Yes, it is possible. You need to use a lot of 'averages'. If you can vary the speed of the machine as you hit it then you will get better data more quickly. Typically a static hammer test will need only 3 or 5 frames of data per location, you might need several hundred if the machine is running. Look at the coherence plot to see if you are anywhere near getting reasonable data. You will have poor coherence at the running speed (etc) of the machine.
Cheers
Greg Locock
Please see FAQ731-376 for tips on how to make the best use of Eng-Tips.
You may have a signal to noise ratio problem. (desired signal being the impacting response, noise being the machine and everything else besides impacting response.).
The more energy you can [safely] put in with your impacting, the better your s/n.
=====================================
Eng-tips forums: The best place on the web for engineering discussions.
Just thinking about it, a much better approach would be to use a swept sine input via an electromagnetic shaker. I think you'll find it very hard to do an impact test with several hundred averages. Then you can use a tracking filter on the response channels and that should solve your S/N problems at most frequencies.
Cheers
Greg Locock
Please see FAQ731-376 for tips on how to make the best use of Eng-Tips.
This can be done relatively easily, depending on the type of analyser that you have. You have to be able to store a series of spectra obtained at known time intervals and present them in a '3-D' or waterfall format.
Keeping the machine at steady operating conditions, take 5 spectra (no-averages) at say 2 second intervals - then start impacting: begin with light taps using a hard rubber/plastic mallet at 5 taps/second) and take another 10 spectra also at 2 second intervals, followed by another 5 spectra also taken at 2 second intervals with no impacting.
Plot the specta in a waterfall format and you can see the effects of any excitation of a natural frequency (unless it is always visible due to being excited during during normal operation). There's a bit of trial and error involved, but once you can see that you are getting results, you can fine tune the procedure.
I have carried out this test with Adre and Zonic book systems, both of which allow you to automatically schedule spectrum capture at fixed time intervals. With a bit of thought and planning you could also do this with a hand-held datcollector system.
Of course, this will only tell you if there is a natural frequency present close to your measurement point and not its location (foundation/bearing housing/casing/shaft or piping).
A better way of identifying natural frequencies associated with rotating equipment is to measure 1X (and nX) amplitude and phase during a shutdown (or slow startup) and inspect the bode and polar plots - for that you would definitely need an Adre or Zonic book and fit the machine with a keyphasor (once per rev tacho/reference pulse)
CSI has an averaging method in their analyzer called Negative Averaging. It does exactly what you are asking about. The first set of data that you acquire is with you impacting the structure, then the second set of data is with just the machine running. It subtracts the second set of data from the frist and you are left with the frequencies that you excited by impacting the structure.
You’re right Don – I had forgotten that name (negative averaging). It’s kind of comical that they decided to call it “negative averaging” rather than simply “subtraction”. There must have been a marketing guy involved in that decision somewhere.
=====================================
Eng-tips forums: The best place on the web for engineering discussions.
I'm sure if you talk to anyone mathematically inclined that they would laugh at the terminology we use in the real world. "How many averages did you use?" 6
No we didn't, we vector summed 6 frames of complex spectral data and divided the magnitudes by 6 to get one average spectrum.
etc
Not that it really matters!
Cheers
Greg Locock
Please see FAQ731-376 for tips on how to make the best use of Eng-Tips.
Good one Greg. I have used the term averages many times in that context and never thought about how silly it is.
Another one that used to bug me was FFT vs DFT. Coming out of school we knew well that the DFT was what we were computing, the FFT was simply the result. If I want to compute DFT by brute force on my computer I can do it and have done it using the definition (without the FFT). But it's a helluva lot slower.
So DFT is the result, FFT is simply the algorithm. But the term FFT is so universal now that we all use it to refer to the result. DFT is gone from the vocabulary and most vib analyst would look at you funny if you used that term.
=====================================
Eng-tips forums: The best place on the web for engineering discussions.
AFAIK, "FFT" is the name given to a particular DFT algorithm developed by Cooley and Tukey (?). It can only be used for powers of 2 numbers of points. It works by splitting the data in half, then half again, then half again etc. and performing bitwise operations. Hence it is ideally suited to low-level assembler level programming and is therefore very efficient. This is why we historically use (or pad-out to) power of 2 points in data collection. Of course this is somewhat redundant as the speed of computers has increased. You can test this by trying an matlab DFT on a record with 65535 65536 and 65537 points and looking at the cpu time. 65536 will use the FFT algorithm and is significantly faster.
