Sign up Calendar Latest Topics Donate
 
 
 


Reply
  Author   Comment   Page 1 of 2      1   2   Next
BaartCM

Member
Registered:
Posts: 9
Reply with quote  #1 
Hi Guys,

I read this paper some time ago and files it away for future use. http://profs.etsmtl.ca/mthomas/Publications/Publications/A19-sassi.pdf

I would like to try using these new parameters in an app I am writing for continuous monitoring of pumps at a large plant in north east England however, the paper gives no indication of the sample rate or sample length. I am collecting 2048 samples at 2560 sps so I can get an 800 bin spectra with 1khz fmax to do ISO calcs and was thinking of determining the thikat and talaf values from this sample.

Has anyone used these parameters? How successfuly? What sample rate and length would you suggest?

I’d be grateful for any advise.

Regards, Steve.
OLi

Sr. Member
Registered:
Posts: 1,478
Reply with quote  #2 
What sensor do you use? All starts with the sensor.
__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
OLi

Sr. Member
Registered:
Posts: 1,478
Reply with quote  #3 

Well I have that Utopian parameter since 20 years.... For normal speeds anyway, Yankee cyls do give lower signal levels but the trend is there and NC spindles similar.
Anyway I have so far limited my compromising to at least measure up to effective 12 kHz minimum for reasonably early problem detection, and we used that for 15+ years in 350 instruments world wide. if you like to have up to 1 year time to plan spares and repair.... 1 kHz is no near enough IMHO my parameter is based on a HP 600 Hz and you don't get much left then, for sure not enough. You get a final warning but no time to plan. On electric motors you should at least go to 1500 Hz but I guess your sensor system may not allow you that and then your selection could be better..... 

I do object to the simulated data, depending on what you use it for. To me it is like cutting a scratch with a angle grinder across the race, it does not look like anything near a real bearing fault. I would prefer the data from the Red Army running down more than 20000 bearings in controlled environment...... 


__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
Walt Strong

Sr. Member
Registered:
Posts: 709
Reply with quote  #4 
Here is a link to one of the authors of the paper you referenced:
https://www.linkedin.com/in/sadok-sassi-627b8033/

Perhaps he can recommend accelerometer and measurement settings compatible with the parameters described in paper. It would be interesting to see if the author has made extensive measurements with the "parameters" since the paper was published and still recommends them. I am not saying they are good or bad for any specific application.

Walt
BaartCM

Member
Registered:
Posts: 9
Reply with quote  #5 
Thanks for the replies,

The sensor I am using is http://www.hansfordsensors.de/assets/Uploads/Datasheets/HS-104-Datasheet.pdf hs-104 from Hansford sensors with frequency range 0.3-10000 hz and 250 mv/g output.

I have already tried emailing dr Sadok using his Qatar University address I found online and am waiting a reply however, I feel his experiences may be on a test rig in a lab whereas you guys, like me, are working with ‘real’ data. A newer paper from Dr Sadok also gives two further parameters. I am putting them all into my software and when my ‘normal’ trends indicate a problem, I will cross reference with these values to see if they could have given earlier warning.

Basically once it’s installed next month, my software will be requesting a short twf (2048 samples at 2560 sps) every five minutes from which it derives my normal trend parameters ( acceleration rms, velocity rms etc.) and also other values like kurtosis, Skewness etc. These values are displayed on the pc up to a weeks worth at a time. Every seven days, a full twf of 8196 samples at 5120 sps is stored (more frequent if the velocity or acceleration alarms are exceeded). There are 16 pumps with 2 accelerometers on each so it won’t be long before there’s a plethora of data to look at!

Regards, Steve
Curran919

Avatar / Picture

Sr. Member / Supporter
Registered:
Posts: 452
Reply with quote  #6 
Cool paper. Like they say, RMS is monotonically increasing with bearing damage, but not very much, while statistical bearing health indicators like pk-pk, crest factor or kurtosis behave less predictably as a defect gets 'rounded out'. Sometimes they decrease sharply, sometimes they remain constant or even increase slightly still.

