AI, Digital Twins, and the Future of Manufacturing - With Laurent Laporte of Braincube

[MUSIC] This is Dan Vagella and you're listening to AI in industry. In this episode, it will be focused on the topic of digital twins and the future of manufacturing. I think most folks have an idea of what a digital twin is. But very few have an idea of how artificial intelligence can help enhance the idea of a digital twin and bring value to the manufacturing process. And that's what we cover in this episode. Our guest this week is Laurent LaPort, who is the CEO of BrainCube, French firm, focused on artificial intelligence in the future of manufacturing. This episode is sponsored by BrainCube. For more information about sponsored episodes and creative services here at Emerge, visit emerj.com/advertise. Without further ado, we're going to hop right in. This is Laurent LaPort with BrainCube. [MUSIC] So Laurent, I wanted to be able to start us off with sort of the opportunity of data in manufacturing. I think there's some folks that are really excited about it. I think a lot of people are quite confused about it. When it comes to sort of understanding the new value of data in this era of AI, how do you like to frame it for business leaders? >> Well, there are different missions of a plant. In the plants, there are a lot of numeric digital equipment for about 20, 30 years. So using data is something that we do every day. We use it for controlling the process, for we have some KPIs. But before the IoT era, before this time, we could never cross the economical information and the physical information from the plant. Now, let me illustrate my point. Let's say you have a nice business, you know how to manufacture your product and you receive some order, but the economy is not stable. Some months, you get access to higher demand. So what you're supposed to do is to set up as a KPI, your tonnage, your speed. You want to deliver more. And you think you can do it just by pushing the speed button. But what will probably happen is you will end up having more issues and troubles and unexpected things because you just try to push more from a system that has not been designed for. There is the other case, that means the demand is dropping, then you are in a situation when you need to deliver at the lower cost. And most of the time what the people does is just shut down their plants for a few days like do start and stop and this are costing a lot. In these two cases, what the people are doing is just running on habits on a started way and they don't really take care of the opportunity of maximizing their cash flow according to the economical situation. So what will be the future of using data is to be able to cross information from external system, describing the economical situation from internal IT system and all the equipment. And two answers to this very complicated equation is what is my optimum configuration to deliver the maximum cash flow. Got it. So an optimum configuration, I guess, is sort of what steers us into the world of digital twins. I mean, as someone who doesn't spend all their time in manufacturing, when I think about a digital twin, I guess I think about models and representations that we could use to determine different experimentations on what optimal could be. Maybe you can talk a bit about how this leads us into the era of digital twin and how to think about that. We'll talk about the twins we can use in the manufacturing. There are basic twins like the asset-win or the logistic flow twins that have been used for a while. The new twins that are coming are really interesting and powerful. Every product you manufacture is unique. It has been made with certain conditions and you can describe the exact condition. But if you look at the situation, most of the data you collect in a plant are not related to the product. The only one are lap test. For every product, you might have a lap test telling you this product is exactly this and this. But how did we manufacture this product? We don't know. On the other end, we have a giant amount of data that are generated in life continuously by the equipment. The digital twin would be, I will use this road data to recalculate the exact condition of manufacturing to everyone product. And at the end, you generate a new data set, which is a higher level of information, where you have a detailed description for every product you manufacture in your plant that tells you how you did it, how much it cost. We're doing it. What was the condition at this time? What was the external condition? And you add the lab measurement at the product level and you have it. And what does it change for you? It changes. If you do a regular accounting system, you are in the business of averaging everything. You have reached your real-time cost. You are averaging your overhead cost. So you just capture where you are in average. With the digital twin, as you have a highly detailed description of real condition of manufacturing of everyone product, you get access to something that is really more powerful than average. You get access to viability. So you can clearly see, visualize or even count and measure the viability of things that are new to you. Starting from this, you can understand how sometimes you can manufacture a product at a lower cost, or how sometimes you can manufacture a product with a shorter cycle time. So you can probably generate from this data set different scenarios that will match to your accounting called constraint. And then you start to close the loop with my first point. And you can start to drive your plant according to the account called situation. And so you can start to maximize your return investment. Got it. Okay. So a much more granular responsiveness to the situation if we have these different kinds of digital twins, which I know we're going to get into some examples of in a moment here. But as you had just mentioned, data has been