Rockwell Automation, Inc. (ROK) Earnings Call Transcript & Summary

Hello, and thank you for joining today's webinar. Before we get started, we have a few housekeeping items. The audio for this event will be streaming through your computer speakers. So make sure your volume is turned up and your speakers are turned on. Our Webinar platform performs fast on Chrome and Firefox browsers. On the lower left-hand side of the presentation, you will see a Q&A box. We encourage you to answer any questions you have here, and we will answer them at the end of the presentation. If you are having any trouble connecting to the webinar, please take a moment to refresh our browser and disconnect from your VPN. If you're still having trouble, please clear your cache. We have instructions in the hand down section of the webinar platform. All of the panels on the webinar platform are adjustable. To resize simply like the corner to adjust or hit the maximize screen at the top right-hand corner of each panel. Today's event will be recorded and will be available right after it's completed. You can access the recording utilizing the same link you used to access this live event. After the Webinar, we will also be sending new an e-mail with the resources from today's events, including the slides, handouts and event reporting. Additional information regarding today's topic can be found in the handout panel on the Webinar platform. With that, I would like to introduce today's speaker, Alan.

Thanks so much, Juliana. Good morning, everyone. I'm very excited to be here presenting this topic to everyone today. The title may be a little different than what you had signed up for or expected, but we'll still be covering accessing data, using the FactoryTalk Historian connector, as well as talking about some of the things that we can do with that, like machine learning and analytics. So don't worry, you're in the right room, and we will be talking about that connector. I'm also joined today by JT Frandsen and Harshi Weerasinghe, other technical consultants who've helped put this solution together. And I'll be here to answer any questions and help them on the way. So I want to start by talking about what is machine learning and setting kind of a base definition here. Really, over the last few months, AI and machine learning has become a super hot topic. It seems like you can't walk out the front door without seeing an ad or hearing something about AI. So I just want to set a baseline here. Machine learning algorithms build a mathematical model based on sample data, also known as training data in order to make predictions or decisions without explicitly being programmed to do so. And really, the key here that I want to focus on the 2 things that are required is the training data and what better place to pull that data from than a FactoryTalk Historian archive. And then the algorithms, which can come from something like a machine learning platform such as ThingWorx Analytics. And we'll see how those things come together and what we can do with that in the next few minutes. So what are the implications or what are some of the things that we might see as a result? Well, we think you would see reduced downtime through more proactive maintenance, moving things from unplanned downtime to a scheduled downtime. We think we might reduce maintenance costs through earlier detection and reacting faster to issues that maybe we might not have been able to catch without these tools. And ultimately, I think we're going to see better product quality and better productivity by reducing scrap and rework and being more efficient with our maintenance planning. So I think we can see some really, really significant business outcomes as a result of these tools. And I do want to talk about how FactoryTalk Historians are being used today. These are very common, lots of customers have been using these for many, many years and have a large wealth of data. And typically, they're using that for regulatory purposes, I'm going to throw everything into the Historian and it's all going to be captured and then I can account for everything. I might be using that for quality reasons, looking back and seeing what happened, what was the root cause, what caused this issue. I might be doing some improvement, what kind of basic analyses can I do to getting some insights to my production. And some of these analyses might be trending. I might be exporting data to a third-party quality tool. I might have some dashboarding going on, and then I might be filling that into like a Power BI or some other sort of reporting tool. So I want to jump into our first question. So let me go ahead and throw that up. How are you using FactoryTalk Historian today? So that should be popping up on your screen and take a second to answer that question. Should be a little select multiple. So if you're getting more than one, click on those. I know typically, I see most people are doing maybe some dashboarding, maybe some trending or probably the most common use cases as well as the regulatory right, I'm going to log everything and have it accounted for. So I can see these coming in. It looks like trending is in first place right now. Troubleshooting and analytics are tied. So let's see, you will get another few seconds to get your answers in. Okay. So looking at these results you're trending by far, the most common was 67%. Troubleshooting and analytics right behind, and dashboarding is actually the lowest here. I do see -- one person said they're doing machine learning, that's awesome, super excited about that. And yes, can't wait to get more into that topic. I got one more question for you, do you wish you could apply machine learning to your archive data in FactoryTalk Historian today/ So for everybody except for the one person that's doing machine learning today, is that something you're interested in? So go ahead and answer that question. So I am seeing a lot of yeses. Few are answering no, you know. I think lots of big content for you, but -- and maybe we'll convince it's a worthwhile endeavor. So just