This is an example of feature generation, forcing novel, complex quantities from simple, known quantities to generate some random THIKAT with the behaviour that you want. The first has a monotonic increase, which can be nice for some monitoring functions, the other appears to have a sharp crest and fall when the defect is mature... also a useful behaviour.

They could be useful, but I doubt they are going to take the world by storm. These statistical parameters only have so much information, which is why most analysts use parameters based on HFNBRIs like peakvue/enveloping/SPM. Wringing them out to try and get more information can only get you so far. But of course, you have lots of pumps that you can try these out on. That is pretty valuable data... for not only you. Try out different parameters. See what gives the best indication.

For the statistical parameters, 5kS/s is probably high enough. Higher is usually better (10-20kS/s). HFNBRI parameters need much high though, as Oli said.
BaartCM

Member
Registered:
Posts: 9
Reply with quote  #7 
Thanks curran919, I can easily sample up to about 35ksps but I have to write it into the firmware of the modules before they get installed. The pumps are in UK, I am in Cyprus. It would be easy matter for me to add extra readings into the modules now - basically, the client pc sends a 3 character message to the module it wants data from and waits it’s reply, it’s ST1 to ST4 to get the short twf 2048 @ 2560 sps and LT1 to LT4 to get the long 8192 samples @ 5120 sps. The modules take all the data first and then transmit it via Ethernet to the pc so there is no possibility of data transmission effecting sampling rate.

I have designed the modules and written both the firmware and the pc software so control is very much in my hands .... until they are installed!

If you think additional sample lengths and rates would be useful let me know, I can add them now, later not so easily. Note though, 8192 is max number I can sample and hold in memory and additional memory is not an option.

Regards, Steve
OLi

Sr. Member
Registered:
Posts: 1,478
Reply with quote  #8 
You need to comply to Nyquist sampling theorem of nominal 2x IRL std 2.56 times highest freq used and good anti aliasing filter.... so to get 5kHz data y need sample rate 12.8kHz or y will aliasing badly... nad be wrong, sensor is ok. 
__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
BaartCM

Member
Registered:
Posts: 9
Reply with quote  #9 
Oli, if you see in my first post, I said ‘I am collecting 2048 samples at 2560 sps so I can get an 800 bin spectra with 1khz fmax’ so I think I have already demonstrated I have considered the Nyquist theorem with respect to my ISO calculations. However, this only applies to doing an FFT, and as my post was to do with calculation of scalar values from a time wave form where the Nyquist theorem doesn’t apply, I’m not sure what you’re getting at.

Regards, Steve.
OLi

Sr. Member
Registered:
Posts: 1,478
Reply with quote  #10 
Well in my book antialiasing, Nyquist theorem apply to all calcs, and sampling not only FFT's and I think it is not only me so I don't know what y are doing either...... Thank y for your kind remark to someone that done this for 30 years that is trying to help y to be a competent competitor I guess...
__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
Walt Strong

Sr. Member
Registered:
Posts: 709
Reply with quote  #11 
"There are 16 pumps with 2 accelerometers on each so it won’t be long before there’s a plethora of data to look at!"

All the more reason to be careful what you are measuring and how you are interpreting it! A lot of data without value is data diarrhea! 

"my post was to do with calculation of scalar values from a time wave form where the Nyquist theorem doesn’t apply"
I generally agree, but for the RMS and other statistical values to be relevant, then the Sample Rate should be much higher than the expected frequency of interest. Consider "seeing" a Sine wave in the time domain; it takes 5 or more points for visual recognition.

I think what you are tying to do is commendable, and I will not start or add to a discussion on bearing fault diagnostics or signal processing methods. Please share your results, since there are more complainers than doers!

Walt
OLi

Sr. Member
Registered:
Posts: 1,478
Reply with quote  #12 

I sincerely test in my lab before applying to real world, it is cheaper that way in the long run experience show IMHO.... and I learned that the hard way, more than once and sometimes
costing 2-4 man year work just to prove a solution is bad but rather that than sending off any in to the real world with aliasing problems or similar. I have seen people do that IRL even in the last 5 years with disastrous aliasing giving huge credibility issues and sure incredible analysis problems trying to analyse aliased data with spooky peaks that do not compute everywhere, go figure what nightmare......  So be careful guys and girls, do not thread to hastily, reality may bite you seriously. I know as I done that been there... But I have not released any aliasing crap, you can be sure of that. Also computing running average on a TWF may look mundane in the math book, not really always so in real world calculations believe me. Don't forget the numerical book also... Also don't forget that decimation also require serious anti aliasing filtering that works IRL a thing that like Analog Devices completely violate in their digital output devices.... just saying and I can prove it by hard work the hard way that we did, it was that took took us 4 man years to understand what happened as we could not believe what we saw.....


