The Goal of Data is Prediction, Not Records
Data on its own has minimal value to a shop floor. What is really important is connected data and relationships across the production environment – even the whole lifecycle of a product or facility. To examine how these connections are paving the way for the future of operations, we have Mark Hindsbo for another episode of the Future-Ready Podcast.
Nick Finberg
00:09
Hello and welcome to the Future Ready Podcast from Siemens. I’m Nick Finberg, and I’m happy to continue my conversation with Mark Hindsbo on operations software and software-defined systems – this time exploring how hardware, AI, and Digital Twins are coming together to reshape the future of industrial operations. In our last discussion though, we explored how adaptable, software-defined production relies on strong data foundations: from connecting legacy equipment to enabling more proactive, AI-driven operations. And one idea that really stood out was that moving toward software-defined manufacturing doesn’t mean less hardware. In many cases, it actually means more: more sensors, more connectivity, and more intelligence at the edge. So, Mark, picking up where we left off – how do you think about this relationship between hardware and software? How does the need for more connected hardware shape the future of software-defined operations?”
Mark Hindsbo
01:56
So, if look if I look at say the drives and so on that we are designing ourselves. We have extremely smart design teams that that look at all the characteristics of that drive. They know exactly what ball bearings, what other components, what the housing looks like. You need to do all of those things to create a virtual representation of your product when you then have that digital twin then you can have sensors on the edge so to speak that take the macro vibration of an overall drive and is able to pinpoint what’s going inside going on inside, right? But if you don’t know that if right, if you don’t have an understanding, full understanding of your system, you can’t do that at all. So, I’ll give you another example that might put it on its edge. But yeah, I talked a while back with a producer of pumps and they are putting these patterns on their seals, and these are nanometer thick patterns Okay. And the way those nanometer thick patterns affect the flow create suction that creates a better seal so that so that the pumps can be you know more tight, more durable, have a longer lifetime. If you want to put a flow sensor into the flow If that is a couple of millimeters at least thick, if not more, you’re now disturbing all the patterns at the nanometer. So, you cannot put a physical sensor in there, right? Without ruining the effects So you kind of have to be able to simulate it. You have to say if I put a nanometer, so with simulation, you can kind of create this X-ray vision if you want that lets you peer inside And then you can use that to then extrapolate outside and say, okay, if this is happening on the inside, this is what I should be able to see on the outside where I might be able to place a sensor. But again, if you didn’t have an understanding of what those patterns should be or what they absolut or what they should look like, then you can’t correlate the two So hopefully that gives you sort of some meat on it. It’s the same thing if you don’t know what machine you’re doing, right? If it’s a metal bending or if it’s injection molding or whatever it is on your shock floor you need that representation and understanding. And that’s also what I meant of data and context. So, I might have all the temperature readings, all the vibration readings. If I don’t have the context that I can attach them to then I cannot diagnose correctly. And if I cannot diagnose correctly, then I can’t act correctly either.
Nick Finberg
04:57
We’ve been talking a lot about software defined and we have these other very core topics around AI and digital twin. Is there a hierarchy between the three of them? Is one more important than the other? Is one undergirding the others? Like how do these interplay?
Mark Hindsbo
05:14
Yeah, I think they’re highly synergistic in multiple ways. So, I gave one example before. If you have a digital twin of an assembly line on a factory and you know what robots are what their capabilities are how their arms can twist and so on You can test for a lot of things as you design your factory. Do I have enough clearance? Can I maintain this assembly line? If something breaks, can I reach it? as the product you know traverses the assembly line, can the robots actually reach it, is what we’re asking them going outside of their capabilities. can the human operator get to where the human operator can the AMR on the shop floor get material to and from the warehouse in the right way? So that’s one where you sort of simulate and you then build your production facility. Then in production, you can then use that model to query changes So, you know, what happens if I now do A instead of B? Does is that better? Is that worse? So, before you or you know, imagine that that has two robots crashing and now destroying each other. You want to discover that in the virtual world before you make the change in the in the physical world. So, if you have an AI who is really good at reasoning but might be ninety-nine percent accurate. Right? You can you can let the AI play around with different scenarios and you can catch the hallucination or the place that they’re wrong with the deterministic model before we implement something. So, you kind of have both design and operate where they feed each other, but you also even have We talk a lot about synthetic data as well. So, when we talk about training an AI and developing a new machine learning model, we might have some data on factories and how they perform. But if we have a great simulation model, we can produce even more data and even more variations. Let me give you maybe a Stupid example, but just to illustrate it, are you familiar with survivor’s bias?
