Industrial AI Beyond the Hype: Context, Reasoning and Autonomous Action - with James Loach

Trend Detection Podcast ·

Show notes

Welcome to the Trend Detection podcast, brought to you by Senseye Predictive Maintenance – which gives you visibility and insights into all your assets, from single machines to full plants to help you reduce downtime, increase knowledge sharing and accelerate digital transformation across your organization.

In this episode of the Trend Detection Podcast, we're joined by James Loach, Head of Research for Senseye Predictive Maintenance at Siemens, to explore what industrial AI means in practice and how it is changing the future of maintenance.

James explains how statistical systems, machine learning and generative AI can work together to monitor assets at scale, investigate potential problems and provide more prescriptive guidance to maintenance teams.

He also discusses why machine context, maintenance history and human expertise will become increasingly important as AI models grow more capable.

In this episode, you’ll learn:

What industrial AI means beyond the marketing terminology

How machine learning and generative AI work together at scale

Why context is critical for improving AI-generated insights

How agentic AI could support more prescriptive maintenance decisions

Why the future of maintenance could become increasingly autonomous

You can find out more about how Senseye Predictive Maintenance can reduce unplanned downtime and contribute towards improved sustainability within your manufacturing plants, by visiting: www.siemens.com/senseye-predictive-maintenance

