Introduction

Connect with Bryan DeBois:

Rovisys: https://www.rovisys.com/

LinkedIn: https://www.linkedin.com/in/bryan-debois/

Lisa Ryan:  Hey, it's Lisa Ryan. Welcome to the Manufacturer's Network podcast. I'm excited to introduce our guest today, Bryan DeBois. Bryan is the Director of Industrial AI at RoviSys, a leading global system integrator for manufacturing and industrial clients. With 20+ years of expertise in manufacturing software and Level 3 information solutions, Bryan excels in implementing AI, data infrastructure, and advanced analytics to boost productivity in the sector. So Bryan, welcome to the show.

Bryan DeBois: Thanks, Lisa.

Lisa Ryan: Please share your background and what led you to do what you're doing at RoviSys.

Bryan DeBois: I started at RoviSys right out of college, so, for 23 years at RoviSys, which you don't see a lot nowadays, folks sticking around for that kind of longevity.

I've worn a lot of different hats at RoviSys. First, I started programming software for manufacturers. RoviSys' focus is entirely on manufacturing and industrial customers. So I started out writing software. We wrote things like manufacturing execution systems or MES. We didn't call it that because it still needed to be named.

But what we were effectively was building custom MES for customers. We did many different things and worked a lot with historians, which I'm sure we'll talk about today. And then, 2019 RoviSys had our 30th anniversary in 2019. The conversation was, what should we be doing next?

What should we be looking at next? And the decision was made to create a new division, an industrial AI division. I was appointed the director of it. And the message was, we know we should be doing this. Figure it out. Figure out what's state of the art. What kind of a team do you need to make this successful? Then find customers, talk to customers, and make it a reality.

And I've been doing that for almost four years and having a great time doing it.

Lisa Ryan: Wow. This was an excellent time for this conversation because we can only look at something online, on the radio, or on television with somebody talking about AI and how the technology is exploding. What do you think are some of the most significant challenges as well as opportunities for manufacturers in this AI space?

Bryan DeBois: Yeah, one of the things we see with early adopters of AI and manufacturing is those big productivity boosts they're looking for. We're seeing the one, 2%, double-digit percent improvements in throughput, reduction, and scrap. In specific processes, that's multimillion dollars right there. These are often 20 or 30 years old processes already pretty optimized. The opportunities AI will give you are those stepwise improvements we're looking for in some of the existing equipment.

We've squeezed all the productivity out of some of these existing lines, so now we can leverage AI to give that big boost. The challenges are not really that different. I've spent my whole career doing these technology projects with manufacturers, so the challenges are similar. One is organizational change management, which you'll hear me discuss today. Ensure you've got buy-in from the top down to the operators on the line, ensuring those operators have a seat at the table right from the beginning with these projects.

The typical issues you have with companies who get excited about technology, and I love that I don't want to go into these meetings and just put a bucket of cold water on their dreams around technology. But I always try to bring the conversation back to use cases.

These types of projects are only successful if you start with use cases. So first, you need to consider the ROI and the benefits the technology can bring. Then we'll start applying the technology to it. But if you go into technology first and say, here's this Whizzbang technology, where can we fit it in, or where can we plug it in? That's typically different from where you'll find a lot of success.

Lisa Ryan: So, give us an example of how to get started. How would a manufacturer discover that first process or decide that they are bringing the process first and then looking for ways to use the technology versus like you said, the technology first?

Bryan DeBois: How do you get to those use cases? And it's a question I get a lot from customers. They'll even say, okay, we want to do digital transformation, or I'll get a call from a director of the transformation, digital transformation, or VP of Industrial 4.0. They're like, where do we even start? So typically, the place that we start with them is to sit down and do a workshop with them. Typically, it's about a half-day workshop. The workshop aims to generate use cases because we strongly believe these projects must first start with use cases.

So by the end of that, we've generated, and it's not us generating it. We try to get in that meeting across the section of operations and engineering and maintenance and quality and management. We work with them to generate these use cases. Part of that is identifying whether these are low-hanging fruit. Are they high-impact and low-cost? Those are typically the types of use cases you want to start with. And now, we can generate 30, 40, and 50 use cases from that workshop. And so, the question becomes, where do you want to start?

Then, we typically do an assessment that will put some meat on the bone. The use case at this point is a couple of sentences. So, we picked one or two of those. We're going to do an assessment that's going to put the meat on the bone. We're going to go in, and we're going to talk to all the people that the use case would impact.

We will examine their available data to make that use case a reality. Then we will give them an estimate of what a project like that would cost. And now you've got an actual executable project. You can put it in your budget for a given fiscal year.

And now we can start. But starting small, building ROI, let that snowball into the next project. In the next project, walk before you run. It's unique in our industry because what we hear instead is you have the big IT consulting companies who come in, and they say, give us 20 million, and we'll solve all your problems, or we'll give you digital transformation.

