Future industrial professionals will look back on the year 2023 as the year that artificial intelligence truly started to scale and reshape plant operations. The increasing integration of software into manufacturing processes plus massive cloud compute power has laid the foundation for plant teams to apply AI to drive better business decisions, from supply chain and resource planning and scheduling to improved physical asset management.
Michael DeMaria is a product manager for Azima DLI, which is part of Fluke Reliability, where he manages the hardware platforms and integrations, diagnostic software and AI tools, and user portal deliverables and business metrics. Michael’s background is in Navy nuclear engineering, but he has been working in the vibration-analysis arena for more than 30 years.
In the following interview, DeMaria explains why starting with a pre-trained AI is critical to successfully using AI for machine condition monitoring.
Below is an excerpt from the podcast:
PS: You know, I remember working and talking with people like you and Burt Hurlock, they were some of my first conversations nine years ago for Plant Services. I think the industry is finally maybe starting to catch up with where you guys were heading in those early days.
MD: Nine years seems like yesterday, Tom. But you're right, technology has been changing so quickly, hasn't it? Where we were when we started, though, you know, a lot of the principles that we've done, from the old DLI days since, like 1966, the ideas of how we approach the market still hold true. But I think what's really changing is how fast things are growing, how much we can scale, that's really novel nowadays.
PS: I was struck this year at the SMRP Annual Conference at how much Artificial intelligence has penetrated the industry conversation. It’s been here for a while, we've had these tools available, but maybe not in their currently mature state, certainly not in the state where you get a lot of presenters at an annual conference, showcasing what they're doing and how they're doing it and the results they are achieving. Are you seeing that too?
MD: Oh, my god, yeah, we're in a world of so much data, and humans can't handle that kind of data. If you haven't been focusing on where AI is today, I think you're behind the curve just a little bit. We had conversations about where is say ChatGPT and the like – and if you haven't used it, you'll have at some point this mind blowing experience with something like ChatGPT, where you're thinking how this is going to really kind of revolutionize the skill sets, the gaps, the efficiency of being able to get through data. I think you're right, like at SMRP, plants have so many more sensors that are tracking everything within. It's not just a “here I have a maintenance team that's just trying to fix problems and fight fires.” It's, “I have sensors on everything about my process.”
And it's weird, I think that a lot of people don't know what to do with all that data. There was a study out not long ago, I remember seeing, that a very small percentage of that data is actually really being utilized. That's where AI is really going to come in: how do I get through that?
PS: Maybe we could take a step back and ask, where do you see artificial intelligence as having the greatest impact in general for businesses over the over the next few years? This could include condition monitoring, but it could include other stuff as well.
MD: I think the biggest thing is going to be that the skills gap the plants have, the resource gap, I think those are two different things, right? I don't have enough people, or I don't have enough skilled people. When we've been talking with customers, that seems to be the biggest thing that we hear back is, I have those constraints, I don't have the expertise that knows what to do with all of this data. I keep hearing “something happens and then we go back and we look at the data to see what might be telling us or give us an indicator.”
The other thing is you're waiting then on something to occur. So I think getting a trained AI system is going to be the key to it all. I can have all this data that comes in and then be able to then utilize the limited staff that I have more effectively. Get more out of less, you know? You actually have so much more data. But I have less resources. So how do I bridge those two things? I think that's where AI is going to take us.
PS: Interesting, that echoes what I observed at the FabTech industry show this year, where it was a combination of robotic automation, but also artificial intelligence exhibitors. This is critical right now to covering the skills gap, and there just don't seem to be enough people with necessary skills on hand as either new plants ramp up or old plants evolve.
MD: Yeah, production is so much more than it was before. Plants are operating around the clock, and this is a skill that is hard to come by. New engineers are not wanting to get into this field. It's not really the trade that people go after coming out of schooling. And then if you do you have this as a skill set in a plant, it’s hard to hard to keep them, right, because they're in such high demand. That's where I hear the most from our customers today, is that I just can't find them, I can't keep them. And we are now overwhelmed with a production understanding of the data that's coming in. You know, it's costly.
PS: You alluded to that earlier, too. Let's focus on that data issue. There's so many sensors available, so much data being collected by plants, not just condition monitoring or asset health, but also business data coming in. What are some of the challenges that you're hearing from people in the field as they try and throw their arms around all the data being collected?
MD: We do have a lot of experience in this one, Azima DLI. In addition to having software that handles kind of volume of data, we have a team of service people that actually utilizes that to get through it.
