How Machine Learning And Artificial Intelligence Automate Process Improvements: A Q&A with Beckhoff’s Daymon Thompson

Welcome to the Smart Industry Podcast, Remaking Industry, where we dive deep into the tools, techniques and technologies that are accelerating digital transformation. >> Okay. Thank you for joining us for the Remaking Industry Podcast today. My name is Chris McNamara, Editor-in-Chief with Smart Industry. We are thrilled to have you join us here today. We are looking at how machine learning and artificial intelligence can automate process improvements. We're connecting with Damon Thompson, with Beckup Automation. I'm going to share his thoughts on all things in that space, including chat GPT, some of the gender AI elements, emerging ways of machine learning are playing into this. It's a very high topic in our Smart Industry Universe and the larger manufacturing world. So we're excited to dive into this with Damon. Once a topic for technology futures, only machine learning and AI are now being used daily in discrete and process automation applications. Most likely in your facilities as well. Running ML algorithms on a standard machine, controller offers a way to automate continuous improvement in process automation, optimations with faster reactions, and for less data transmitted than purely cloud-based approaches. On the programming side, AI tools such as chat GPT are now being used to assist in control system programming to suggest entire code segments, complete code and debug it. So there's a wealth of opportunity in this space. And a lot of us, I think, are just getting started or seeking out information about how to get started or how to scale out initial efforts. So we're thrilled to be connected with Damon Thompson with Beckup Automation. Today, Damon, welcome to the podcast. Yeah, thanks, Chris. Tell us a little bit about who you are and what you do with Beckup. Yeah, so Damon Thompson, I've been with Beckup for a while, actually 12 years, and I am the Director of Product Management. So I lead that team and also the advanced applications team. So we're doing kind of proof of concepts for, you know, next generation machines and technologies. You know, you're knee-deep in this stuff. It's an exciting time, huh? I mean, it seems like this is as big of a buzz phrase or as big of a trending topic as I've seen in a number of years. Is that accurate? Oh, man, totally. And not just in our space, but also, you know, just in the general news, right? I think everyone has an ML conversation recently. Yeah, yeah. And it's interesting to see the overlap with applications in our homes, and then applications in the industrial space, and to see where those things coincide. It's very exciting. And the capabilities seem, I mean, it seems endless. I mean, every day, you know, I'll use the phrase open a newspaper, which I know is antiquated. There's a new story about a new application and how it's revolutionized in this field, and it's changing the way we work in this capacity. It's just really, really in stock. So let's talk about that. Emerging ways that machine learning in AI or automating process improvements, specifically. What are some of the most common ways? Totally. So I want to break it down a little bit. What we, like you said, you open the newspaper. I love your knowledge. It's great that every day we see that there's something in the news. And sometimes I think there's a misconception, too, where people think, oh, this is what ML is. But I think there's a lot of use cases that people don't necessarily think about. And then industry, the way I kind of break it down is there's just four different ways. The main way that I think most people think of it is, I'm going to collect a lot of data. I'm going to put it off on a server. And then there's some really nice fancy package that's doing some analytics there. It's doing machine learning. And a lot of the focus of what people think is predictive maintenance. And yes, of course. But that's kind of taking the data off, running it off on a server somewhere and doing its thing and then notifying somebody if there's a problem or potential. The other one that I think people don't, so number two, that people don't really think about is running machine learning algorithms in the machine control, actually kind of in the PLC. And what that can really bring advantages to the, not only the machine builder in making something innovative, but also, you know, on the process and what that does for the end user. The second one or the third one, sorry, out of the four, is actually doing machine learning on the control, but not necessarily in real time, but doing things like real time, or I should say, machine learning combined with vision and things like that, but really on the control. And then the fourth one that's not really on the control, but it's more on the engineering side is using large language models like ChatGPT. And that's going to bring some very interesting things for programmers. Excellent. And what process is it we talking about here specifically, Damon? Yeah, it's a good question. So it could be applied to anything, just like you mentioned earlier in every industry, you hear news about it being innovative. The machine learning is a set of algorithms and neural networks. And it doesn't really care some of the data that gets processed into it, what industry or process. It could really be from anything. So packaging or, yeah, really anything, even entertainment. Okay. And then you talk about improvements being enabled by these approaches here. What improvements are we talking about? What are most common improvements? Or