Sign up Calendar Latest Topics Donate
 
 
 


Reply
  Author   Comment  
BaartCM

Member
Registered:
Posts: 15
Reply with quote  #1 

Guys,

I am working on a project for a client in Malaysia who is looking to offer IIoT services monitoring many facets of industrial processes. I have been brought onboard for my Vibration Analysis knowledge. My client is also working with a Machine Learning (ML) company who will be looking to examine the data for ‘features’ indicating the likelihood of certain events but prior to being able to do that, they need to ‘learn’ by feeding into their algorithms, data which has a known outcome as well as data which led to no outcome. That way, the system can look for features in the data unique to certain events allowing that event to be predicted in the future.

From my perspective regarding vibration analysis, I suggested to the ML folk that it may be an idea to ask if VA data sets which had led to a known outcome could be obtained from the many excellent VA Engineers globally in exchange for these contributors being given some form of access to the ML system once it is up and running and they indicated that they were open to this idea so basically, I am asking you for data. Loads and loads of data!

What we really want is time waveform data. Let's say you have a fan on which you detected an inner race defect which was verified when the fan was repaired and you have acceleration TWFs from the bearing for the previous 6-12 months then this data could be fed into the system to look for features indicating that this exact defect was occurring and you also have several other known outcome problems with similar data sets, then these data sets can also be examined for features giving us a hit list of different features to look for in your data to indicate when a specific problem is starting to occur, possibly much sooner than we could detect them by analysing the spectras.

Obviously, the more data we have, with different faults, speeds, etc etc the more accurate this system can become and the more information you can supply about each set of data and the outcome, the better but even if you don’t have exact speeds, bearing details etc and simply know that this data set culminated in a bearing failure, that data is still useful.

If any of you would be interested in collaborating with us on this project, then please get in touch with me. I will then forward your details to the ML folk who can let you know how to get your data to them.

I would be very grateful for any assistance you can give and thank you in advance. Please contact me on s.ferry at efftek.co.uk

Steve.

OLi

Sr. Member
Registered:
Posts: 1,874
Reply with quote  #2 
That is why that road have not worked since the computer capability came 20+ years ago, there are no compatible, standard format data available so you can't train the ML w/o waiting the normal 3 year apprentice time you train a human AI and then you need data from all machines and they need to generate the standard faults. So that is the reason in my view there are no commercial systems based on that type of AI that works, correct me if I am wrong. Sorry for that. Those existing 4-5 that I used including my own are fixed rule based and that works to satisfy the 80% out of the 80% most common fault criteria and I used it in 350 units for 15+ years....... You only need 50-60 rules to cover that and the speed of the machine to fix that if you think about it. Sorry for the bad news. I once 10 years ago a 3-letter corp announcing a system directly connected to the work order system, still waiting to see that. Only system I seen that worked like that was a "hi vibration alert, generating manual analysis to vibration dept. work order" system.
__________________
Good Vibrations since early 1950's, first patented vibrometer 1956 in the US.
http://www.vtab.se
Shurafa2

Sr. Member
Registered:
Posts: 162
Reply with quote  #3 
Steve,

Good luck.

I understand that your ML team are trying to "train" the machine and this is very foundational. I worked on related projects and found that the wider you go, the less control you have on the quality of the outcomes. I suggest starting small.

A part of the difficulty that you may want to pay attention to from the beginning is the data type to be used in the ML. Most of the solutions I had a chance to see are based on trendable operating parameters like pressure, temperature, flow, level etc. These systems do not really look at vibration spectra simply because they are not geared towards that. They use the vibration readings as received on the DCS just like other operating parameters. 

Those programs that focus on dynamic signals and provide you with suggested faults to verify and investigate are usually pure vibration analysis software. They use the rule packages based on the experience of the developer and/or user. These are best generalized if you have a fleet of assets of the same model and similar operating contexts.

Calling for input from interested endusers can help if you can unify the data file formats (which is not easy). Some may send you pdf files, word files, native files that need commercial software. I'm interested to know how would these go together in the big data for feature recognition.

I'm sure if you give more details, more people may join.

