Skip to main content

Dear all,

In an on-line vibration monitoring system for a VFD controlled pump, I have observed a large variation in measured vibration data [RMS velocity and acceleration]. Now in order to decide the health states or for trend analysis, what should one do to deal with the fluctuations. Can you suggest any smoothing technique or indicator insensitive to such fluctuations ? 

Thank you 

Regards,

Akashri

Original Post

you won't have any good result on trending different speeds, I would trends within same speed.

you will have to program something to filter and record a different speeds.

also for spectrums do order based setup and your fail frequencies will be aligned, a bearing BPFI, PBFO, ftf, etc will always be the same non synchronous peak.

Last edited by fburgos

r.m.s. is a statistical method, basically the variance of the distribution (since mean zero vibration signal).  

One can try a moving a average to 'smooth' the data (you get a lag).

If you have something that an do Wavelets.  Try de-noising using the Wavelets.  This can be good if you have regularly spaced data - may not have enough points to make it of any use.  Similarly, empirical modal decomposition, EMD, might work if you have enough data to make it work.

 

The moving average is simple.

A well known principle of condition monitoring is to compare vibration data at the same operating condition (speed and load). I have not seen any analysis software that automates this process. The simple way is to define vibration limits based on the highest vibration operating point. Other methods could be implemented depending upon the features of the monitoring system or if data can be pulled out for special processing. What is your system?

"also for spectrums do order based setup and your fail frequencies will be aligned, a bearing BPFI, PBFO, ftf, etc will always be the same non synchronous peak."

The problem with using orders is that the amplitudes will still vary depending upon speed and load. Setting alarm levels based on highest vibration operating point would be necessary. Narrowband spectrum with selected band levels or envelope alarm could improve the quality of alarms compared to simply using overall levels.

Walt

Last edited by Walt Strong

agree order based its not for alarm limits, but your peaks will be in line. at least it will be easy to spot any new nonsynchronous peak at different speed, at scada level your IT crew could figured out something,

for example save [time,speed,amplitud] that could be later filtered and ploted by "speed" or more complex scenarios like generate an average and alarm for each speed

Hi Akashri,

if you monitor pumps using MCSA (electrical waveform analysis), the speed of the pump is taken into account. MCSA system calculate the load and speed of a pump as part of the overall analysis, compensating for deviations in patterns that emerge as a result in changes in load and speed, distinguishing them from changes in patterns that are a result of upcoming failures.

This improves the accuracy of the analysis and eliminates false positives.

In addition, the information about active power and speed - inherent in MCSA-applications - allows for the display of a workpoint of a centrifugal pump relative to its Best Efficiency Point, offering insights to optimize the performance and lifetime of a pump, bearings, and seals. More information here: https://www.semioticlabs.com/r...real-time-pump-curve

Happy to schedule a call to discuss the technical details! 

best regards,

Simon

Thank you Simon, for your response.

It seems to me that we can prefer MCSA over vibration analysis for electrical faults identification and power quality analysis. However, for early identification of mechanical faults, vibration analysis would be more effective. What is your opinion ? I would also like to know about the effectiveness of MCSA for VFD operated pumps (if speed is varying during current and voltage measurements) and two-pole motors (where rotating frequency is close to line frequency) ?

Regards,

Akashri

Hi Akashari,

MCSA detects mechanical failures on both the motor and the driven equipment reliably and consistently by analysing the data based on artificial intelligence algorithms - weeks or months in advance. Please see the attached picture. 

*High-end vibration systems will generally detect mechanical failures earlier in the process - but will miss a lot of electrical failures. 

*Vibration systems, when there are 2 or more identical bearings (for instance), will localize the exact bearing that is failing, where MCSA will detect that a bearing is failing (butnog tell you which one). 

*VFD-operated pumps can be monitored just as effectively. In fact, MCSA calculates the speed of the pump when analysing patterns, and will take the load and speed into account, reducing false positives. 

