Generating maintenance confidence using Gen AI and Predictive Maintenance - with Adam Cartwright
Okay. Hello Birmingham. Are we, uh, interested in a presentation here in generative AI and how that can be used in smart maintenance? Hopefully that means you're in the right place. Okay, that's good. Right. So what I'd like to ask, first of all is how many of you of you here have heard of generative ai? How many of you have used it for fun? Keep your hand up. Okay. And how many of you actually use
it in your day-to-day work? Okay, brilliant. That's really good. So what I want to show here is actually how generative AI is in the reach of every person who's in a maintenance team in a water company, and practically how they can use that. And a bit like Ruth earlier, I'm gonna be a bit brave and have a demo to see how that works. So fingers crossed, this will be fine.
So if we all crash and burn, then please I'd like some sympathy. Okay. So first of all, what is the challenge that we're trying to address here? Ampage we know has a huge capital program that's coming, but also within all those draft determinations, there's also quite stringent demand in targets around opex efficiency, particularly when it comes around to maintenance of the existing asset base.
So here we can see with the United Utilities for example, they're aiming for a 20% reduction in reactive maintenance costs by shifting towards proactive maintenance. Now in every single water company the number will be different, the phrasing will be slightly different, but the objective is more or less the same. How do I use digital technologies and advanced analytics to help me more effectively manage my assets to reduce risk and reduce operational costs? And so a question I'd ask you is to say, how confident would you be in your ability to reduce those maintenance costs by 20% in the current hub? Do you sit there feeling confident that that's something you can do? Or actually is that yet another very stretch and demand and target on top of all the other ones that you've got? If we look at this from a maintenance team perspective, what does that actually look like and feel like? There's three points that we draw here from our experience in other sectors and what we know about the water sector. The first thing is dashboards does not equal insight, particularly when we move into the realms of rotating machinery and away from single asset sensor points, you suddenly have lots of data that all comes together. What does that actually mean and what do I do with it? The second one is lost knowledge. Paul, in his earlier talk, discussed about this, what some call the the grace, the silver tsunami, the number of people that are retiring from the industry.
A lot of tacit knowledge is being lost that people know about how assets and how a particular asset works. Oh yeah, Jim, he's worked in that area for years. He knows always, always about that pump station. Let's give him a call. Unfortunately, if he retires, what do we do or if he's left the industry? And the third one is the challenge of sharing knowledge and collaboration.
How is it when you look across your whole company, you make sure that your best experience and your best people are sharing their knowledge with everybody to lift the level for everyone. So actually these are challenges that all your maintenance teams will face in every single water company. And our experience has been from speaking to most of them, none of them yet have a systematic, proactive, what you could describe as best practice maintenance operation. Definitely some hotspots of best practice and really good things going on. But in general it's never systematic anywhere. So the good news from that is it means you've got opportunity to get ahead and what we can in particular learn from this presentation today is how you can leapfrog.
What we'd like to reassure you is that actually other sectors have solved or at least made big progress in this area. And so water can learn from what they've done. So let's take us through that journey that other sectors and water needs to go on.
So the starting point is condition monitoring. You need to bring all your data together to know what data you've actually got. If you've got no data, don't expect to do too much with clever analytics, it becomes really hard. The next level of evolution from that is once you've actually got your data together, you can actually do some condition-based maintenance where you can say, let's look at what the asset is telling me now. I've just breach breached a threshold now so I need to go do something now. Okay? This allows you to be more quickly responsive to the issue, which means you're likely to have a lower impact or you might be able to get slightly ahead of a problem.
The next level after that is how do you look what you could call predictive maintenance, which is to say, in the future I'm about to have a problem. So you can definitely be ahead of the problem. You can definitely be there before it causes an issue and fix it in advance of time. Okay, this now means you are bringing in some quite complex analytics and artificial intelligence technologies, whereas condition-based maintenance, you don't necessarily have to have sophisticated ai. You can do quite a lot of stuff with some quite simple maths and thresholds. What we're now gonna talk about is how does generative AI come in? Which is how do you make that conversational predictive maintenance? Because the more complex the analytics are, the challenge is you need more skilled people to have that insight to interpret it.
