FFT Interpretation 4

[MachineryWatch]

I positioned a photocell speed sensor, that generates a TTL signal, to read the passing vanes of a cooling fan on an engine driven generator. The speed was 3600 RPM (60 Hz). There were 14 vanes on the fan. I ran the signal to an FFT analyzer. The FFT is 0-1000 Hz with 3200 FFT lines.

The FFT shows 60 Hz harmonics with a dominant peak at 840 Hz (60 X 14). The FFT also shows sidebands spaced 90 Hz from 840 Hz (750 Hz and 930 Hz).

The "Vane Pass Frequency" at 840 Hz is certainly expected. The 60 Hz harmonics I guess would be a result of slight differences in blade spacing that repeat each time the rotor makes 1 complete revolution.

What could cause the 90 Hz sidebands? Torsional rotor vibration would cause Frequency Modulation, correct? Would this appear as sidebands in the FFT? Is it proper to assume that a good TTL signal would have no Amplitude Modulation?

What further processing could I do on the captured waveform to try to extract torsional vibration information from this signal?

Thanks for your help.

Skip Hartman

http://www.machinerywatch.com

 

[rodar]

Hello
It sounds like the engine speed is being modulated at
60 and 90 hz. One thing to check is the photocell speed
sensor. Is is retroreflective type. In other words is there
any difference in the pulse start and stop time based on
the position on the fan as it passes. I am thinking of a
fan blade with more dirt on one half of its blades than on
the other half. For an engine you might expect some modulation at the rotation frequency but I am not sure what
would modulate at 1.5 times rotation speed.
If the TTL signal comes from a line powered device with
inadequate power filtering there might be 60hz patterns
in the amplitude of the output signal. You should be able
to check this by setting the sensor output on and do another FFT with the signal. For a more detailed analysis
you probably need more than 14 points / revolution on the
sensor. If you could mechanically connect a rotary encoder
with 100 or more points per turn that would give a clearer
picture.
Hope this helps
rodar

 

[bridog]

Do I understand you correctly that you are sampling a periodic waveform with a 840Hz fundamental using a sampling rate of 1000Hz? If so then you are sampling at less than the Nyquist rate and that can cause aliasing problems that can result in undesired frequency artifacts like what you're describing.

bridog

 

[MachineryWatch]

Sample rate was 2560 samples/sec resulting in a 0-1000 Hz FFT frequency range.

Skip Hartman

http://www.machinerywatch.com

 

[maximizer]

Hi Skip,

The "sidebands" that you are observing could be the result of sampling a square wave signal. The Fourier Transform of a square wave signal is a sin(x)/x function. Typical of this sin(x)/x function is that it has a main lobe centered at the frequency of the square wave signal with "sidelobes" and nulls to either side of the center. You can improve that by using so called Windows that are weighting functions. One such window is called a Raised Cosine Window (cos^2 Window). There are many other Windows known.

It seems to me that you may have also some aliasing problem because your sample frequency is only about 3x, which in general is not enough to sample a square wave function (that has many harmonics). Unless you apply some lowpass filtering you are aliasing some upper harmonics of the square wave signal into your signal of interest. I recommend that you sample at least 2x the 5th harmonic, that is 10x of the fundamental, or use a lowpass filter that limits the fequency to perhaps 2x of the frequency of interest.

Regards

Max Edelhauser

 

[MachineryWatch]

Thanks for the responses. I am interested in other's opinions, as well, so please continue to comment on this. Max, I guess you are referring to antialias filters? My system does have very good antialias filtering. Or are you suggesting that the sample rates you recommend be used even when you have effective antialias filtering?

Rodar, the situation that led to the capture of this data was a request for vibration diagnostic testing to be performed on the engine driven generator. While the customer did not think he had a torsional vibration problem, I am aware that many torsional data acquisition and analysis systems simply demodulate an encoder signal. These decoder signals usually have greater than 60 pulses per shaft revolution. I did not have an encoder but I thought it would be interesting to try to do a "quick and dirty" method to see if any worthwhile information could be extracted from the data.

