Practical Machine Learning for Predictive Maintenance

hi Terrance O'Hanlon's CEO and publisher reliability web and uptime magazine and if you haven't noticed this industry is moving fast as we learn more about the industrial Internet of Things artificial intelligence machine learning predictive analytics there's so much going on not to mention the security concerns and connecting to the edge or connecting to the cloud and we're trying to do our own learning and bring you along with that as we learn what's working when it comes to enhancing the liability and asset management so join us for the AI and IOT technology showcase session now what happens is we're teaching a little bit about the basics on how to use IOT and AI to advance reliability and asset management as well as getting small technology solution providers as well as the large technology solution providers a chance to do a short presentation so we can quickly get up to speed with all the choices all the innovations that are currently available to us to get us up into the cloud to do so in a safe way and to do so in a way again the focused on advancing reliability and asset that's our focus so I hope you'll join us for the AI and IOT technology showcase [Music] all right thank you everyone for coming I know it's 2:00 p.m. hopefully you can stay awake I know the steak tips those are really good at lunch when my name is Simon I'm from a company called Pettis ins and our goal at peda cents was has always been to simplify predictive maintenance and so the last couple of years we've been working with a few of our customers to experiment with different ways that would make this process easier and more efficient and around that time too we started noticing machine learning being very prevalent in a lot of different applications so naturally we wanted to see if any of those techniques that they used translate over into PDM so today I wanted to share with you some of the experience experiences that we got some things that we've learned and hopefully that will help educate you as to whether whether machine learning is the right fit for you as well and that pretty much sums up it spells up spells out May outlined for my talk I wanted to address these three major questions today first of all where is machine learning used today and how did it do it but importantly why is it even practical for them to use machine learning and secondly I want to tackle the question of like is is machine learning useful for PDM or not what kind of techniques do that is being used will be useful in the realm that we're in right now and then lastly I wanted to demystify some of the basic concepts of machine learning but don't worry it's it'll be at a high level it's just for it's so that you can understand what your role is in making sure making a machine learning program and be successful and if you guys have any questions I'm gonna try to leave about ten minutes at the end of my presentation feel free to jot your questions down I'd be happy to address them over at the end now if you have if you don't know what machine learning is machine learning is a type of programming that differs from traditional programming in the sense that rather than having a person manually program a function for instance like a rules-based system machine learning is what creates that program now even though machine learning is very good at programming it's still up to us humans to figure out what types of questions we want to answer and there are a lot of good questions being asked out there you may have heard Tesla Google uber are working on self-driving vehicles they use a vote video and image recognition algorithms to try to get a person from point A to point B as fast as possible while simultaneously ensuring absolute safety to both the passengers and also to pedestrians that are around you may or may not have heard about the Google's go playing algorithm it's it's really good at playing this board game that it's interesting because only a couple of years ago the best algorithm that was available we're talking like post deep blue for chess era the best algorithm that was available for this couldn't even beat an average player much less the grandmasters but now we we have an algorithm that can beat any person at this game and that's just impressive and then in the kitchen you might see now that they have ovens that have cameras in them for whatever reason and allow you to see allow the machine to also see what type of food is being put put in and it learns how to cook your food properly so the Millennials don't have to do household chores anymore so some applications are more interesting than others probably the most most profitable area of where machine learning is being used is in the area of internet advertising perhaps you've heard of Amazon and Facebook tracking all your activity on their sites if you didn't know that surprise okay surprise they're doing they're doing just that they're looking they're tracing not just what you click but also how you scroll through the their website what video what images you're looking at and everything you're doing on that site now if it makes you feel any better it's almost certainly not a person that's going to be tracking your activity it's going to be a machine learning algorithm and that's because Amazon and Facebook have millions and millions of users you probably most of you probably use them and it's going to be very impractical for to have a human manually try to build this digital profile of each person just so you can better target that advertising and that pretty much is the reason why why machine learning is being used today one of the main benefits is for scalability right but how to take all this the wealth of data that's available right now and make some useful insights out of that data and I want to give you an example in credit card fraud for instance do any of you know how many credit card transactions happen every second it's about it's about ten thousand credit card transactions having every second around the