used in manufacturing for a while. And processes to pull data from machinery, et cetera. I mean, this is not entirely new in novel. However, these new kinds of digital twins and this more swift responsiveness to real time probably involves not only more kinds of data, but a really a different way of treating it, a different priority of treating data. I know you're pretty big on emphasizing the importance of data quality in making these, the benefits of this digital twin kind of a reality there. What do manufacturing leaders need to understand, I guess, about what's changing with how data is treated and collected? We are today in a world where the new technology of crunching the data are based on learning principle. And so everybody can understand that if you learn from a data set that is not accurate, correct, or that doesn't describe what you want to understand, then there will be no good answer. By building a virtual twin, what you do is you don't really use your data. You use a new level of information. And that is really not information. It's just an opportunity to build information. And by doing this, you can access to information you cannot really measure. Like the real real time costing, you can build some predictions, predictors inside your description. So there's a new granularity of details that you can capture through this. And so it's a basic of building your data sets where any kind of learnings you want to do out of this will be possible before it will probably not be possible to do something more than a small and local study of a detail of your process. But what you want is not to additions, some mini studies everywhere in the plant, but you want to add the optimum in the world picture, the one that will deliver big improvement. And when you think about what that's going to demand of folks who are leading manufacturing, what kinds of new staff, new processes, new instrumentation to kind of change culture and change the environment to be able to prioritize data in this way, what are some of those shifts, these underlying shifts that are actually going to unlock AI's actual value, the value of digital twins. I think there's probably a lot that's going to need to change. And maybe you've got to take on that. 10 years ago when we started this business, I was discussing with the process engineers that were in charge of solving the problem. And they were all telling me the same story because I don't have any information about what happened, like the last Saturday night or whatever. The guys are spending 80% of their time to try to go and discuss with people and collect and call for data from different systems. And they were spending really a huge part of their time just to get access to what was existing as information. And then they had to provide answers, solutions or ideas about them. When you have a fully connected plant and you refine your data and generate virtual queens that are accurate, you save 90% of this time. What you get to do is just to access to your platform and start to ask question to this level of information. So it totally changed your journey inside your job, inside your mission. From spending your time chasing information that will be inaccurate because more of them will come from human. You can get information from facts. And so the next problem will be how smart you are to be able to use this information. So sometimes it will be easy. You draw a graph and you can see a tendency. You can see some points that are strange and you can start to dig and you can start to figure out what the situation and sometimes you don't see it through a simple graph. So you need to go into deeper analysis and try to find the interaction between the key impacting factors on your problem. That's where you will need to be augmented by new systems of machine learning or advanced analytics. But you see before you didn't even think about using this type of system. You were just trying to understand what really happened. Today you have a huge detail refined description of what happened. So the questions are totally different. You start to see what should I do to never be in the situation again. And so you start to deliver potential actions that your team or the plan can execute to reduce the risk of being again in unwanted situations. I think a lot of these points are coming through more clearly as we get into some real examples and walk through how data was collected, why and how was important and how it can be interpreted to adjust to more real time scenarios, get beyond these averages. When you think about the favorite examples that you have to illustrate these points, what are a couple that would help folks who maybe are or are not in manufacturing or listening right now really understand where you're headed here. Okay, so let's talk about, for example, tire manufacturing. When you manufacture tire, tires are a very complex product. It comes from 100 ingredients you are assembling and curing probably 20 to 25 subassemblies. And when you look at it, it comes from trees, you know, you need to get gum from trees. And we consider that a tire is a long process of four to six months to be finished, the time to collect the gum to transport this to you to do all the subassemblies and things. So, okay, let's say you have a machine, which is a laser tester that test the conformity of your tire at the end of the line. And this tire is pretty unique. It's coming from one of the three or four hundred machines of the step before. And it's coming from assemblies from the one of the three of four hundred machines before and so on and so on. So, it's a unique assembly of a long logistical complex flow. Today, we can collect data at every level. So, we have for a tire between 5,000 to 10,000 unique descriptors. When you are chasing for a few percent of rejected tires, I mean, a good plan today will reject point something percent of tires. But there are manufacturing 30,000 