Transcript

So the job of this first part of the system is to find things of potential interest in a fast, efficient, clean way, and then basically to push high quality triggers forward to the upper brain. Hello and welcome to the latest episode of the podcast powered by podcast. I'm really pleased to be joined by my colleague James Loach, who is a friend of the podcast. He's been on many times, but I think it's been a little while since you've been on, James, so welcome back. Yeah, I know. Yeah, very happy to be back again. Fantastic. And today, we're gonna talk about industrial AI, which is the term, if you listen to any or see any materials from Siemens, you know, industrial AI is very important to what we talk about. We're gonna go beyond the sort of marketing speak, Speaking as a marketer as well, we do talk a lot about industrial AI, but James is actually doing the real stuff and the real implementations of it with our, our customers and, and behind the scenes. So, yeah, it's gonna be a great conversation, but I guess before we, we start, James, it'd be good to, for you to introduce yourself to the audience and remind us what your role is in Siemens. Yeah, no, no problem. OK, so I'm head of research at Siemens Sensai Predictive Maintenance. That's a role I've had for many years, 8 or 9 years now, since this was a very small startup. My role, yeah, head of the research group and part of the product leadership team, so responsible for developing all of the algorithms, the analytics, and the conceptual structures that those things sit within. Fantastic, fantastic. Thank you for that. First of all, let's start with that term, industrial AI, uh, and what it actually means in reality. So, not that there is a marketing reality, but there's also what's happening on the ground. So, what's your sort of perspective on that, James? Yeah, I guess maybe I'll put across a very simple view of this. AI is obviously this sort of suite of technologies that we all know about. Industrial AI, yeah, it's just applying all of those technologies within an industrial setting. This makes it automatically not one thing, right? There's all kinds of different things that AI. It's very suitable for everything from supporting coding all the way through to agentic decision making or generative tasks or simpler things like summarization or how you mind things in disordered data, all of these different kinds of things, but those applications of the technology start in the industrial world all over the place. OK. And it's a huge area of opportunities and what everybody is working out at the moment really is how best to apply these technologies. What are the situations where it works really well? What are the situations where we have to wait. Yeah. Uh, I guess just to cover first of all, because we all know, and you'll know obviously better than most people, the explosion in terms of generative AI, but of course, Senseii, we talk about Senseii specifically. AI has been the backbone of Senseii. From the beginning, right? So do you think it's almost been forgotten about these more traditional, I'm gonna say traditional AI technologies. Has it been forgotten about in this new age that we're in at the moment? Yeah, I don't know, it's interesting. There's a long history of using machine learning, which is what you're talking about there in the pre-generative AI period to do lots of interesting tasks and S I, for many years before generative AI was doing things with. This kind of technology, but there have always been issues with scaling, basically, right? Being able to use this technology for particular applications but not being able to push it out, yeah, at scale, right, across applications or different machine types or all of that sort of thing. And one special thing that AI is letting people do and definitely in the case of Sensai, is to take what now seem like sort of traditional technology as a foundation. And then, yeah, using AI to make the most of the things that you were doing before and in particular for addressing that sort of generalization problem or other aspects as well, communication of results and things like that. Yeah, but many systems do use a combination of these things and S I also uses a combination of even older stuff, right? We have a set of statistical systems, we have a set of machine learning systems. We now have a set of AI systems. And one thing which is nice about Sensai is that these things do are quite complementary and work nicely together. That's really well explained, actually, the sort of foundations of Sensai with AI. And what I'm wondering is, obviously, generative AI has opened a world of opportunity. So, what was simply impossible all those years ago when you started at Sensai, which is now possible today? Yeah, So Sensii has always been about trying to look at sensor data from all different kinds of sensors attached to all different kinds of machines and do that at scale over hundreds or thousands of assets and try to find things that the user would be interested in that essentially help them with their maintenance, right, so typically developing problems and all of that kind of thing. So traditional techniques, if you apply them well, do allow you to do the first part of that pretty well, right, which is to be looking over all of this data, identifying things, essentially really unusual things or things that seem to have some implication, triggering a system and getting those things to the user. But then it becomes very difficult. So we can, in the past were able to identify that a particular machine seemed to be having an issue, and we can describe what's going on, but It was very difficult for us to really say what that particular issue was, and we had to make best use of the techniques available to us at the time, which were things like mapping the data that we were seeing to matching the data before previous failures in the past identified from maintenance management systems, these kinds of techniques. But what generative AI really lets you do in Sensei is to take those triggers, so the understanding that there's something up, pass that to a part of the system. Which in our case, we may touch on a bit, is highly agentic and very smart, that can take that trigger as an indication of what it is. There seems to be something going on, and then to call on all kinds of supplementary information about what it might be. Everything from the documents to looking at images of the time series to actually getting stuck in and doing some data analysis and changing the view of the data, whatever, all of that kind of information, bring it together. And