A Practical Application of AI for Manufacturers and beyond: Automated Inspection & QC

Welcome to the Industry 4.0 Club's fireside chat. The club's purpose is to engage the power of diverse worldwide talent to accelerate the global evolution to Industry 4.0. We accomplish this by hosting and supporting respectful, thoughtful, and interdisciplinary open conversations and creating quality content on wide-ranging Industry 4.0 topics to expand the reach of our community. My name is Mike Ungar, and I am the CEO of the Industry 4.0 Club. I have over 40 years of leadership and experience, with most of that being in manufacturing at Michelin. I'm joined today by my co-host, Mike Yost. Who's chairman of the club. Mike, would you introduce yourself, please? Sure.

Thank you, Mike. Appreciate it. Thank everybody for being here. As Mike said, my name is Mike Yost. I've been in industry for about 35 years, the first half of that on the industrial solution provider side and the last half being on the nonprofit side. So I'm the volunteer chairman and really excited to be here today and learn from Wilhelm about what he does. So thanks for being here.

Thanks, Mike. So today's topic is entitled A Practical Application of AI for Manufacturers and Beyond: Automated Inspection and QC. We're joined today, as Mike said, by Wilhelm Klein, who's CEO and founder of ZetaMotion. Wilhelm combines a deep technical background with a focus on human-centered design and societal impact. Under his leadership, ZetaMotion develops AI-powered quality control solutions blending innovation with a pragmatic, empathetic approach that enhances both technology and human value in industrial operations. Welcome, Wilhelm. Glad to have you.

I'm delighted to be here. It's lovely to join both of you, Mike and Mike. I'm very much looking forward to this conversation today.

One comment, Mike, before you jump in there. For those listening in on LinkedIn, if you would like to ask a question of us or Wilhelm during the broadcast, please don't hesitate to put that comment in the LinkedIn comment section, and we'll be glad to pass it on to Wilhelm. Mike, to you.

Sure. Well, welcome everybody. And Wilhelm, just one— before we jump into the topic area specifically, just wondering if you could give us a little bit of background on what led you to found the company and Absolutely.

Well, I'd probably have to start very early when I first took apart my parents' vacuum cleaner when I was about 12 or so. And also my love for science fiction that was passed on by my mom. So growing up with a vision of the future, what it might look like. talking a lot about artificial intelligences, artificial consciousnesses, and so on. I think that very much got me on the path towards technology, but then also marveling about what is intelligence, what is consciousness, what makes the lights tick in our brains, and could that be transferred elsewhere as well. And through various different pathways, I found myself doing a PhD on technology ethics, actually, which had some connections also into what is now a very well-known field. But at the day, back in the day, nobody was talking about it— AI ethics, robot ethics, and so on. But I was also, yeah, still on my technological journey. I was a director at the time at at the local hackerspace in Hong Kong, the first maker or hackerspace in the city. And yeah, I got familiarized with the topic and got in touch with my co-founders there. And when it was time to finalize the academic path there with a PhD and to see where else we could go, having the opportunity to apply or to be part of creation rather than examination of these kind of topics and starting a company, that sounded like a very interesting direction to go. And that's what we did. We got started as many classic deep tech startups do with very cool tech.

Yeah.

In search of a good, of a good, Yeah, a good application. And we first started on what was motion capture and 6 degrees of freedom tracking, but relatively quickly pivoted into a different direction. And again, this classic startup story, the way that came about was, yeah, 6 degrees of freedom tracking, motion capture. We got some very nice early traction in the public large event space, tracking crowd reactions and so on. None of the creepy sinister stuff, but rather allowing interaction and so on. But that was 2020, which was the worst possible timing to get into anything that had to do with large crowds and so on. We quickly saw that if we dropped the 6 degrees of freedom tracking and just doubled down on the AI, we could land in the field that viewers will be seeing up here in the corner, which is in the quality control space. So that's a bit on my background and then also a bit on the company then.

I'm just curious if you ever put your mother's vacuum cleaner back together or not.

I did.

I believe there were several screws missing in the end, but it did its job. It did its job.

