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. [Operator Instructions] Today's event will be recorded and will be available right after it is completed. You can access the recording utilizing the same link that you used to access the live event. After the webinar, we will also be sending you an e-mail with resources from today's event, including the slides, handouts and event recording. Additional information regarding today's topic can be found in the handouts panel of the webinar platform. With that, I would like to introduce today's speakers, Richard Resseguie and Chad Brunner.

Hello everyone.

Hi guys, how are you doing today?

Doing well. Thanks, Jayraj.

Doing Excellent.

All right. Well, thank you, everyone, for joining us. So today, we will be happy to host this webinar and really get you a deep dive into our topic around operating closer to design using artificial intelligence and machine learning. So we've built really an application with Logix AI that targeted towards OT personnel, powering the controls engineer and those who are close to the process to leverage machine learning and use it to optimize their process. So we'll be diving into that topic today. And so with me is Chad Brunner. So a little bit about my role. I'm a product manager here within Rockwell Automation's software and control business unit and I oversee a suite of products targeted towards leveraging machine learning at the OT layer. And with me is Chad Brunner and Chad, please I'll turn it over to you to introduce yourself.

Yes, sure thing. My name is Chad Brunner. I've been with Rockwell for nearly 19 years. I'm currently a project engineer in our marketing team. In that role, I support a pretty large portfolio of products, including Logix AI. I'm here to cover a brief demo a little later in the session. So just thanks again for taking the time to listen.