another few seconds here. All right. So we're seeing the vast majority here are interested in applying machine learning and trying to glean some insights from that Historian archived data. And I think we're going to make that really easy and show how we can do that. Of course, we got 3 easy steps here, and this sounds maybe a little basic, a little simplistic, but when we actually get into our demo and talk about how we do this, I think we'll see that it's not as complicated as maybe it seems. And really, the first step is accessing the FactoryTalk Historian data, and we'll do that with the FactoryTalk ThingWorx Historian connector. We're going to configure and apply our machine learning models, which is really easily done through ThingWorx Analytics. And then using the ThingWorx IoT platform, we will turnaround and enable some significant business outcomes. So taking the results of those models and then turning that into action to impact your production. I'm really proud of the tools that we put together. We've worked with data scientists who have years of experience in the industry. We really tried to boil it down to tools that are made for a productions engineer, quality engineer, someone who doesn't live and breathe analytics. But at the same time, we've captured some of the most important tools, the things that you need to see and made it easy to see the right information. And we have also made it easy to pull data out. So if you say my data scientist wants to export this and look at it in their own tools, we can make building those datasets really easy, too. So very excited about the tools we put together. And then kind of here's an overall architecture. You see at the bottom, we start here with Plant beta, so whether that's sensors, motor controllers, PLCs, anything that you might have on the plant floor or other devices that are being streamed into the Historian? We then use the Historian asset framework to kind of build the data model and provide context to all that information. The FactoryTalk Historian Connector then pulls all that information into ThingWorx. And then using ThingWorx, we can create those data sets, pick up the information we want to look at, build those models and then turn that into action. I'm going to go through each of the pieces a little bit more in detail right now, and then we'll jump into the demo. So Historian hopefully, you may be familiar, but if not, it really collects data, really designed to optimize data storage and collection, so that you end up with large ballooning databases, you can just get the information you need, provide context. So using the asset framework, we can organize that information into a way that makes sense. I don't just have a list of thousands of tags that I might be looking at. And I could do some basic analytics, some [ Means ], maxes, some averages, and we can actually use that to make building our data sets for machine learning easier, and we'll talk about that in just a little bit. The next piece is ThingWorx Analytics, which provides our machine learning platform. So by bringing the data into ThingWorx and using ThingWorx Analytics, it gives us all of these algorithms available. So if these look a little scary, don't worry. Very easy to implement, and you can combine them to get the best results and build some really good models. And then finally, the piece that ties it all together is the FactoryTalk Historian ThingWorx Connector, and this lets you really easily access the data that's stored in that historian data archive. Pull it into ThingWorx, bringing with it all the context in the model that you have in the asset framework, and we can use that to build our models and compare data, explore our data, and we'll see what that looks like in the demo. And so finally, how does it work? How do we bring this all together? Well, first, we select our data. So we'll pull from the asset framework. We'll see pull-in machines that we want to look at and make sure that we have all the properties assigned and we'll see how that works. We'll do some basic visualization in exploring that data. So the first step is just looking at the distribution, looking at the data that's there, understanding what does that look like? And then finally, we'll build some models and see what kind of predictions we can get. Yes, we'll get more into what that process looks like. And then finally, I want to show some examples. Just kind of tease a little bit, so we can do sort of descriptive what happened, looking at a trend and seeing the history. We can do some diagnostic analytics. So we start to say, why did this happen? What causes were correlated? Are there patterns emerging in the data? And then finally, we can do predictive analytics. So based on previous performance, what might I expect my values to look like in 30 minutes, 60 minutes and start to get some look ahead or some predictions made there. So we do have one final question before we jump into the demo here. So yes, what areas would you like to apply machine learning? So we've got some of the ones that we kind of thought of anticipating breakdowns, spotting quality issues, increasing production and yield. And we've got another field. So if you've got another, feel free to type that in the chat or in the questions Yes, any areas that you'd be interested in kind of getting some more insights, seeing if you can get a machine learning model to spot some patterns that may be you can see on your own. Anticipating breakdowns looks like the number one, give this just a few more seconds. All right. We have 5, 4, 3, 2, 1. Okay. So it looks like, number one, here anticipating breakdowns, given that kind of preventative maintenance, spotting issues before they can take early to a downtime situation. Increasing production or yield is the next one, and I think we've got some cool stuff that we'll address that. And then spotting quality issues also in there, but a little bit further behind. So with that, we're going to go ahead and jump into our demo. So let me get this pulled up. Juliana, can you confirm that is sharing okay?