Yeah. Well, that's fantastic. So you pivoted to this, to computer vision quality control, which is really great. So can you talk for a minute, what kind of problems do you solve for manufacturers?

Yeah, so this is really not a unique problem in general, but our particular capabilities find unique niche and unique applications where we solve a lot of problems. In general, it's quality control, it's inspection, it's the job that in many companies is the most hated. We see in many cases people are rotated through because nobody wants to stay on the inspection tasks for hours on end. But it is absolutely crucial, right, to make sure that quality is maintained, not just at the end of the production line with the end product, but ideally throughout the production line and throughout the production process. Because the earlier you can spot issues, the better you can increase the efficiency and the better you can stop putting value into a product that might already be doomed, right? Because you already have some critical error. in there. And classically, this is done in a manual fashion wherever possible, but in many cases it is not possible at all because you have a very fast-running production line and you can only do samples in the end. But here with automation, with the technology similar to ours, or with ours of course, you can introduce inspection at earlier stages at speeds or at environmental contexts where you wouldn't be able to put a human. And you can, yeah, you can imagine having a superhuman who never gets tired, who never makes any mistake, and just, yeah, does an excellent job with superhuman eyes at extremely high speed.

That's I appreciate what you said. I mean, I think building it in upfront really makes a lot of sense. Are you seeing more and more manufacturers today asking you to help them early in the process as a result of that? I mean, that thought process is really important, I think.

Yeah, so the first impression is often that it's purely about replacing human labor for what is already in place, and that's really not the case. What we see as the much bigger opportunity is with this, having what you might call a digital passport or a digital twin of the product as it moves through the production line. So you can completely correspond what is happening with the processes you have on the line. So the earlier you can spot, the more you can increase your efficiencies, right? So this is Because end-of-line inspection, you often get what in the pharma industry is called a batch killer inspection, right?

Right.

Because you do the sample inspections and then you have too many critical errors and then the entire batch is dead and it's gone. In some industries, you can't recycle it, it's just lost. But if you could inspect every single product early on at every step and you can do correlation of causes and effects, then you can see that something that might not even be an actual defect yet will result in a defect at a certain level of probability later on if no further actions are taken. These kind of analytics we can do if you have very high accurate tracking and inspection at different steps of the production line. And a lot of manufacturers are looking into that. They see that there's a lot to be gained, a lot of data is left on the shop floor unused, and much good could be done with that, not just for the bottom line, but also in terms of the waste management, in terms of the resource management, energy being wasted. So it fits very much into Yeah, not just ROI improvement, but also improving their sustainability.

So, Wilhelm, you mentioned, you know, starting the company in 2019, 2020, which obviously was a difficult time for all of us. But obviously, it's been a time of rapid technological change as well. So how have you seen the technology change? How's that impacted the applications, your company? As someone who sits right in the middle of the AI world, how have things evolved over the past 5 or 6 years?

Yeah, the speed is quite amazing. But then at the same time, you really have to differentiate between application level and what's in the media, because a lot of the things that we see in the media have actually been around for a very long time. They just haven't been scaled to the same degree. And they haven't had the same exposure or had not been translated into a format that was so easily usable. Of course, the big LLMs and so on, the things that are really transforming a lot of sectors right now, they are something truly impressive. But yeah, it's not that things have gone, yeah, explosively fast.

Yeah.

But indeed, there have been important trends that one shouldn't miss out. So we're always making sure that we are remaining at the cutting edge and we incorporate whatever comes through. So if you look at our particular area of application, which is visual AI, which is computer vision, surely the latest tech and generative AI was quite the important step. So if you look at the kind of synthetic data that we are using, this kind of approach, if you go back about 5 years or so when we got started, the main way of doing synthetic data was using either just completely manually generated synthetic data using Photoshop or similar tools, or the more advanced or sophisticated approaches were using game engines, something like Unreal or Unity. And now this whole approach with Stable Diffusion is really entering the scene as well. But it comes with a lot of challenges because you need to have full control over the model. Otherwise, you introduce a lot of risk to have too much hallucinated data that contaminates what you're doing.

And are there advances both from a hardware perspective and a software perspective that you have to keep up with?

