From time to time I find sidebands at spacings that don't relate to any know causes. Here is an example.

This is an overhung fan on spherical roller bearings. Sidebands around running speed (3545) harmonics spaced at about 636 cpm which is about .18 orders.

Any ideas?

sidebands.doc (50 Kb, 86 downloads)

Try to Autocorrelate this TWF. Out of curiosity. The "dirty" one looks like showing every 5th peak standing out. Not that it explains anything...

David

Any belts, gearbox or vfd?

In absence of above, here are my wild guesses:

The most likely possibility seems like process oscillation. Perhaps a control valve oscillating. Check for pressure or flow oscillation at same frequency.

Or possibly, clearance or lack of load/preload in rollig bearings creates some strange frequencies sometimes.

Least likely, is it mounted on spring isolators? I have seen a fan mounted on spring oscillators that had sidebands like that at around 250 cpm spacing, which was approx the vertical bump test resonance of the machine sitting on the isolators. I can't exactly explain it, but that's what we saw.

I'll guess it's direct coupled because of the speed. As EPete suggested, process variations may be a leading candidate. Does the motor show any of the same sidebands? Can you take a current spectrum on the motor? Sidebands are the result of amplitude modulation or frequency modulation, so if it is load related, you may see them in a MCSA around LF. Dirty filters or restrictions on inlet/discharge? David has a good suggestion in using autocorrelation. It sometimes makes AM easier to see. The vertical isolator natural frequency is a good one too. Any broadband excitation (impacts or flow?) could excite the frequency and it may may modulate other major forcing frequencies (carrier).

Always interesting to see the stuff that stumps you (me too ,) Danny.

I agree with Bill that taking electric current data will allow you to determine if SBs are speed (variable torque) related. But it may not be the case. Then 1 stronger impacts out of 5 impacts (where ever they are coming from ) will be responsible for the modulated pattern.

I also suggest taking higher resolution spectrum and using lower Fmax to measure the spacing more accurately. You may see fundamental frequency of modulation. It could be impacting at 1/5x. Does Peakvue show anything?

David

David,

The Autocorrelated twf is attached.

Pete,

No vfd's, belts, gearboxes or springs. Like Bill supposed, direct connected. Ac Induction motor, 2-pole, 60 hz, Falk Grid type coupling, Linkbelt spherical roller bearings in pillow block housings, overhung rotor, known slightly loose anchor bolts/grout. (I'm don't know the specifics of the bearings because they installed completely new ones after an all-out crash a couple of years ago. If you are not going to listen to your vibration analyst when he tells you that you are headed for a crash why would you waste your time telling him about new bearings? )

Bill,

The sidebands do show up on the motor but to a lesser degree the farther you get from the fan. A low fmax spectrum is posted. (Thanks Bob Cecil)

I've already been stumped by most of the stuff that is actually in the books so I'm moving on to the other stuff.

David,

I do suspect looseness (see the PV data in the attachment) so 1/5 is a possibility but the sideband spacing in the low fmax reading shows 636/3589 or .177.

Something that I noticed when autocorrelating these twf: they repeat at 3531 rather than the running speed of 3589. ?

There is almost no chance of getting motor current analysis.

This fan is the combustion air fan for the afterburner that serves 4 big rotary kilns. One of the kilns (the biggest) was out of service during these readings so there is a good chance of irregularities in the process. I'll keep an eye on it and see if it settles down when the process does. It just raised some interesting questions.

Thanks everyone,

autocorrelatedtwf.doc (92 Kb, 43 downloads)

It'll be interesting to hear what you eventually find out.

I think the difference between 3589 running speed and 3531 read from autocorrelation twf might just be accumulated numerical errors.

Let's assume when you selected the time waveform peak you selected one extra time sample (it looks awfully zoomed in and you don't have a peak select/interpolate for time peaks like for frequency I'll bet.... we don't have that in Emonitor). That would explain the difference.

Fmax = 120000 cpm
Fmax = 2000 hz
Sample Rate 4000 hz
Time Betw samples is 0.00025 sec
Speed = 3589 rpm
Speed = 59.81666667 hz
One revolution at 3589 is 0.016717749 sec
If we add one extra time sample (0.25 msec), that period becomes 0.016967749 sec
This is 58.9353378 hz or 3536.120268 cpm

One (old) method for looking at this is to count a number of cycles and look at the delta time for n cycles. Divide by n (n cycles - don't over count the first peak).

Danny,

One more request; A PV AC plot.

Danny,
Have you acquired any high frequency data. I noticed that your PV data is a 2000Hz HP, That tells me that whatever you are seeing there is actually occuring in the frquencies above 2000Hz. Try acquiring a 5000 Hz fmax, 6400 line spectra and see if you have any turning speed modulations occuring in the upper ranges. You may even have to go higher than that, but it is a starting point. I have found that if I see it in peakvue, but not regular data, I can start acquiring data in the frequency ranges above the HP filter to at least tell me where what I am seeing in Peakvue is coming from. You will probably see your waveform amplitudes seem a little more in line with the PV WV also. I believe that if you autocorrelate your PV waveform, TS impacts is what you will see.