__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
BaartCM

Member
Registered:
Posts: 9
Reply with quote  #13 
@OLi firstly, I have no aspirations to be a competitor, I’m doing a job for a company and want to do it to the very best of my ability. Secondly, I started doing vibration analysis in December 1985 so I do have a little experience and understanding and my understanding of an alias is that if you sample 1024 data points at 200 sps, you will get a spectra with an fmax of 100 Hz and 512 lines. If you had a frequency of say 105 hz in your sample, the fft spectra would show a peak at 95 hz which is an alias of the actual 105 hz frequency. It does not exist, it just seems to exist. Therefore the fft spectra is not a true representation of the actual vibrations. Yet, the time waveform IS a true representation, you can’t see aliasing in the twf, only in the fft spectra. In reality, we don’t use 512 line spectras, we use a sampling rate that gives us the fMax we want at 400, 800 etc lines which is why we multiply by 2.56 not 2. In addition to dropping the top part of the spectrum where aliases are much more likely, we can employ digital of analog filters to remove any aliases.

Let us not forget either that the fft spectrum is actually only the ‘best guess’ of what the true frequency based spectrum would look like, do an fft on a square wave for instance and you get harmonics of the square waves fundamental frequency. These harmonics don’t exist either, they too are aliases but sometimes they are aliases we want to see because they can indicate looseness because a slack bolt is turning what would be a nice sine wave into a pseudosquarewave (I’ve just invented a word or at least used one that wasn’t in my iPads dictionary lol).

I appreciate that selection of the correct sampling rate is CRUCIAL to getting meaningful information from your data. Hence my question “what sample rate and data length do you suggest”

So far, that question remains unanswered.

@Walt, I don’t intend to interpret these values until such a time as the velocity rms or acceleration rms and the spectras indicate a definite problem. Then I’ll look back and see how helpful they may have been. Until I understand them, they are indeed ‘runny poo’

Best wishes, Steve.
fburgos

Sr. Member
Registered:
Posts: 425
Reply with quote  #14 
i´ve been thinking on this because the data in the "wave" have to be representative for whats happening in the machine.

since this is a two part questions

"sampling rate".... this will give the resolution of the waveform more sps more points in the wave better but youre limited by the memory.

"lengt of the wave".... I would say between 3 and 5 times the lowest frequency tipically this will be FTF.

first you already defined that you will collect 2048 samples points, then your sampling frequency will give you the lengt of the waveform needed.

I will asume a FTF of 0.4x, the inverse is 2.5, every 2.5 shafts revolutions my cage have one full rotation, for to have 5 rotations of the cage i will need 12.5 rotations of the shaft, this is the length of the waveform you need, what are the pump speeds? 

your current selecion 2048 samples at 2560 sps, will be 0.8 seconds length waveform.








BaartCM

Member
Registered:
Posts: 9
Reply with quote  #15 
@fburgos Thank you for your input. I like the angle you're looking at this from

The pumps are all two pole (or before somebody corrects me, the motors are) and we have 50Hz supply so a touch under 3000 rpm so your 12.5 revolutions will take roughly a quarter of a second.

Now, I have already modified my firmware so that I can collect 2048 samples at either 4000 sps or 6000 sps but to collect 2048 samples in 0.25 seconds requires a sampling rate of 8000 sps which would just need a few more lines of code and an extra command in my comms protocol. I think I'll do that, Im only using about 32% available program memory anyhow, its the dynamic memory for variables which Im short of!

Cheers, Steve.
Previous Topic | Next Topic
Print
Reply

Quick Navigation:

Easily create a Forum Website with Website Toolbox.