How to Use Machine Learning for Predictive Maintenance

Рет қаралды 41,802

Күн бұрын

▶ C'mon over to realpars.com where you can learn PLC programming faster and easier than you ever thought possible!
▶ You can read the full post here
bit.ly/3eD2Knz
⌚Timestamps:
00:00 - Intro
00:14 - Motor vibration example
00:47 - How do we know when the vibration is unusual?
01:54 - Normal operating condition
04:28 - Webinar Registration
=============================
Here is a basic example of how machine learning can be used for predictive maintenance. You don’t need to be an engineer to understand this. It’s very basic and fun and you can understand it very easily.
We have two vibration sensors here. Vibration sensors are usually used on rotating equipment such as motors, fans, pumps, gearboxes, and so on.
Let’s say we have our two vibration sensors installed on an electric motor. We can use these sensors to measure the vibration of this motor.
In normal operation, we have normal vibration for the motor. But, when there is something wrong with the motor, we’ll have an unusual vibration.
Now, the question is, how do we know when the vibration is unusual? What should we consider an unusual vibration?
Well, to do this, these days with all the advancements in Al or Artificial Intelligence, we can use simple machine learning techniques to figure this out.
To do this, I can simply measure the vibration for sensor A, at a random point in time, and write it down. Next, and at the same time, I can measure the vibration for sensor B. So at one point in time, I measured the vibration for both sensors.
I will repeat this measurement one more time. The values are a bit higher but still kind of in the same range.
One thing to take into consideration here is that it does not matter when I measure these values. I mean I could measure these values at any random point in time.
But the only thing that matters here is that the measurement for both sensors should happen at the same time. That means I need to measure the values for both sensor A and sensor B at the same time.
Ok, I will repeat this process for measuring the values for both sensors a few more times. By doing this I can get more data points about the vibration of the motor in normal conditions.
Now, why do I measure these values or data points, you ask?
The reason that I measure these values is to be able to come up with some sort of model for the motor vibration in normal operating conditions.
That means, using these data points, I can now have a pretty good understanding of what the motor vibration value could be approximately when the motor is operating in normal mode and without any problems.
Now, let’s say that one day, and again at a random point in time, I see that the value of sensor A is 8, and at the same time, the value of sensor B is 2. This is clearly an unusual value.
How do I say this?
Because I’ve already measured the vibration of this motor several times and I have LEARNED that when the motor is operating in normal mode and without any problem, the vibration values should usually fall in this area, right?
This is the model that I have developed for the times that the motor is operating in normal mode and without any problem.
Now that I see a value outside of this area or outside of this model, I can easily say that this new value is not normal and can indicate that there might be something wrong with the motor.
This is how I can use machine learning to detect the unusual behavior of a machine.
So this was a simple example of using machine learning for predictive maintenance.
=============================
If you want to learn more about predictive maintenance make sure to sign up for RealPars first-ever live training event on Oct 26.
For this live training session, we have teamed up with Edge Impulse to teach you how machine learning can help with predictive maintenance for industrial applications.
To reserve your spot, simply head on over to this page: bit.ly/3gmqigD, and enter your name and email address to get all the details.
=============================
Get a RealPars pro membership: learn.realpars.com/bundles/pro
=============================
Missed our most recent videos? Watch them here:
realpars.com/sequential-funct...
realpars.com/laser-sensor
realpars.com/u-remote
=============================
To stay up to date with our last videos, make sure to subscribe to this KZfaq channel:
bit.ly/realpars
=============================
TWEET THIS VIDEO: ctt.ac/0dWJV
=============================
Follow us on Facebook 👉 / therealpars
Follow us on Twitter 👉 / realpars
Follow us on LinkedIn 👉 / realpars
Follow us on Instagram 👉 / realparsdotcom
#RealPars #maintenance #MachineLearning

