In setting up peakvue points in the db, should I stick with only 1 average as CSI recommends? Talked to someone else one time and he uses 4 averages. Says he gets good results. Just wondering what the majority out there uses on # of averages for peakvue.

Just one.

I agree with Lawrence. One

Been using 3, but trying 1 in some places. No problems yet.

Why would you use only 1 average (besides "that's what CSI recommends")?

In line with what Rusty asks? Why wouldn't one average?

Almost (or all) any meaniful measure can be averaged with advantage. If the measurement were random then certain averaging (with certain types of random) could cause problem. Hopefully, this can not be the case (random measurement).

As it it is well known spectral averaging helps to reduce amplitude of a random signal in the spectrum.

As was stated by CSI in the case of PeakVue, having even minute shaft speed variation during the measurement will be detrimental to spectrum quality. Don't understand why...

As David_G pointed out, CSI says that averaging will smear peakvue. I used 4 averages for years, but have used 1 average for most of my peakvue data for the past year. I got good data both ways.

David E

Hi all,
I used 4 or 6 averages before and i was getting mostly some junk readings only. But when i changed to 1 average, Peakvue readings become more reliable.
The primary reason for this is that the Peakvue time waveform has equal importance to the PeakVue spectrum. Analysers only store the final time block captured,no matter how many averages have been requested for the spectrum unless synchronous time averaging using a trigger is invoked. In some cases, the number of averages was increased upto 6 with the result that the bearing fault frequencies that were very pronounced in one time block data were eliminated by the averaging process because the machine was slowly but continously changing in speed during the averaging. I also agree with u that multiple averages will improve the signal-to-noise ratio in the spectrum. It is better to use more no. of line to reduce PeakVue spectral noise than averaging.

quote:

As it it is well known spectral averaging helps to reduce amplitude of a random signal in the spectrum.

Averaging helps to estimate the mean and the variation. The only way spectral averaging, not triggered time domain, can reduce the amplitude of any frequency is for that frequency component to have some of the average components smaller than the largest ones.

Statistically, with 1 or rather no averages you have 0 confidence in the measurement amplitudes.

If one can't average it how can one compare two from different time periods?

I used four averages with Peakvue for years and have experimented with it using one, two and three averages. I really can’t see the difference either way. Try it!

With PeakVue you are sampling at 100 kHz and storing the highest value in the period that is the inverse of the sample rate. If the speed changes it can shift the peak from one period to another not giving you as good of an averaged value in the spectrum. The impacting will still be in the waveform. As far as averages, what reduces the noise floor and gives a better averaged valus in the spectrum is time. I use one average with a little more resolution (time) to get the time needed to get a good spectrum.

Jim,

I understand that what you said conforms to the preaching of Dr Robinson, the father of PeakVue. The Noise Floor is by definition Random Noise that can only be made statistically more accurate by averaging. Using more Resolution (time) means that you are including more shaft rotations. How can the random noise floor be "smoothed" or be averagred? If shaft speed is unsteady, then I would think that peaks correlated to shaft speed could get "smeared" as well as if by averaging. Please explain, because I don't mind being wrong, if I can learn something new or correct.

Walt

Walt,

I am no signal processing guru by any means but I remember being taught that the noise floor can be reduced by either more averages on by taking more resolution for the spectrum. When I hear this it tells me that it is time that reduces the noise floor. I have experimented using both ways and have decided I would rather take more resolution and less averages and have the reduced noise and better separation of the close peaks in my spectrum. As far as smearing in the PeakVue data if the speed changed enough it would have the same result if you took more averages or more resolution. The main difference I see is that PeakVue only holds one value per period. This is why I always use a lower f-max with PeakVue than regular data. A lower f-max gives a longer delta time to hold the peak and it is more likely to be in a bin that will be more repeatable. It will almost always have the fundamental defect displayed in the spectrum so I only have the f-max set to 3 or 4 orders of the highest defect frequency I expect to see. I probably do more diagnostics in PeakVue with the waveform and autocorrelated waveform than anything else.

I don't know about peakvue in particular but I agree in general that increasing the frequency resolution is a more effective way to decrease the noise floor than increasing the number of averages.

Think about it this way. If I had a "flat" noise floor (same in each bin) from 0 to Fmax with X lines. The noise floor is Y. The digital overall is X*Y^2 (neglect the window factor).

Now what if I double the number of lines (2X) with same Fmax. The overall stays X*Y^2. To preserve the overall, the magnitude (level of the noise floor) has to decrease to Y/SQRT(2). This preserves the overall as 2X * (Y/sqrt(2))^2 = X*Y^2

Back to 1 average for PeakVue...

Does anybody know if in the PeakVue the peak value in a time period of T = 1/Fmax is always stored in certain time spot of period T (say, at its beginning)?

If yes, why then PeakVue spectrum is more volnurable to speed changes then a regular spectrum (provided Fmax, LOR is same)?

David,
PeakVue stores the highest peak sampled at 100kHz and stores it in a period that is the inverse of the f-max*2.56. This amplitude value will be for the entire period so any defect in the period with the highest amplitude will be stored. Where there can be a problem with a speed variation is if the defect impact moves from bin to bin. More time either by averages or increased resolution makes this more likely to happen with a speed changes. If this is a problem you can reduce the f-max to increase the length of the peeriod as long as the f-max is high enough to see the expected fault frequency. The good thing about PeakVue is the fundamental defect frequency us usually seen (except on some bearings you will see 2X ball spin defect as the fundamental, so you don't have to have the f-max high enough to see 10X defect frequency. This is mainly to help identify the defect in the spectrum. The more bins that have the defect out of the total number of bins in the waveform will make the defect frequency stand out more in the spectrum. This is the reason it is better to judge the severity by the waveform amplitude.

Jim,

While we are picking your brain, what about solid particle contamination in a bearing that possibly does not produce a uniform frequency from one period to the next? Would it be advisible to take more than one measurement to see how repeatable the spectrum is? On very low speed bearings (say below 10 rpm), many folks are not inclined to wait long enough for several shaft revolutions or to see if bearing is contaminated but not defective. What is your technique to address this issue?

Walt

Walt,
I assume you are talking about particles larger than the thickness of the oil film. If they are suspended in the oil and move around I would expect to see high amplitude energy in the waveform with no periodic content in either the spectrum or the autocorrelated waveform. A lot of times with just a regular lubrication problem you will see cage frequency in both, not always. If the particles are rolled into the race it would be periodic just as if the bearing had a regular defect on the race. If you want to see defects on very low speed bearings you must take the time necessary to get enough revs of the shaft. I have a gearbox running less than .4 rpm and it takes over 20 minutes to get each reading.

Thanks Jim,

My point was that it can take a long time to get enough data to determine if peaks are periodic of random, so both the wavform and spectrum should be viewed.

Walt