Yes, so good.

Okay. Awesome. So here, we got kind of this analytics accelerator build. These are some tools that JT, Harshi and I have been working on to integrate the Historian Connector, bring in ThingWorx Analytics and really bring this all together. So we're going to open up this menu, and we're going to start by defining our things. And this is where we're going to use the asset framework to pull in the data and organize the data that we have and make sure our data is available for analytics. So you can see here, I've got this asset tree. This actually mirrors the asset tree in the asset framework. Just want to open my demo. I'm going to select my mixer. So I've got a fairly simple tree here just for the sake of the demo. But if I had multiple plants, multiple maybe cells within that and several machines or assets that I want to track, all that will come in here. So I think I've got some more complicated ones, but we'll keep it simple for the demo today. So you can see I select that -- see that I've got all of my properties, all of my tags come in here. I can set upper control limits, lower control limits and sheetability is going to be our kind of goal that we're looking at today. And so you can see I've got my upper control, lower control and a target value side as well, and we'll see how that plays in. If I didn't have one of these created -- so here's an example of an asset that doesn't have an associated thing in ThingWorx created. I could use the services to create that as well, and these are services that are coming from that FactoryTalk ThingWorx Historian connector. So now that we have our things to find, we're going to be looking at this mixer for the rest of our demos. I see that it's created. I see all my properties that I want to look at. So from here, let's go into explore data. Now you can see, we got explore to compare, practice machine learning and then anomaly. So we'll take a look at each of these. So again, I'm going to go in and select my mixer. Let's take a look at this sheetability property again. So here I can define my start and end times. So we'll just take a look at the last 24 hours. I can define a sample interval and sample type, so this is where we're using some of those Historian tools to get average mean, max, some other ones in here, but we'll just look at average for now. So I can see my distribution. I'm certainly not looking at a normal distribution here, but probably these are good values that are within range and what I want to see. But I also have a few outliers appear at the top and then some way down here at the bottom. So I know that maybe we've had some issues or some areas of my data where that sheetability has deviated. And then I'm going to show another one here, just to maybe show a little bit different example of the distribution. Just again, a different value, I'm kind of seeing different values for that average. And that's good to know that I've got sort of that distribution. We also have a performance chart here. So this is going to show my raw performance data without any sampling. I'm going to see every data point that has been captured in my Historian. But if I wanted to smooth this out a little bit and maybe capture every 10 minutes, maybe every 1 minute, I could do that here as well. So I got options for looking at my data and seeing that in the way that I want. Another nice feature is the ability to export this data. So if I get a data set looking the way I want and maybe I can't do the analytics that I want to here. I can export that data set, I can bring it into Python, bring it into -- or any of these tools that I want to use to analyze that data. We made it really easy to create these data sets and pull them out, expand on stuck with these tools, where you need something more. So from there, let's look at compare data. I can see one point, but maybe I need to see multiple to really get a better understanding of what's going on here. So I'm going to pull up my mixer again. I can see I got all my properties. So let's take a look at sheetability, average and RPM. Execute this query. And so I've got each of my lines here, I can choose how I want these, maybe want them overlaid. If you want to see these stacked, so I can start to understand what are some of the patterns and how are these values moving together with each other. If I want to take one off the trend, I can add or remove it and look at just the values I want. And then if I want to zoom in and get a closer look at a certain section, I can do that, too. And again, if I want to take this whole data set and look at it and something else, I can hit the export button, and I'll get that data set in a CSV file. Now for the really exciting part I'm sure that everyone's been waiting for. Let's look at some models. Let's do some cool stuff here. So again, looking at my mixer. I'm going to start by showing a diagnostic model. And really what the diagnostics do, is give me correlations and combinations. So correlations are how strongly are my other values predicting the behavior, the movement of my goal. And so if I'm looking at sheetability here, I can see that material pressure is very strongly correlated and most strongly correlated. RPM is not quite as strong, but still up there. And I see that my Zone 1 temperature, it's really not showing much correlation at all. And so going through and doing this process, I would expect to kind of iterate through and build multiple models or run multiple diagnostic jobs. And so if I'm seeing that Zone 1 temperature isn't really providing a lot of value here, I can remove that and maybe add in some other tags or other properties, and start to kind of identify what is the most important -- what has the largest impact and what's kind of extraneous information that maybe isn't so important to our model. The other thing I can do in here is show combinations. And these really are going to represent patterns of my data. So I'm seeing a lot of correlation or a lot of groupings of certain datapoints. I can see that here. So for example, where my material pressure values between 224 and 225, with a motor RPM right about 13, 13.5. I can see that, that makes up some grouping of data, and I can see the number of examples here. So another one might be material pressure here, again, 224 to 225 and then my zone temperature, just about 69. And again, I can see that a number of examples kind of fall into this category. So it helps spot patterns or groupings that are large in this data. So once I get a diagnostic job that I like, I say these are probably going to be