The hardware is not scaling as quickly. as the software. There are constantly new types of hardware coming out, but it's relatively easy to adapt to these. The magic is really usually in the software, is in the models, in the AI. But we see that we can, or that most of these AI models can interface with quite a lot of, quite a large range of hardware types, including even very, very old ones that might already be in a production line or in a given inspection context.

Yeah, that's what I was going to ask. So if you already have infrastructure in place, systems you've done in the past, and then there's breakthroughs, you know, I guess maybe what would be some of the breakthroughs if somebody has You know, a system has used one for years and, you know, now, you know, with an approach that's AI supported or driven, what could people expect today if their mindset is a few years old of what you can get out of a, you know, a traditional type of visual system?

Yeah, you would usually see quite a lot of limitations to what the system can do. You would be told a very rigid set of parameters that it can check for with probably not the highest level of accuracy to expect and very, very little ability to upgrade or to grow. And that we really found to be quite the struggle for a lot of manufacturers. We had a conversation not too long ago, and I think we even put a version of this in a kind of anonymized fashion on our website where we were talking to a steel manufacturer who had spent a lot of money on an inspection system to inspect their large metal sheets, which is ultimately going to be building cladding. And it was working quite well, adequately for the task it was given. But they saw their production line evolving because they followed their own market demand, and they started to put embossed patterns on these metal sheets. And the vision system they had in place flagged everything as a defect. All of these patterns, all of this is wrong, right? Everything is defective.

Yeah.

And the manufacturer went back to the vendor and said, can you please do something because your system is now rendered unusable? And the vendor said, nope, sorry, that's what you got, that's what you paid for, that's the specs we gave you, and there's nothing we can do about it. And that, of course, is rubbish, right? How can you work with something like this? The reality on every shop floor is that production lines are evolving systems. Products are evolving. There's always going to be changes, and you need a system that can adapt quickly. And for that, you need a system that can scale with you, that can grow with you and co-evolve with you. And some of the most advanced AI systems can do that. Now, the older generations will still be struggling because you tend to have this very long training cycle until you reach high levels of accuracy. So especially if you have a lot of variations, if you have quite a lot of product designs or changes in some of the properties, it can be quite the challenge to have a system that can inspect all of this and can respond quickly. So that's why it's quite important to choose one wisely to make sure it can adapt as We have a question from the audience that kind of comes to what you're talking about here.

Valentin asked the question, does the quality control process update the digital twin? You mentioned earlier the idea of a digital twin for the product. How do you— it sounds like the answer to this is yes, but can you elaborate a little bit more?

Well, we're not We are generating the twin essentially, or we are generating part of the twin. So what we would be doing is we would be determining what is the correct sensor to control or to inspect a given parameter for a given product.

Mm-hmm.

And then we assess that parameter accurately, and if it makes sense, near real time. And make that available for either via our own platform or interfacing with an IoT management system or PLM system, whatever someone might already have in place. So what we do is we generate that primary data about the product, and with that data you can generate an accurate digital twin. Okay, that makes a lot of sense. And that gets updated. On every stage, right? The more inspection stations you have, each time you run through, the current status of that product is updated, of course.

No, that makes a lot of sense. I appreciate that. I think that what's exciting when we think about AI and the dynamic nature of this is quite interesting, right? Because we were talking before we got on the broadcast about some products that are natural, like rubber. And, you know, from moment to moment, they're not going to be exactly the same. And so that idea of a dynamic model is really, really interesting.

Yeah, we also briefly talked a little bit about what makes our approach unique, and that is our synthetic data approach. And synthetic data sounds a bit foreign, right? So the analogy that I like to use in explaining what we do and how it works is really, if you just imagine the process of onboarding a human quality control engineer, right? When you do that, you don't hand the engineer 5,000 fully annotated pictures with the labeled defects and a label of what that is to teach that engineer, right? Instead, what you do is, let's say you are inspecting a mouse, right? That's the job. You would give him or her one or 2 example mice, and you would show, okay, look, there's a scratch here. And this scratch is okay if it's like 5 millimeters long in total, as long as the radius of the defect area is no bigger than 3 millimeters.