It doesn't work. We've gone into many of those customers where they've spent a couple of years on a couple of million dollars and only have a little to show. And this is a much more bottom-up approach. It is a much more practical approach to get to the digital transformation future they're trying to get to.

Lisa Ryan: I like the bottom-up approach because if you walk in, there are so many people working in manufacturing, your hourly employees that are terrified of this type of transformation because it's immediately going to lead to them losing their jobs, is what they're thinking, so, getting their ideas, their processes, and doing the slow, "okay. Let's show you how this works. Now we can move to the next one." But when you're starting to have those conversations with employees who will be impacted somehow, how do you build the trust? How do you get them involved in the discussion?

Bryan DeBois: Being the director of industrial ai, one of the most common questions I get is, are you putting people out of work?

And the reality of it is that I've been at this now for over 20 years. This is the first time I've seen one of our projects lead to a labor reduction. Lisa, these companies need more people to work. They're desperate for people. The last thing they will do is lay people off because of a technology project.

They're going to reallocate those people to higher-value tasks. And honestly, those folks don't want to be doing this. I talk to people all the time who spend four or eight hours a week building Excel reports, and that's different from their job. They don't want to be doing that. That's not fulfilling work by any stretch of the imagination.

So if we can automate that and I can pull that data from some of those systems, I can automatically generate some reports that free up four to eight hours a week for that person to do higher-value tasks. And that's how these always go. So you do have to build some trust.

But one of the things we're going to talk about is some of the more advanced AI that we're doing. Part of it is that you must extract some of the skills and strategies from your best 10- and 15-year operators and engineers. You must extract those strategies from them; they're excited to discuss this.

Let's say you've got Bill, and he's a 15-year operator and has spent his 15 years being able to run that line. He's as good as anyone, the better, the best in the business, and he can run that line as well as possible. So who's he going to talk to about that? He's, if he goes to the bar after a shift, his buddies don't want to hear anything about it. If he goes home, I can guarantee his wife doesn't want to hear anything more about how he will squeeze in a little more productivity. But I sit down with him with humility, right?

And I sit down with Bill and tell me everything about how you got so good at running this line, and he's excited to talk to me about that. He's thrilled. Getting people to buy into this is more accessible than people think. And to get excited. I've also had situations where folks were close to retirement.

This happened with a pipeline customer, and they were the best production scheduler that the pipeline had. And we were going to build an AI system to try to build production schedules, and this person. And they saw this as their legacy like this was a way for them to pass on their work.

This is what they've spent their whole life getting good at. And they saw this as an opportunity to pass that on before they retired. It is not as complicated as people think; it does not lead to massive layoffs or anything like that.

Lisa Ryan: I love that legacy language because people want to be a part of something bigger than they are. We realize that people, at some point, will age out of what they're doing. We lose so much of that company history of that expertise as our tenured employees walk out the door. What great language to use, and like you said, to be able to talk to employees who are excited about what they do and nobody else on the planet understands it.

So you've mentioned a couple of times industry 4.0, the fourth industrial revolution. What exactly is that? What does that mean?

Bryan DeBois: We can go through the whole history of it, but in terms of Industry 4.0, this is the marriage of digitalization to the existing technology, approaches, and equipment we've had. Part of the challenge with defining Industry four's point is that we still live it. It's hard to say where this is going to end up now. In my role, I've got a narrow focus on AI. When we're done with this industry 4.0 era, we'll find that the destination is artificial intelligence. The goal was all the work we're doing now regarding data, digital transformation, collecting that data, and correlating it. I believe in an AI future because that's the only thing. We can manage that data and get great insights and analytics. Still, until you can get those big wins and those big stepwise improvements that we've seen with the transition of every other industrial revolution, that's where AI will end up being the outcome of this one.

Lisa Ryan: In your experience, what are some of the best practices for integrating a new data pipeline into somebody with existing infrastructure so that they can avoid causing a lot of disruption in their operations?

Bryan DeBois: Yeah. And every customer's different, but we have a playbook we go by. One of the first things is that we implement historians. As I'm sure you know, historians and time series databases use it on the plant floor to collect data. And it uses lossy compression to store massive amounts of process data. So that's typically one of our first plays. So if a customer needs a historian, we're going in, and we're saying you have got to start there because if you're not collecting the data, you can't do anything else—all the ideas and visions you have in your head of analytics and AI and all that. You can't do any of that if you don't have the data, so we have got to collect the data first, and historians are the most efficient way to do that.

Suppose they have process historians, of which most of our customers at least have a couple. Now we're talking about an enterprise historian. We will pull all that data into one single time series database. Often nowadays, it's living in the cloud. That's fine. That's no problem.

And then, finally, we're going to marry that process data to the transactional and relational data coming up off the plant floor. That's the panacea. That's what customers want they want to see. So, if this transaction happened in Miami's system, what's the process data associated with it at that moment?

And now we can start to do some interesting. Everything I just described causes zero downtime. None of that would cause any disruption within a plant. We can hook up a historian without having to stop the line. Because it's a passive data read, it's not impacting anything happening on the line.

