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.
Listen to Michael DeMaria on The Tool Belt Podcast
PS: Your example of the overabundance of alarms shooting off reminded me of a presentation that was delivered at the SMRP conference, and I believe it was by a Saudi Aramco team, where they had worked with some machine learning AI type software to cut through all the alarms that were that were being sent by a program where the baseline, I don't know if it was standards-based set baselines, or if it was simply the corporate baseline. But in any case, the baselines were set so generously there was simply an overabundance of alarm showing up. And they had to have a tactic to understand as you said, what were the most relevant alarms? What was the likelihood of the costliest failure? What could they learn from these alarms which wouldn't make them wait for the catastrophe and shorten that time to action?
MD: When you set your thresholds, and you do this for say the overall data, as you mentioned, or just kind of process point thresholds. But also it applies to narrowband analysis, which we're doing, which has a lot of data. You would have to pick which side you want to would fail on. You can't be 100% accurate 100% of the time. You have to plan for, if you have inaccuracies, which side of the line are you going to be on, a false positive or false negative? And you most certainly don't want to be false negative, you don't want to have a machine failure without having any kind of warning about that.
We do a statistical average baseline representation of what's normal and healthy for the narrowband analysis, and then we have this pattern recognition to find every deviation from that, and then it weighs it. How does that data look from the low range, the high frequency range within a location on the machine, or all the different orthogonal axes on that location, to across the entire machine train? How does all that data come together and generate a pattern?
Our tooling then deciphers that, try to come up with, what is the actionable item? And when we set those thresholds, we might overstate those to say, “hey, there's a problem, you should certainly be aware of this.” But it certainly could be overstated. We’d rather overstate the issue than to completely miss the issue or understate it. So yeah, we do find, I think our numbers from that data set that I’m talking about, the more than 450,000 tests, we track all of those false positives, all those false negatives very closely. How can we improve upon the accuracy of the system? I think we have about a 3 or 4 percent false positive rate, meaning that we've downgraded a fault down a level. So we've overstated it, but we needed to downgrade it or steer the course just a little bit.
From a nuisance perspective, though, I think this is where you would also have to consider where people think of AI as 100% autonomous, or it's 100% accurate all the time, and nothing that has to be manipulated in order to make it work. That's not really the case.
PS: That's interesting, because it's the kind of concern that would arise with someone trying to implement it for the first time. They might not understand that this is not just a turnkey autonomous thing, like you're saying.
MD: I guess the way I always talk about it is, you're not going to do a multi-million-dollar repair on a machine without having some human eyes that looked at that data at least once. And it's the same case here, that false positive rate that I mentioned, that's from our perspective, from our services. The customer doesn't ever really see that. We've intercepted that, we track it heavily because we want to improve upon the system so we can retrain it, we can add new datasets into it, we can help define the patterns. I think that’s another really good cool thing of AI: how do you fill the gaps in order to understand what peaks might be coming from.
But we've steered the course, we've looked at the data, we've looked at other factors that come in, like other process parameters, or things that somebody sees on a machine, “hey, it's also hot,” which helps steer the course of what you're going to do. And then we adapt it so we can fine tune the end result, so a customer just sees what's really true about what's wrong with the machine.
I think the point is if you are putting a thing in place, you shouldn't ever think about vibration being a completely autonomous tool as of yet. It'll certainly get you started, it certainly creates warnings and alerts. Ours for example will give you a first pass of analysis. But you would generally not leave it just at that. You would want some human element as part of your workflow. And that's back to our at number one challenge that we talked about, right, plants that just don't have those resources.
PS: So you mentioned vibration sensors before Michael, when you were talking about the role of AI to process data. In general, for those who might be thinking about moving into more proactive maintenance modes, how have you seen wireless vibration sensors applied specifically for condition monitoring? And how do they work with AI analysis software?
