Signal Processing Artifacts

[MachineryWatch]

I have been collecting vibration data on a paper machine press section that has abnormally high amplitudes at around 28 Hz. It is believed to be a resonant condition. The vibration sensors are accelerometers. The sample rate is 320 samples/second. This yields a 0-125 Hz FFT. When I record at this rate for 10 minutes and then look at a "waterfall" plot of 200 line FFTs the amplitude at the 28 Hz frequency of interest cycles high and low in amplitude with a period of just over 4 minutes. If I use 100 line FFTs for the waterfall the plot shows the 28 Hz amplitude still cycling, but now with a period of just a little over 2 minutes. The period's dependence on the number of FFT lines chosen leads me to believe that the amplitude cycling is not real, but rather, an artifact of the signal processing. Beyond this how can I explain what is going on? Can someone help me understand the reason for this?

I notice that when I use very high resolution (3200 line FFT) with the same frequency range I do not see the cycling in amplitude at 28 Hz, but the time block required for this spectral calculation is nearly 30 seconds.

Thanks in advance for your help and comments.

Skip Hartman

http://www.machinerywatch.com

 

[VEBill]

Does your measurement system include an appropriate antialiasing filter - something like about 100Hz low pass?

The likely answer is that such a filter is built-in to the system and is automatically invoked; but it might be worth double-checking to make sure.

 

[VEBill]

Re-reading your original post, it seems that you might be referring to variable artifacts in one fixed data set. If so, then my previous question about anti-aliasing filters is not applicable. Sorry.

How about the FFT software's 'window' setting? The FFT 'windowing' is how the start and end of the data set is weighted to reduce artifacts that might be caused by the sudden start and sudden end of the data set. Hamming is one well-known FFT window filter; there are others. Maybe check that setting...

 

[MachineryWatch]

Hi,

Thanks for your comments regarding anti-alias filtering and FFT windows. The system I am using does have a very good anti-alias filtering that is automatically invoked and is dependent on the sample rate chosen.

This data was captured using the "Hanning" window. This window is typically used for most machinery vibration diagnostics. At least where transient signals are not involved.

I actually had the same results capturing the data 2 different ways. First I used storage of sequentially captured waveforms and averaged FFTs. Second I "streamed" data to the harddrive, essentially making a digital recording, using the same sample rate I had used for processing the original FFT data. I saw the same cycling of amplitude using both methods.

Since my original post the machine speed has changed slightly changing the dominant frequency from near 28 Hz to near 31 Hz and the cycling of amplitude has almost completely disappeared.............

Using a much higher sample rate and lower resolution the "real-time" appearance of the peak in the spectrum shows it varying up and down erratically from moment to moment, but not the slow cycling of amplitude described in my original post.

The slow cycling has to be an anomaly resulting from the digital sampling of the data and the frequency involved. I am just not sure how to describe what is happening.

Skip Hartman

http://www.machinerywatch.com