Getting Started With Machine Learning For Predictive MaintenanceSeptember 10, 2019/Posted By : Agosto/ 0 comments /Under : BlogPredictive maintenance is a growing buzz word in the industry, but how many companies are actually making progress? Some companies are reporting a reduction of equipment downtime by up to 50 percent with predictive maintenance using IoT. The main takeaway is that you can save a lot. On this episode of CloudUp, we’ll be getting past predictive maintenance as a buzzword and get into the what, why, and how your company can make progress.Meet the Speakers Rick EricksonCo-Founder and Chief Cloud Strategist Mark BroseVP of Software EngineeringTranscriptMarkWhat we’re seeing is there just hasn’t been quite as much pick up of it as we would have expected, so a lot of interest, a lot of value, perceived value, but it’s moving a little bit slowly. I think what we’re seeing, even for single-use cases, you’ve got clients that are seeing hundreds of thousands of dollars that can be saved even with single-use cases so, the take away there is there’s a lot of stored value here in doing something with predictive maintenance, so definitely worth taking a look at and seeing if there’s any value you can add to your company.RickOn this episode of Cloudup, we’re gonna be getting beyond predictive maintenance the buzz word and focused on the what, why, and how you can make progress quickly.RickSo welcome to Cloudup, the series where we explore the coolest things built on the cloud today brought to you by Agosto. Today’s topic we’re gonna use machine learning to predict maintenance events and there’s a ton to dig into here so let’s jump into it. First off let’s talk about how companies perform maintenance today. Is it proactive, is it reactive, and what’s the impact of that?MarkYeah, we definitely see companies doing both reactive and even proactive maintenance. You know I think the more advanced companies we’re seeing are doing proactive maintenance in the sense that they’re really doing maintenance on a schedule right, so preventative maintenance. Over time you can develop some experience with when you should replace things, but what we see there is that’s not optimized, right? So that’s not, in some cases you’re gonna replace things too early, in some cases you’re gonna wait ’til it fails, so it’s difficult to get it just right. So there’s some cost to that. You do it too early, you’re losing some material usage that you could’ve gotten and at volume and scale, that can be a lot of money and it you wait ’til it fails obviously you might have, if we’re talking tires in a fleet, trucks out of service, maybe has an accident, liability issues, so there’s cost of waiting too long. What we can get to with predictive maintenance using machine learning is you get that a lot tighter. Not saying we’re gonna ever get 100%, but you’re gonna get a lot tighter inside that window so what we’ll be able to save on material, but also prevent more of those failures from happening. That’s to me is the big impact.RickYeah, so going back to your point about just reacting and not being able to predict the impact of failure, not good, right? Sounds kinda like how I rolled in high school with my Malibu. Let’s talk a bit about how using predictive maintenance can help avoid some of those unexpected costs.MarkAgain, it’s really being able to optimize that time window. We’re doing that by having a lot more data. We’ve got kind of a wider range of data that we can take advantage of so we can use somewhat static data like making models of a tire, a piece of equipment which tells us something about how it’s constructed, and there’s some predictive value to that, to how long it’s been running, what conditions it’s operating under, to real-time telemetry data like temperature, tire pressure, vibrations, that kind of thing. All those things together then really can be used to build a good model.RickSo these are key attributes that ultimately humans can even understand the basis of those key attributes, but I imagine that at scale when you have millions and millions of events, it’s really hard to understand what’s happening, how to use that data and create classified information that fits into categories that we understand.MarkYep.RickI imagine by using machine science and Cloud ML, we can use some of that information to train models so how does that all work?MarkYeah, so that’s a good question. We find still in machine learning there’s a lot of value still to human input and the primary value of that is in this area we call sort of feature engineering. It’s the fancy word for knowing what data elements will be predictive of failure. So it’s still helpful to have domain knowledge to sort of pick those data attributes that should be included, but then what we can do is we can take advantage of the machine learning technology to take that data and create the algorithm for it based on machine intelligence, so it’s not something the human has to spend all this time engineering the algorithm. Their focus is on getting high quality data in place that we can use to be predictive. With the cloud, the cloud really brings to us the platform to run all that on so a lot of time the data scientist that’s working in this space won’t have infrastructure background or development background. So what cloud platform can give us is a lot that just as a service so we don’t need to spend a lot of time or having broader skill sets in that team to build out these models.RickSweet. So now we’re gonna take a quick break to hear from our sponsor and then we’ll come back and do some jamming on the jam board.SarahMore and more we’re seeing organizations wanna be really strategic with their success and part of that means they’re moving to the cloud. At Agosto, we’re seeing a big uptick in clients using Google Cloud platform for their online business operations. So whether you’re thinking you need a boost with