MD: I think wireless is most certainly the future, there's no doubt that that's where we're headed. It’s at least 10 years now that we've had wireless sensors on the market. We haven't though, we started our wireless vibration sensor journey only about maybe three years ago, and there's definitely a reason for that. We were waiting for the technology to be able to use our diagnostic engine, and plants are operating more around the clock, so I also want sensors on my machines, giving me more frequent diagnostics in what we did with monthly routes or quarterly routes.
So most definitely, it is the future and the technology is still rapidly changing. There's a couple of things here, when you're heading down that route of, what is the role? I think the role is keeping plants aware of where's the risks? How do I steer the course and the like. But I don't think you should think of it as, hey, this is the technology today, and this is what I'm stuck with. I think every year we're going to see improvements across that that technology. Where we're at today will be probably different next year, drastically different three years from now.
I think companies have to really kind of learn how to adapt to that, how do I manage it? It certainly creates new challenges. You know, I'm putting sensors on machines, how do I manage those. But then back to that volume of data, right? So now if I think about I had a route-based collector that was capturing data once a month, and now I have a sensor on my machine that's capturing data every day, from a human's perspective that’s 30x more data that you're trying to get through. It's not sustainable, and that's where AI is really going to come in: how do I meet those plant’s expectations for planning for repairs, and knowing what the risk is? How to mitigate that risk? So yeah, wireless is certainly the future there.
PS: You're touching on a topic which I was going to ask you about too, which is about scheduling repairs, that planning and scheduling function, but also spare parts inventory. I've noticed this past year especially, a lot of CMMS vendors are talking more and more about modules they're adding on or features or adding into their systems to help automate that process, especially dynamic scheduling in the field. Maybe not necessarily within the four plant walls, but to help manage field personnel for sure. How are you seeing AI redefine what plants can do when it comes to improving their scheduling options and managing their spare parts?
MD: Yeah, spare parts is interesting for me in that it's coming to be more of a talking point now, because we've done this for decades. If you think about the big contracts that we've managed with, like the U.S. Navy and other maritime industries, where you don't have the luxury of having a huge room full of spare parts, and you need to keep your ship afloat.
PS; You mean they can’t just go to the big spare parts warehouse at the bottom of the submarine?
MD: Exactly right. The ship depot in the middle of the ocean.
PS: Right, all those floating depots in the middle of the Atlantic and Pacific, sure.
MD: So yeah, I think we had an approach to this that said, we need to find emerging faults. It's something that's years in the making, and our system has about five levels of severity. The least is nothing, no problem, this machine is healthy, which is still super valuable for a plant knowing where my risks are or my readiness. Then a slight issue is a marker for an analyst, it's my beginning stages, I can start trending it, start getting a sense of how rapidly something is progressing, what the root cause might be.
Then we have an emerging fault. It's something that is maybe many, many months or a year or so in the making. And that's always been a paramount feature of how we approached vibration analysis, that that fault feature extraction from known, normal, unhealthy machine, run through that pattern recognition to say, “hey, this is an emerging issue.” And it speaks to my ability to be able to schedule the repair plan for what parts are needed, make sure I get those on order.
I was in the Navy for eight and a half years, we did have a huge spare parts inventory, but you have bearings that are sitting on a ship that has a lot of vibration out at sea, so your probability of having a bearing that has a preexisting fault is kind of high. The more you transport it or if you don't have really solid storage practices, then you still run a risk. So waiting until the machine tells me that I need to have a part on order, and then having time to get that part on order, I think will evolve considerably significantly with now more AI tooling and the like.
From a vibration perspective, because analysts generally need to get through manually a lot of data, you generally look for the bigger problems, bigger takeaways from the data. And you don't necessarily manipulate the data in your view to try to find those very early, early faults. This is where AI comes in. Our diagnostic system sees those little subtle changes, and then you can track it. You trend those, you get a progression of time, and then you can start making decisions about hey, this is where I need to steer the course and get parts ordered.
PS: Let me ask a wrap up question for you in general. At the SMRP convention in October of this year, a lot of first timers of the convention would have heard a lot of presentations on AI. And let's say they were in reactive mode, or they're in route-based mode, they're seeking to evolve their maintenance programs to something more proactive. What are your thoughts right now on what to tell people as they're moving into more proactive modes, especially when it comes to AI? I think there's going to be a fear of missing out if they don't get it right the first time. When you come across most customers, what's the best advice for them?