world and for a credit card company to trace all that using people you would have to hire millions of people just to track whether a particular transaction is fraudulent or not so they don't do that instead they use a machine learning algorithm an anomaly detection type algorithm - that basically pulls information about your purchasing habits such as like what what types of items you purchase whether you like to purchase things online or offline mostly locations where you do your purchasing and it a great that data across your entire purchasing history to determine whether a particular transaction you made is fraudulent or not in and in another area is oh yeah and once once something is tagged as fraudulent that's when maybe somebody from their fraud department will take a look at that transaction to determine then is it actually a fraudulent piece of activity so another area that's that's commonly where machine learning is commonly used is image recognition as humans are really good at recognizing the difference between a jacket or a dress but machine learning algorithms now can actually look at images and pick out particular characteristics from an image to determine the difference between say what makes a sweater vest the sweater vest instead of say a formal vest okay and and once that this algorithm is trained it can classify thousands of images very very quickly having a person do go and do that would take an enormous amount of time and it wouldn't be very practical either and you can see that this is very useful in e-commerce for instance where where you have a lot of images of different products and you want to help customers like us look try to find the product that you're look we're looking for and you know I already made five found a face book once but in a more nefarious use case could be face book looking at your facebook photos and trying to determine whether you like turtlenecks or not now just like how the consumer world needs scalability so will the industrial world and right now the industrial world is there's new technology out there to add more and more sensors instrument different parts of your airplant and with with all but having all that sensor data is going to be useless unless there's someone there to interpret it now even if you have a monthly program where someone comes in to look at all this data it's you're going to be missing out on the wealth of data that's coming in on a daily basis everyday data is coming in there's some may be something that you could take action on earlier within the month and the other the other thing that that is hard for humans to do is to examine the relationships between multiple sensors simultaneously for instance one machine might be might be operating improperly or abnormally but it might not because the machine itself is is is in an abnormal state but because something may be farther upstream may be causing the machine to behave abnormally and so so that's the point I'm trying to make is that having all these sensors it's just like.how it's difficult to to manage a bunch of credit card fraud it's very difficult to manage all the sensor data all at once what it will be helpful is to have a machine learning algorithm help pinpoint the analysts to specific areas to spend to focus their time on and so that way energy can be focused on the areas which which need focusing on and let machine learning deal with the machines that are currently looking healthy now whether you should use machine learning or not will depend upon a variety of factors but the I think the base a good basis for any sort of decision making is to use what's called the 4m approach have any of you heard of the 4m approach I can't take credit for it I learned it from another conference actually it's very easy to remember it's just make me more money and and so the basis of any any sort decision makes should come down to whether it it contributes to the bottom line and for instance in a lot of reason why people a lot of the reasons why people do predict in maintenance is to reduce unplanned downtime for instance if you have a power plant an average one maybe about 250 megawatts with in the cost of power about $40 per megawatt hour if if the machine sorry if the plant encounters some unplanned shutdown of a week we're talking about one point sixty eight million dollars lost now people who are in the power industry are very familiar with this concept and as our people who are in oil and gas instead megawatt hours it's the number of barrels per day but this is also applicable to manufacturing for instance the number of gadgets that you're putting out each day or even in pharmaceuticals say if if if an HVAC equipment goes down it causes it requires a two-week experiment to be rerun again for instance that's something that that happens sometimes and that that can be really expensive so when you're evaluating this all we look at how much how much downtime you're you're seeing on average in your past and whichever predictive maintenance program you use whether it's machine learning or not see how evaluate how the costs how the costs and the benefits add up in in your calculations so with that being said I want to talk a little bit about the benefits to using machine learning to help you decide this even if you had only just started instrumenting your plant with sensors you can get some benefits very early on and one of those ways if you remember the credit card fraud anomaly detection a similar technique can be used here which takes multiple sensors and aggregates that in to create some some anomaly score or in this case I've called it a health score and so that you own you have one single trend that tracks the health of a machine and instead of instead of having to look at like maybe you've outfitted the Amish machine train with ten different sensors instead of having to look at ten different sensors you only have to look at this one single health trend and once an anomaly is triggered that's when that's when you would call in an analyst to come take a look at that