All right. Thank you, Chad. So what are we going to cover today? We'll first just begin by giving you a little bit of background and setting the stage around introduction to analytics, how we're utilizing it and what we're really focused on. Then after that, we'll go more into Logix AI, what it is, how it gets deployed, how it works as an application? And then go into some implementation example. So we'll dive you through how to use it. Now where do you actually implement it? What does it create value? What use cases can I leverage Logix AI with? And then after we show you some implementation examples, I'll turn it over to Chad to walk you through a live demonstration of the product and show you how to use it to optimize a particular use case with the PID feed forward with implementation. And then lastly, we'll walk you through some key takeaways of our discussion and focus on the Q&A section. And just as we use the platform here, feel free to enter questions in the Q&A as we go throughout. So with that, first thing to really focus on here is when your production is out of spec, what you do? You're working with a particular asset. The asset is going to degrade over time or you're struggling to stay at the particular set points that you have specified. So the typical process, if we go through it, is a reactive thinking process. Okay. We're seeing a problem who's the expert, what's the problem? What tools do we use in order to do this? What we're really trying to do is to reverse this paradigm to help you understand the performance of the equipment and model the equipment itself and model that particular process and create a sensor for the outcome that you're trying to achieve. So where analytics comes into play, if we look at the formal definition, discovery, interpretation and communication of meaningful patterns and data, we can leverage the data directly producing the asset coming from the controller in order to model its behavior, train against that data and give you some insight around the expected performance and the idealized performance that you want to reach. And so we're leveraging analytics here. Analytics as per the definition, relies on the application of statistics, computer programming and operations research to quantify that performance. What we're really doing is leveraging machine learning here to understand the variables that are going to impact that overall performance that are going to influence your process, determine which variables are truly contributing versus not contributing to that outcome that you're trying to model and then to build that model, which is really an equation under the hood around the process of that equipment. So if we look at optimization, that really occurs at all levels of the plant. We can have equipment optimization, line, sight or enterprise, and there's really different tools and different outcomes that you're trying to achieve, depending on which level you're implementing. With Logix AI, we're really focused on the equipment optimization layer. So we want to focus on one single asset, maximize that performance by continuously monitoring its behavior using real-time data in order to adapt the model and tune the behavior of the model to that asset. And leverage our application to model the behavior and use it to predict a particular outcome that we want to ultimately optimize. Now of course, there's other ways to optimize here as you move to line and sight and enterprise. There are different tools that are available within the Rockwell portfolio. One of them might be Pavilion8 or MPC. But today, with Logix, as I mentioned, we're really going to stick to that equipment layer optimization. So let's dive a little bit deeper here. What is Logix AI? How do you use it? So Logix AI is, as I mentioned before, machine learning application, it empowers the OT personnel. It's really a new code machine learning application that is running directly on the edge. We've built the application directly into this machine learning embedded compute module. That's really a card that plugs in directly into the ControlLogix chassis. And it is connecting directly to the controller in order to take in the controller taxes inputs and then model a particular outcome, a particular output. And we'll dive into use cases around boiler, around perfect fill. But if I'm trying to reduce product give away with perfect fill, I can model that particular behavior and use it in order to achieve that outcome. So the machine learning model trains directly at the edge. It runs on this card in order to get the data to train, to adapt and then ultimately to predict or to do an inference on the model. So it essentially uses the model in order to predict that outcome. Now very recently, we did release Logix AI 2.0. We have a brand-new version of the product that's now available. And with the 2.0 version, we've really focused in 3 major areas. One is a new and improved algorithm based on prior feedback and experience with Logix AI. We've improved the overall model consistency and stability when you're training. We also enhanced the ability to handle unbalanced data. So meaning when you trying to -- when you're in the process of training, if you have too much data in one area of the process, Logix AI will maybe throw away some data points in order to have an overall better model over the entire outcome that you're trying to predict. And we also increased some of the configuration around the data input. In the past, we had kind of a fixed amount of data that we were training against. And now we've changed that in order to make it configurable. So we recommend 10 n square where n is the number of inputs. So let's say you have 2 variables that you're using as inputs 2x2 is 4x10, 40. So that would be the amount of data points that we would use in order to train. So in addition to that, we've also improved our model quality indicators. So our confidence level is now modified our squared value. We've also added net absolute error in order to understand the magnitude of the difference between the prediction of the observation and the true value of that operation. And then lastly, we've looked at data preprocessing. So as I mentioned, we are always targeting Logix AI towards the controls engineer, towards someone who may not have the deep machine learning background. And so what we've done is we've added a flow to automate data cleaning and data preprocessing directly in Logix AI, so that when you first train your model using Historian data, we will do some automated data cleaning. So that's going to be removing empty columns, removing linear correlations in the data, any missing data points duplicate variables or constant value columns are all going to be treated and preprocessed. And then we'll give you a summary of that outcome, and we'll even let you download the cleaned up file. So it's just a little bit of what's new with Logix AI 2.0. There's a lot more there. So please do refer to our release notes that are available on PCDC or don't hesitate to reach out to us if you have any