confirm that there is an issue. Explain to the user what it is with some kind of confidence and go through and say what we think you need to do and what's likely to happen if you don't do it, right, and provide all of the evidence for that. So that last bit completely unimaginable before generative AI. You would just be surfacing objects of context. So you'd say, oh, here's the thing, it's similar. To that occasion where you had a bearing problem, this kind of thing, right? But yeah, we can be massively more sophisticated now. That's, you know, some amazing opportunities very much into the prescriptive side of things. So it's not just this is happening and this is what we spotted, but we suggest you're, you're doing is a lot, is stronger maybe than it was before. Yeah, taking the super intelligences, which are now available for hire and connecting them up to all of this sort of information that we have and ways of interacting with that information. I guess to balance up a little bit as well. So, is it, what areas are still, like, what challenges do we still have? What's just as hard as it was before, maybe that the moment it generative AI hasn't solved. The way I like to look at this, we just described a little bit of. About how Sensii works, but many other systems have the same kind of form, right? Especially, I think, industrial AI systems. Did you have some kind of live information about what's going on? This can be time series, uh, different kinds of vibration data. You can imagine even images, video, whatever, right? You've got something which is information about what's happening, OK? And then you have information about the setting. Say like what the machine is, the history of the machine, you mean, engineering drawings, all of this kind of thing, right? And so you're taking the data that you're seeing and you're trying to make sense of that data fundamentally with what you know about the context, all of that additional information. And I think ultimately, what generative AI allows is to do that kind optimally, right? So given the data you have, given the context you have, you know. Like what is the most clear and confident things that you can say, right, that are most useful to the user. So there's a sense in which you set up the system properly, you can solve that. But the situation that I guess Sensi is starting to touch, and I think many of these applications will, is you have limitations on that input information. So the AI is letting you do. Ultimately perfect reasoning over the input information, but your input information compared to the amount of information that's out there in the world is typically very small, right? And then this is ultimately going to set the limit, and there'll be a point where it makes no sense to keep upgrading the models in these kinds of systems without doing anything about the information, right? And so there will be a lot of differentiation. In systems in how they approach this problem of getting context. Need to get information about the machine, what the machines are doing, what's happened to the machine in the past, how the machine works, and specifically what it's doing now, right? Yeah, and it's interesting because we, you know, in our discussions before, I think we knew where this podcast was going to lead, and you mentioned context. The first mention with 12 minutes in, so I thought it was gonna actually be front and center. But you mentioned it there about it's all very well changing models and getting better models, but if the context isn't there. So that, what you're saying maybe, James, is actually that will become the competitive advantage. What context can you bring to these models to, to, for the outputs to be improved? Yeah, I think so. And yeah, so, I could have mentioned context in response to all of the questions earlier, later question. But yeah. Yeah, it's what you bring as an organization, all the people developing the software to the software, and in the case of Siemens, you can think about all of this collective industrial knowledge, not all of which is available to these standard foundation models, and then, yeah, all kinds of systems and tricks and techniques and ways of getting context basically out of users' heads, right? And as we said, technical stuff about the machines and different kinds of data. And you can imagine. And there's now generator AI is also helpful to this, right? So again, it's become a very easy task, for example, to photograph nameplates on machines, pull the text out of that and, and use it, right? So that could be done at some level before generative AI, but stuff like this just becomes very accurate and very easy, right? But yeah, it is exactly as you said, I think this will be differentiate lots of these products, how they approach. That context problem and data collection problem on the input, and then the second area really would be on the output. Yeah, how you're getting this stuff to users, how you're interacting with customers or, or stuff like that. Yeah. And it, I think you touched upon the elements there, James, or maybe we could spell it out clearly for your, from your point of view, what information would an AI system need to understand what's happening with the machine? In an ideal world, which we're not, I know, but if we had every bit of context available, what should we be feeding into the system for it to provide the, the best output it can do? Yeah, I think there's probably a massive difference here between what's sensible and what the perfect real world is, you know, you know, in the real world, you, no, no, I just mean in a perfect real world, it's you're interested in the physics of things, right? You know, the complete digital detailed digital twins, right? So very great detail so that the AI can essentially have access and know that information and then also test its hypothesis against the models derived from that information. So really, of course. In an ideal world, this can be absurdly detailed and, and impractical, right? But if you're going to being more practical, yeah, you need to know the, the, the, the, the basic things, right, from what the, the machine is, right? It's, it's make and a model and what it's doing down through the details of its construction and deployment, different components inside it, all of that kind of thing. Yeah. And it, it's strange to talk about things as obvious as the make and model of the machine, but in, in many settings, yeah, getting this information users and getting it into the app requires work even for basic things. Yeah. And of course, you want to go all the way down into the physics