what is maybe an improvement that might surprise people? Yeah, so some of the improvements, it's really like its data. So what people are doing with it that we've seen in real time is a series of inline kind of quality control. I mean, I could give some examples about that. Or optimizing processes, for example, motion control, optimizing how it moves and what the motion paths are. Also optimizing just product quality by making changes in line while the machine's being run. And there's of course the predictive maintenance aspect to it that can be run in real time. Also just making predictions on, let's say packaging, for example, how well the operation's running in real time to make adjustments. Okay. You know, as we said a few times here, it seems like this is ubiquitous now. We heard about it for the first time. We heard about CHAT-CTP for the first time. Three months ago, and now it's ubiquitous. And it's already changed the world now. That's not true, obviously. How widespread is the adoption of these tactics that we're talking about in the industrial space? In terms of your perspective here, we still at an investigative stage. Are there front runners really leading the efforts here? What's the maturation level with these tactics? Yeah, it's a great question. So I broke up kind of the four ways. The way that I mean of like, let's collect data and go put something off on a server or the cloud and do predictive maintenance with ML. You know, that seems to be pretty understood from the, you know, the OEMs and end users that I that I talked to that seems to be out for a while. And people get it and they're like, yeah, okay, I can buy a package. But some of these other technologies are running it on the machine really doing real time. Or, you know, hey, how can I use CHAT-CTPT on my engineering side of things? Or how do I do on machine predictive? Not necessarily all that widespread really. And honestly, I think it's because people that just have a little bit of a hard time with seeing an application and going, oh, I think I could do something. I could solve that with machine learning. Or even a machine builder saying, you know, I have this really complex algorithm. If I could wave a magic wand and improve this, I would. But man, that seems to be a ton of efforts try to build a program around that or an algorithm to try to make that consistent. And it will, at some point, we'll get those programmers and developers to think, you know, hey, this is really hard to implement in a traditional if then else, you know, maybe I can do this. But if that, that kind of smells like machine learning application. And then they will kind of become to the realization, hey, I can reach for that technology. So at some point, this gets baked into the process earlier at the development of those algorithms or even, you know, considering potential programs or developing strategies, this just becomes kind of the ubiquitous tool implemented at that early stage, huh? Exactly. And the next part of the adoption is obviously understanding, you know, demystifying it a bit. Like, what is it all about? How can I do it? And usually it's just not in the traditional machine builder, controls engineers, wheelhouse. It's just it's an emerging technology kind of on the data science space. So, you know, helping the adoption, you know, we've actually implemented or employed some data scientists that come in and we say, hey, we have a process expert, a machine expert. I know the process. I know the machine. I know what I'm trying to do. And they get coupled with a data scientist that the data scientist helps identify. Okay, well, here's the data I think we should pull in. And then we'll go through the training process of the machine learning algorithms or the machine learning model and then re-deploy it back to the machine where the the process and machine expert says, yeah, okay, that's good, but we need to tweak the outputs a little bit. So, honestly, we see in some regards to help the adoption, it's a little bit of a team sport. And rather than, you know, one person trying to take it on. Yeah. And you find yourself playing educator in that respect to enable end users to recognize potential gains there? Exactly. Yeah. We see that we come in and say, hey, you know, if you could wave a magic wand on your machine, what would you do? What could the number one thing you want to optimize? And then we can help with some of our experts say, you know, we think there's something here. We think we can do a little machine learning model. And then if we have to bring in a data scientist to help out with some of the detailed machine learning model training, great, but absolutely. And we have hopefully we're helping bridge the gap a little bit about the kind of feels like we could do something with machine learning here to optimize that. Yeah. Next one. So we're talking about kind of the hypothetical situations here. Give me an example. Give me a use case of a recent win on this front. Something coming up. Oh, man. Interesting. I love talking about these examples because you're totally right. The abstract from machine learning could do anything. You know, it's great. You know, Alexa uses machine learning. Your iPhone uses machine learning to really like, okay, great, but tell me how I can use it in industry. Like what really what is done? So I really like talking about these. I think it helps solidify for people. So I'll go through a few a couple. The first one I mentioned motion. So we did an application where we had product coming down a conveyor belt. It was kind of randomly spaced. And we were secretizing it to do an