My mundane post about setting the overlap on a 2120 seems to be evolving into a very interesting discussion on data averaging and the reasons for and against it. It is one of those topics that I have never really considered. It was taught to me to get 4-6 averages, so that's how I do it.

So far we have some quite convincing arguments against averaging:

http://maintenanceforums.com/eve/forums/a/tpc/f/3751089011/m/1781012043

Now that we have officially changed the subject, would anyone else care to comment?

This message has been edited. Last edited by: Danny Harvey,

I have so many times sat in the van and watched vibration signals 'real time' and capture that signal.

Averaging helps pull things out. 32 averages are optimum; 16 are very good and very repeatable but 4 works well for PdM route programs. Averaging may be done from an internal/external clock or event or key phasor device as well.

Averaging is a tool - is it one that is needful for your function at the time for your particular application? The taking of a 'snap-shot' with high LOR is not the same thing as averaging. There's a difference from time passing and averaging.

I'd love 32 averages and have had to use > this for some applications, but time is money. 16 is good, not always practical for routes.

No averages - Did you get a peak or valley?

With many measurements I not only like the average (mean value or median usually), but providing the variation (or std, or variance) for the measuremnt often aids in understanding. This can be done by plotting the average and an error bound or variation around it (point for average, line for variation).

This doesn't seem to be a standard feature for spectral plots. I think it would add value in a number of cases.

Averaging prefered - measuring variation of spectral peaks prefered.

Sam mentioned watching the real time signal (spectrum?). This is the type of value a skilled analyst can provide by looking at the data. The variation can tell much.

A plot of peak hold vs a plot of average shows this to some degree.

For in-depth analysis variation could provide valuable information, but for the most software types it is not an option, so variation becomes transparant for an analyst anyway.

From this prospective, for a survey program, a better resolution instead of 4-6 avgs provides more benefits IMO. At the same time spectral peaks variability will be clearly reflected in the spectrun as sidebands or mounds around those peaks.

David

A simple waterfall plot presents the variation, also.

Bill and Sam, not sure if you read the other posts, but there is a little more explanation over there.
You know if you are using 50% overlap with 4 averages, then as far as shaft rotations are concerned you are only getting 2 averages of totally new data. If you use one average and set up your reading to get 50 revolutions of the shaft, then you are getting the same averaging if you set up to get 25 revolutions of the shaft with 4 averages and 50% overlap.
Here is a comparison using some formulas provided by that well-respected friend.
A 2000 Hz Spectrum at 1600 lines and 4 averages takes approximately 2 seconds resulting in a binwidth of 1.25 Hz. In 2 seconds a shaft running 30 Hz will make 60 revolutions.
A 3000 Hz Spectrum at 6400 lines and 1 average takes approximately 2.13 seconds resulting in a binwidth of .468 Hz. In 2.13 seconds the shaft will make aout 63 revolutions.
So for an extra tenth of a second I am getting 1000 more Hz of Fmax, with a resolution that is 2.6 times better, and still the same number of events averaged.

quote:

If you use one average and set up your reading to get 50 revolutions of the shaft, then you are getting the same averaging

1 average = averaging ???

Taking a long time sample to obtain resultion gives the equipment a greater chance to have speed variation during the sample.

So, 3000 Hz is 100 times 30 Hz. I've been bashed a few times in this forum for considering high speed equipment (whatever that definition is). Do we really need to go out that far?

Binwidth does not equate with abilty to distinguish or resolve sepearate closely spaced frequencies. Another question to ask does the bin width suffice for the job?

1st avg = 100%
2nd avg = 50%
3rd avg = 50%
4th avg = 50%

So, 400 Hz w/400 LOR = 1 sec + 0.5 S + 0.5 S + 0.5 S = 2.5 Sec.

An average equals an average; two don't become the same event - each avg is an event or seperate operation.

If you have an analyzer that allows centering on a given or specific frequency then you can have that frequency in the center of the bin or filter window.