Okay.

And you spend a little bit longer and you go through the defect catalog and about an hour later or so the human says, okay, I get it, I understand now, right? And if tomorrow you change your production line or you introduce a new variant where the mouse has now lots of triangles on it because it's a new design that your customer has asked for, that the human is still able to inspect. Because he or she has understood the deeper meanings, the semantic principles of the quality parameters that you're looking for. Now, if you try to do the same thing with a classic machine vision system, it's going to freak out because you're changing from something like this to something that has triangles. I have never seen this before. All of this is wrong. All of this is defects, right? So you need a system that works like a human because what works, what happens inside the human brain is extrapolation. And what we mean when we say extrapolation, essentially we mean synthetic data. The human mind imagines, dreams up variations of what that would also look like. So when the human says, I have understood it, he has imagined the different ways how this could also look like. Now you can have a system, a clever AI system that mimics exactly that approach. but at a superhuman level and synthesizes these variations as well. And then you can have a system that can adapt like a human would without the requirement of having, again, thousands of pictures.

Well, I guess, first of all, thank you for the question and remind folks that if you have questions, feel free to throw them into the comment field and we will interject them here as we can. But you talk about the human, and I guess it brings up a question from my perspective of humans using the system, and are they expected to know AI? Do they have to keep up with the evolution of the technology, or is it pretty much a self-contained box that they just have to know how to operate? What's the interaction like, or the expectation from a user perspective? of what technical capabilities they need to have?

That's a wonderful question, thank you very much, because it's a perfect cue as well to another big pain point and annoyance. Because a lot of the tools, especially things that you can download online, a lot of these platforms tend to externalize the difficulty and the challenge to the user, to the manufacturer, and make it your fault when you don't achieve the accuracy that you're looking for. It's because you didn't upload enough data. It's because you didn't do enough labeling. It's because you didn't do enough training and whatnot. Or even worse, you didn't select the right model and so on. And these are the application types where you really need to get pretty deep into these topics, and you might even have to hire an AI engineer to do this. What we are aiming for is really Something that is very easy and intuitive to use. All you have to do in order to, let's say, onboard a new product variation is you upload one scan, as in you load the new product into the inspection station, you go to the dashboard and you say, I'm onboarding a new product, you click scan, it runs through, the system assesses it, captures the scan, you have the opportunity to enter whatever additional parameters that you have for this product. Let's say you have your customer doesn't care about a certain area on the product, so you can just mask that out and the system won't assess it at all. But otherwise, you don't have to do any magic. You don't have to know much about it. And when we're getting started, it's really the same way I explained the system mimicking what a human would do. We are mimicking the same process when we learn from the manufacturer in order to teach the system. So you would also spend some time with us and explain it to us the same way you'd be explaining it to a human, and we then translate that to the computer, to the system for the initial onboarding. And then after that, we are not even necessary. They can just do it from the inventory page.

Yeah, that's a really— that's, uh, thanks for taking us through that explanation. Um, I want to just pivot a little bit to talk about sustainability. Um, you know, we have sustainability— a lot of companies today, in fact, any really— most leading global companies are focused on sustainability. Um, how does Zeta Motion help a company achieve its sustainability goals?

Yeah, I alluded to that a little bit in the beginning, but we are— the personas that we see getting in touch with us tend to be CTOs, tend to be production managers, tend to be innovation managers, but also a lot of people who are responsible for the green efforts at manufacturers. are responsible for driving development towards improved sustainability, decreasing the environmental footprint, decreasing their carbon footprint, because they know that so much can be gained by simply assessing more about what's happening on the production line.

Mm-hmm.

Because there's If you increase the efficiency, if you increase the quality, you don't have to reinvent the wheel, right? You don't even have to change much about your product, and you're still making big steps towards a greener manufacturing process and a greener product in the end, simply by eliminating the waste that might otherwise be happening. And especially when you have products that are being produced at very high volume and very high speed, and cannot easily be recycled, that can just make such a huge, huge impact. Because if you have— imagine you're producing at, say, 5 meters a second some continuous product and it cannot be easily recycled at that, and you— and it takes you, let's say, 30 minutes to do an inspection of a sample. Within those 30 minutes, you will have produced a few thousand additional products, all of which might have to be scrapped because of that delay. And it's not because the workers aren't the best at what they can do, right? It's just beyond what is possible, what is humanly possible, right?