the properties that are most strongly correlated. I can then create a model here and I would do that -- I'm going to show how I would do that, but kind of pull [ it ] through the network trick, and we're going to look at [indiscernible] ahead of time. So if I say I want to look at sheetability, let's say average, RPM, the Zone 1 temp. I can pull all these properties in. I could select them all if I wanted to, but we'll just look at these for now. I'm going to say sheetability is my goal. So that's going to be the property that we're trying to predict. I'll select my sample type here, so again I can do average, then max, [ set deviation ], setup a sample interval. They can do intervals to forecast and prediction frequency. So how often should I be looking for new data. It could be 10 minutes, it could be every 1 minute, just depends on what I -- how time-sensitive this prediction is. And then I set my start and my end time for my training data set. I can check batch optimization, which I'll come back to that in just a second. And then you can select my algorithms. So all those algorithms that we mentioned earlier, we can choose to apply those to our data and see if we can get good results. And actually, if I want to ensemble these or if I want to combine multiple, you can check them and then I've got a number of ensemble methods here. So fairly simple to build models, and I think a lot of this is going to involve kind of an iterative or sort of a trial and error process where we try different algorithms, we try different sample types and we get a model that gives us the results we want. I won't tell you that you immediately are going to apply a neural network and see amazing results. But we do make the tools available and lets you take away your data, try to balance, really giving you access to sort of the raw tools without making it too intimidating, right? We don't have to open up -- don't have the right python code to build a neural network. We can do that with ThingWorx here. And then let's take a look at the results of what we would see, once we create that model. So I can see here, my real values, I can see the features that I'm looking at. So the rotor RPM is going to be the highest weight. Zone 1 temperature a little less. So you can see these kind of correlate to those signals or those correlations we're looking at. And let me see if I can reload this? I can see the distance from my target here and see how far away my predictions are. Let's see. Should have a prediction [ model ] here. So let's see if we can load a different model. There we go. Okay. So wasn't loading in properly. You can see here, I got my real values in orange and then that predictive value in green. And then over here, I can see my distance from target. And so when we set up the target for sheetability, really, what we're predicting here is now how far off from that target are we going to be, and then how can we bring that closer. And so I mentioned batch optimization, and here is an example of what that model would look like. So as we're predicting the difference between target and real values, we can make recommendations on how to bring that closer together. So if I want to bring my sheetability target closer to that value one that we had set up initially, here are the recommendations for bringing that closer in line. So looks like if I could increase the pressure just a hair and increase my rotor RPM, I would expect that target value and that real value to come closer to that target value. And that's an example of the batch optimization model there. So the last thing I want to show, kind of the last type of model is anomalies. And these are models that will observe the behavior of a property over time. And so I can say, looking at my sheetability over whatever time interval to establish normal behavior, I can make sure that I can build a model that will spot when something is out of the normal behavior. So this line down here shows that there's no anomaly and you can kind of see the behavior of this data. Looks fairly uniform. And let me see a shift here, something comes down, it takes a little bit, but then it says, 'hey, something is outside the norm. There's an anomaly here, right? Maybe something is off, your door got open. The motor amperage is higher. We're running at a higher speed. Something is off, but I am saying this is different than what I would expect my normal behavior to be, and then could tie this into something like a CMMS system, maybe I'm using fix, I can automatically generate a work order and have someone go out and check this, before I make a whole batch of product in this anomalous state. And to create this, it's really straightforward. Again, I open up my asset here. I select my property. So in this case, we're monitoring sheetability, say enabled. Click on my description, calibrate the certainty. So maybe if I'm closer to 50%, I'm going to get a lot of false positives. And as I raise this up, I'll get fewer and fewer. So again, there may be some trial and error here to find the right settings for this, correlates with what you're trying to catch. But I can create this anomaly alert. It will take a little bit to watch and calibrate and train. But once that's done, I'll have this anomaly alert -- they will look just like this. And I think that's about it for our demo. So for questions, if anyone has them in the chat, and I'll go ahead and jump back into -- back into our slides. So for some frequently asked questions and kind of hit some baselines, it does require the Historian connector version 5.01 or greater. We do support FactoryTalk Historian version 6 or greater as well as OSI PI Historians. This is built on ThingWorx 93. So anything 93 or newer would be just fine. There is some pricing available, just working with an ISSC to get that sorted out, and it does require services to deploy. Some building blocks are acquired, but hopefully, there shouldn't be any issues there. And then again, this does all build on the FactoryTalk Historian and asset framework. So if you don't have that set up, we can work with you to get Historian asset framework set up and start collecting data. So some next steps would be setting up a workshop. We could work to identify use cases. You can work with Harshi, our data scientist to think about what types of models would be the best to build, what are the highest priority things we would want to analyze or predict. We talk about defining objectives and aligning stakeholders and then getting a delivery team engaged. So with that, I think we'll move into questions.