tires today. So, it's still a cost. So, what there will be capable of doing live on time is to use millions of tires you manufacture. And to immediately understand that when you use this type of gum at this temperature on this machine and you assemble this with this bar, then you augment the risk by 10 to get the reject a tires. If you understand my point, it's highly complex system of information that was left over everywhere. They were in Thailand or in Brazil for the raw material. They were at your supplier for some subassemblies. They were inside different level of machines and things. And it's all this is past history. So, the visual twin of this tire capture the history and describe exactly what happened. And then you will become really powerful because things you probably cannot use at the machine level if you are smart enough and you spend the right time. Here you can really put a weight on this. You can say this is really impactful. This is not impactful. This exists. This doesn't exist. And you do it on time. So, you can save a lot of money by crunching data and sending back to the field information that they will use. And we are calling it. We're not talking about big things. We are talking of small adjustments that at the ends are probably 90% of loss. So, the only way of improving today when you are already at a certain level, when you are already reduced all the ways the physical, the organizational ways, when you are facing a level of complication that without a new system, you will not be able to address digital tween and machine learning will open you new big, big opportunities for improvement for sure. That's what we have been doing for 10 years. So, we have been doing this for 10 years. We have been doing this for 10 years. In thinking about what those myriad factors could be, some of it probably thinks they are already staring at. For example, they are looking at their machines and they know what temperature they are. They are looking at which machines stamped out, you know, the shape of the tire, which of the many stamping machines. I don't know how tires are made by the way. But you know, which of these many machines kind of made the actual mold, let's say. Then there are probably other factors like maybe even where we source the materials and there is maybe even the temperature of that cargo boat while it was on its way over. I mean, it almost seems like there are so many moving parts. Does it take, is it kind of a bespoke in different system per client, per business to really sit down and think hard about what are the factors that could influence our results because it seems like there could really be a lot. You're right. That means most of the time a plant is just a part of the problem. They are having the opportunity to solve some of it. Let me tell you a story about a giant European plane manufacturer. They are building a new plane and they are using for the first time ever a new type of material just to maintain the wings. So it's it's a giant nine meter long pot in a metal that is used only for this plane. So you can imagine we don't have a lot of historical knowledge about how this material will end up and so there is one company that is building the raw material that this is a foundry. Then there is another company that is processing and a different company that is processing this raw material and then the end customer is machining this part and sometimes they have the price. I mean, the part is not exactly what's expected. The part is twisted or the part is longer than expected. So the machining company is blaming the processing company and the processing company is blaming the foundry that says, hey, I'm doing the right thing. So what we offer that case is to build a virtual queen all along the chain. I mean, we in a platform because a platform is a kind of a giant start out for storing data. We could get data from any system from all this company and recombin all this and recalculated the conditions over process that last six months where we included as you mentioned the storage, the temperature outside the transportation conditions with some IoT devices we could do this. And then there is no guilty. There are just a combination of factors coming from the first supplier to the second supplier to the final users where some storage condition, some machining conditions, some pollution inside the element of addition inside the room. All these were small, small, small spot of a giant problem and we solved it for the good and we have been awarded by a quality award by the corporate of this company because they thought it was genius in few weeks to be able to leverage data instead of having a big fight in between companies and suppliers and customers as you see the form away of something the problem. There was no way to solve the problem if you don't understand the answer underlying interactions between all these steps of processes that are made in different organizations. Got it. Yeah, and I was actually going to ask for a second example, but that was a good one to sort of illustrate this point. It sounds like companies need to be prepared to be sort of nimble here to take into account what they think the major factors are but be able to look at new anomalies, ask new questions and then integrate new sources of data to be able to continue to sort of update where these anomalies come from and how to improve. It sounds like that's probably just something folks have to prepare for here. Yeah, it is. I remember in the early days of what we were doing when we were just the first one starting to pull that out of pins and putting them in a cloud, even if the world cloud was not really existed at this time. I'm talking about 2008. My first problem solving in paper was was funny. The guy we're having a historical problem for about 25 years, which was a defect a surface defect on a product. And after discussions and things when the problem becomes