COMPOSITES LOUNGE - Das Online Experten Netzwerk

Ilkay Özkisaoglu
Since 04/2021 147 Episoden

#142 sensXPERT about Deep Learning in Engineering and Manufacturing Industry

16.09.2024 12 min Staffel 5 Episode 99

Zusammenfassung & Show Notes

Exciting discussions with Dr. Phil on industrial machine learning, algorithms, and deep learning in today's podcast episode.

Dr. Phil Gralla and I discussed industrial machine learning, deep learning, and algorithms. We emphasized the importance of understanding the use case and the machines in use before delving into advanced technologies. Not everything that is labeled AI is actually AI. Many other techniques, such as decision trees and support vector machines, are valuable. It's crucial to choose the right approach for specific applications rather than following trends. In traditional industries like composites and plastics, there may be apprehension about adopting AI, but it should be seen as a tool to support human decision-making, not replace it. Utilizing available data can lead to significant efficiency improvements, as demonstrated by a case involving robot utilization.

Dr. Phil emphasized the significance of statistics in decision-making by machines and the diverse tools available for different applications. Applying AI and machine learning in manufacturing is about supporting efficiency, not replacing human workers.

Stay tuned for more insights from our material characterization journey! Overall, the key is to understand the processes and make informed decisions using the available tools and data.

If you want to discuss this in person, then let's meet at JEC Forum DACH 2024 in Stuttgart on October 22-23, 2024 where Dominik Riescher of sensXPERT - Optimizing Plastics Manufacturing will be available for meetings. See Outro for details on JEC Group's format in collaboration with AVK.

Composites Lounge will be there, too.