All right. Thank you for the presentation. We do have -- still have some questions coming in, but maybe we can start with this one. Can I use data from my historian archive, or does it need to be in the asset framework first?

Yes, absolutely. You could use data from the Historian archive. Do you want to add something to that question?

Okay. You can say something now, JT.

Okay. Awesome. Yes. So that's a great question. So with the way that the analytics accelerator is set up, it is important that you have the data modeled in the asset framework, before you start working with the data in this tool, and that's just because we wanted to ensure that seamless user experience, right? So the data has been modeled in the asset framework and then you're starting -- you can roll forward after that with this nice data model to quickly get these analytics gains from there. So that's a great question.

Okay, super. So let's move for this one. What database are used? Can I connect -- port data to SQL?

So I mean we could port data to SQL, absolutely. SQL as a source for analytics data, it's not something we support today, but it is on the roadmap. So we'll be adding that feature soon. But as far as getting data from the Historian archive or from these predictions and then sending that into a SQL database, that's absolutely something we can do. I mean, we support all the common SQL, MS, SQL, Postgres, all the big ones.

Okay. Great. Let me pick one here because we have a couple -- maybe -- all right. So maybe we can move on with this. I don't know much about analytics. How can I be sure I choose the right algorithms?

Yes, that's a great question. So in the models, there are some stats that basically tell you how good the model does at predicting the outcome or predicting the value. And so if you don't know where to start with the models, I don't mind just trying one or maybe trying a couple of them ensemble, and then you can look at that and see how close is my model coming to actually predicting values and sort of through iterating and maybe trying a few different options, you could see if you're getting closer or further away from making a better model. And maybe the best way to start and just start trying to build some. I would say if you feel completely lost, we do have Harshi and some of the data scientists who could work with you and kind of advise on a good place to start. So if the type of problem you're looking at is better than -- or is better done by a neural network or a linear regression, we can work with you to advise on a model to start with.

All right. Okay. Here's a good one. Are there online demos available?

So we do have this demo hosted in Azure. If you would want to get access to that and get in and play around with it a little bit. It could work with an ISSC or get in touch with an ISCC and we can work on giving you access to that demo.