questions on that. So as I mentioned, everything happening with Logix AI at the automation layer. So everything from training to predictions, to those model quality indicators are going to be available there so that you can control when you train and when you adapt the model and do it in a closed-loop fashion directly by using the controller. And we'll walk you through those steps. But with Logix AI, we have a dedicated interface to configure the model, configure the prediction and the outcome. You then identify the data. And you undergo the training using those controller tags within Studio 5000, in order to describe when you're trained and when you're predicting. So everything is automated at that on control layer and that really makes Logix AI an ideal product to use when you're dealing with fast use cases happening in line. That's why we'll be focusing today on kind of that perfect fill use case or that boiler application. And the key aspect is Logix AI is really there to help you understand your process. As you're setting up a particular outcome, whether it's that steam flow and the boiler and you use a set of variables that are associated that you think are going to be influencing that steam flow, you can leverage Logix AI in order to tell you which variable is contributing and not contributing. So after you train, Logix AI is doing an analysis over your data, it's looking at the variables, and then coming back to you and telling you which one is contributing and which one is not contributing. That really helps you understand your process and understand which data is being streamed and how that data affects that overall outcome that you're looking to predict. And so where we've applied Logix AI today, auto entire doing that prediction around the perfect splice location of that raw rubber material before you go and roll it into the tire shape. That allowed us to use it in order to decrease those out of tolerance equipments and decrease or increase the overall efficiency and production output out of the tire building machine. The other one is in dryer use case. Where predicting the moisture levels and dehydrating pellets that's decreasing the overall waste products. It's also decreasing any kind of operator input with the equipment. Or with roll products around that prefect cut kind of use case, where I'm making maybe toilet paper and I'm trying to figure out where to make the preparation or I'm trying to figure out where I'm going to cut the roll to apply the glue in order to roll that cardboard roll and roll the cardboard paper around it. So Logix AI is great for these high-speed use cases happening in line where I need to figure out the exact perfect sizing and reduce that product give away. And we'll dive more into 2 exact examples and show you exactly which variables we use and how we implemented the product. And we'll also turn to chatting just a little bit, that will give us a live demo of how you would use it. But those are just a few examples of Logix AI and where we've applied it successfully. So going into the products and how we utilize it. There's 2 main workflows in Logix AI. When you first go into the application, you'll be greeted with this interface. So as I mentioned, everything is hosted on to that card that works directly in the ControlLogix chassis that card is a compute model and it host this web-based interface directly on the edge. So there's 2 workflows. You have the ability to create a live prediction model on the left-hand side there. That connects directly into a Logix controller in order to take the Logix tags and determine which variables are inputs and which one is the target output that you're looking to model? And then on the other side of it, you have what's called an experimental model. This is when you're trying to experiment with new use cases and build the model for the first time, you can leverage historian data with Logix AI in order to create a model, look at that predicted outcome, make sure that you have a good model out of it using a CSV data input and then convert that model to a live prediction. This is also a great example to use when you have a use case where the output is maybe something you measure after the fact or it's something that's hard to measure and you can't measure it live. You can use it in the experimental model in order to build that model and then convert it to a live prediction model and then use controller inputs in order to then get you a prediction of what the outcome would be. So a great useful tool there for balancing these of live training models and also models that are created against historian data in order to help predict that outcome. And so once you build the model, there are really 3 major steps that we'll walk you through. One is you first bring the configuration into the UDT controller, then you train the model, and then you start making the calculations and predicting. So here is the example of the experimental data training. As I mentioned, you train using historical data. So we provide a CSV template that's available on the product that you can download. That template as a particular format where you first have the outputs and then you have the correlating set of inputs that are to the right of that. And then from there, you go and feed the template and upload it directly into the interface. We will go through and clean the data. So that's the feature that I mentioned that is brand new with Logix AI 2.0. Now where we added this data processing step, where we'll remove those empty columns, we'll remove linear correlations in order to do what we really require as the basics for preprocessing data for machine learning application. And then afterwards, you name the model, you begin the experiment, and that's where we'll train the data and give you then an output on CSV file that will include not only the real data set but also the predicted data set that's added as an additional column after you train the model. And with that will come the outcome of which variables that we found to be contributing versus not contributing to that particular outcome that you're trying to predict. And we'll also give you the model quality indicators. ER square, the mean absolute error will all come out through that CSV download after you do the trend. So that's the workflow around training with historian data using CSV. Next, we have the other side of that workflow, which is against live data training. And so at this point, you have an entire workflow that is working directly within Studio 5000 in conjunction with the Logix interface to build the model and then run the training and do the calculation. So we first have that connection to the L8 or L7 Logix controller in order to import the controller tags, you then configure that model prediction and I'll walk you through those steps. You then bring in the UDT into Studio 5000 and that will bring in those tags in order to train the model and calculate the prediction. So this is the workflow that we have here for that first step. When you connect to the Logix