operation on the product coming down the belt with like a rotary track system. So back off XTS system. And and they were, you know, not evenly spaced. And we had multiple movers on this track system. And so the traditional way of programming that would be, okay, when I'm done synchronizing with the product, I just need to zoom basically zoom back into the into line so that I'm ready for the next product. So just go at a high velocity high acceleration to make sure I'm available for the next product coming down the conveyor. But if there's a lot of factors there, you know, what if you have extra movers already sitting in line? What if the product has a bigger gap than the last one did? You know, there's several things we can bring in. So what we did was train a machine learning model, a little neural network to actually say, you know, if there's already a lot of these factors in place, then generate me out the best motion profile I can to minimize acceleration and deceleration, which uses less power and less acceleration and deceleration also means less wear and tear on on the equipment. So just doing that, you know, instead of doing predictive maintenance, you're actually running the machine in a different way that you don't have as much maintenance or you don't need as much preventative maintenance. So that's kind of one on a motion front. You know, maybe one, what do you want the quality side? So we did an application welding application actually and they used a thermal imaging camera and it was a robotic welding application. And with the thermal image coming back, they could see exactly how well it was welding and how the weld bead was going. And so we could do things like, you know, move closer or farther away from the material or slow down and speed up the robot, the process and really change it so that the weld seam was as optimal as possible. And doing that with machine learning in real time. So it's constantly reacting to based on what I'm seeing in the weld, what exactly are the best parameters for speeds and feeds and material flow and everything else to make the best weld possible. And that was really good. Similarly in a carbon fiber application, we're building a carbon fiber tubing and we were laying down the carbon fiber and as we were doing that, we took in a lot of things like, you know, what is the pressure being put on the material itself? What is the pressure on the webbing of the line coming in? There was, I don't know, 10 or 15 different variables we brought in. And it would do a prediction score on what the tensile strength of the product coming out of it would be. So because we're collecting this in real time and making predictions on how to change, we could optimize the tensile strength. And we wouldn't have to do destructive testing to figure out whether those new process parameters really worked. I mean, we'll do it on the fly based on the training. So so a couple of kind of interesting examples. I'll throw out one more real world example. So it's in the in CNC mold making actually, to mold when you're doing mold for plastic injection, you want that that metal of the mold to be as absolutely as smooth, meter finish as you can get it. So that when you when you do the plastic and plastic goes against it, the surface of the plastic is extremely smooth and it doesn't take any kind of secondary operation. So what we what you know, mold maker's been trying to optimize this for a very long time with different tooling and speeds and feeds and trying to figure out how to do it. But the contour of the mold that they're milling might have something to do with how the tool performs or maybe I need different speeds and feeds. So one application we worked on was watching a few different things like the spindle speed, the vibration in the in the machine itself, vibration on the spindle head, several factors. And it brought those all those back. We did machine learning model in real time to predict what the optimal process parameters should be. And we fed those back into the CNC to change things like speeds and feeds and even coolant actually at some point, coolant rates so that we could get that as smooth as absolutely possible. Okay, so those are a few examples. Maybe one more. Just as you do, we love use cases because it's often times you talk about these things in the abstract and people can't really figure out how does it supply to me or how could I implement this in my work space or any type of case studies are great. Okay, good. I'll give one more. Like I said, I love talking about this, too, because I know how it goes to sit in a presentation or somebody talking and they're like, you can do anything. Okay, great. But I want some concrete examples. So I've been in that presentation, David. I was in the same one. Yeah, exactly. Exactly. Yeah. So, okay, another one. So another one we did was, yeah, you could think of a packaging application really high speed flow wrapper, but not just flow wrapping. You're flow wrapping. Yeah, you can think of it as ramen noodles. And, you know, for what we know, ramen noodles, you also get an insert in that ramen noodle packet that is powder, right? It's the spices. But this one that we were doing is really unique. It had that not only the powder, but two liquids. That was an oil and there was a spices that was a liquid form and those also got put in that high speed packaging form. And so what would happen is, you know, doing, I think they were doing 150 packages a minute, something very, very high speed that occasionally one of those inserts would quite get inserted enough. And when you when you sealed the package and then