Right, that's great. No, it's— when you think about that, it makes a lot of sense. It kind of goes back to what we talked about earlier about moving the inspection up early in the process and making sure that at every step we're producing good product, huh?

Yeah, indeed, indeed. And of course, in the end, if the end product is better and lasting longer, that also contributes, right? So there's many ways how this kind of improvement of knowing what is happening, of having an accurate image, of an accurate digital twin of what your product does can really help you on the sustainability front as well.

Well, Helmut, I know we're getting toward the end of the talk here. I'm wondering what you've learned along the way, right? It seemed like a good idea 5, 6 years ago to jump into a startup and do this. So what have you learned as a leader of a tech startup company?

I think one of the biggest lessons is really quite fun because it's such a juxtaposition because the tech that we're talking about is so sci-fi and is so— yeah, it sounds so, you know, oh, it's all about robots and the Terminator and so on. But ultimately, it's such a fundamentally human challenge after all in this area of application as well that you just cannot underestimate or you cannot overstate how important it is to run a human business that applies AI.

Yeah.

Because it's all about human expertise. It's all about capturing what has in many cases grown over decades in experience and know-how on the production floor with these products. And yeah, you have to talk to these people and you have to give them a tool that truly enhances what they're doing. So another image that I like to use is that don't think of some robot coming in and taking over the job. Rather think of being handed an extremely clever magnifying glass that allows you to stay in control, that allows you to keep or maintain the positive aspects of everything that you've built, all of the expertise that you have, but do so faster and more efficiently. So it's really a helping hand rather than something that takes away. But again, all of this is only possible if you arrive with empathy, with understanding, and with respect. So yeah, despite everything being about AI and sci-fi, it's really the human touch and the human connection that counts.

That's great to hear. Appreciate you sharing that.

Yeah, that's really great. We're about ready to wrap up, so maybe Mike, a final thought, and then Wilhelm?

Yeah, I guess just a couple of things. I mean, I certainly appreciate Wilhelm's perspectives and the closing point on the people and how important that is. I mean, that's really central to what we believe at the club, why we do what we do. And so I think that's fantastic. And also his recognition of sort of how things have evolved. And, you know, I guess when we step back and look at what we do here and we keep getting smart people like Wilhelm showing up and talking about these new advances and new technologies and stuff. It just reaffirms to me that there are smart people out there that have empathy, that know what they're doing. And so hopefully people are, I guess, feel welcomed into or safe stepping into an environment and having a conversation with folks like Wilhelm. So I appreciate everything that you've had to share here today.

Thank you very much.

Yeah, Wilhelm, any final thoughts on your end for the group, for the audience?

I was delighted to be here, and if there's any questions, I can easily be reached on LinkedIn. If you search for Wilhelm Klein, Zito Motion, you'll find me. So if there's any question, if there's anything that you'd like to follow up on, don't hesitate to reach out and send me anything.

Well, great. And so Again, thank you so much for joining us. It's really been great to hear your insights about AI, quality control, sustainability, leadership, and most importantly about people. To be honest with you, I think it's— that's really great. So the Industry 4.0 Club would like to encourage you to take action to advance the deployment of Industry 4.0. Read an article, attend a workshop, develop a smart manufacturing roadmap for your plant or your company, or really challenge your organization to move forward. With the deployment of Industry 4.0, we'd like to know what actions you're taking to integrate Industry 4.0 into your operations. So please reach out to us on our website or on LinkedIn. The club routinely hosts conversations like these on LinkedIn and YouTube. If you'd like to join us for a future conversation, let us know. Our website is www.industry4.0club.com or and or follow our LinkedIn page. So. We, we'd be happy to hear from you. And with that, uh, we'd like to wish you a good rest of the week and a good day. Again, thanks, Wilhelm, and thanks.

Thanks, Wilhelm.

Thanks, folks.