Пікірлер: 43

@marcfruchtman9473 Жыл бұрын

Interesting video. I like the motor vibration example. Note: There are already sensors that record data and that have alarms which are reported to maintenance technicians. I can tell you that having a machine learning algorithm tell you that a particular expected value is out of normal will not really provide any additional information unless it can also tell you more than the sensor information can provide. In that sense, if a certain set of equipment has similar profiles to other sets of equipment and the algorithm can compare its current data to a known set of issues with similar data readouts, then you might get something like a list of possible faults explaining the data. This would generally require multiple erroneous data points from several different types of sensors thru-out a system. Tho many systems already feed that information and collate it so it can be viewed by a technician. It would be helpful to have a system that can monitor a system 24/7 and provide something more than a list of alarms. But, overall, I get the impression that such a system would offer no more additional redundancy than the last person to look at it. Since that will be where the blame falls.

@realpars Жыл бұрын

Thank you for sharing that, Marc! Truly appreciated

@deepakmalviya6629 Жыл бұрын

Very helpful information about motor vibration sir 🙏

@realpars Жыл бұрын

Thank you so much!

@glennedgar5057 Жыл бұрын

I could be wrong, but bearing failutes do not exibit large vibrstional noise till the end. As a result an FFT analysis that looked for frequency of bearing failures would be a necessary addition to the test. The FFT analysis works particularly well if there is a homing mark sensor present. The allows the raw data to be aligned in time. Virtually any ARM based micro controller could due this analysis. In addition there are microcontrollers with built in WIFI. LORAWAN interfaces could be used for remote facilities

@realpars Жыл бұрын

Bearing failures can occur in many ways. Bearing fracture may not generate as much vibrational noise as, say, a bearing misalignment. Certainly, vibration can detect bearing issues prior to total failure. FFT is more likely to allow the technician to see wear and disintegration over time, which would be most valuable in predictive maintenance.

@narciso.glauber Ай бұрын

Awesome! Congrats

@realpars Ай бұрын

Thank you!

@aminokamad1681 Жыл бұрын

Awesome video Sir 🔥🔥🔥

@realpars Жыл бұрын

Thank you, Amino!

@kevinwikram1245 Жыл бұрын

Thank you

@realpars Жыл бұрын

You're welcome!

@eugenrullz Жыл бұрын

This kind of graf usualy comes in the spec sheet of the motors. For example a dumb protective relay can be set manualy to give a warning signal when the current [A] value is 10% bigger then the nominal current [A]. Not to be "that guy", but this is not "machine learning or AI". This kind of dumb relays were used since before the AI term was "the next big thing".

@realpars Жыл бұрын

Thank you for your comment. This example shows how a simple analytical analysis can be made using two sensors. In most cases, we would use X - Y - Z axis sensors to identify a "volume" in which the data would normally reside. Outside of that 'volume" would indicate a problem. This is a very simple application of AI. In order to understand more complex AI problems, it is first necessary to understand this concept of data collections and how they are collected.

@kiranchoudhari4717 Жыл бұрын

Thanks you team rearparls

@realpars Жыл бұрын

You're very welcome! Happy learning

@slowtrigger Жыл бұрын

Ehh, you really need an FFT to do AI vibration based CBM or at least work with band filters to get relevant velocity and envelope/G gSE overalls to understand what the possible failure modes are. Defining"normal" operation conditions it's not trivial if you want to avoid false alarms and quickly lose the confidence from the maintenance and operations teams.

@meghnaanilkumar3568 Жыл бұрын

Please do a video on loop checking

@realpars Жыл бұрын

Thanks for your suggestion! I will happily pass this on to our course developers.

@charliechimples Жыл бұрын

Great vid, thanks for sharing. 🐵

@realpars Жыл бұрын

Our pleasure!

@visitjpt6555 Жыл бұрын

Can you provide the exact link for the course together made with edge impulse for using machine learning with industrial automation.? I cannot seem to find it in realpars web site.