cut the the foil between the packages, you would nick the corner of the one of those packet packets inside the package. And what happened, of course, if you nick a packet of oil is eventually it would leak out during transport and packaging or putting a box, all like I said, so it ended up on the store shelf. It's just this bag of mushy noodles that nobody wants to buy. So you would think, hey, that's pretty easy. We as we're sealing the bag and we're cutting the film, you just look and see if there's any additional torque, but it's not that easy, right? So it could be that the cutter has worn out over time. So it takes a little more force to even cut the foil. Maybe the coil foil actually has a little bit different thickness to it. There's just all sorts of factors, especially running at that speed. So we actually spent a while, but we trained on machine learning model that watched this and would predict if we had nicked one of those internal foils of the packet while we were seeing the overall package, which was pretty impressive because there's the very, very minute differences that tell gathering multiple different data points. And it was very successful. We actually deployed it across 50 packaging lines and they were extremely happy with how the quality at the end ended up at the end customer that no longer getting soggy packages to have to be returned and dealt with. So saving a ton of money. Awesome. Well, I've made a hundred of those types of ramen packages for my daughter and I have yet to have a problem with any of the oil packets. So you're doing good work there. Thanks. Five specific examples there, which I love. It's really great to see, again, these things, these concepts put into play. Let's back things up after that now. Talk about a little bit of strategy that these, you know, what lessons can be shared from these type of projects can be kind of extrapolated to everybody there. Is it, is it the breadth of applications coming from these techniques and technologies? Is it the ease of implementation? Is it the accuracy enabled by, you know, machine learning and AI in the modern smart factory? What lessons can you learn there? Yeah, I mean, I think the the big overarching is, I mean, I think machine learning can enable some things into the process, whether it be, again, on the machine builder side or the end user side that really has advantages. The key is identifying those things and being able to say that that could be a machine learning thing. I could implement something that gives me some advantages here. Or that seems too challenging to take on. I don't know, maybe the machine learning a machine learning project could actually implement a result there. And I, it can be used in all kinds of different cases, right? So again, predicting quality instead of having to actually do destructive texting testing. It could really change the process and what businesses are doing. So it's, the takeaway is, I guess, think big picture. Think, think, you know, magic wand. What is, what might be possible and explore it? Because you'd be kind of surprised what I think advantages or what really innovative things can be done. Excellent. Interesting. Process versus discrete camps. Any difference with this approach on those two sides of the going? Process versus discrete. You know, we're seeing on the discrete side where it's identifying where we would run things like high speed real-time machine learning inferences so that we can make predictions, you know, on the fly, like we were talking about the applications, or implementing a machine vision and running the machine vision machine learning really on the same machine controller within PLC. So those are pretty, those are easier, I guess, to identify the high speed approaches. The process industry is also looking into this, absolutely. Depending on which part of the process, they're taking a use ML, but wrap it in a lot of boundaries. So you're, you're kind of writing checks and balances for the ML prediction, just to make sure that things are really safe. We're not opening vows, we're not supposed to be those kinds of things. But it's similar, similar implementation, just slightly different approach. But I think both both are really finding advantages of looking into it. Okay, excellent. We can't talk artificial intelligence in 2023 without looking at generative AI and specifically chat GPT. How does that come into play in this conversation here? Oh, yeah. So, okay. I like this question because like we hear all the time, the general thrown out in the news is ML, AI, and they're kind of different things. So I had, there's different algorithms coming into play like, you know, whether it's, you know, train or untrained machine learning models, or whether it's, you know, neural networks for doing vision. And in this case, which at GPT, they're actually called large language models behind. And it kind of scrapes languages, pulls together all kinds of words to make, you know, good responses and things, not necessarily looking straight up at raw data or images. So large language models. And so what's happening on the large language model side is because it's text and it's language, it can really be used for generating even PLC or control code, generating algorithms. If a developer's not really sure how to get started, they can even go to such at GPT and say, hey, I'm new to PLC programming. Can you help me get started with controlling a conveyor with three zones? And please, you know, write that instructor text and give me instructions for of what each piece of code is doing. And it'll really generate something like that. So instead