Two Problems Standing Between You and Industrial AI at Scale

Welcome to this new IT/OT Insider video, this one is on industrial AI, not just about the buzzwords as you'll discover later in this video. So I'll be talking about challenges which is holding industrial AI adoption at scale back. And we first created this content for our pitch me and Willem did a pitch in September in Vegas at the ETLS conference, which was very interesting because we were talking to an IT first audience there. So we wanted to figure out a way to explain industrial AI and by doing so, we actually came to the conclusion that there are two major things which is holding back adoption at scale. To explain those concepts, first of all, need to explain or re-explain the concept of the digital versus physical twin. So as you see here on this slide, on the one hand, the right side, we have the physical twin. The physical twin is what is out there, your lines and your reactors and your belts and your motors and your robots and whatever. all the things we've engineered to perfection which is running sometimes already for decades, sometimes it's new, but the physical reality out there. The digital representation of that physical reality is what we call the digital twin. Now the word digital twin is misunderstood, misused already now for many many years because in reality there is not just one digital twin. We have several ones. Digital Twin consists of a variety of applications, your SCADA systems, your historians, your maintenance systems, your lab systems, ERP simulators, you name it. So all different systems, they all serve a particular purpose. They are built for a specific thing. operators, engineers, technicians, are using one or several of those digital twins to bridge the digital to physical world. So they use a SCADA system or they use the engineering system or they use the lab system and they bridge from the digital world to the physical world or they made the bridge from the physical world to the digital world. One of the problems obviously with these sets of digital twins is that As you might imagine, are often not connected. They are often standalone systems. And even if we would be connecting them, you'll see that other naming conventions are being used. They work on other parts of the data. they have one system stops at a certain level and then another system goes into much more detail. But it's often difficult to link those systems to each other using a very unified way. We have our operators or engineers bridging the digital and physical world. They can't do that for every control decision being made. Luckily things are automated. Already in the 70s and 80s we started automating our plans as much as possible. And automation is basically where if there is something we can describe in a deterministic way, in a way where you say, okay, if this happens, then do that. If we measure this, then change that. If a combination of these factors occur, then you should do that, et cetera, et cetera. So these very deterministic ways to automate our plans. that is being introduced. And this is basically where we take away the majority of the work from these operators, from these engineers in the simple, well, it's not always simple, but I would say in the describable automation steps we've been taking in all these decades. Where do operators, where do technicians come into play? For all the things which are outside. those described limits. So for things which are non-deterministic, for things which are unforeseen, for example, an alarm sounds, when there is something regarding safety or changes or you name it, once the deterministic definitions of when this happens do that, when you measure this then do that, when we go outside those boundaries, that's where the operators, where the engineers. where the technicians step in. And so the next question here is, where does AI fit in today? Because I think this is clearly the status quo. Everybody working in operations, everybody working in manufacturing understand that this is the status quo. So if we look to AI today, then I dare to state that AI AI applications are bolt-ons, as you can see here on this slide, on those different elements. So that means that bolt-ons on your sensor data, making it easier to search through data, making it easier to find certain patterns. Bolt-on on your production data, for example, to optimize planning systems. bolt-ons on your quality data to find quality problems easier, et cetera, et cetera. So on the side of the digital twin, we find those bolt-ons and they can be really, really clever bolt-ons, but they are still bolt-ons and they do not replace that entire part of the ecosystem. We also see AI bolt-ons more and more on the automation layer. Now not to replace the... deterministic automation part because that's something you really don't want to do. But for example, to make configuration of these systems easier as coding assistance, as ways to easily help developers in their work, to make it easier to name tags, to make it easier to create new screens, to make it easier to create new logic, but as an aid, as a help of a technician, for example. but also for an operator, might find, you find more and more or you come across more and more systems where these AI bolt-ons, they help for example, to analyze certain events that happens. But again, bolt-ons. And the question is, what do we need to do in order to really advance the application of industrial AI in our plans to maybe go to that, autonomous, dark, lights out, manufacturing type of things we've been talking about. Is that far-fetched or not? Well, I think it's rather far-fetched, except in certain industries where again, the industry as such is very deterministic, but in most of our plants, it is indeed far-fetched. And why is that? And to understand that, I'd like to introduce you the three steps. which we have to go through from an autonomy perspective. We call it here three steps towards the virtual operator. And a virtual operator is an AI system which is able to replace our current operators. And is that actually possible? So step one is where we are today. It's clearly the assistance phase. perfectly available. There are a lot of applications which you can use for all parts of your digital twin and even on your automation layer, depending a bit on the vendor or the vendors