*2-Pole motors are indeed tricky. There are some failures that are difficult to detect in 2-pole motors using MCSA. 

Please let me know if you'd appreciate a call to facilitate a discussion!

Best regards,

Simon

 

Attachments

Images (1)
  • Non-exhaustive list of detectable issues

Simon,

That fault table makes little sense to me about "how far in advance we can detect this?". What is the basis or reference measurement that establishes when the fault actually started and the weekly or monthly clock is running? For example, if a motor rotor bar cracked, then why would it take weeks to detect? Why does it really take "months" to detect pump Cavitation? Most of the faults listed are either present or not when using portable survey or online monitor. The severity (amplitude or fault cascade) may change over time, but the detection is nearly real-time and certainly not weeks or months.

It is good to know that at my age that I can still make a fault detection and diagnosis much faster than artificial intelligence!

Walt

Hi Walt,
Thank you for noting that the table is unclear. The period mentioned (weeks/months) indicates the typical remaining useful lifetime of an asset after the first detection by the system. So the MCSA-will detect in real-time, weeks or months before catastrophic failure. 
Whilst I am not familiar with your age, I completely agree that you can detect failures better than an AI-based system! In my view, the best condition monitoring "tool" is a trained condition monitoring specialist, using years of experience, domain knowledge, and common sense, to monitor equipment. Add the distinctly human quality of creative problem solving to the mix, and you have the ultimate monitoring machine: A maintenance engineer specialised in condition monitoring. 
But across the globe, some 300.000.000 motors power pumps, compressors, conveyors, etc. power the global economy. And there are not nearly enough trained condition monitoring specialists available to monitor them all. And that is where technology comes into play: The combination of high-quality data and powerful algorithms can monitor at a very large scale, with a high detection rate and very few false alarms.
The combination of the two is the most powerful: A condition monitoring tool to monitor at scale, alerting specialists to upcoming failures so they can focus their attention on assets that are failing - and not continually spend time verifying that an asset is operating perfectly fine - which is the most common state. 

"The period mentioned (weeks/months) indicates the typical remaining useful lifetime of an asset after the first detection by the system. So the MCSA-will detect in real-time, weeks or months before catastrophic failure."

That description makes more sense to me than the table title!

How is the projection of time (weeks or months) to failure made? Is there any definitions of machine "useful life" and "catastrophic failure" that could have many variables involved? What about the detection of two or more faults at about the same time? How is the presence of background vibrations from possible identical or other machines affect fault detection and "life" estimation?

Walt

Hi Walt, 

My answers are here! 

  • How is the projection of time (weeks or months) to failure made? 
    The remaining useful life of an asset is estimated based on experience and assumes an asset runs 24/7. Because most assets will run less than 24/7, we under-estimate the RUL by design: Better safe than sorry.
    By experience, I mean that we have run tests in our workshops, in laboratories working with a large steel conglomerate, and by monitoring since 2015. We have empirical data as to how long in advance we typically detect a failure. However, we describe remaining useful lifetime in broad terms, acknowledging that it is an imperfect metric and should be regarded as an advice, not a definitive answer. Note too that many factors have an effect on the RUL for any system to be highly exact here. 
  • Is there any definitions of machine "useful life" and "catastrophic failure" that could have many variables involved?
    What do you mean by "many variables involved"? 
  • What about the detection of two or more faults at about the same time? 
    That's possible. If multiple failure mechanisms are at play, we'll observe multiple failure frequencies in the data. There are situations conceivable where failure modes that have very similar patterns will occur simultaneously. In those cases, they will both be reported as the potential culprit. 
  • How is the presence of background vibrations from possible identical or other machines affect fault detection
    The effect is not zero but does not impact detection and classification negatively to an extent that it hurts the value of the analysis and in that sense mostly a theoretical issue. 
  • and "life" estimation?
    We have not researched this, but as a general statement, I would say that background noise does not help. I'd think in effect it does not hurt either. 

Add Reply

Post
×
×
×
×
Link copied to your clipboard.
×