The ends of the fan blades were visible and the reflective tape for the photo cell seemed to adhere well. I also was able to fabricate quickly a fairly rigid mount for the photo cell.

I have a little previous experience with looking at the encoder signals. I know that when ever the encoder signal is produced from a magnetic sensor on a gear or from an optical sensor looking at stripes placed on a rotating shaft the shaft rotation harmonics show up in the FFT. My understanding is that this is due to mechanical runout of the gear when using a magnetic sensor and some inequality of the reflective tape spacing when using optical sensors.

Too bad I can't paste the waveform and spectrum here for all to see. The TTL waveform actually appears to have some amplitude modulation at 40-45 Hz.

Finally, the engine/generator had significant 1.5X RPM vibration (90 Hz). I guess that even though I had the photo cell rigidly mounted, the vibration of the photocell could be contributing to the 90 Hz sidebands evident in the FFT.

Max, your comments on the windowing to be used are interesting. My standard vibration analysis system does not have the window options you mentioned. If you use the specified window for the FFT calculation does it only remove the sidebands that are a result of performing an FFT on a square wave?

I have a program that I can export this waveform to that may have the FFT window that you mentioned. If it has the correct window I will run the data through that program and see what I come up with. I will let you know the results.

Skip Hartman

http://www.machinerywatch.com

 

[MachineryWatch]

Max,

My other FFT program does not have the window you mentioned (Raised Cosine Window (cos^2 Window)). I guess I do not have the option of trying your windowing solution.

These are the Windows available to me:

Bartlett
Blackman
Flat top
Hamming
Hanning
Kaiser
Parzen
Triangular
Uniform

Thanks again for your help.

Skip Hartman

http://www.machinerywatch.com

 

[Highspeed]

Hi,
In my opinion, if you are willing to look Torsional rotor vibrations and used an optical tacho on the blades for that, the usefull signal will be among the noise.
Basically with this setup, you have 14 valid samples per turns and by derivating them, you'll get 14 speed informations per turn thus preventing you to see more than 7x base frequency regardless theFiltering, FFT windowing and approach
Further more the optical sensing mode will probably give you informations that will also deal with the physical geometry of your blades and there, it will be hard to differenciate them from the real "torsional rotor wibrations" you are looking for.
My recommendations if you want to do a quick and dirty testing will be to use a simple accelerometer.Not going to the expensive B&K range you can get usefull signals from simple piezzo shock sensors (see Murata) .
You can hook that directly to the audio input of a sound card and then run FFT analysis using CoolEdit or some other sharewares
Good Luck
HS

 

[Highspeed]

Add on..
Rethinking to your setup and assuming the following:
Your Opto TTL triggers perfectly at the edge of your blades that are perfectly spaced.
Then your vibration (changes in speed)information is contained in the timing between edges of your TTL signal
Assuming 60 Hz / 14 blades gives 840 Hz or 1.2ms beween edges. To get it more accurately , you could try to feed your FFT with the time delta between 2 edges instead of the ttl pulse but that is an other story needing some numerical xformations.
Cheers
HS

 

[maximizer]

Hello Skip,

In regards to your two specific questions directed to me:

1. Yes, I am refering to anti-aliasing filters. The filter bandwidth has to be set such that you block all frequency components that you do not want aliased into your spectrum of interest. If you sample 3x only you need a very steep (high order) anti-aliasing filter.

2. The raised cosine window is better known as the Hanning Window. Any weighted window is better then no window. To be correct, there is no such thing as no window because you are selecting a finite number of samples. That by it self constitudes a window; but it is uniformly weighted and is therefore rectangular. Thus, you actually convolving two rectangular resposes (the sampling window and the TTL pulses), wich make things not easier. Perhaps even better sampling windows than the Hanning Window are the Hamming and the Blackman Window, with the later is perhaps the best window because it has much lower stop band attenuation than all others.

Hope that helped somewhat.