Regards- Ali M. Al-Shurafa
Noknroll

Avatar / Picture

Sr. Member / Supporter
Registered:
Posts: 837
Reply with quote  #4 
Hmm, sounds like being asked to contribute to our own redundancy.
OLi

Sr. Member
Registered:
Posts: 1,874
Reply with quote  #5 

Dream on I worked on that for 48+ years and not succeeded you can argue that training people also do that but also untrained people work in this business and even more so now with the IOT buzz quite a few totally ignorant entity's pop up and pray magic stuff they have no idea about. If I can I train those too, or I let them shoot their feet if they don't care about reality but it do nag and reduce the tech reputation I think..
Collecting sound only? One mike per machine and some AI should be ok shouldn't it`, In a saw mill? No it is not enough to find bearing faults in time, there are university studies 30 years ago on fork lift gearboxes for loading ships....... Before doing some fancy system do a literature study as you are supposed to... Not nagging this specific case, just ranting on the general IOT hype at least before the virus.
CSI Nspectr II was largely expanded with much user data where users sent in data and descriptions of special machines and faults and suggested rules.
All rule based systems are generally not used since they require so much data to be entered from startup and the more specific you get the more wrong answers you generate, almost like a human w/o the gut feeling "AI" or actually experience, part from mine as I cheat and only limit myself to the obvious faults and only need the machine speed....... Surprisingly many are happy with that as a help. You can not in as I believe do lab faults either and train or simulate in software, they don't look like real world cases and AI will not recognize real faults IRL but some day somebody maybe invent a miracle? I hope it is me. This is only my view but I have worked on the subject for a long time as I like to spread this technology also to smaller factories that have less specialized staff and fewer employees but still wants a reasonable uptime and maybe even a planned maintenance.

 


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

Member
Registered:
Posts: 15
Reply with quote  #6 
@Nocknroll, this is the previous post you made, 5 days ago

‘Plant was down today so got the opportunity to do a not running bump test, there are 3 natural frequencies one of which is 73 Hz. Pump was running at 70 Hz when 15 mm/sec was recorded. Also checked distance from bottom of motor feet to sole plate and found 2 mm difference from front to back feet. So still looking to do a running soft foot check’

Do you really think that an algorithm to look at your data and ASSIST you to analyse it will make you, and properly qualified, conscientious analysts like you redundant?
WWST

Sr. Member
Registered:
Posts: 86
Reply with quote  #7 
The point of ML/AI is to greatly expand overall coverage and focus the efforts of experienced analysts. Walking around has value, but could be done by well-trained operators for visual inspections. We have expanded to over 300,000 points taking temperature every 15 minutes, overall vibration every hour and FFT (11.2 KHz, 8K lines) once per day and on alarm. We have 7-10 remote vibration analysts covering this 24/7/365. 90% of the issues are flagged by the algorithms, written up by a certified level 2/3 analyst and re-checked after investigation and repairs. Maybe 2% need investigation on-site with an advanced capabilities analyst. The number of analysts to do this as walk-around would have been much higher.

The truth is at least 90% of the issues written up by a walk-around vibration analyst in a typical plant will eventually be automated by wireless sensors ML/AI routines. We have been doing this for 4 years now, and that is my best guess. I would not encourage young people to be vibration analysts nowadays unless they want to be part of a very niche career in the future. There will be fewer people needed every year as walk-around analysts are replaced by sensors and remote diagnostics.

Let the flame throwers start now...
OLi

Sr. Member
Registered:
Posts: 1,874
Reply with quote  #8 
What rate of bearing faults do you have compared to all "others"? 75% bearings or more? About what type of machines is it? Any failed w/o detection? I have to ask I am so curious.
Constant operation machinery or variable speed, variable production machines?
Do you describe a rule based system or a AI- inference machine based system? In that case, how did you train them?
I know rule based systems work that is not surprising.

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

Sr. Member
Registered:
Posts: 86
Reply with quote  #9 
"What rate of bearing faults do you have compared to all "others"? 75% bearings or more?

>50%

About what type of machines is it?


General rotating equipment found in paper, chem, gen manufacturing, extruders

Any failed w/o detection?

Yes, of course. It will never be perfect and 100%. We have learned a lot in 4 years, but it keeps learning and getting better. Most failures are "sudden" keyway, set screw or other hard to trend/predict failure modes. Imagine if you could collect data every 15 minutes how good you could be. Then imagine doing this on millions of points like we will be doing in a few years.

I have to ask I am so curious.
Constant operation machinery or variable speed, variable production machines?

We have both and wireless tach to trigger and give speed reference.

Do you describe a rule based system or a AI- inference machine based system? In that case, how did you train them?

Mixed rule based and Guided ML. Big distinction between Guided ML and AI. Trained by experienced analysts (like me) working with data scientists.

I know rule based systems work that is not surprising.

Guided ML works better in my opinion and scales much more efficiently.
Previous Topic | Next Topic
Print
Reply

Quick Navigation:

Easily create a Forum Website with Website Toolbox.