controller, you have the controller tags on the left-hand side, you can drag and drop those controllers, you drag and drop the variable interest. You also set boundaries, Logix AI to have some understanding, for example, if you're modeling the power of a motor, what's the min and max RPM or what's the min and max power that you would typically reach. So that when you fall out of back of those boundaries, we can provide a warning and make sure we exclude that data from the training process. So that's the overall configuration. Then afterwards, you review that model definition. This is also where you can configure the data sample size, the amount of data points that you bring in. And then lastly, we move over to the controller. You import the UDT and this is where you have the variables to train and calculate and view the overall outcome of that training. So with the training variable, you first need to do that, that you have 2 modes. One is default training where you first create a brand-new model, a brand-new structure of the model. So when you're first discovering your use case discovering the process, it's great to use that one option. Because that's where we'll create a brand-new model equation. And then beyond that, you need to adapt the model against to lay this data and make sure that you're counting for model drift. And so that's where we have static training. Static training will use the existing model that you've initially created and then adapt it without changing the model definition or the model structure. So the best way to think about it is, if I'm looking to have an outcome that is predicted here, what we'll initially do with option one is create an equation. Think of it as ax2+bx+c. And then after we discover that equation of ax2+bx+c, option 2 will keep that equation as is and then just refresh the coefficient of a, b and c. So that's really the distinction between the 2 new models, the discovery versus model adaptation. And then after you train, you have option one there, which will create the calculation and get you that prediction and that outcome. And one of the benefits of being into the controller and having all the variables available as tags is you can build face place around it and surface the outcomes or even when you train and when you adapt directly into the HMI to put this in the hands of the operators. And that's something Chad will cover in his demonstration. So we've gone through a lot there. But what I really want to do now is focus in on, okay, I showed you how to configure. I talked about what Logix AI is, now I want to go into a real-world example. And dive much more deeply into how you would implement Logix AI. And I want to talk about 2 use cases and one of them is really around the boiler. So boilers are really common unit of operations that will be used to heat water and provide steam into a production process. Over time, that boiler performance is going to degrade. So that's going to result in energy losses. You're going to have varying outputs. You're going also to require more operator contributions are more operator tweaks in order to maintain that boiler and require more frequent adjustments. So where we utilize our Logix AI is around leveraging plant personnel to improve that boiler performance. So we operate in a more stable energy-efficient process. We're also controlling that steam moisture content closer to what we're trying to achieve. But -- and remaining within those operating limits, and we're really reducing the overall operator interactions around that prediction of that steam flow. So here's your guide to how you would implement Logix AI for that boiler. You have a particular outcome, which is the variable of interest here, which is our steam pressure. And then from there, we're trying to model that steam pressure, and we're utilizing controlling process variable. So those are things that we have ability to control, ability to use as inputs. That's going to be your fuel -- feed rate if you're utilizing a gas boiler, boiler that's burning fossil fuels. You're going to have the temperature, you're going to have flow rates of air, water, the overall vapor consumption. So those are some of the inputs that you can use there in the model in order to model that outcome of that steam pressure. So we developed a soft sensor out of this. Based on that variable of interest, you train the model using that real-time data in line, and then from there, you can deploy it to make a real-time prediction of that -- that's steam moisture content based on those operating conditions that are coming in. And then you interpret the prediction to improve the overall boiler steam flow control through that feed forward operation. And that's something Chad will explain us how we utilize this in a feedforward manner. So the result there is a consistent steam generation, reduced that product variability. You also have monetary savings due to increased boiler efficiency and you're reducing that manual operator in invention and that overall operator error. So that's just one example on the boiler. The other example I want to highlight is around perfect fill. And so there, it's kind of a simple ROI to think about. If I'm looking to fill a particular packet let's say, with detergent or, let's say, it's with ketchup, for example. I'm trying to make sure that I hit the minimum amount of quantity of product that I put into each amount. And I'm trying to make sure that we're reducing the overall product giveaway. So the closer I get to my minimum number that I'm supposed to having in my packet, the better it will be. And so forth, that you can use Logix AI to help inform the amounts of that product that you're going to fill. It's a high-speed application. It happens in line over time, you'll lose accuracy. And so leveraging a Logix AI model to train against latest data set will help you reduce those inaccuracies and reduce the product giveaway. And so our solution there is we identify a variable of interest, which is our product fill level. And from there, there's going to be controlling process variables, which is going to be the flow rate, the level of product, the raw product that's in the tank. Could be also impacted with the temperature of the room, the product viscosity. So those are just a few variables that can be used as inputs to predict that product fill level. We developed that soft sensor. We again trained the model using real-time data and the overall outcome there is to reduce that filling error and minimize the product giveaway. So those are just a few examples. As I mentioned, we talked about how Logix AI works, how you can use Logix AI and configure the interface to create that outcome, that variable of interest, the correlating inputs. And those are just 2 examples of how you can use it. But now what I'd like to do is turn it over to Chad to give you a real-world example of how he's applied it and show you a live demonstration of Logix AI to give you a better idea of how you would use Logix AI yourself. So with that, I'll turn it to you, Chad.