so critical at the customer level that they decided that they had suspicion about what it was. And it was due to probably a moisture problem coming from the fresh air that were you think that was having a lot of problems. of variability because of temperature, outside temperature, and things like that. So they were ready to invest a massive amount of money to put an air conditioning system in the final step of the process. And we crunch data and we just show them that they were having 32 different suppliers of fiber, you know, long fiber, short fiber. And the problem was coming from the mix of some of them that were not just going well together. So just by changing the rule of which raw material we mix, we totally solved a problem that was lasting for 25 years. And the guys were investing in the giant and air conditioning system. So we'd have not solved the problem. So you see, even at the early days of this, when you have the right approach, when you know how you can treat this data coming from implants, they come regularly from 10 to 20, 30, we get a plant with 100 and more data sources different. And I'm not talking about individual machines. I'm talking about different systems, automation, IT labs and things. So when you know how to do this, you do it for the good. You do it for ever. Everybody can benefit from it. Being at the product level, describing how we make a product can concern the financial guys, can concern the process guy, the quality guy, the production guys, everybody can get something out of this product digital twin. No take in there. And I think that hopefully for the folks tuned in here, this is sort of maybe eye opening in some regard of thinking through what are the source of data we might want to capture. What's the kind of, you know, time series and sensor information that could help us upgrade the systems that we're running. So Laurent, I know that's all we have for time. But I appreciate the examples and you jumping on to join us here on AI industry. Yeah, thank you for the opportunity. I really appreciate to talk about this. I think we, we are here to help the people to understand that we can go beyond that just measuring the performance, doing averages and we can really leverage all this American formation. That is today the new waste of the industry. We knew that there were physical waste and the lean manufacturing system know about it. We know we can do better without team on management level. But today it's foolish not to consider it's using the data we generate is not a priority because it's not costly to collect. It's not costly to store. And you can give to your people tools with AI that augment their ability and their mission daily that works and that has been proven today just to finish up because we're talking about digital twin today. Yeah, yeah. A company has bring cube as more than 2,500 different products on light that generate digital twins every day. And we're talking about big millions of product a day. And so it's not that difficult to do it today. We even have solution when the customer can do it themselves. So this is the next focus of every corporate leader. If they want to improve their operation, they should start to think digital twins of the product. So once we've collected the data here, what does it look like to actually turn this into to translate that into a digital twin at the level of a product? It seems like there's a lot that goes into that. Yeah, I want to explain what we do with this road data. If you take a process that is a long process, let's say something like people understand like to manufacture paper, we need to transform wood. So we have some lots of food coming in that we cheap. And there will be a long process where we will refine them bleach the chips and get some pulp. And from the pulp, then we will finally go on a paper mission that deliver paper. Okay. So as you can understand, this process will take two or three days. I mean, the fiber that will finish into a jumbo roll of paper will have a life through the process. And what we are measuring is not theirs. The measurements we have are capturing the real status on time of all the equipment of the plant. So what we have to do is to go back in the history and to model how these fibers went through the different equipment. And they say, oh, yes, these fibers were in this bleaching machine, like 14 hours ago, and this day four hours. And so I need to find what were their exact condition at this moment of time, where I made this the transformation on this fibers that are my product now. So you can imagine there is a complex city of circulation, recirculation. We have some process depth that are batch and some other process depth that can be very quick and different. So we walk with the local team to map this flows and to define every process depth that generate lag in the manufacturing of the product. And so we create this dynamically. We use the real measurement. We use the level of the tank. We use the consistencies of the product. We use the flow, the flow meters and everything that is available to describe what's going on. And when you have done this model, you can process your historical information and generate the virtual twin. And you can keep on doing it live. And that's where you really recalculate, regenerate, uh, informations that describe how you manufacture your product. And then you really transform. It's like a refining process, you know, you transform your oil into something you can use gasoline in the case of oil and for us is information as virtual twin of the manufacturer product. It's really at the end, a data set that describes what you want to know, but you could not measure. And it's a perfect data set for doing machine learning. So that's all for this week's episode. Be sure to stay tuned next Tuesday for our next normally scheduled episode. You're on the AI and industry podcast. I look forward to catching you that.