0
$\begingroup$

I am working with a public dataset (mentioned in few papers) of vibration measurements of broken and healthy gearbox under various load. Most of the measured channels posses the same parameters no matter whether the gearbox is broken or not - almost same FFT, same histogram of the raw data, same cepstrum, same IMF (obtained by EMD).

However one channel carry a significant difference in vibrations from the broken and the healthy gearbox (see picture). I checked some papers and books about this topic (fault detection in gearboxes, FFT analysis of gearboxes etc.), but they seems to represent way to nice and clean results far from the dataset observations I have. So I am still quite unable to understand why the FFTs looks different in this way. Few notes about the dataset:

  • frequency of measurement / or speed and size of gears are unknown (I guess something for the FFT x axis - it has no real meaning)
  • the broken gearbox suppose to suffer with a "broken tooth"

My question:

Can you form any hypothesis, why the frequency spectrum is changed in the way displayed in Figure in relation to "broken tooth" fault?

enter image description here

$\endgroup$
3
  • $\begingroup$ Hi matousc. You'll probably get a better answer if you plotted those in log scale, and if you added the labels on the graphs (I don't know if the data is available to you). $\endgroup$
    – NMech
    Apr 23 at 9:06
  • $\begingroup$ @NMech Thank for comment! The label is in graph title. Axis labels are impossible to obtain (I do not know the sampling frequency, so it is all relative). If I plot the Y axis in log, it looks like random noise - no peaks in there. This is imho the best information I can obtain with FFT :-( I know it is not much. $\endgroup$
    – matousc
    Apr 23 at 9:25
  • 2
    $\begingroup$ If you don't know even the basic details of the system (number of gear teeth, RPM, etc) you are wasting your time trying to understand the data IMO. $\endgroup$
    – alephzero
    Apr 23 at 12:09
4
$\begingroup$

Vibrations are the manifestation of things occurring at some frequency. This can be one per rev from imbalance (nothing is ever perfectly balanced) in gears spinning, one per tooth from each tooth changing where it presses against another tooth, one per specific tooth if a tooth exhibits an anomaly differentiating it from other teeth, specific interactions between two specific teeth, frequencies of components as they vibrate and turn energy into noise and heat. Overlay all these and you get a plot like that.

Note that the per-rev frequencies depend on how fast you turn shafts. Modes, etc can vary depending on how you affix and agitate something.

While it is possible to design models for a specific gearbox (you'd have to precisely measure every part), it is not cost-effective. Much of the vibrations and interactions result from deviations from the nominal values. What you can however do is track a specific gearbox's vibrations over its lifetime. You look at it when it is new (and good) while spun at a specific rpm. Then you look at it again when spun at the same rpm.

You asked for a hypothesis on a broken tooth so here goes:

Assuming your gearbox still functions with the break (Depending on the specifics of the break and gearbox, you could expect the output to stop altogether or turn the output differently. Such cases are better detected with other sensors):

What you might pick up include broken bits that are now vibrating on their own, a possible reduction in your one per tooth (this should be more apparent in an fft of the phase-aligned difference between good and bad rather than a difference of fft), and an increase in your one per rev due to imbalance (probably the main thing you can look for in data like what you showed).

$\endgroup$

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.