of file new, you're bringing in something from chat GPT. And it may not be perfect, but you bring in something and you go, okay, this gets me a really good start. And so it's really, really accelerating engineering efficiency. Because like I said, you're starting with file new and trying to figure something out is even writing, right? That's, as always, the worst is blank page. So it's, it's, it's really accelerating how codes being developed. Excellent. Let's talk about your team back off. Give me some examples of how you're assisting clients with these techs. And then we touch on, you know, those five use cases we look at were great examples. You touched on a little bit more earlier about your role as an educator with clients. What does backup do in this space? Yeah. So specifically on the chat GPT side, we actually have shown at the show recently at Automate and actually at a hundred of our MESA or International Automation Show, where we actually implemented an interface to chat GPT within our development environment. So it's, you can even right click on a PLC program and say I'd like to optimize this code. And with the help of a large language model behind an open AI, it actually optimizes your PLC code, makes it more efficient to run or it compresses it. Even if you maybe have a partial code, you say, well, I know what I want to do. I have some variables defined for inputs and outputs and a little comment section in my code for what I want to do. You can even just right click on that program in the PLC and say auto-complete. It sends it out and chat GPT will do its best to write the rest of the program, the rest of the algorithm. And again, 100% perfect. No, but it gets you a really, really good portion of the way there. We're even in the development environment, like I mentioned, you can go over and say, you know what, I really need a small piece of code that can run the access and do kind of a motion profile like this and you kind of type in what you want and it can generate out code and drag and drop it over into your PLC to accelerate engineering. So we're looking at how we can continue to integrate large language models into the software even through our documentation to say, hey, how do I set up this network? Give me a step-by-step instructions and it goes out, scrapes our documentation and says, oh, I see what you're trying to do. Here's step one, step two, step three with maybe even some screenshots and some help. Even in the engineering tools, the chat GPT, large language models and machine learning, they're going to continue to be filtered into every tool we use, not just in the controller, but in really every piece of software, I think. Yeah, exciting stuff. Last question for you, Damon, big picture. Look into the future here's a peer into the crystal ball. What is the near future of machine learning and artificial intelligence in the manufacturing space look like? And what most excites you? What are you most optimistic about? Yeah, so I really see, you know, short-term, more and more people are going to jive into the investigations. What can this do for me? What can it be in this discovery? I feel like it's almost like the beginning of the IoT or industry 4.0 or everybody said, okay, well, what can we do? There's an investigation phase and I think that's where we're at now. Everybody's looking at, you know, what can be done? How does this help me in my process and my machine? And that's near term, I think really we're going to get their longer term. I think at some point pretty much any machine that an end user buys, there's going to be a little piece of ML, even if it's not known, buried in the in the process or doing a little algorithm and to do something interesting with the process or the machine, and the machine builders are going to be looking for those things of how they can utilize this to become more innovative and honestly more competitive against their competitors in the market. And I like I said, I'm really optimistic that at some point there's going to be two, twofold on the engineering side that the development of PLC code and development of machine controls is going to be so much faster because you can you can ask for those kinds of things. And on the machine learning side or sorry on the runtime side, the machine control side, I really think that we're going to do things that we never ever thought possible 10, 20 years ago because of machine learning. And I think that like I said, there's going to be some huge majority of the, I mean, especially the front runners and the innovative machine builders that will latch onto this and they absolutely will put it in the control and it's not a, it's, you know, we've heard this before, it's a little cliche, but I really believe it's not an if, it's when and it's who jumps on it. The first is kind of going to kind of be the front runner and be out and from can't be the most innovative in the market. Yeah, and you're there to help him with it. Yeah, I mean, absolutely, we'd love to be a part of that and help even identify some of those applications. If it's, we're trying to solve this, not sure, is this a good fit? Yeah, we'd love to be there for that. Excellent. Damon Thompson with Back Off Automation. Thanks for joining us here today on the we make an industry podcast. Very exciting stuff. Yeah, thanks, Chris. That was fun. I always like talking about technology and innovative stuff. Me too. Very, it's, I mean, it just, it couldn't be more cutting down. So it's very exciting. We want to thank our listeners for joining us here today. As always, we encourage you to go out and make this smart day. [Music]