All right. Thanks, Richard. Richard did a great job laying the foundation here. So I'm going to just step in. I'd take probably 10, maybe 12 minutes to work through this demo. As Richard had explained, it is around PID feedforward. Anybody who has been set to any of my labs, perhaps at a tech ad or automation fair knows I love cookies. So a lot of my demos are around the cookie-making process, whether it be mixers or conveyors. In this case, we chose an oven. So we have 2 identical ovens just to kind of frame this up. And we have a PID instruction configured to control the temperature of that oven -- for each of those ovens, and we have feedforward with Logix AI enabled for 1 oven; in this case, oven 1; oven 2, we do not. So you'll be able to see this as we work through it, the comparison between the 2 and the response for that we get with the PID feedforward implementation. So with that said, I have a very basic diagram here of PID to help me reference some of these variables. So I want to just kind of highlight them and call them out. So as I'm talking about the PV or the DV value, you know what we're talking about. So in this case, our set point, of course, is our desired oven temperature. Our process variable or our PV is our actual oven temperature, if you're on the right-hand side. Our control variable, our CV is a gas burner valve position, so 0% to 100%. And then our disturbance variable, our DV is our line speed, and you can see at the top here. So in this particular simulation, as our line speed changes, it can disrupt the PID controller based on new product coming in that perhaps is cooler or moving faster, it can try to pull the temperature of that oven down. So the gas valve would have to respond to be able to compensate for that. On the flip side of that, if we slow that conveyor down, we could have a situation where that oven begins to heat up beyond our set point. So we need to make some quicker changes and adjustments that CV value to be able to compensate, to avoid any quality issues or perhaps equipment damage. So with that said, I'm going to step over here and share my screen. So I love technology when it works, let's make sure it works here. Hopefully, everybody can see what I'm looking at. With that said, we're going to cover 3 different pieces of software. We're going to start in Google Chrome, where we're going to be investigating the web page of the Logix AI appliance. We're then going to step over and talk about Logix Designer and the configuration that needs to happen to be able to bring that model online. And then, of course, we'll finish up and view SE just to visualize and see the impact that the PID feedforward use case has on our 2 ovens. So on this screen, this is the main screen for Logix AI. You saw some screen shots in Richard's presentation on this, but we're going to be able to click around here and take a look. This is where you'll go and you'll create or import new models, you'll conduct your experiments or look at your existing models that are currently running, and be able to dive in and take a look at perhaps how they're configured and their current status. We're going to be focusing on this one right here. So this is the oven 01 feedforward. As you can see, it walks you through the steps really nicely. They're really relatively easy to set up. The trickiest part is always picking the proper variable of interest as well as. The actual creation of the prediction is very, very easy. In this case, we're all the way through to the review steps because this is a functioning model that has been trained and has been calculating. So we'll take a quick look at the configuration of this prediction. Only jump in here. There's 4 main steps at the top. Really, the primary work that you're going to do is going to be in steps 1 and 2. There's, I think, an acknowledgment in step 3 and then finalized in step 4. So really, really easy. This is our model name. Our controller slots right here, in this case, I should have mentioned this earlier, you'll see behind me. I have a control Logix rack. There's a controller, an L8 in slot 0 and my Logix AI appliances in slot on 1. In this case, we're talking to an LED5e. Logix AI supports L7 controllers as well as LE controllers and all associated firmware with those 2 platforms. So I just want to throw that -- a little bit of information out there. Our select tag method is right here. Mostly, you're always going to be loading from the controller. That means on the next screen, it's going to pull all the tags out of the controller and give you the option to be able to select any one of those tags for your model creation. You can also import from file, and this allows you to define exactly what tag names you want to use on the next screen, you could use this perhaps if you don't have those tags currently in your controller or if it's -- maybe it's in a test environment and you're going to be moving it out to something else later. So you would only be able to -- you'd only see the tags that you have defined in that text file. Of course, we have our prediction name down here in a description. We'll jump to the next screen. This is where all the magic happens here. Again, our development team did a great job making this super easy. At this point in time, it's pulling all the tags out of this controller. Should take a few moments. Once it's done, you can navigate through all of these tags to create your prediction. And the way that you do this is as you drag and drop tags, from the left window here over to these windows here on the right-hand side. So let's talk about these variable types quick. So a variable of interest, every model or prediction you create has one and only one variable of interest. And this is the primary variable that's under test and one, we're trying to build everything around. Okay. In this case, we defined it as our CB value for our PID, which is our gas valve. You've got balance here, you will need to define a bound, lower and upper for each variable that you move in here. Those bound help put up guardrails when you're training to make sure that if you get any variables that are way out of bounds, it actually caps them. We know that if you put junk in, you're going to get junk out. So we want to make sure that we put up some basic guardrails to prevent that from happening. I always recommend plus or minus 10% of the normal operating range of each variable. It's a good place to start. Here, we have our input variables you're going to have at least one input variable and no more than 20, right? So at least one but no more than 20. And in this case, our 2 inputs are our line speed, which is our disturbance variable. And our oven temperature, which is our process variable. All right. And then on the next screen, again, a system review a click and finish. So