and engineering as deep as you can. But yeah, so if you want information at different levels about what the thing is, and then you need to know what it's doing. And then you need to how it relates to other machines around it and the processes, all this kind of stuff. And then you want to know the history of things, which is often extremely useful. And yeah, that, that's the kind of thing that's in maintenance engineers' heads. They get to know their machines over time, they understand their quirks, they know what they did to them last month to fix X, Y, Z, all of this stuff. Yeah, it's super useful, right? So it's the details of what is and then over time, pushing that back over there. It's. narrative histories of your assets. I guess the, yeah, elements, and if we focus again, you, like you've mentioned that phrase, what's in the, the maintenance person's head, I guess a lot of that. Yeah, we, we capture this the best ways that we can, and we're constantly thinking of ways to improve how we do this, and it's very tricky, right? So traditionally in Sensi, it has been, like you say, feedback on alerts, people telling us oh it was this and this was useful, this wasn't, all of that kind of thing, and then maintenance data, so the actual hard interventions. On the machines, but it's very clear that maintenance engineers have a lot more in their head than that. Yeah. And a lot of that is implicit knowledge, a feel for things and what have you. And it. Those things can even be things that they find themselves difficult to express in words. But yeah, so it is, it is as you said, and we're aware that this is only, as always, quite partial and try to think of ways to get more information without people doing too much work. Yeah. But of course, we should point out everything we've said, you know, in an ideal world and what's practical and what's not, but, but you can get started with, there's a minimum you can get started with and still get impressive results, right? So it's probably a minimum. What's ideal and then what's, when you said about going into the physics and, and that side of things, which is above and beyond, but yeah, there is, it's, yeah, there, there is a minimum starting point you can get, get started with. Yeah, and, and generative AI helps massively there as well because you, you know, have a system there which has this knowledge of the whole world and everything in it up to a certain level. Sensor itself is designed to produce interesting results out of the box, so to spot things happening that, that might be of interest to the user. OK, just out of the box, yeah, I mean if you bring those two things together, you can immediately start giving a certain quality of prescriptive messaging. Even just knowing that it's a motor, the system can understand what the time series are, and see what's happening in them, and it knows all about motors and all about time series and can start to say, yeah, at a certain level immediately. Yeah, that this could be a bearing issue or something like that, and then its ability to do that, yeah, it improves as it gets a better handle on the specifics of your motor and what it's specifically doing. And like we're saying, it's history and all of that. But yeah, very important to sense that it's a system that is designed to work outside the box, do something useful and valuable to you, and then try to be a sponge for additional information, additional context that helps to ratchet up this performance over time. Exactly. Yeah, so it can build up that context over time, of course, as well. Yeah, that's what you're saying. And I guess, as well, so, we talked a little bit about the machine learning foundations of Sensai, but what I thought might be interesting to people as well, because, again, Egentic AI is a hot topic, and a lot of software platforms and whatever industry are moving in that direction. But how does that sort of split, if you wanna call it a split between the machine learning side and the Agentic AI. Side, how does that, I guess first of all, why, why is a split like that and how does it work? How does it benefit the end user having that split? Maybe is a good way to. Yeah, so the division you're talking there is quite a hard division in the sensii system that we alluded to earlier, and we often talk about this as the lower brain and the upper brain because in some sense, it's mapping onto these different functions in, in, in animal brains. But the first part of the system is this statistical machine learning part, and that is designed to operate at massive scale, maybe monitoring tens or hundreds of thousands of time series at a customer and doing it every time these data points come in. Yeah, so this is just basically a huge, huge amount of data. And what you are looking for in that first part of the system is things of potential interest. Basically, and it would work very well. OK, to have a, an AI agent full of data analysis tools and all kinds of stuff analyzing data from every machine at every single time point, they would be prohibitively expensive to do that. So, the job of this first part of the system is to find things of potential interest in a fast, efficient, clean way, and then basically to push high-quality triggers forward. To the upper brain, yeah, which is the expensive reasoning part of the system. OK? And if we make that division very efficient, we only pass through very few signals into the agential part, and it means that we can spend a lot more resources on the agential part to make that very, very good. OK? Yeah, so there's this, it's basically a filter. OK, to make sure that you only call the expensive agent when it's worth its effort to do that and you get a return, return on the resources. OK? Yeah, and the, the benefit to use it here is purely one of one of efficiency. OK, that the intelligence, which is expensive, gets directed precisely to where it's needed and not wasted on lower value, low-value stuff, right? And the output of the intelligence insensi, the agentic part, is the thing the user sees, right? It's composing and structuring and writing though. Yeah, writing the output. OK, and we obviously want that to be as good as possible. Yeah, so it's interesting we're talking about this lower half and upper half of the brain, and I was wondering, like both now and into the future, whether is it the 50/50, if we talk about percentage terms, like a 50%, both are equally. Important in what Sensi produces as an output, and is that the case now, or, and will that change in the future as well as we go to lead more into the agentic side of things? …

Full transcript (WebVTT)