you're using at that layer. But yeah, you can see it as I explained more as digital assistants to get work done more effectively. And the only thing it needs is gets access to a part of your data set. Where we clearly want to go towards in the years and decades to come is to autonomy. Autonomy means that the function, the role of an operator is being replaced by clever algorithms. But the problem with that is that at a minimum, that requires a full rethink of your planned control and safety concepts. And I would like to make the analogy to self-driving cars here. So, self-driving cars, we've been talking about that also for decades. And the problem with a self-driving car is not the fact that we are unable to control the steering wheel, control the direction the car is going, the speed, the braking, et cetera. The problem with self-driving cars is the fact that the environment where the car drives is non-deterministic. At any point, at any time, at any point, road conditions will be different. An intersection will always be different. There will always be unforeseen circumstances. And that is the complexity of the self-driving car problem. So as we did with those self-driving cars, it makes sense to first go to a collaboration step for industrial AI and also for those cars to make in a car it could be your your lane assist or your adaptive cruise control or your sat-nav or whatever. But in our manufacturing context, collaboration means that we have some kind of a virtual coworker. So a coworker where the real operator can work with, that's what we then call that virtual operator, who helps us, recommends actions, for example. But it has no direct control of the plan because it can't, right? It can suggest stuff. but it can't have that direct control again, because we do not or the system doesn't understand the full reality of that physical twin. But it needs context. It needs to be able to get to that data to analyze that data. So if we summarize that into our graph we had before, challenge number one. in the road to have to implement that virtual operator, in the road to autonomy in the end. Challenge number one is having an integrated digital twin. It's not about replacing the specific applications or maybe not all of them, but at least having an integrated layer where all data is available in a structured contextualized way. We talked about that in previous videos as well. And when we talk about the industrial data platform, when we talk about unified namespace, et cetera. But it is a requirement for everything we do with AI. If we don't have that fully integrated digital layer, yeah, well, then again, we would be doing local optimizations. And the complexity is there. The technology is available. The complexity is... probably more just the amount of work it takes to integrate all these systems, not just from a connectivity point of view, but for example, also from a data governance point of view. The data modeling problem is a real problem. Also there we talked about that as well, also in previous videos. I'm sure that there will also be clever AI tools which helps us with that data integration part, the data modeling part. But let's be honest. you need to build that model from scratch. It won't just magically drop out of the air. So challenge number one, having that integrated data layer. And then challenge number two is if we would like, if we really want to have that virtual operator, so which is able to act as a core work. So to create suggestions to... Yeah, maybe in the end starts acting on small problems and later on on bigger ones. Again, think about your car and how that's also evolving or have been evolving over time. that's operator. That's the model and that's the AI if you want to call it that way. It needs to consume data from the integrated digital twin. But this is problem number two. It needs to understand the physical reality as well. And we're not there yet. Yes, there are startups today who start working on creating that understanding of the physical twin. They do so by, for example, parsing energy, energy, but I'm saying engineering diagrams. They do so by parsing literature, literature, literature, et cetera, et cetera, creating very clever models about that. But this is a very, very, very tough, tough, tough challenge to solve. and we'll be working on that part for many years and probably even decades to come. But that needs to happen. If we don't understand the physical reality, we will never be able to cope with those undeterministic things, stuff we haven't described. And as long as we don't do that, there is no replacement of the human operator. Full stop. So with that, I hope I gave you a bit of additional insights in the problems around industrial AI. I think we're making great steps. Again, it's important here to understand that the data foundation is everything. It's so important. But it's also important to understand that there is no silver bullet yet. There is no magic solution. It is simply impossible to use some historical data to understand that physical twin. It can't. It's impossible. I've seen companies trying it and they all fail. We need to step into understanding that physical twin and that's not something for end users to solve. It's also not something for data solution providers to solve. Those are very, very specific problems, very hard problems. And unless we get those solved, we will remain in step one and a little bit maybe in step two. Was that a bit too pessimistic? Was that a bit too negative? I think it's more realistic. That's maybe as a final note, now that I think about it. Obviously, I wanna look forward. Obviously, I wanna understand what is possible in the future. But sometimes, I see all these claims and... It's good to take things step by step. I hope this small explanation helps you also in understanding why it's important to take these things step by step. Thank you very much for watching. If you aren't subscribed yet, make sure to do so. If you go to itotinsider.com, can find our blog or podcast, et cetera. And for those who want to understand more what it takes to architect a modern digital organization. both from the organizational side as well as from the technical side, there is our ITOT Academy. So you can go to itot.academy to find agenda and pricing. Thanks for watching and until we see each other again, bye bye.