particular that the data that's coming from a machine and oh and this is actually better than using alarms basic alarms but I'll explain to you a little better why later on in this presentation another benefit just like we are able to classify different types of clothing we can classify four different types of faults too and a machine for for vibration thoughts faults they exhibit themselves in very distinct patterns as many of you know and machine learning is very good at recognizing those patterns and so once the bachina algorithm is trained it can actually provide some sort of diagnosis and this can be integrated into your work order system for instance to to help you make sure that the inventory for spare poor spare parts are available for whatever whatever the diagnosis is long term you get even more benefits from having machine learning once you have enough data you can provide a very accurate estimate of the remaining useful life room of a machine and this is useful for doing maintenance scheduling or doing financial forecasting and that and that remaining useful life estimate can will change it will change in response to varying conditions and a very another very neat thing is a process optimization now this would use a similar technique actually to what what self-driving cars would use it'll take all the sensors that are positioned around in your plan and try to decide how to best optimize the next action for the plant and whether you want to optimize for efficiency or optimized for throughput and I probably would say that you probably wouldn't want to completely Skynet your plant but by having this spit out a recommendation for you it can help it can you can then allow the operators to decide whether the action that the machine learning is recommending is a good recommendation or not and of course there has to be a feedback system whether an action is being taken or not and that that allows the machine learning to overtime gradually learn how your plan operates how and and what can actually optimize for certain certain things now one point that I wanted to re-emphasize here is this is pretty long-term for both of these it's going to take a while you have to be patient for a machine learning algorithm to even achieve something like this and that means that you have years of integrating machine learning into your workflow at your plant before you can get reap benefits such as that so after I've since I've talked about the benefits now I want to go over briefly the some of the basic machine learning fundamentals and in how to make a machine learning program actually succeed now you may have heard that machine learning and data science be those terms being used interchangeably and that's because data is very very important and in not only just the raw data but how you convert and adjust that data to be submitted to a machine learning algorithm will be will be important in creating something successful so I'll talk about data and how how machine learning looks at that data but then I'll talk briefly about unsupervised learning and supervised learning typically referring to both anomaly detection or class vacation and it's just to help you peer into the black box a little bit so you can see what it takes to make successful algorithm work program work so the first step is always to collect very good data and to decide what kind of data to collect think about what analysts would need to collect in order to say to determine the health of a motor pump drivetrain right you might want to take have current transformers - to take a look to diagnose any sort of electrical faults in the machine the motor maybe you want to also have that current so you can evaluate the speed of the machine especially if your motor is running on a VFD vibration if you heard we're here for the last talk of course is a great a great source of information for getting the heartbeat of the machine and as is the differential pressure to evaluate whether your pump is operating correctly now even more important than the amount of sensors that you have on a mounted on your machine is the quality of that data bad noise from bad cabling or having bad poorly calibrated sensors are all going to contribute to garbage data and garbage in is still gonna be garbage out so even for things like vibration right vibration is challenging because not only is you don't just mount it you have to think about how it's mounted whether it's mounted axially or radially in addition also the type the type of adhesive you use to mount that sensor on or if you're just using a magnetic mount that'll that'll change the frequency response of that sensor so if we take a look at just vibration for instance this is a typical vibration waveform and when an analyst would do is you look for characteristics in that waveform but also pull out certain characteristics which I'm gonna call features and features are the term used the machine learning in this case we're here maybe we could pull out the RMS which is the rooming square like the average vibration level also you could might pull out the crest factor which shows how peaky or sharp the the points on the waveform is and both of these both of these are net are features that can go into a machine learning algorithm but of course that's what I wouldn't be enough for vibration there's a lot more data here usually a vibration analysts will also look in at the at the spectrum the frequency spectrum it's a transform from the time domain into the frequency domain which creates a histogram that shows the energy that exists at each frequency bin and from here you can look at the running speed harmonic and then the harmonics of that running speed and each one of those items can be entered into the algorithm as features so if you take all that data and you place it into a chart you'll get a matrix that looks like this where the columns are the different features that you would decide to input in this case RMS crest factor only next level as an example and the rows correspond to different data points at