we're not going to go any further. I just want to show you how this is constructed here. Again, that is 95% of the work you need to do to actually create a prediction. Once that prediction is created, it's going to bring you back to this screen, which case, the next step is going to be to integrate it into your Logix Designer application. And you're going to do that by saving the cell 5x. So when you press this button, it's going to automatically save the cell 5x right here, which is a handful of UDTs as well as a controller scope tag that you then need to import into Logix Designer, and we'll do that here on the next step. I should say it's already done over here. I'll just show you where that happens. So this is our Logix Designer application. I think most of you probably have imported UDTs or components in the Logix Designer, very easy, you would hit import, and then you would grab that all 5x that you just downloaded and you would import it. Now when that import takes place, you're going to bring in 5 UDTs, 4 of them are going to be standard across all predictions that you import. Those for our Logix AI additional commands, calculations, module training -- our model training and training, okay? It's going to be these 4. You're also going to get a fifth UDT that's going to be the same name as the prediction that you just created, in this case, of oven 01 feedforward, okay? So 5 UDTs and you heard me mention that you're going to also get a controller scope tag. That controller scope tag is going to be of the same name as your prediction and of the same data type name that was just imported in this case of oven 01 feedforward'. So if I were to scroll over, you would see that this is of that data type, okay? Now Logix AI does not live in the IO tree. So if I zoom in here, you'll see that there's no Logix AI appliance in this IO tree. It communicates solely through this tag for that prediction. So again, you'll have to do this import for every prediction that you create, but it communicates to this tag right here, right? So I've expanded this tag out. There's a couple of tags in here. I want to just really zoom in on. And Richard did this for us a little bit earlier as well, but I want to just call him out. And that's our calculating our train. This is where you're going to issue commands to do a train or calculate to our Logix AI prediction. When you do that, it is a snapshot. It's not a continuous train. It's not a continuous calculate. You certainly can do that, but you have to write Logix in your controller to be able to handle any automated training or automating calculate process. Very easy to do. We give you access to it. But just remember, when you issue -- when you place a one in the train or one in the calculate, it is a single snapshot of the data you've selected as the inputs and the variable of interest, okay? And as it feeds through its algorithm. The statuses will come back in the calculations and training buckets down here, I call them buckets but these structures, we expand these and we'll zero in. If it's hard to see on your side, I apologize, I'll read some of the values. So we have our 4 calculations, the response and success, and then we have actual value, predictive value and relative error. Those are the important pieces. And then on the training, we have our absolute error, relative error and our confidence. So pretty straightforward. It's all zero here because I just fired this up this morning. It's already been trained. I have not trained since this morning, so this was not updated since then. So if you see all 0s, don't panic, I just haven't done a training today. So that's why. Okay. So I think we've covered everything there. Let's jump over. How am I doing time wise? I think we're doing all right. View SE here. So let's take a look at the overall response and what's happening. So I've got a 10-minute cycle that happens for ovens. Now just to highlight some of what we're looking at here. This blue line at the top is oven 1. That is our process variable temperature, okay? That's the PV for oven 1 with PID feedforward enabled. The red line is oven 2 without PID feedforward. And you can see down here on the bottom right corner, oven 1 and oven 2 key parameters. As this is cycling through, you'll see these variables jumping around and changing and watching the PID respond. The green line here is our disturbance variable or line speed, okay? So we're kind of going up here. You'll see some bigger jumps and smaller jumps, and at the end here, you'll see a full range jump from max to min on that disturbance variable. And of course, you see it reflected pretty heavily here on the response for ovens 1 and 2. You'll see a little blip here on oven 1, but the PID feedforward value adjusted very quickly for that radical change in that disturbance variable, where in oven 2, the PID is doing the best it can, and it's trying to respond as quickly as possible. I will say, I mean, you could probably go in and tune these PID instructions to get them a little tighter. Absolutely could be done. We could probably clean this up a little bit. But I found that if you have big swings or a pretty critical disturbance variable that can impact a fairly great deal, you might get into an under damp scenario where you start getting some oscillations overshoot with the response to that PID. So these -- I should have said this earlier, both of these PID instructions are tuned exactly the same. The only difference is we have feedforward on for oven 1, okay? So if I turn feedforward off, you'll see these 2 responding exactly the same. So I want to make sure I covered everything here. I believe so. So what I'm going to do is step back over, and we'll summarize in the slide here, so I'm going to stop sharing. Hopefully, you guys can now see the slide. I'm going to continue forward as you can. Okay. So we have our PID instruction here. Our feedforward value, you can see going into that input pin. We've got our response that we just looked at. So using soft sensing model, in fact, analyst Logix AI is a great use case for feedforward -- integrating feedforward into a PID loop. Bear in mind that this was a fairly linear response for the disturbance variable, where I think Logix AI really shines is if you have more than 1 disturbance variable or more of a complex relationship with those disturbance variables and how it impacts your process. That's where I think Logix AI can really, really shine in this use case. But bear in mind, this is not the only use case for Logix AI. This is just one of them. There's a lot of opportunity out there. This is just another 200 toolbox. And hopefully, this helped kind of get an idea of when you would go to that toolbox and grab this tool out. So I think that's it for the demo, and I'll pass it back to you, Richard.