Nick Finberg
07:30
Yeah, I survived thus my actions were the correct actions.
Mark Hindsbo
07:35
Yeah. Or there was this there was this study under you know the Second World War where they analyzed all the bullet holes on the returning planes oh yes to figure out where you should reinforce it but in fact they came back with those holes having survived so what you want to do is you want to analyze the planes that went down And those are the holes you want to reinforce, not the holes on the planes that came back. And if you think about our factories, we have a lot of data in the normal operating range. But we don’t take our factories and run them dangerously close to breakdown or beyond breakdown. So, you don’t have much data in a lot of different areas that might be outside of normal operating range, which is a I guess sort of analogous to survivor’s bias. So that’s a place where you can use but it but you want your AI and your autonomous processes to be able to act also when there’s an emergency or when something is going out of spec. So, you can then so it’s not just sort of your It’s your AI feeding one way, it’s also your simulation feeding the other way and creating better models for tomorrow. So highly synergistic.
Nick Finberg
08:45
Yeah. It sounds kind of like what we or what companies do for training autonomous driving of That’s exactly it. We have this 80% very easily. We know the rules of the road, we know the speeds, everything. But we don’t know if something goes wrong, what happens? What do we do?
Mark Hindsbo
09:03
And there’s been a lot of studies that show and you kind of need both, right? In the in your analogy, you actually need actual cars that drive a number of miles as well, and you need that real data. And then you need synthetic data that can then complement it. And so far, a lot of the research has shown You can get to one level of accuracy 90% for your physical data. You might be able to get 90% only with your simulation data. But then when you combine them, you might be able to get to 95 or 97, right? So almost always when you combine the real and the digital world in these scenarios, you get a better, more accurate result out of it.
Nick Finberg
09:44
Could you talk a little bit more about how this is expanding the design space if you’re able to test a little bit earlier and try out maybe wild ideas first in simulation or even with a first order model of some kind, whether that’s AI or a more deterministic style.
Mark Hindsbo
10:03
Yeah, it definitely. So, I in general, especially when you come to complex machineries, jet engines, cars, you name it, building a prototype is extremely expensive. Millions of dollars, many months. six months maybe. And then you can produce one experiment. Typically, with the simulations we have today, you can produce thousands, if not tens of thousands. of simulations. So, you can do crazy things. One of the things that that I’ve talked about for years now is Sometimes as practicing engineers, we are taught not to ask too many questions because they’re expensive. If I ask you know fifteen what if questions and I have to build this foot physical prototype every time right it’s fifteen million dollars or if I’m on the shop floor and I’m asking a million questions and it’s stopping my production every time I ask the question, very quickly you get, hey, think first. Right? Think very carefully. Only ask two questions. Only ask the most important questions. But that’s fundamentally wrong. We should be like small children. We should be asking a million you know, what if questions. But we can only ask those what if questions if the cost of asking those questions and the speed of which you can get answers to those questions come down. And that is really one of the great places where both for product development but also for manufacturing development, the digital twin and the digital you know prototype really allows us to ask orders and orders of more questions than we could before.
Nick Finberg
11:50
It reminds me a little bit about some information I was looking at around software-defined automation specifically and the split in the engineering stack so that you can do those questions without impacting the live operations. Yeah. could you talk a little about that from your perspective?