different measurement times this is also where you would include a lot of other information to your machine learning algorithm if they exist such as blade pass frequencies gear mesh frequencies if you are able to get that and I wanted to stress this because the customers who have seen the most successful of machine learning are the ones who are willing to put in the effort to find out some of this information so this information is not always available on the nameplate on a machine so getting that min from if you're able to get that information then the machine learning algorithm will be able to track that information as well so I'm one of the so once the once you take this data now we're it's in a form that can be submitted to a machine learning algorithm this 2d matrix of data that I am showing you right here and what the machine learning does does what this data is essentially creates the scatter plot of that data so on the left side over here this is this is actually data from one of the air handlers at one of our customers sites and over here on the left side I plotted the 1x harmonic versus the RMS and you can see it almost basically creates a straight line because it makes sense because RMS the most of the waveform is comprised of the rotating speed meanwhile if I plotted say the One X versus a 2x I get almost no correlation it's just the blob of points but what a machine learning can do is not just identify the relationships between two different features it can look at three different features maybe it can plot one x versus rms versus two X even though I'm not visual creating a visual for you here it's what it's doing and it does even more than that for dimensions five dimensions as many features as you want you know it'll be able to analyze the relationships between multiple features simultaneously now here's a here's where I need to give a warning making making this work it's not going to be as easy as just picking all the features you can possibly think of and throwing it at the algorithm unfortunately because what happens is if you especially if you don't have enough data so having an overwhelming amount of features is going to cause the algorithm to fit the data in a way that may not make sense and I think it's best illustrated by an example say you wanted to make an algorithm a machine learning algorithm predict what type of car you would like alright but you're not a millionaire so you only own two cars and say both of those two cars are tan so if you use color as a feature into your machine learning algorithm obviously the machine learning algorithm is gonna think that you like tan cars even if you just happen to pick those since they were on sale or something now if you had 40 cars a saree fit are maybe in 50 cars let's say you have 50 cars and 40 of them happen to be tan then the machine learning algorithm would be pretty accurate in knowing that you have poor taste in car color choices but that's the point I'm the point I'm trying to make here is that it's it's not as simple as just throwing any feature you want it's important to have domain expertise in order to in order to pick the correct features that will make the machine learning program operate correctly and so if someone if someone say from last Tigers mentioning uh some from Silicon Valley comes up to you and says that they have a stock picking algorithm that picks good blue chip stocks and they say oh you can also use this for your application I would probably ask a few more questions about it so actually given so given this this now that you know that how machine learning looks at these data points I can kind of explain to you then how this is actually better than just using basic alarms because say over here I have on this left side I have one feature plotted against another feature and this could be anything this could be temperature let's say and this could be pressure now let's say that the normal and normal operating conditions of this machine the data points form this odd l shape these are all normal conditions but the the relationship between pressure and temperature just happens to be this way now if you just use alarms alarms will just set it on the singles on a single sensor right so you'll have one alarm here and you'll have one alarm here which forms a box and boxes are not even typically how sensor relation data relationships are anyway if you have a red point over here this looks anomalous to us because it's not fitting into the shape the alarms is are not going to pick that up and then similarly if we look over here on the right side we have maybe two different operating conditions this is very common some machines will operate in two different states and and you'll have one cluster of data here and one cluster data here which is healthy but then some of these other points are not healthy the alarms can pick up they say this orange point here which is outside of the boundary but there are other points that that these basic alarms just won't capture now machine learning is capable of drawing a more complex boundary it can draw a boundary that surrounds the points more accurately and you can adjust how how tightly you want to fit the curvature of the data and also where the offset is and so to give you see a little bit how this works I want to talk about anomaly detection one common algorithm for knowledge section is one class support vector machines and say over here I have some data again it could be any features it's not really important and I have the blue points here are the data early on when it's very healthy and the red points are every other piece of every all the other data points that I want to evaluate the health of what OC SVM one class has support vector machines is is a distance based algorithm so once I supply the algorithm with the set of blue points here the the healthy baseline it'll draw a boundary around it but notice no note here that even though I've drawn it in two dimensions this plane is actually in multi-dimensional space it's a hyper plane or a hyper ellipsoid