I am trying to find out percentage of analysts surveying large number of machines who have set up and are using narrow band statistical alarms (enveloping).

Question:
Do you use statistical alarms ?

Choices:
On all machines
On some machines
Do not use at all

 

I use multiple systems. On one system for one plant I have set up - the bands are spot-on and provide machinery condition by a type of banding or failure reporting from amplitude analysis.

So with some systems I do not use at all as I do not have the necessary confidence in the system to rely on its diagnosis.

I'll vote "some machines"..............

A big picture seems to be emerging but more entries are needed for better accuracy.

Maybe then we can discuss factors preventing people from using statistical alarms. It is hard to believe that 10-12 wideband alarms are better. Of course I do realize that prior to establishing a statistical alarm one has to accumulate at least 12 readings and this is one year time.

Thanks,
Dave

I do use overall statistical alarms as well as fixed alarms. Unfortunately AC frequency drives are in extensive use with most of my customers this tends to bring up many false alarms due to drive harmonics (2 X supply frequency and higher)

David,

That's a good point. In my case, I just have for some equipment 3 o 4 readings and I have been thinking to "experiment" using 6 spectral bands as a solution, because I have a lot of motors in my CBM program.

Regards
ZaB

quote:

Originally posted by David_G:
A big picture seems to be emerging but more entries are needed for better accuracy.

Maybe then we can discuss factors preventing people from using statistical alarms. It is hard to believe that 10-12 wideband alarms are better. Of course I do realize that prior to establishing a statistical alarm one has to accumulate at least 12 readings and this is one year time.

Thanks,
Dave

David,

I start with 5 or 6 readings, or use my category to lump several different machines together if they are all running about the same levels for that point. Then, as time goes on, I re-run the statistics. I believe that is much closer than just setting a generic overall. I also use narrow band alarms along with the overalls. It is especially attractive on those machines that run down in the 0.0XX ips range, and trend steady.

Back in 1996 or 7 a man I had a contract with once told me his expectations of my job.

"I want you to tell me what is going to happen before it happens, and be 100% accurate. If you do that, you can stay here till you want to quit, if you can'[t do that, you won't be here long"

I spent too much time analyzing and second looking at stuff, but I stayed. Catagories, Narrow bands and statistical alarms were a major part of that contract.

Dave

I have tried it a couple of times. I find I spend more time tweaking and setting alarms then looking at data.

I like looking at data. I think I have less risk of missing something that way.

You can spend a lot of money and have someone else set up your alarms. But, I am back to not trusting any alarms that I did not set up. So I look at the data. If you set up the database right you can go through a lot of data fairly fast.

I also like looking at the data instead of relying on alarms even set statistically. For people monitoring large number of machine this is not an option though.

Here is an example of input data which, if used for statistical calculations, probably won't be very useful since StDev is relatively large in comparison with the Mean and therefore Alarm will be set unrealistically high ( one can possibly adjust it manually).

One more concern... Even when utilizing statistical alarms, it is still beneficial to look at the trend pattern as oppose to confirm absence of penetrations, so again, one has to look at all machines, points, and bands.

Large_StDev.doc (36 Kb, 21 downloads)

One trick I have used in setting up statistical alarms for TWF amplitudes is run the numbers (mean & standard deviation) and then throw out the values that exceed +/- 3 sigma (standard deviation) and re-run the numbers. If you are limited in the number of data points, getting rid of the outliers with definitely tighten up the standard deviation.

Also you may want to run the statistics for the individual machine and as a group of identical/similar machines to see how different they may be.

I tend to use the statistical and narrow-band alarms as a "flag" to drive me to look at all of the data.

Gentlemen,

I thought getting rid of outliers was common practice before generating the statistics.

With Odyssey, it's only a click away, and you have three methods with which to get rid of them.

Dave

quote:

Originally posted by RRS_Dave:
Gentlemen,

I thought getting rid of outliers was common practice before generating the statistics.

With Odyssey, it's only a click away, and you have three methods with which to get rid of them.

Dave

Curious if you or anyone knows what those methods are doing?

I was specifically talking about TWF alarms based upon peak acceleration. Can you set up statistical alarms in Odyssey for TWF? I have exported the data to Excel and performed the necessary calculations to develop my alarm limits.

Steve,

No, I do not know if you can set them for TWF. I have never went that far and tried.

However, to explain the other methods of getting rid of outliers, I give you these:

For the Grubbs Test: http://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/grubtest.htm

For the Median of Absolute Deviation method:http://en.wikipedia.org/wiki/Median_absolute_deviation

For the manual method of looking at the trend and saying, " I believe I do not want that point to be included in my statistical calculation, even though ole Grubbs or the MAD did not get rid of it" I right click, select "Storage Flag", then select "No Stat" from the drop down.

Those are the three methods, each works on different looking distributions, differently. When Grubbs may not find any outliers, MAD may find 2 or 3, and vice versa.
Ole manual finds them the easiest.

Dave

Good info Dave.

I'm familar with changing the Storage Flag. Sometime I even change it back to Archive when Grubbs selects it for no aparent reason.

Steve,

I have read the Grubbs thing about a zillion times. Every time that rascal makes a pick that I don't understand.
I too have went the way of changing to archive when I don't agree.

Dave

I tend to use statistical alarms on all constant speed machines and sometimes on some variable speed machines providing that they don't vary too much.

Admittedly, statistical alarms will generate a lot of "false alarms". By that I mean they will highlight data that requires no action. I am happy to perform analysis in this way as opposed to not looking at the data at all. At least I make the decision myself as to what is acceptable and what isn't. Together with magnitude alarms I feel that I've got everything covered. I only analyse machines that are in alarm and I don't miss any failures.

A spectral alarm will often detect a fault condition long before the trend goes into alarm.

I am also using Odyssey. Typically I base my alarm on the last 12 months worth of data (using the archive filter). If I feel that this isn't a true representation of how the machine normally runs, I may choose a wider range on which to base the alarm. I decide myself which data to choose as "No Stat" as I feel that the both the Grubbs and Absolute Deviation methods exclude data for no reason that I can fathom. Providing that the data outliers are excluded, then a meaningful alarm can be generated and updated quickly and easily. I don't generate alarms based on categories. They are always based on the individual point.

I use statistical alarms on like machines and fixed speed. Never had good results with variable speed machines. I have some machines that change speed constantly. No luck with narrow band alarms on these machines. Only 10 of them so I look at them every month.Does anybody use alarms on machines that change speed while collecting data?
dnk

quote:

Originally posted by DK:
Does anybody use alarms on machines that change speed while collecting data?

DNK,

If data acquisition time is 0.8 sec = 1600 lines/ 2000 HZ, then speed change may not be a factor even for spectral analysis itself ( the peaks won't be smeared very much). But if during one survey speed is on average say 1700 RPM and during another - 1400 RPM, then it may become a problem for statistical alarm calculation.

I have not tried it myself but, assuming speed variation is not large ( 10-15%), couldn't orders be used for statistical calculation and be valid?

Dave

This message has been edited. Last edited by: Dave_man,