Just a quick summary here and our key takeaways. I want to thank you all here for sticking with us through the introduction, overview of use cases and demo from Chad. But really, our key takeaways here is Logix AI is really targeted at OT personnel looking to leverage machine learning at the edge in order to solve particular use cases with one target outcome that you're looking to optimize. So if we look back to our perfect fill example or our boiler example, what Logix AI is really doing is it's training against that data set, looking at the variable space that you're providing. So the set of inputs that you're providing that you believe are contributing to that outcome. And then it's doing the feature engineering behind the scenes to determine exactly which variable should be used, how it should be used into the overall model, which is ultimately an equation that is produced and that is adapted every time you train. So yes, we tell you that temperature might be influencing the process, but it could really be used multiple times in the overall equation that gets produced underneath or the model that gets produced. So you may see it as 1 over temperature, temperature square. So that's all being handled by Logix AI in order to create a model of that piece of equipment and of the outcome that you're looking to understand and optimize and build the soft sensor for. So we're trying to predict hard-to-measure manufacturing parameters, replace those manual testing steps with soft sensors or leverage that soft sensor to help inform how you can best improve the process in a feedforward manner as Chad just showed. And ultimately, our goal here is improved production efficiency. We want to decrease rates, reduce waste, decrease the amount of energy that's being used. And really increase the throughput and product quality out of the equipment that you already have. So with that, I really want to thank you all for the time that we spent going over Logix AI, helping you understand how you can leverage it for your process. And with that, I'll turn it over to [ ]Jayraj to go over our Q&A session. Here we've seen a lot of questions come throughout so we're really excited here to answer them and work through the Q&A.