and now now what it can do is evaluate points based upon whether it's inside of that boundary or the distance away from that hyperplane boundary and as you as you Inc as you look you can base that the health of whether it's moderate severe based upon adjust based upon that distance now if we look at the machine health score this is this health score is just a plot of those distances you can see over here that it starts off as green and with it fluctuates between yellow and green before finally ending up in the yellow area and if we take a point from the healthy area you'll see it in a spectrum that the vibration spectrum looks very healthy and if you look at the unhealthy portion you'll see a very high one one times peak because remember this health score Oh oops wrong button this health score is taking into account all a bunch of features simultaneously the 1 X 2 X the RMS like I've said another nice feature that some machine learning algorithms offer is the ability to a pinpoint the features that are that are contributing most to an unhealthy score so for instance that point that I had pointed out earlier the one X level was very high and the the algorithm that you're using should point out that that one X level being very high later on perhaps the the the noise gets a little bit worse the algorithm can point out to you hey there's anything there's high one X and high noise maybe there's some sort of cavitation so now I'm gonna want to just skip over this to talk briefly about classification okay so classification is really useful in identifying different states that a machine is in now this particular machine this cooling water injection pump that we are looking at I've taken the multi-dimensional data I've used the technique to reduce it down into two dimensions so that you can see it plotted here and you can see a few distinct clusters in in the location where this pump is located in the winter in spring it has it doesn't have to work as hard as when in the summer and fall so the values end up being a little bit higher but these are both normal operating conditions of that of a pent pump and that's something that ban alarms can't necessarily capture meanwhile these over here are actual anomalies I think this one was actually a this was actually misalignment I believe so anyway if you if you can provide a labeled a dataset to your algorithm you can train it to be able to determine these different states and this is useful for whichever at any any sort of machine that operates in various states or clusters for instance machines with VFDs will regularly switch speeds or places machines that have varying loads so I'm going to skip over just a couple more things and it's for the sake of time and just give an example of how you actually how you would actually train a machine learning algorithm to recognize something like unbalance here we have data coming from a machine where there's actually unbalanced and we did confirm that there was imbalance on this machine and the way to train of the what a data scientist would do is take a portion in that data as input it needs to contain both healthy and unhealthy data data with unbalanced and that would be submitted to the classification algorithm too and it would train create that function that rules based function that I was mentioning to determine whether something is healthy or unhealthy lastly I would take the remainder of the data to test the determine what the accuracy of that is and Vibrations science like I said shows up in very distinct patterns and if that's really nice because it helps the algorithm we found we find that classification our algorithms tend to be very accurate in classifying some of the major faults such as misalignment or unbalance so I'm going to take the remainder that the last bit of time before questions to go over the key takeaways of that I wanted to I want you to see through this presentation and the first is to recognize the advantage that machine learning offers which is that you can look at multiple sensors in the relationships between them simultaneously and having and especially with a bunch of new sensors being placed in your plant having that will be very useful in helping your analysts focus in on the particular particular areas that that need more a further examination and if you instead have to hire a bunch of people do that it would take an enormous amount of people to to evaluate all that data the second thing is that I wanted to stress the importance of data of course the more data you have the more accurate your algorithm becomes and you can use leverage some technologies that are out there but be careful because you shouldn't just necessarily just use the cheapest sensor available because the data the quality of the data like I said is very very important so making sure that and making sure that data quality is good and coming in properly is why I'm gonna I say machine learning is kind of a journey it's not necessarily difficult to implement machine learning at your plant but it's not going to be a one-time thing it's something that has to be integrated into your typical workflow so if say a machine goes out for repair the machine learning algorithm should have visibility into that say say you switch out and in teller on a pump but the but the machine learning algorithms still thinks that it has the old impeller it's not going to be able to properly use determine false in that category so so yeah so this is it's a journey and but if you if you stick with it and you have the patience to make sure that it that the machine learning algorithm is getting all the information it needs you can get some of the benefits that that come with it that like I've discussed such as process optimization and I'm gonna stop right there because I I think I'm well hoping you guys have some questions you wanted to ask okay so these are tapes so be patient now I have to bring the mic if you have questions okay I got a couple I got one over here first hello I'm Preston Johnson with allied reliability group I'm a