Bill,

quote:

1 average = averaging ???
If you take 1 average then you are averaging the peaks that are in that time period within the FMax or at least I am thinking that is the way it is.

Taking a long time sample to obtain resultion gives the equipment a greater chance to have speed variation during the sample.
Whether you are using 1 average or 4 or 16, if you take the same amount of time to take the reading then shaft speed variation would be the same.

So, 3000 Hz is 100 times 30 Hz. I've been bashed a few times in this forum for considering high speed equipment (whatever that definition is). Do we really need to go out that far?
I was just using that as an example. For most faults you don't need to go out that far.

Binwidth does not equate with abilty to distinguish or resolve sepearate closely spaced frequencies. Another question to ask does the bin width suffice for the job?

I am really confused at what you are trying to say here. Binwidth has everything to do with resolving closely spaced frequencies.

1 average is really no averaging.

All the data from one collection (i.e. time ensemble) are converted by the FFT. So, the length of time in one conversion makes a difference. Every try doing a high resolution FFT on a coastdown?

Besides lines of resolution (bin width), the window choosen for the analysis affects the ability to resolve frequencies.

The FFT is an estimate for the spectrum. One gets a better estimate with averaging.

So what are we seeing with a spectrum that takes 2 seconds to acquire. If the shaft is turning 1800 cpm and there is unbalance then the heavy spot will be recorded passing the transducer 60 times. Which of those 60 peaks will we see in the spectrum? Is the amplitude reading that we get at turning speed not the average of all 60 peaks? I am sure that they are not all the same amplitude. Does it just pick the highest?

quote:

Which of those 60 peaks will we see in the spectrum?

That's not how a DFT works. You get the extra resolution with the extra time in the sample.

One way to think of it for a discrete FT is breaking the time series into a sum of sine and cosine functions, which is the solution of equations for the coefficients (which in turn are [after taking amplitude of like sine and cosine terms] are the spectral amplitudes). The more data one has the more terms up to the frequency limit one can calculate, i.e. greater resolution.

The windowing function works to reduce the waveform to zero at either end so that the sample is repeatable over time. One of the rules for a fourier transform. So data that is not close to the center of the waveform (timewise) has a reduction in amplitude based on the type of window used. Averaging helps to alleviate this reduction. If the waveform event doesn't occur in the center of one sample it probably will in some other sample being averaged.

Sean

Without a trigger! In some analyzers a pre-trigger on the 64th window or selectable is performed in conjunction with antialising filter.

quote:

If the waveform event doesn't occur in the center of one sample it probably will in some other sample being averaged.

But if it is not in all the averages, then the amplitude will be reduced.

If you have an interest of an event, then you should trigger that event. You may use a clock or some other devise. A good example of triggering and filtering to a specific event such as 1X for balancing is common but not inclusive.

It seems that there are not a lot of people in my corner on this one.
Remember when portable data collectors and analysis software came out, hard drive space was at a premium. It took a lot of space to keep all this data that we acquire. I think that is one of the reasons for averaging. You could sample a longer time period by averaging out the data. That would get the repeatable events in the longer time without having to keep the extra data, saving hard drive space.
With hard drive space realtively inexpensive now, we have the ability to keep a lot more data.

I think the good from averaging is the reduction of random events and enhancement of things that exist in all dataframes and primarily that´s what I would like to see, I hope. So "noise" is reduced and "real stuff" is popping up more. Or else I have been looking at the wrong universe the last 20 years. I can´t see that long measurement time w/o averaging will do the same magic since the averaging process isn´t there. On the other hand you maybe don´t need that, there may not be so much random noise? Olov

So you are saying if you select 1 average on the set-up you are not getting an average spectrum of the data contained in the length of the timewaveform. Then if 1 average is not an average then if you pick 4 averages you are only getting 3?

An FFT, as usually implimented, produces the RMS averaged spectrum of the input time waveform weighted by the window function, if any.