This is the problem we're gonna take away by using the generative ai. It means there are no challenges around interpreting what comes out of the system. So here's some examples about how other sectors have used this in their daily work. So a big global automotive manufacturer, a return in investment of less than three months because they reduced their downtime by 50% on some of their key robots. In particular, this is now rolled out across uh, their global manufacturing operation. You can see over 10,000 assets are being monitored in the system.
And this is exactly the system I'm gonna show you now. So the key message there is it's not just a short ROI. Okay, it's automotive. So the cost of downtime is different. Key point there is scale over 10,000 assets in the system live being used every day. A paper packaging company here.
This one I think is an interesting example because again, it's a lower value product. This is more like the world that we are in in water and wastewater. And the key point here though it's lost in the screen a little bit, is actually this is about retrofitting sensors onto existing assets. This isn't the fancy world of robotics and automotive. This is the humble world of corrugated paper manufacture and how you can retrofit sensors and use that data to do something clever. Another one then is the global steel steel making producer ROI in less than six months.
And there the key lesson is they use the stranded data that already exists in their PLCs. The water industry is filled with data in controllers, in PLCs, in isolated systems that are not bringing that data together. So in this case, by adding some senses but mostly using the data they already had, they were able to then deploy the system.
Okay? So the important thing is water. We can learn from these sectors. Okay? So what I'm going to bravely attempt to show in a live demo at the moment, and I will need some interaction from you to actually ask questions of the generative AI to see if it actually works is the system we're going to show you. It's called Senseye Predictive Maintenance.
It's actually the result of a Siemens acquisition quite a few years ago from a company based in the uk. But since it's been acquired by Siemens, we now have a number of generative AI products. And we've used this as one example to bring that generative AI um, into this. So we've lost all the words on the bottom there, but the benefits that you're looking at are improved uptime, um, precise and correct maintenance. So you're channeling the right person to the right place with the right tool.
You're reducing risks and costs 'cause you're avoiding that massive cost of um, on kind of on-call reactive maintenance. Instead of planning ahead maintenance, you're capturing the knowledge and actually by driving around less, it's more sustainable. Okay? So these are the benefits that our customers realize and this is how it works.
The generative AI bit, which is gonna impress you all, I can almost guarantee you that only works because underneath that there's a really clever artificial intelligence engine that analyzes the data that comes on the assets. So in this case, what you do is you take a range of process electrical and physical data from these assets and the AI learns within at least a two day period what normal looks like for that asset. You then use that to be able to do anomaly detection comparing the normal which you've defined and the user can have a role in that saying if this is normal or not, against what's actually being observed.
So the visual way to think about this is I have a fingerprint for what normal looks like based on how all these different data points interact and I'm now looking for variations from that normal. What the tool does is it then aggregates together all those different types of anomalies you can have, whether that's trend violation, thresholds being different to what was forecast or degraded signals of a time. And it integrates that into a single attention index. And then it rag rates that quite simply. Okay?
And you'll see this in the app in a minute. So this is what we're going to look at. So the important thing here just to recap, this is about scale of systems. This is a cloud-based application or it can be on the edge if you want. You can see the attention index here, which is the rag weighted color of those little cases against the assets.
And then you can see the actual time series and how that's all flagged there on the right. Okay? So the headline kind of concept that we're talking about here is we're moving away from this quote here, which is from the CEO of of our business unit. In the past we had to speak to machines in their language with our copilots. We can now speak to machines in our language. And this is the generative AI piece, right? This is how we type things in and it's gonna give us an answer. Now let's all just pray to God that this actually works when I show you, okay? So what's the copilot going to do? It's gonna give you instant answers to your questions.
It's gonna retain all the knowledge that comes in from the cases that you submit and that you learn from. And also it means because it's all stored there in the system, it means that knowledge is available for everybody. Okay? Fingers crossed everybody. Let's hope the conference wifi holds up. Okay, so here we go. Here is the system.
I've got a few wifi bars. Oh hang on a sec, we don't have it on the screen. There we go. Now we do. Okay, there we go. Okay, so here we go. Here's the actual system itself. You can see here we're in our demo with drinking water treatment plant asset.