platform lead in intelligent monitoring when as from a vibration analysts perspective which I've had some training there I would look at a number of vibration sensors on something like a motor pump I would calculate ten or fifteen condition indicators off of each one of those sensors and then I guess I'd be looking for some combination I'll probably have a hundred and fifty or so features how do I mean if I'm getting started could I use some how do can you just start feeding data in call it normal and then machine learning will find an anomaly that I can then go identify is that a reasonable approach yes some somewhat like if I go back to the this example here right you would you would have to indicate a healthy baseline in this case yeah and and when an anomaly is detected it's going to be detected based upon a a number of features simultaneously and so so so the short answer is is yes and it's nice because like you said it's the even though the anomaly detection is not going to say exactly what caused it it can highlight certain features those in in your case probably either like certain bins in the frequency domain or certain sensors that seem to contribute most to that anomaly and then you can like focus in on those particular sensor that data and see if you can come up with any good diagnosis my name is Gopal with oSIsoft PI system my question is how are you converting all these features into a single health score so people will do that differently it's um so the data that we take we we actually don't just simply take this like the chart that I'm showing here it's we don't typically just put this data in directly we'll will actually do some we'll look at the features that the vibration analysts will look at we are a company right now mostly focuses on vibration so then I'm going to talk and in that realm mostly and we have we have the data typically for just a single machine train and we'll take a bunch of features from the spectrum and we typically do some sort of feature reduction technique which I won't go into here but that that data is then fed into the anomaly detection algorithm was there it was there an add-on question to that no sorry can you speak to that feature reduction technique I think we can talk about that we can talk about offline if you it's like okay hi there a good presentation by the way thank you very much so my question relates to one of your first slides we showed the whole process from loading of a fuel to to do whatever was power plant or something and it illustrates the fact that even though you have many features but there's a time delay between them when when you correlate two features that they don't happen at the same time how does that work right that's a good question yes that that well the time is going to be of course very important the time stamp is going to be very important in [Music] coming to coming to a decision based upon that hopefully you can we can get to a point where the data is coming regularly enough to where that wouldn't play as much of a role but yes if the time if the time doesn't isn't happening simultaneously it's not going to if the measurements across the plant are not happening like at a fixed time point then yes that would that would change things you're absolutely right so it's important to make sure that that happens I have a question and can you set those parameters yeah you keep well depicted that depends on the system that you're using whatever product you're using to collect data across your plant there there are wireless technologies out there which use like a cloud based cloud controlled measurement interval and so whenever the sensor connects it it gets it receives information oh I need to collect information at this particular time and then it can then collect it that at that particular time great I got one here first I'm Abdullah Andy is it similar to the transpose control concerning presenting the process optimization sort of sorry no the petrochemical industry they are using advanced process control to do a better optimization for the plants and they are collecting many data to get one output I know this is from Machine point of view but from post point of view there is something similar like a PM transpose control oh that's really interesting I'd love to talk to you later about that I don't think I don't know as much about that in particular it's certainly possible I am NOT yeah I'm not sure sorry sorry about that we can talk later about it I love to hear more about it okay my name is matthew furry with foot tetra back in sweden i wanted to find out from you in the company that I work for we do cotton packaging and we're also looking into machine learning to try and understand you know the remaining useful life of certain components that are high cost contributors and one of them is in induction eating that we that we use to be able to seal the packages well what I see demonstrated of you is only as a single area of collection when it comes to the features vibration do you use other data sources as well - from other monitoring that you have on the equipment to be able to gather all these features to be able to make a conclusion as well we've looked into that so I just wanted to find out if you've done any investigation with your company with customers from before and so well I don't think you I you probably don't know too much about our company our first product is not a sales being I'm just telling you that our first product was a vibration sensor Wireless vibration sensor so most of the focus like you said has been mostly on vibration we are looking into other integrating other sensors into it and and we're yeah so we're we're looking into that but we I don't have any exam both of that as of yet yeah but I think that it should overall I think that it it should work we might you might have to do some sort of clever feature engineering with that as well

Channel: Simon Xu

Views: 12,180

Rating: 4.8978724 out of 5

Keywords:

Id: 9jPxWzrbLLI

Length: 41min 17sec (2477 seconds)

Published: Mon Mar 19 2018