All right. It looks like we've got a pretty lively discussion and a very nice presentation. Thank you very much, guys. So the first question that I -- that I would like to start with, and I would like to encourage everyone to give there questions using the Q&A books. First of all, I wanted to ask the question that came in from Rick. Is the historian data stored on the card or can it reference an external database?

Yes, I can handle that one. So the historian data, it comes in through a CSV data that you feed into the card. So we're not connecting directly to an external database. You download the template from the interface of Logix AI. You drag and drop the historian data into that CSV format. The first column will be the variable of interest. That's the target outcome that you're looking to model. And then the subsequent columns to the right of that are going to be the inputs that are going to influence it. After when you do start the training, we will store and maintain the CSV file in the back end. And you can also use the historian data then to further inform the model and retrain and readapt the model if you need to. Chad, anything you want to add to that?

No, I think you did a good job explaining that one.

All right. Next one is how long does Logix AI take to train a model?

Yes. So that really depends on the amount of data and that configuration that you set. But typically, it's maybe a few -- in some cases, a few seconds. It depends on the amount of data. But Chad, how long would you say in your experience for the use case that you just showed that would take to train the model?

Yes. So that's a very good question. And it's kind of a dynamic answer. So I mean, for PID feedforward, for this particular use case, it took more time to gather the data than it did to actually train the model just because we had to let the PID settle and then capture a training sample and then change the disturbance variable, let it settle again, and that just took a little bit of time. The actual training itself is probably going to be process dependent. Some people may take -- the process might be one sample every minute. It might be one sample every 3 seconds. It might be one sample a day. So it's going to be process dependent, but the actual physical training aspect once you issue a training command is relatively quick.

Thanks, Chad.

That sounds pretty interesting, and I would encourage everyone to also ask their questions because we're all here live to answer those for you. Of course, you can ask them after the webinar as well. And we'll be available -- the platform will be open and the presentation will be also available on demand after the webcast is gone. So Edward is asking us, does this AI module require the use of Rockwell programming and [indiscernible] blocks or can it be used with a cast program, which does not use Rockwell programming blocks.

Chad, do you want to take that one?

Yes. So it's a Logix tool, right? It sits in the Logix racks. All data that goes into an operational model has to move through a tag in the controller. So that -- I think that probably answers that question. So it's very integrated into the Logix platform.

Okay. We have some questions that came out during the presentation as well. They were answered in line by our speakers. And so now is the time to ask any questions that you would like answered live. And there is a point here that came up just now, is there an upper or lower bound on the amount of historical data you can put in the CSV?

Yes. So if you're referring to the quantity of data, it's we take upwards of 10,000 data points and up to 20 inputs. So that's a 10,000 rows by 20 column matrix that's being brought into the CSV plus the variable of interest, which is the output that you're looking to predict. So it makes it a 21 by 10,000. And then if you're referring to the bounds of the data itself, we use the min and max within the CSV to define the bounds of the lower and upper that's expected within the actual client. So let's say you have a power column, the min value and the max value is going to be like a plus or minus 5% of the min and max that's coming in from the CSV to determine the balance.

Prefect. So we are pretty much approaching time. So I would encourage you to give us your last questions because we will be answering only a couple of those until the end of webcast. So the next one, Bruce Honda is asking us, can the model continuously train based on process drift.

So there's 2 approaches there. One is I mentioned the static training, which allows you to adapt the model and train it live in order to account for process trip. The other approach that we've seen customers implement is 2 models that are working together. So one is constantly predicting while the other is constantly sort of training and adapting itself and then you have the ability to switch back and forth on the predictions where the other model takes over and gives a prediction. And then you bring the first model, let's say, back to training. We are going to bring in a new metric after the 2.0 release. We're working on that now, which is really going to be looking at the model drift as a metric itself and looking at the latest streaming data and looking at how the model is performing, I guess, that latest streaming data to give a better view of when you should retrain and when you need to adapt the model so that you don't have to use this kind of 2 model work around approach that we look at today.

Okay. Perfect. Well, that pretty much puts an end to the question of what we have lined up for now. So I would wrap it up by saying that I would like to thank everyone for attending this webinar and taking part in this lively discussion and ask all of those interesting questions. In case we didn't answer your question during the live Q&A session, don't worry, we will be doing that by e-mail. And the follow-up e-mail, you will also get the answers to all of your questions that you asked. We would like to thank you for attending today's webinar and in an effort to keep improving and providing the topics that have a value to you, we kindly ask you to participate in our brief survey that will pop out right after you exit the webinar. And if you would like to speak to a representative for more information, you can make a request in your post-webinar survey. There's a question there that if you would like to be contacted by someone of Rockwell's employees in order to give you some more information regarding what we spoke about during today's webinar. So we would be happy to hear from you during this survey. Thank you very much, Richard, and Chad. I'm happy to end this webinar and wish you all a very good day ahead.

Thank you, everyone, for joining us.

Thank you, everyone.