@realpars Жыл бұрын

Hi there, Thanks for your comment! We have had a webinar on this with Edge Impulse, you can still watch that over here us06web.zoom.us/webinar/register/WN_FXGK6v6FT7Sb8QC9vpu7vA Other than that, we have this video and the related article over here realpars.com/machine-learning-predictive-maintenance/?Blog%20Post& Hope this helps! Happy learning

@jumeldipancaputra87 Жыл бұрын

Next, please explain how this can be aplied with PLCnext and C++/Python

@realpars Жыл бұрын

Thanks for your topic suggestion, Jumeldi! I will happily go ahead and pass this on to our course developers.

@badrraitabcas 10 ай бұрын

I’m a big fan of realpars and your videos have been very helpful, don’t get me wrong. However this explanation wouldn’t be very useful in practice. One of the issues is that class imbalance (number of failures vs not failures) would be very large, making the ai hard to learn. On the other hand if you have many failures it’s going to be a problem for production and using preventive maintenance would be more efficient. Therefore one should use a predictive model which predicts the vibration of both sensors while taking into account the previous recordings into consideration (in chronological order) as well as other info like plc controls and use domain knowledge to analyze these anomalies (given that you trust your model). This would be more cost efficient and in practice works much better than classifying the data points individually, which in my opinion is a useless oversimplification of the process, and wouldn’t work unless you’re trying to create an AI dataset rather than running a factory.

@realpars 10 ай бұрын

Thank you for your valuable suggestion. The analogy you offered is truly captivating. As mentioned earlier, our videos are tailored to a beginner audience, aiming to simplify complex processes. We appreciate you taking the time to share your insights. For further enriching your learning experience, we recommend exploring additional resources from one of our trusted partners at the following link: edgeimpulse.com/ Wishing you an enjoyable and insightful learning journey with RealPars!

@MuhammadShahzad-dx5je Жыл бұрын

#awewome

@tommays56 Жыл бұрын

The motors can RUN for decades so keeping the sensors working and gathering the data becomes and enormous task

@aj6365 Жыл бұрын

Yah as control tech, I swap sensors alot but vacuum and motors rarely, when they go bad the circuit breaks indicate that

@bitebonumbere1426 Жыл бұрын

Thanks

@realpars Жыл бұрын

Thank you so much, Bitebo!

@askilash7855 Жыл бұрын

I want to learn PLC and DCS system siemens PCS 7 software anybody know pls teach me

@realpars Жыл бұрын

Hi there, Thanks for your comment, and great to hear your enthusiasm! In that case, our course library would be a great option for you. Our course library is filled with over 500+ courses, covering a variety of topics for automation engineers. Including a start-to-finish course on PLC Programming. Feel free to have a look around learn.realpars.com/collections Additionally, you will have access to our Technical Team - consisting of high-ranking engineers with a lifetime of experience - who are here to help you out with any questions you might have along the way!   Hope this helps!

@nbrodbeck3839 Жыл бұрын

“I’ve just picked up a fault in the AE35 unit. It’s going to go 100% failure in 72 hours.”

@nikitajha4802 Жыл бұрын

Please explain how you did it and which algorithm you applied as I'm not much familiar with deep learning is it possible to do it without deep learning using simple ml models please please help I have a major project on this topic

@balamus Жыл бұрын

How is this machine learning, as in what they call AI machine learning? Looks more like principal component analysis to me.

@realpars Жыл бұрын

@user-yb5cn3np5q Жыл бұрын

I remember the time when statistics was still called statistics. Am I too old?

@edgaromarreynatorres3940 Жыл бұрын

Yeah IA is quite similar to statistics

@yayser 8 ай бұрын

This example was so simple that I thought I can do the same using ordinary calculus. By the way, why did we need two sensor? Why not Just one?

@realpars 8 ай бұрын

Thank you for your inquiry. Depending on your specific design, it is advisable to consider the use of two sensors, with one serving as a backup. Alternatively, in some cases, you may opt to monitor both the front and rear ball bearings, depending on your design requirements. Utilizing dual monitoring sensors can enhance the reliability of your collected data, providing more accurate and robust results. We wish you productive learning experiences with RealPars