Running all of our numbers for 2023 of what is that volume, I think the number came in at about 450,000 machine tests that we're analyzing this year so far. We have a small team of analysts, but our utilization rate is astronomical. I think it came out to somewhere between 1,500 to 2,500 machine tests per month per analyst. It's about five times the industry standard, and it's that ability to do that is having kind of tooling, these types of AI enabled diagnostic tools. We certainly have some experience in how we handle this, but I think it speaks to the challenges that that people have in this.
The reason why we can is because that AI-enabled diagnostic software kit is already pre-trained. When people think about AI systems, there's a kind of a misconception about what it is. You hear about, say, ChatGPT, and I can put in a question and it spits out all kinds of answers for me. But that comes from a tremendous amount of training, to have it know what it's supposed to talk about. If you look at what Google or Microsoft or the like have done, OpenAI have done in getting all of this data sourced into the AI tooling in order to do what it does, that doesn't necessarily equate to anything across any industry.
You still have to have a trained data set, and with vibration and predictive maintenance, generally speaking, there's not a lot of data that's out there, trained data that's out there, tagged to different machine types and faults that have been identified to understand how those patterns work. We've had this luxury, and how we can get through that 450,000 machine tests is the fact that we have been storing data since the early 1990s, when our first software came out in 1990. We had been capturing that type of that tagging, and we've been training the pattern recognition to understand how these are the different nuances of how data comes in and how you can get a result out of it.
So I think the challenge that people have is there's a misunderstanding about what's necessary to get an AI to work, and that is a trained system. You can't wait for a fault to occur, and then say, “hey, hope we don't have this one again, because that was very costly.” You have to already have something that can give you those faults beforehand.
Another thing I think people see as a challenge of this world, especially when you start talking about IoT sensors, is I see the industry, the maintenance group kind of downgrading the type of expectations out of a program. If you think about what happened in the early 1990s and 2000s when vibration analyzers really kind of came out, you had this uptick in a skill set. You had the mostly Level III (analysts), some Level IV, definitely Level II analysts that would be out in the plant capturing data, troubleshooting, understanding the science behind vibration, and the like. But then we got these wireless sensors out there, and people jumped on having sensors across the plant. But that technology of the wireless sensors isn't the same as that high-end instrument that people would use to go around and do route-based (maintenance). So now all of a sudden, what we've come to is less quality of the data that you're capturing about your machines; certainly there's capabilities of it, but it's not on par.
Plants have become a little bit more accepting of more of what we used to have in the late 1980s, kind of that go/no-go, kind of a hand-raiser, I have the checkbox, I'm doing a vibration program. It's not that technical as it was 10 years ago, it's more generally speaking a solution that we had, say, 30 years ago. I think the technology will change – sensors are constantly evolving, so you'll get back up to it. But right now there's this challenge of understanding what is it that the systems are capable of doing, and if that matches what I really need to get out of my program.
PS: Okay, are we looking at things like existing proactive maintenance strategies, this route-based predictive maintenance changing as a result as AI gets integrated? Or are we looking at AI being more part of the master program, not replacing old modes of doing things but more as an additional resource to us?
MD: It kind of speaks to that previous point where, what are those strategies, right? So when you're looking at your maintenance program, what is it that you really want out of it? And there's not one answer that fits all, there's certainly a lot of variation depending upon how you operate the types of machines that you have, how integrated you are. For example, a service provider, like an OEM has a different approach to how they do predictive maintenance; they're more monitoring, so having some sort of sensor or devices that are installed on the machines to be able to say, “hey, something's going on” and then you can go in and you can troubleshoot. That's certainly a strategy that works.
But a plant who's operating wouldn't necessarily want that. We hear a lot of this type of approach where now all of a sudden, you got a lot of nuisance alarms, a lot of the boy crying wolf, there's just so many things that are happening that I can't filter through all those. So a plant’s maintenance strategy might be, “look, only get my attention when it’s absolutely critical that something's going on.” I might have a different system and approach, and this is again where software can really kind of come in, scrub through that pattern of data and say, “hey, look, I have a problem, this is the problem, this is what needs to be done, this is urgent, I need to make sure that I can mitigate those issues right away,” mostly a maintenance department thing.
And then there's other plants that really want to understand what's emerging, the time that it takes me to get my process in line, the supplies into a queue, and how it manages through my operation is costly and time consuming. I need to understand emerging faults, so I can plan for downtimes and repairs and the like.