MD: You said something that has been has me thinking, and that's that the fear of missing out. That's interesting, I hadn't thought about that before, but that is certainly true. AI is certainly advancing that technology, and as we were talking about the beginning of this podcast, there's a lot of AI tooling out there. And if you haven't played around with it, you will at some point, it'll be there.
It's something that we certainly have to realize that it's not going away, and it's something that I think we need in our industry to handle skills shortages. One of the one of the cool things about AI that I felt was fascinating is how easily it can adapt to a skill set. If you ask ChatGPT a question, and it gives you a very technical answer, you can task it to tell it to you like you don't know anything about anything and it'll adapt. It really has a sense of how to steer its answers to cater to what you're really looking for.
The other takeaway that I have of it was that you're not going to get that answer the first time around. You can't think of it like it's a Google search and it's going to just spit me back a couple of things to reference. Its approach to the world is really, how do I enable you to your skill set, to what your expectations are? If you're coming into this world, I think my first advice would be to really understand what is your objective? What is it that you want out of your program?
We have that conversation just with wireless sensors alone. What is it that you want from your wireless sensor? Do you want it to just tell you an alarm? Do you need it to tell you what's an actual result that you need to do on your machine? And it's not one solution for it; I might have a machine that's certainly important for my production, but I'm not going to repair it. I'm just going to throw it away. It's a small throwaway machine. But I would certainly like to monitor that machine to know what's going on. Is it presenting a risk? So my objective on that machine and how I approach it with IIoT sensors and AI might be different than my big compressor that's super critical to our operation or a huge turbine super critical to our operation.
PS: Or the chiller which can take out an entire batch if the chiller goes bad.
MD: Absolutely, yeah, and does the technology that we have available meet those objectives? Again, the same wireless sensor that you would put on that throwaway little oil pump down here for your gearbox isn't necessarily going to work well on that big chiller, or that big turbine or compressor. So what is it that you really need out of it? What do you have available to you to accomplish that? And then that goes back to do I have the staff that I need? Do I have tooling? Do I understand how to use that tooling?
The other one was how frequently you need the data. This comes up quite a bit. If you think about a mechanical fault, anybody who's gone through any of the vibration certification programs, you think about mean time between failure and the like as a mechanical fault. So usually route-based once a month is totally adequate in that regard. But then all of a sudden, we've shifted to plants wanting that level of diagnostics run on the machines every day, and so there's a mental shift between what is it that you're trying to do? You'll have vibration experts wondering why you want so much data, but then you have management saying, no, I really need to understand the risk. How do you bridge those two things? And that's where technology is really going to come in.
And if it's a process change, most certainly that can create risk very rapidly. So are there other things that you want to incorporate within your predictive maintenance regimen, beyond just vibration analysis? Are there other things that help you steer that course? And then again, that's just more data that comes in that you need tooling to be able to do something with.
PS: Your answer also made me think of the relationship between the average plant and say a condition monitoring technology like vibration, where people may look at AI right now, since it's a little bit sensational, and have a fear of missing out. People don't normally think of vibration analysis as fear of missing out. It's entrenched as an approach and a technology which will help you when you're ready for it. So to your point, what are you trying to do with your monitoring program? What are your goals? What are your critical machines? Those questions, as you said, they're not really going to change once AI is in the picture. It's just, what is the role of AI in this toolkit you're going to use?
MD: That's right. It becomes another tool that you have. It's a tool that's there, learning about it, what it's capable of, what it’s limitations are. I think for all of our maintenance professionals, what we're going to have to go through, is just like when we went from a broadband meter to a narrowband meter back in back in the 1980s. I remember those days well of walking around a machine, all the machines in the plant, using a broadband meter and having a single number that was your deciding factor of whether or not you need to take down the machine and figure out what's wrong with it. We've gone a long way since then, and then I think the new AI world just is going to bring us into another level of that.