And what you can see here, like we talked about before, is a series of different cases on the left rag rated according that attention index highlight in the different type of issue that's in place, you can see the actual asset that's been flagged. And here we go. We can go to the case. So if we look at this for example, um, a micro, the micro sand mixer here at the top, then actually you can see from the very start you don't just have a button that says something. The case is given to you in natural language where the copilot has explained what's happening. So here we go. Look the copilot, which is that little symbol when it raised the cases, it says actually you've got a threshold violation that started here seven days ago.
A rise in trend under a fall trend. And if you go through, you can go back and you can have a look at all of these. So you can see here in this case for example, look there was anomaly detected there. Another there in this case it said look, the trend has also been flagged and this is why this has happened here. So that happened eight days ago. You're looking at the drive end vibration motor current.
So all of these different aspects are being brought together and this is what has caused the notification to be raised. Now the challenge then comes, remember we talked at the beginning, oh I've got all these time series. You might be there also thinking what the hell do I do with all these time series? Yeah, what does that actually mean? What do I do? So I'm gonna need a brave person here If I put in co-pilot, what question should we ask the generative ai now that this notification's been raised, okay, we unsaid what is the likely, likely root cause of this problem? Okay, you'll also see my bad typing.
Okay, here we go. Okay, there we go. So the likely root cause of this problem appears to be barren degradation. This is suggested by the threshold violation and rise in trend in motor non-driving vibration.
The degradation could be due to insufficient lubrication, contamination or normal wear and tear. It's important to check the greasing schedule and lubricate the barren if necessary. If the behavior does not return to nor barren does not return to normal, consider a placing the barren. Okay, what would you want to ask next? Okay, if you just heard that follow up question. Okay, What should I do next? We can also see if it picks up my bad typing. Okay, here we go. Here are the recommended next steps.
Check the greasing schedule, verifying if the bearing's been lubricated according to the maintenance schedule. Lubricate the bearing monitor the vibration level. So you might want to raise a job. Consider the bearing in replacement. So taking these steps should help address the issue and prevent further problems with a micro sand mixer. Okay, what any other questions? Now anything else that we might want to know about this? How long can I sweat the asset in this current condition? Okay, here we go. It is difficult
to provide an exact timeframe for how long an asset can continue operating its condition without further inspection and monitoring. However, operating a degrading bearing can lead to increased wear and potential failure, which could cause unplanned downtime in more significant damage. To minimize risk, it's crucial to monitor vibration form immediate maintenance and plan for bearing replacement.
So this is where we come to the limits of actually the generative AI actually being able to say, when is an asset really going to fail? And this is where you need the input of the experts. We might want to bring somebody in to say, who can help me understand this? So you could ask this question, who could help me understand how long the asset will last? Okay, so these questions haven't been planted by the way. So if you're there thinking, oh these just make one up and ask me and let's see what comes out. Okay, here we go. To get a more precise understanding of how long an asset can last in its current condition, you should consult with maintenance engineers, reliability engineers, vibration analysts, or the original equipment manufacturer.
Okay, some more questions. Do you wanna, do you wanna test it or is there anything else that you think might be useful or we could look at another asset? Ruth, Did you ask? Oh, okay. What is the time from alert to failure for assets of a similar type? Now the challenge here will be some will be 10 depend on what other assets of this type are in the demo account and I genuinely don't know.
So let's see what it says. Okay, so because it hasn't got enough specifics in here, it's given you a general one 'cause it doesn't know the answer. Whereas if it had enough in there, then it would give something. So it'd say actually it depends.
Yeah, you've got kind of a bit of a, a fluffy answer there. Okay, some other questions. Should we look at a different asset? Oh okay, here we go. This is the man who likes the difficult questions. So the question there was understanding the impact of that asset failure on the whole process, if that's this asset fails.
So at the moment the copilot only looks at individual assets in beta. We have ready to be released later this year, the system level one, but we don't have that at the moment. So in terms of looking at the impacts of failure is asset specific, but that will come. Okay, should we look at another asset or do we have any more types of questions we could ask? Welcome to those at the back for those who missed some of the presentation.
What this is is this is a generative AI tool that's linked into an AI tool for predictive maintenance. And the purpose of the generative AI is to help guide a user through uh, the process for how they respond to a notification that's being raised. So the idea is that with a decline in skill level or in some cases new people coming into the job, how do they know in an easy way what to do next? So we've just selected another asset here with an ozone gas pump.