Good luck,

Max Edelhauser

 

[MachineryWatch]

Max,

Thanks for clarifying this for me. I do a lot of vibration analysis and some acoustic analysis. It seems that most analyzers and data acquisition systems built specifically for these uses have done a pretty good job in the area of anti-alias filtering. I understand pretty well the purpose of the Hanning Window in terms of reducing leakage, and also how it may influence amplitudes of frequencies that are not bin-centered, but I do not have a good understanding of most other Windows, other than Flat Top.

There is a pretty well-known signal processing program on the market designed to be used with PC sound cards. I use it mostly for instructional purposes and experimenting with digital filtering on data I have acquired with other systems. I have quizzed them on their lack of anti-alias filtering for data acquired through the PC's sound card. They insist that the sound card actually samples at a much higher rate than "indicated" and that this means the Nyquist frequency is much higher than "indicated" so that aliasing is not a problem.

I have been skeptical of this explanation and therefore have not relied heavily on their software. Is what they are saying possible? In the realm of of sound reproduction, it seems to me that aliasing would tend to produce tones that do not exist. Yet music and other recordings made with the sound card do not seem to suffer from unwanted sound, generated within the system. What am I missing?

Skip Hartman

http://www.machinerywatch.com

 

[rodar]

Hi Skip
I believe most FFT analysers use this oversampling method
to create an anti aliasing filter. The reason is that as
you adjust your instrument to different sampling times
it would call for different AAF cut off.
What is done is incorporate an high frequency analog AAF
that protects the high sampling frequency from alias then
prefilter with an appropriate digital low pass filter to
create the actual samples used for analysis.
Rodar

 

[MachineryWatch]

Rodar,

Thanks for the response. So do you think that I should not be concerned when this company says they do not need anti-alias filters?

Skip

http://www.machinerywatch.com

 

[IRstuff]

You should

> do a recording and see what the sample rate is
> determine if there are noise or other features than might, directly or through aliasing, affect your analysis
> determine your maximum expected frequency

Then, you can determine if an AA filter is necessary

TTFN

 

[rodar]

Skip:
The PC sound card must have some kind of filter before
digitizing but as IRstuff says I would test it. Do you
have a sine wave generator?
If so set an appropriate level and feed into sound input
and tune the generator up and down the band while looking
at the FFT output. If you can find the sample rate for your
card then look aroung Fs/2 to see if additional frequencies
appear on the software FFT.
Rodar

 

[owiecha]

Analyzing the setup of your original experiment I came to the conclusion that the signal you are interested on is the period, not the output waveform, which is basically square and therefore the FFT will show a square wave spectrum dependent on the basic frequency and duty cycle, and some secondary effects derived from the fact you are generating 14 pulses per revolution (from slightly unevelnly spaced blades) and the aliasing effect due to the lack of antialiasing filter. You have two ways to go: Treat the signal as frequency modulated and use a very steep filter with cutoff frequency just below the second harmonic of the basic signal (2x14x60Hz). The spectrum you get shows the variations of the carrier (14x60Hz). The other solution is to use a LeCroy spectrum analyzer, where you can create an intermediate signal derived from the original one and then run the FFT. In your case the intermediate signal is the period of each sqare wave. In both cases the exact spacing of the blades is an issue, so the use of an optical encoder is highly recommended. Of course you can build a circuit that converts period into amplitude using a high resolution encoder, a monostable pulse generator with well defined high and low levels, and a low pass filter.

 

[MachineryWatch]

owiecha,

I did have a very good antialias filter that was set just above the 1000 Hz maximum frequency used for the FFT generated spectrum (sample rate was 2560/sec). I agree that it is important to have this filter. I also agree that any variation in blade spacing will produce some unwanted results. I believe the undesired results of the slight variation in blade spacing will produce the 60 Hz harmonics that I saw in the spectrum since these blade variations will repeat themselves at a rate of 1 time per shaft revolution.

I think the modulation generated sidebands that appeared in the spectrum spaced at 90 Hz were due to either the engine's high vibration levels at this frequency because the vibration was causing the photocell to "shake", or it is from torsional vibration.

Skip

http://www.machinerywatch.com