Mark Hindsbo
12:12
I think it is, as you said, exactly straightforward. Because as you’re optimizing something in in live production If you if you have to implement that every single time, there is a time cost. There is also a risk of being ruining versus being able to virtually run tens of thousands. of simulations and choosing the one that turns out to be the most accurate or have the most in impact. So, it it’s really rather than I mean a lot of what we do and that’s not bad we should continue doing that but your kind of when you’re when you’re optimizing an operational system traditionally you kind of wander down the Pareto curve, right? You make a smaller change; it hopefully gets better. Then you let that operate for five for a little while. Then you, you know, make another change incrementally and you work your way down. to some you know local. Approaching limit. Exactly. Where whereas as our you know digital twins and our simulation lets us be a little more wild and take larger steps and take steps in multiple directions all at once and see where that you know ends up. Then we of course just have to be careful because our digital twins, let’s be real, they don’t capture everything There might be slightly higher humidity. There might be a little bit of dust in the air, whatever it is, right? There might be a crack propagation in some metal or whatever it could be that’s not captured in your model. So, you’ll you know the models do not They do not take work away from engineers or take away from being a good factory operator. They just allow someone who understands the system to ask more questions quicker but I you know, I hate to say it, engineering sense and common sense will still be needed. Yeah.
Nick Finberg
14:08
Okay. This is a lot of really interesting ideas about what is happening in in operations and in the industry right now. if you could kind of visually look at the future, what do you think it’s gonna look like? How is it gonna be different is in say 10 years from factories today? if you’re on an operator on the shop floor, what’s it gonna look like?
Mark Hindsbo
14:32
Well, I think it what it will look like is you will all almost have you know superhuman powers in how much change you can do on this on the shop floor. So instead of understanding one machine in detail and being able to optimize that machine, which is where we typically came from, right? One operator, one machine you might now be able to really understand and use your sort of your digital tools, digital twins, digital AIs. to help you understand the cascading effect of multiple steps coming together. And you’ll be able to troubleshoot and take corrective actions and so on in context. whereas today you would be limited in your scope. So, I think it really gives you a much broader scope. And the output of that, as I said, is we will get factories that are much closer, that are much more adaptable, that can customize products to individuals in in the extreme with personalized medicine. instead of doing a batch of ten million of the same drug, right, I can get a batch of one for me. That’s what it’s gonna enable. It is in in sort of the mid to long term.
Nick Finberg
15:50
Alright, and the last question I have prepared is actually from Rainer. So in in our discussion, I asked him a question, why Siemens? And he wanted me to ask you the same question. We’ve talked about this the entire time, but why Siemens? What why are we the ones to help customers do this?
Mark Hindsbo
16:09
Yeah, so I’ll break it down. I think it is because we do scope fundamentally the digital and the physical. So, we have pure digital software tools for product development, and we have actual factories and hardware PLCs, and we scope that entire life cycle from product ideation. to product creation and shipping it out of the factory. I think that is very unique. The second thing is we are an industrial player. we live this space ourselves. We serve the industry; we have decades and decades of industrial experience. We have hardware people, software people, mechanics, shop floor people, everyone. That I think is very unique as well. And then we have this innate curiosity to reinvent ourselves, you know, as I talked about, that that Siemens is not the Siemens it was 175 years ago. I think if you have a mindset of what I’ve been doing for the last fifty years is the right way to do it. that you know that you have to have respect for the legacy and what that taught us, but enough irreverence to question yourself and continuously reinvent yourself. And that’s a tricky place. It’s easy to it’s easy to fall on either side of that spectrum, right? And be a pure explosive inventor and say the past sucks or be in the other’s side of things and say Hey, the past has taught us ever anything we ever need to know and it’s unchanged. Yeah. So those are the three things for me. I’m spanning the digital and the physical. The fact that we are an industrial player within the space together with our customers, uh, and then the sort of adaptability.
Nick Finberg
18:03
Well, awesome. Thank you so much for taking the time, Mark. This has been an awesome conversation, and I really look forward to talking with you again.
Mark Hindsbo
18:10
It has been wonderful. Thanks for having so much fun together.
Nick Finberg
18:13
And a big thank you to the audience. We’ll be back with even more great discussions soon, including some upcoming recordings from Hanover Mesa twenty-six We wanted to get some of these wonderful conversations from the show and bring them back to a wider audience. But until then, maybe check out some of the other episodes on the channel. We have some focusing on AI in greater detail. as well as some conversations out of the pharmaceutical industry, but we’re constantly adding more. See you next time for the Future Ready Podcast.