So we do those types of projects all the time. And again, it's one of those things: that data infrastructure. I often end up being the bearer of bad news in these meetings because the customer brings me in, and they want to talk AI, and I want to talk AI. So we start talking about all that, and then we get to the point, and I ask them what your data infrastructure is.

We've got a bunch of disconnected skids and a little data logger on each one, but nothing's networked. So I'm like, oh boy, Okay, guys. Uhhuh. Let's pivot now because we have to talk about a data infrastructure project to pull all this data together because the AI can't go out and know where any of that data is.

It doesn't have any idea. One of the challenges with getting over the hype of AI with folks is showing them that these AI algorithms are dumb. They need everything spoon-fed to them. They need hundreds of thousands to millions of rows of clean correlated data.

If it gets to the hundred thousandth row and there's a date missing, it just bails on the whole thing, and you have got to start all over again. So there's a lot of care and feeding that it takes to even get to that point. But yeah, again, those data infrastructure projects are zero disruption, and we can get those done relatively quickly.

Lisa Ryan: How long, how much history would you have to? Collect before starting on the process?

Bryan DeBois: So it's a good question, and it depends on the problem you're trying to solve. But I know that one of the challenges I face is going up against these vendors in this space. So we're a system integrator.

We don't have any products of our own. We integrate other vendors' products. We're independent, so we are not married to any one platform. But one of the challenges we face is we're up against these vendors who are going into these meetings with the customer and getting them excited. So there's a lot of hand-waving around that question. Oh no, you'll be fine. You'll be fine. And particularly for things like predictive maintenance. Unless you've had that specific failure in the historical record, it's almost impossible to teach this system to predict when that type of failure will happen.

You need a lot of data, and you need, and again, depending on the project and the problem you're trying to solve. If it's predictive maintenance, you need those failures to be in the historical record. And you can build a predictive maintenance system that can know, so let's say you've had three major failures, and different things caused them all. You can build a predictive maintenance system to predict those three things. But you've got no way to predict the fourth one you didn't anticipate and have never seen before. So there are definite limitations to that.

And it's actually why we don't lead as RoviSys. We don't lead with predictive maintenance as the primary driver for these projects. Instead, we like to focus on predictive quality and autonomous AI. And autonomous ai. That's the future. That's the real stuff.

Lisa Ryan: And what does that mean?

Bryan DeBois: So autonomous AI is it incorporates deep reinforcement learning. Deep reinforcement learning came onto the scene in 2016. It came out of Deep Mind, a Google spinoff, and they created a novel way of doing machine learning. And they created this algorithm that could learn by doing. They started by making Alpha Go, and then they went on to beat our best grand Master at Go. A guy named Lisa Doll beat him, Alpha. Beat him four to one. And then, they went on to create Alpha Zero, a chess program. It beat all our chess Grand Masters and our best chess software.

Then they made Alpha Star, which beat our best StarCraft players, so they were onto something here. And it learns instead of by data, Hundreds of thousands, millions of rows of data, like the traditional supervised learning approach. It learns by doing. And so we're building simulators of certain parts of their process for customers, and then we let this autonomous AI loose on it. It learns, it plays, and it learns to get better and better. And so it'll run for hours, but by the time it's done, it's like a 50 or a hundred-year operator. And it knows all the ins and outs, and then we feed it. We call it lessons. So we're providing scenarios like, okay, what happens if this equipment goes down?

What happens if this piece of equipment is not running at it? It's full throughput. What happens if you've got a hot batch from a customer, your biggest customer that comes in, and you have got to prioritize that over everything else? So it's learning all these crazy edge scenarios and things and becoming a human operator.

And so autonomous AI, we've got several customers implementing this now. And what we're seeing is this is the big step change improvement. This is where the whole industry is. 

Lisa Ryan: Obviously, there are so many advantages of ai from everything we've discussed. On the other hand, it can also be extraordinarily scary. So what are some of the things that manufacturers need to look out for and prepare for on the dark side of AI?

Bryan DeBois: Yeah, there are only two kinds of ai, at least the RoviSys gets involved in. It's supervised learning, predicting a value, predictive quality, and predictive maintenance. That's where you give it a bunch of data, and it will indicate a value. And all it can ever do is predict one value. Here's the number of days until that piece of equipment goes down, or here's what the final quality of that batch will likely be. It's going to predict one value.

What you do with that, Is up to you. Your action based on that prediction is up to you, the operator, the engineer, the supervisor, and whoever's seeing that prediction. There's nothing scary about that. You're getting some new insight. The autonomous ai gets a little spookier because Many of our customers see this as a long-term way to train.

It's, we call it, a brain. A neural network trained through autonomous AI is called a brain. Many of our customers see an opportunity to hook that brain directly into the control. That's something that we always have to hit the brakes on, and we say, okay, with these projects, the first thing we're going to do, we have to the brain and the decisions it's making. But the second thing we do then, once we validate it, is to put it into production in a decision...