Cepstrum - what is it good for?

[GregLocock]

Back when I was young and naive I came across Cesptral Nyalysis or whatever silly word game they want to call it. Since then I have occasionally tried it on real signals with strong harmonic structure, and usually found it to be less than helpful. I suppose it is fair to say that I usually have the luxury of looking at variable speed data, so a waterfall or campbell's diagram is helpful and easy to interpret.

Anyway, here's an example of a cepstrum, showing how the presence of noise and pure tones seem to render it less than powerful. The blue signal, which is the useful red one with some noise and tones added, fails to strongly identify the correct harmonic spacing.

Sorry about the axis labelling, apart from the first the x axis is in bins, and the y axis is in whatevers.

So am I missing some trick here?





Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[TPL]

I remember it just as a B+K marketing/brand differential thing - their own spectrum of a spectrum technique, claimed to be be good for analysing gearbox vibration. Others that are of similar value: Kurtosis (for finding rolling element bearing faults), full spectrum (marketed by Bently Nevada - brings nothing new to the party except confusion) and my own bugbear - Peakview: overhyped, loved by some, hated and misunderstood by most and responsible for too many wrong 'calls'

I'm not suggesting that these things don't work, it's just there are easier (and often more reliable) ways of achieving the same thing.

 

[hacksaw]

the cepestrum (faint recollection) is the fourier transform of the spectrum and occasionally used to separate multipath components

 

[SomptingGuy]

Fon't forget the Wigner-Ville analysis.

- Steve

 

[btrueblood]

Been awhile since I looked at it. Have had similar success to yours in finding useful problems for it to solve, Greg.

 

[SomptingGuy]

... I always found the term "lifter" amusing, as it's slang for something else in these parts.

- Steve

 

[hacksaw]

the WV approach is for non-stationary signals, the cepestrum is specific to cases where multipath is involved not noise or non-stationarity, agree with GL finding that the benefits in the case of noisey data are marginal

 

[Tmoose]

Everybody probably already has a paper copy of this -

http://www.bksv.com/doc/233-80.pdf

The claimed insensitivity to sensor location sounded interesting to me when doing PDm routes with my mag mount accelerometer, but gearboxes are such cranky sammiches that I soon decided the tools should be oil analysis was 75%, and vibration analysis 25% on them.

 

[GregLocock]

Thank you. The collective wisdom of the NVH community has voted cepstrum off the island.

By the by the speech analysis boys call a very similar but totally different analysis cepstrum, they take the IFFT of the log of the magnitude of the orignal FFT, turning it back into the time domain. Matlab seems to offer both options but you need to read the documentation of their various functions carefully.

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[amanuensis]

Suppose your signal has a lot of harmonics. These harmonics give a sort of periodicity to the spectrum. When you gonna apply the log function to this spectrum, then the shape of the signal looks almost like a sinus function or a triangular function.
Then when you will apply again the FFT, you will obtain frequencies that correspond the periodicity of your original signal.