AI, machine learning, or you just wanna build something new and fast, Agosto is an award-winning partner who would be happy to help you with your needs. We’ve been Google’s Cloud partner of the year multiple times. We’ve got a few other awards as well and we would love for you to get to know more about us, okay, can’t take anymore, by visiting agosto.com. You can get a free guide to your company’s strategic plan for heading into the cloud. We’d love to see you there. Now back to the show.RickWelcome back to Cloudup. In today’s deep dive session, we’re gonna focus on how you can get started on predictive maintenance andMark’s gonna take us through our approach.MarkYeah, so definitely where you usually wanna start is pretty basic just framing your business problem. Essentially that’s just thinking about use cases that make sense for your business. Pretty simple approach, we essentially start out with ID’ing the business area we’re focusing on. An example of that is gonna be, so let’s take the tire example we’ve talked about a little bit. You’re spending too much money on recovering from failures of tires in our fleet. So just framing up that’s an area you wanna focus on and then what we wanna do is dig in a little bit deeper in that and create an actual business question that we’re trying to answer. So in this case, that’ll be something like can we use predictive maintenance with machine learning to better predict failure than we’re doing right now with our preventative maintenance program? What is a more specific thing that we’re trying to target?RickAnd this is really an impact right? So we’re trying to frame this up as high value, low risk, but a problem that’s going to impact this organization in a material way that’s considered financially viable.MarkYep, but it’s a high-level problem, it’s a framing of the problem in a way that we can target, is it something we can do? You have to think about here you’re typically thinking about well, what does failure mean? Tire case is pretty easy, tire blows up. But with a machine, does it just slow down, does it actually shut down the line, you have to frame what is failure for us, what is the thing we’re trying to detect and some sense of what’s different than what we’re doing today? That leads you to selecting a business metric. Say we’re trying to target tire failure. Business is really our success metric that we’re trying to get to. Say we’re trying to get to, for us right now we’re predicting at about 70%. Can we get to 90? So we have a target. In this first phase, you’re just trying to narrow down the focus of what it is that you’re trying to do.RickAnd percentages obviously matter when you frame it up like this, but in some cases, if the impact, even if it’s 75 or 80%, so let’s say it’s 5% or 10% better, in some use cases that can be really impactful.MarkYou can turn that into we’re moving from 80 to 90 and you’d be able to turn that into dollars. That might be for you 50 grand, maybe it’s like hundreds of thousands of dollars depending on what that looks like for you.RickAnd so the way to think about this is I’ve got humans today that maybe can react to a problem, but when I’m using a system that can handle massive amounts of information and then predict an outcome and I can alert on that prediction. You can also think about sort of downstream as I’m framing the business problem, what’s the ultimate impact if I can go from this sort of narrow use case to something that adds scale?MarkAbsolutely. You should think about this is a place to start and a lot of times what we’ll do is we’ll look at a whole bunch of use cases. You might start out with maybe 10 or 12 ideas, run this process through with those 10 or 12 ideas and out of that we’ll get to hey, if we can hit these success metrics, these things logically bubble to the top, so we’re looking for hey we found some high value potential here and then what we’d move into is how do we start to test that out and see if we can actually do it.RickCool. All right.MarkSo that’s framing where we start and then where we’re going to here is really we’re trying to get to somewhat of a iterative process of learning. We’re gonna frame this up as kinda think about this as a circular pattern here where we’re doing a continuous process and what we’re starting is where all the really needs to be is in data prep. So data prep is really, talked about this a little bit already, but it’s essentially pulling in data from a bunch of different places and doing some data exploration, making sure we can suss out what are predictive variables? So again with the tire example, we’re talking about things like make and model of tires, talking about how long the tires have been running, we’re talking about temperature, tire pressure, all these kinds of things. We’re pulling that together and then this is the place where domain knowledge is important. You build this stuff up, you explore it, you do some visualizations maybe, you’re doing some graphing, you have some thoughts on what’s predictive, but in this phase you’re really spending time determining whether it really looks like it’s gonna be predictive and in this case, you’re sometimes maybe you’re creating some absolute values, you’re looking at median values, or moving averages of data of variables, so this is a part where you’re spending probably the most of your data engineering and science time really is pulling that data together and shaping it in a way that you can use it.RickAre we trying in this phase do we try to understand what a representative amount of data or a relationship means to humans yet, or are we just trying to make sure that the data is in a form that’s sort of repeatable and consistent?MarkIt’s a little bit of thinking about what you can actually operationalize and then from a model-building perspective, it’s about throwing out things that aren’t useful, shaping things in a way that look predictive, it’s really