Pete,

I was prepared to say that I would never find out what the cause is but I paused to look back to see if this was truly a one time event and found that this started in August and has increased since then. Amplitudes have just gotten high enough for me to notice them. If it continues to increase, we may find out more. If it goes away when the process returns to normal, then I will probably never know.

I don't think I fully understand your explanation for the difference. I'll think about that some more and probably have some questions.

Bill Foiles,

From the autocorrelated waveform posted above:

11 peaks in 183.79 msec

183.79/11=16.70818 msec=.01670818 sec=59.851 hz=3591 cpm.

The explanation is in Pete's post I suppose. I just don't understand it yet. I'll get there.

Bill Kilbey,

The autocorrelated PeakVue twf is attached. Any comments on this or the other autocorrelated twf?

Viberscott (I'm sorry, but I don't remember your real name),

No, I don't have any higher frequency data. We are only at this site once a month and there is no one on site anymore who knows how to work their 2120. I will see that special attention is given next month.

I posted an autocorrelated PeakVue twf and what I see is running speed impacting but I'm not that familiar with interpreting autocorrelated waveforms. I'm hoping that you experts will clue me in on that.

David G.

I don't see the pattern you are describing with modulation every 5th impact. The variation in amplitudes that I see are not high enough that I would suspect modulation every 5th impact. I'm not saying that it isn't there, but sitting here with the cursors rather than just the one copy of a plot, I cannot see it. Do you see anything in the autocorrelated waveforms?

Thanks all

autocorrelated_pvtwf.doc (28 Kb, 19 downloads)

I believe what you will see in high frequency data will be something is generating a high frequency fundamental frequency(similar to gearmesh or rotor bar frequencies) with turning speed sidebands. The trick will be finding out what is generating that fundamental frequency.(I guess that's what keeps our job interesting

Taking a number of cycles to find the period is a form of averaging. This method has been fairly standard for a long time. One used it with O'scopes.

Another use is doing ringdowns from a bump test to determine log dec. One uses a number of cycles to improve the accuracy.

In theory in a perfect world with perfect resolution one cycle should give you what you need, but taking an average over a number of cycles improves the estimate - These are all estimates.

I guess I have a knack for making simple things complicated (sorry). I was just trying to say that a tiny error (0.00025sec) in selecting the endpoints of your cycle can cause your error in computed frequency. As Bill pointed out, that tiny error in selecting the endpoints gets spread over more cycles and has less effect when you use his suggested method.

The AC PV WF shows very clearly it is amplitude modulation as the 1x pattern is changing periodically. Now back to the hard part; What is causing the modulation?

It's not just that "the" error gets spread out; it's the accumulation of errors in the final sum that helps. If all the errors were on one side or the other, the average estimate would be far off, also.

Sorry for another detour, I just wanted to respond to Bill's comment.

Bill - I think you and I are both right in explaining the benefits of your approach (averaging more cycles), we are just starting with different assumptions.

You were assuming the source of the error is within the signal itself. In that case, statistics tells us that the averaging process reduces the standard deviation of the estimate as you say. (by the square root of the number of averages)

I was assuming the source of the error is associated with the selection process of the endpoints.

Under my assumption, whether we select one cycle or 10 cycles, we are still selecting two endpoints so we will likely end up with the same time error. That same time error has a smaller effect when we spread it out over more cycles using your approach.

As an illustration (for everyone's benefit... I know Bill is familiar with this), let's compare Danny's approach of using 1 cycle and Bill's approach of using 10 cycles (for example)

For round numbers, assume that the error in selecting endpoints is 0.01 seconds, and the true cycle length is 0.1 seconds (corresponding to 10hz).

Danny's approach sees one cycle as 0.11 seconds (0.1 actual plus 0.01 error).
This approach computes F = 1/0.11 = 9.09hz.

Bill's approach sees 10 cycles as 1.01 seconds (1.0 actual plus 0.01 error).
He computes T = 1.01/10 = 0.101.
This approach F = 1/T = 1/0.101 = 9.9hz.

Bill's approach is a lot closer to the actual (10hz) because the 0.01 second error was spread out over 10 cycles instead of just 1.

Danny,

What I see in the autocorrelated TWF is an obvious fact that there is modulation occurring at 326 CPM.

But we are talking of demodulated data, meaning that there is impacting/friction taking place and being modulated ( similar to inner race defect ). So, modulation is not due to load related speed variation. For unknown reason it is rather impacting with 5th and 6th impacts being highest within the modulation cycle.

Could it be impellor-housing contact or flow related issue due to eccentric impellor? How many blades are there? 11?

Can you show normal spectrum with higher Fmax (at least to see 11xRPM?

David

For n contiguous cycles, measuring the total length of time and dividing by n is equivalent to picking the n+1 peaks and averaging the period for each of the n cycles. Some cycles will measure long and some short, no reason to assume a bias here.

It should work by picking n cycles at random, contiguous or not, and averaging the periods, too, but this seems much more difficult.