give a good result. I'll have a look at your items and see if i can find a better way to solve this $\endgroup$ – VegardKT Nov 24 '17 at 11:44
  • $\begingroup$ PSD definately gave me a much "friendlier" result. I can clearly diferentiate the faulty signal vs the normal one here as well. I'll see if i can apply PCA to it, Thats fairly new to me though, so if you wanna give me a quick rundown i wouldn't mind... $\endgroup$ – VegardKT Nov 24 '17 at 12:42
  • $\begingroup$ It's hard to know how to approach it without knowing your background. But it basically works like this: What PCA does is basically to project $X$ to a lower dimensional space ($XW$) such that the variance explained is maximized. In your case, I would suggest to use log-power features instead of the raw power. Check this video for an intuition of PCA youtube.com/watch?v=kw9R0nD69OU $\endgroup$ – Juan S. Castaño Nov 24 '17 at 17:33

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.