kind of all of that. You hear terms around data engineering, data normalization, that’s kinda what’s happening in this phase.RickOkay.MarkWe mentioned a little bit earlier sort of your feature engineering and you’re really pulling together the things that look like they’re gonna be valuable.RickOkay, and throwing those things out that maybe won’t be.MarkExactly.RickSo it’s not just getting everything, it’s also trying to be smart about what you are using because we’re trying to again, move relatively quickly to get to some validation of the business problem.MarkThat’s right. In some cases, you’ll already have this, like in a data warehouse. In some cases, there’s gonna be some work to get the data. This is the foundation really. We have to have good data and lots of it.RickCool.MarkSo then from there, we move into what we call model involvement. So here, this is really where the machine learning part kicks in and sometimes there’s straightforward answers to what type of ML approach we’ll use here, but you may experiment with a few different options and kind of see how that susses out. There’s a little bit of data science to this part, but here’s a lot of our letting a machine now use all this data that we got and building models from it. And so we’re doing this typically with a subset of our data as we’re engineering models that look like they’ll be valuable.RickOkay.MarkThis is the machine part and then we’ll move from there into an evaluation and review. In this phase, what we’re doing again we’ve taken a subset of our data to build a model, now we’ll run that model against a test holdout set of that data. So we’ll have trained a model and here we’re running that against some test data to see is it actually predictive? Sometimes the model may get overfit to the data you used if you have other data that doesn’t look exactly like that data, maybe it isn’t quite as good as you thought. You wanna do some work here to make sure that it is as predictive as it looks like it was when you were building that model.RickOkay, so you’re actually looking at how well this information, this model is predicting the outcome that you’re hypothesis is expecting?MarkThat’s right, is it performing as you thought? At this point we’re all operating here very much in the cloud, cloud platform tooling, usually a couple of data engineers, data science people can do this all this work. Here already, you might have some iteration. If this is bad, you may go back here and like, all right, we need to take a new approach. That can happen.RickSure.MarkBut ideally, from here then we’re moving on to looks like we’ve got something that’s useful, now we’ll get into deploying that model. This can be straightforward if you got a lot of the things in place, but this could involve okay now we’re putting sensors in the field, maybe we have some edge processing that we need to do, so we might engineer a mini data center close to that edge where the data is.RickSo you’re gathering more data?MarkYou’re actually gathering more data, but definitely creating the infrastructure to wherever the feedback loops that we want are available. So put a model out there and now what do we want to have happen? If we’re predicting tire failure, we wanna alert somebody that hey the tire’s about to fail so somebody does something. So here is the work we do to get the model out there, get whatever tech deployed to do the alerting. If we’re alerting a driver directly, maybe it’s notification immediately in the vehicle or maybe it’s to a dispatcher that lets them know, depends how we wanna architect it, but this is when we’re operationalizing the whole thing. And in the first rev of this, this may be a small group, you’re gonna wanna not affect your whole fleet ’cause when you go to the field you inevitably figure out things you didn’t think were gonna happen. At this point, you’re again validating all this stuff works in the real world.RickOkay.MarkAnd making tweaks that will fit for your business.Rickwhen we go back to this first process, which is framing of the business problem, understanding these key components, defining some sort of targets of success that are important for the stakeholders in your organization, and then we start iterating on this sort of loop of processing. How long does this usually take us?MarkWe can typically do a small run of this in a short time period. You might be able to do something as fast as four to six weeks. A lot of it will depend on where you’re at with your data. Sometimes there’s work that needs to be done to get that data in place so that can take a little longer so that data engineering piece, that will drive a lot of how quickly the rest of it will go.RickIn our experience, we typically will run workshops and help the executives and the engineers that we’re working with at organizations that our organization partners really help understand what’s possible so that they’re not stepping into mistakes that typically because of inexperience and not understanding this process real well, that they won’t do that. They’ll make good choices and we’ll help them so that they’re focusing on the right problem with the highest value and ultimately is gonna have the highest impact to their organization.MarkThat’s it.RickAll right.RickThanks for watching this episode of Cloudup where we focus on the coolest technology delivered on the cloud. We’d love to hear your feedback comments on how you can use predictive maintenance in your own industry, how you’re making progress in this space and if you have questions or challenges that we can address here around predictive maintenance, leave a few comments, leave a few notes asking questions, and we’ll give you some swag. Thanks again.SarahCloudup is brought to you by Agosto, a leading Google Cloud platform partner. Like this episode and subscribe to our channel on YouTube to learn more. We would love to help you out. Visit agosto.com to learn more.