Maybe one thing we could ask it is what content should I put into a work order for this asset? Should I put into A work Order for This? Okay, so here we go. Now it's telling you these are the things that it thinks someone should actually do. If you're gonna raise a work order for this, what you could then do is you could ask how urgent is this? If you can spell it right, here we go.
And then you can say how urgent it is depends on these other patterns. So it's telling you you need to go and consider these other things. Okay, so the idea here is to give you a feel for that. Maybe just that a hands up thing. Thumbs up.
Were we impressed by that? Yeah. Okay. It's pretty cool, right? Um, and the, I am standing in for the person that was supposed to present this, who actually can't be here and the words they said to me was, Adam, it's designed for any idiot to be able to present this, so you should be fine. Um, so this is exactly the point that we're going for here. So the question then is, okay, we've got this opportunity in water and wastewater to benefit in the investment that's been made in a tool designed for sectors that can afford to pave it, right? Water cannot afford to build this type of system by itself and make it scalable.
But automotive can Oil and gas can, pharmaceuticals can. So the opportunity for water and wastewater now is to say, okay, how do we learn from that and apply it into our systems? So as an example for this, the next step would be to say, how do I take enough assets where you've got, um, likelihood of a failure or issues of some kind, say a treatment works water or waste water. And then you evaluate that for three months and you see what it can find and you see how the actual maintenance staff on that site manage and engage with the tool.
'cause part of our experience is one of the main reasons condition monitoring or predictive maintenance programs fail is the lack of buy-in from the staff that use it. The amount of sites that we go to. And there's a bit of old Siemens kit, bit of old IFM kit, a bit of old xylem kit or anybody's kit there where someone once upon a time got really excited about condition maintenance, bought some kit, stuck it on an asset, didn't know how it worked, or they moved on to a different role because they were the innovative one and the stuff was never used and never picked up. This means that everybody can use it and importantly it's not replacing the person, it's enabling them to act more effectively.
So I'll stop there. Um, if anybody would like to learn more, please let me know afterwards. But otherwise, any questions? Okay. Um, so the data sources a question there is what data sources have you used to train the model? So this is a asset agnostic, um, ai it's not about recognizing failure modes, it's not a model based system in that sense.
So the more data you have, the better it will be. What you'll find is if you don't have enough data, then actually you won't have a good enough, um, fingerprint for what is normal. So in this case, for example, if I switch back to, yeah. Can we see this as an example of the type of things that you'd want to have pressure, current torque vibration.
Classically, in a condition monitoring world, you've got three categories of data, physical process, and electrical. And the more of those you can get and bring in, the better your model will be. So the question was could you build a model just for that asset? So the way this works is instead of having, so say that automotive company that has 10,000 assets connected, you don't have 10,000 AI models. Yeah. As soon as you start to do that, you are in a very, very difficult world of AI model management.
So instead what you have is a generalized model that allows you to monitor any asset in effect. So using the same AI model, you're just having a separate deployment of that for each asset. And then that allows you to think about how you classify an asset. So in a pump for example, you could have a motor asset, but you might also have a drive asset. Yeah. And you can do that in that way.
The important thing is here is you can learn for each asset with only up to two days worth of data and then see where you go from there. Okay, my time's almost up. One more question. Oh, go on Bruce. Oh, there's a microphone. I'm next to the speaker. Um, how often can it be? Is it, uh, is it reinforced learning like every night or is it Yes. Yes. Okay. So the AI's continually learning and defining what normal is for that asset.
You've paid for your difficult questions with your help. Thanks for that. Uh, how do you use, um, domain knowledge? How did you ingested, uh, domain knowledge into this is a Train, uh, okay, domain knowledge. So in this case it doesn't need domain knowledge, but what you can do is in here there's a knowledge fab, and actually I should have mentioned this, this allows you to then ask questions where it interrogates data sheets or schematics or anything for that where you can then ask it how that works.
So I actually, um, this again is another feature that's being released next month. I haven't downloaded the video, but what this allows you to do then is if you've got your own procedures, data sheets or anything in there, you can then ask copilot and say, okay, what should, uh, what drive processes say about this or what was in the actual design specs about how this should be performing? And the generative AI will then read the PDFs or whatever that you put in there and then give you an answer based on the documentation for that specific asset. Really good question 'cause I forgot to mention that.
Thank you very much.
2025-02-25 21:12