Transkript

Deep Dive into Industrial Machine Learning I'm here to have some deep dive into deep learning, decision trees, and algorithms. With Dr. Phil, I had very inspiring talks the past days about mathematics, algorithms, deep learning, deep machine learning. Phil, I would like to ask you to explain our community before we dive into your laboratory equipment here. Because community, you have to know, we have set up a few things, and the machines are working right now. Crystallization / hardening of the composites takes place within fifteen minutes. And we thought we use the time to talk about algorithms and machine learning. Now, Phil, one of the questions is, I'm reading everywhere now machine learning, deep learning, everything is machine learning, deep learning, neural networks. Tell us, and give us some insights from an expert perspective: is everything that we are presented with deep learning, really deep learning? Or is it that people maybe do some simple decision trees and just label it AI? That's a very good topic. So, most likely not everything that you hear deep learning is actually deep learning. And, we have a saying here in the company from us, we say everything is statistics, and that's also the most important part about this one. If we look into how do machines make decisions, and that's the core what we're looking at, is statistics. At the end, it is statistics. And we have different way to do it. We have very classical things, decision trees where information is taken and based on this information, you either go a or b. And then you keep going. You have this one decision. Okay. What is the next data you do again? And, this is already some kind of AI, but it is not deep learned. You might make simple steps deep learning if you want. The other parts is for vector machines, here for over 20 years being used a lot, still. And they're not as fancy as deep learning, but they have their advantages for not using as much data. And that's also a point where Don't get me wrong. Deep learning is very exciting, especially for a mathematician, if I look at Chat GPT, if I look at Image Generation, it is cool. I really like it, but it might not be what we need, especially not if we're working with data that's very different. If we look at our cycle data, if we look at material data, you cannot go online on GitHub and just download a sample of a million annotated data and start training. You most likely have to do it yourself. And then if I go to a customer, be it like aviation and tell you, you have to produce one hundred wings for a first start, and I need half a million to actually train a model. I'm pretty sure I'm not gonna sell you this algorithm anymore. So there is something where it is a lot of times a good decision to actually take a step back and look at the whole tool set that we have. And there, we go back to regression models. We have auto regression models, with random syncs. That's something we are using, especially if you have a little data for analyzing and making predictions in the future. We have some typical decision trees. We have ensembles where we use a lot of different machine learning algorithms that are all good in their own way, and together they form a good decision. You can compare this if you are getting sick, and you're not only going to one doctor, but you have a general doctor, and he says, you go better a heart doctor, and you're gonna go out to a heart doctor, and he checks, and he said, like, yes. But we need someone who's a surgery. So you have to, they're good to go the next step to a surgery doctor. And this is techniques that we can use in the same way where we have a more general model that points us into the direction which one to take next. None of them is perfect, but the combination together makes it. And that's something where we, a lot of times, look at it, and also we make our own models, because this is a process we take. We look into the material, into the information we have, and then choose models that actually are appropriate for our application and also feed the requirements. And that is something always keep in mind. We do not need a model that can tell us and translate languages because that's not what we're doing. We do not need a model that can make images. We are not doing images. We need a model that understands composites, thermoplastics, and thermosets. And if it can only do this one, I'm happy. It doesn't have to talk with me. It only has to fulfill a job. And that's something to always keep in mind why I'm very nice to have a hype, and it makes my work much easier talking to customers if they come to me requesting all this. But don't be surprised if it is not deep learning, but some other technique instead. Doctor Phil, now the composites industry and the plastics industry are very, very conservative industries. With conservative, I mean, they have capital expenditures, CapEx investments in plants and operations and machines. It could be. I'm just just throwing this out of my own observations. It could be that engineers that are operating these machines, setting up these machines. They are looking for a fast output, higher output that they are a bit scared, challenged with all these new forms. I will call it support, because it will not take over the job, then and make the job better, for the human being. It will support the operations. What would be your one tip for people that are scared about applying AI deep learning or even simple decision trees in their manufacturing? So one thing is, like you already pointed out, the it is a tool to help you, and you probably have used tools that are much closer to machine learning than you expected beforehand. Computers do it all the time, making decisions, even though they don't use deep learning, And if you have something like, PID controller inside, that is already mathematics making decisions for you to have this, like, a control of your machine. And we are comfortable with this. And I think that is something to keep in mind, your machine is already doing this, and we just apply it now on more steps. And the first thing you should have is get rid of this idea that machine learning AI has to be like a human. It doesn't have to be, And if you get rid of this idea that we're trying to replicate a human, but instead take it as another tool, another step like a PID controller, I think it makes it much easier to feel comfortable about what is happening and also understand why decisions are not made the same way as I as a human would make it, but it's this helping you to control a very difficult process. That's, really a great conclusion. It's the mindset of the engineers has probably to adapt and understand. We are not replacing any workforce, but we are supporting a more efficient production. And this is all what sensXPERT is about their process optimization. And it helps you understand the process because even if you have 20 sensors and you see exactly what's happening, it's overwhelming. It's something that I noticed before, and people were asking, why do I take so long with analyzing some data and make some nice plots? And I said, because I have to extract data, if you look at what we are collecting, we have a hundred graphs, a hundred plots for each cycle. No human can understand that easily anymore. We have to make it relevant, to make it understandable. And this is already a tool which we use without realizing how much mathematics, how much of the machine learning is already behind it to just make sense of all the information that is being gathered. So, Doctor. Phil, I have one final question on this small excursion, which I thank you a lot for highlighting this The last question I'm having for this part is, I'm assuming, just assuming, seeing walking around the manufacturing processes of our customers' inspection. You know, they invite me to look at the production, and I have a lot of insights And I'm seeing everywhere machine control rooms. I guess there's a lot of data today already available. Now, if you think that all this data is available and nothing happens with it, tell us what is the chance or the opportunity that is missed by not utilizing this data? That's a very good question. And I do have an example about this one, which is actually not related to sensXPERT, but work I'm doing. And when I was student, we had a company we're working with, and they were asking for, we need a new robot, which takes very heavy parts from one machine to the next. And they had all these information already when machines are done when they're ready. And no one actually had a clear look at it. So we started looking into it, and the first conclusion was, like, you do not need a second robot. He wanted one, but actually a lot of your machines are stopping not doing something. So by just more efficiently sorting where the robot is going and where he's putting stuff, we already had a 40% increase of production. And then you can talk about, let's add some robots. But, that's something where we, realize that, the efficiency sometimes is already there, but it just gets lost. No one has a whole picture about the whole process. Also here we at sensXPERT, we look at only the curing of the material, but there is more steps involved coming afterwards. So if you want to have a whole process optimization, this is only one step. And, understanding how machines feed into each other, understanding early enough, when something is happening, like maybe a part that you have is not perfect getting. There's some misfunctioning. Maybe it breaks early. Maybe it didn't bind correctly or it didn't get hard enough. All of these things can have many causes. And some of the stuff you might already monitor without realizing it. So it can help to make use of all these collected data by applying proper tools. Wonderful. A nice conclusion. Thank you so much for listening to our short episode, and now we are turning back to our inspection of materials here, and I'm really keen to see what the result is.