Podcast: Why industrial AI is moving closer to the edge

TW: Would you say that that's most important in energy or those sectors we've talked about? Is it equally important across all of those sectors? Do you see energy in particular is leading the way in deploying AI this way? In my experience, energy has such a high stringency rate when it comes to reliability requirements, like five nines reliability, that they are most often quickest to embrace things like advanced vibration analysis techniques to detect anomalies and fault modes early. 

AF: I think it's a common theme across industries. AI has enormous potential to change the way that businesses and companies operate. Obviously there are some industries that maybe are slightly more advanced, perhaps things like, as I said, for doing DER optimization within the energy and renewable sector. But in manufacturing, companies are looking to see how AI can improve their operations, optimize their production lines, and support greater yields within the process industry as well.

TW: Let me ask you a general question about governance. How do you see the governance of edge AI evolving over the next five years? Is this a case where we’ll be able to help the AIs govern themselves? Is this a case where humans will need to step in with more sophisticated techniques to achieve the outcomes?

AF: The main I think will be to move from a more of an ad hoc model oversight, which is probably how people are rolling out AI initially to start with, to a much more formalized, auditable process, which really mirrors safety critical engineering. You know, AI is now going to be treated as a first class safety critical component within the system as they get involved in more sophisticated types of use cases, from monitoring to possibly interacting with the plant, making recommendations about set point changes and stuff like that. So, you can expect things like the life cycle management of these types of technology will have to be formalized. So, you can expect things like, as I said, very sophisticated life cycle management: how you train your models, how you version your models, how you deploy your models, update the models, rollback the models. Lineage tracking – you need to be able to have traceability on how the training data sets that you train the models on so that if there's any problems, you can go back and look at how the models were created in the first place and detect problems and stuff like that. 

And also things like the other things that you'll see as common, there'll be continuous monitoring of the model outputs against what the expected behavior is, formalized security certification. So these are things that some companies are adopting as I said, as back best practice, but this will be much more formalized, I think, across different areas. And certainly in terms of there will be, I think there will be better governance and formalized processes for the adoption of these technologies.

TW: You know, that leads me to a roles and responsibilities type question, Andy, to take a step back from the tech for a second. So many maintenance and reliability departments get rolled up through operations, so that combination helps maintenance work with the operators effectively to get things done and when they want to stop the line to work on it. Given what you're talking about, these new governance responsibilities, what would you say is the role of asset managers, the maintenance folks, the reliability folks, when it comes to supporting edge AI operations and to drive this improved governance?

AF: I think the asset managers will become the stewards of the AI-enabled equipment, and I think they're going to form the bridge between the data science and the operations. So this is a kind of a new discipline in terms of, what the operators are going to starting to have to become experts in. They're going to be involved in everything from ensuring data quality. (So data quality is a key thing in terms of how the models perform and run and how models are retrained, et cetera, et cetera.) They are going to be key to basically in terms of the monitoring process as well, so validating the model outputs against the expected real world behavior, and things like integrating AI insights into things like the maintenance strategies. 

TW: A lot of the conferences that I've been to which focus specifically on the maintenance function and the planning, work planning and scheduling, focus on large language models which help collect better data from the field techs to plan the work better and develop better work plans. I like that you're talking about the practical governance part of this because yeah, it's going to require differently skilled maintenance professionals to who are going to, they're going to inherit this asset, and that's what the AI is. In addition to the physical machine, it's all a kind of asset which needs to be maintained and taken care of.

AF: 100%. I think with the AI rollout, it's certainly becoming like a continuous process monitoring as opposed to basically, say, a one-shot testing type process. You need to constantly monitor for ongoing assurance. As I said, we mentioned before, you have to be monitoring the predictions against the real world outcomes. You need to be looking for things like the models can drift over time, so monitoring the models for drift. And also, as I said, we mentioned the life cycle of the models. The models inevitably need to be improved and redeployed over time by using new additional data sets, new data sets. In terms of the rollout, we need to constantly monitor the output against expected results, looking at the confidence levels that the models are producing.

And then from a governance point of view, depending on what type of AI you're using and what AI is used for, they may have to have guardrails around the AI, which may, in many cases, that's a human in the loop type of guardrail. You know, you can build constraints and boundaries into the AI, so it can't move within expected limits. But for example, if the AI is making recommendations for different changes to the system, optimization of set points, which affect control loops and stuff like that, a lot of times you have to have a human in the loop or certainly, to make sure that the predicted forecasts are within the expected bounds of the system and that a human approves those changes.

TW: We'll get you out of here with a couple of safety questions. The first one's more of a macro question, Andy. What are some of the biggest risks that plant operators should watch for when AI is influencing decisions closer to the assets and processes?

AF: I think there's three broad categories of risk. So I think there's potentially a safety risk, of course, is probably one of the most critical things. If the AI is making predictions, or driving outcomes that could potentially affect the actual physical equipment, then of course there's a safety risk there, there's a potential to maybe cause some physical harm if the equipment goes wrong. So as I said, many of the cases with the type of systems we've been discussing so far, the AI isn't directly affecting or driving the equipment. It's more about monitoring, making recommendations. But there's, of course, there's a safety risk there. 

Another type of risk is what we call systematic or cascading risks. If the AI is making changes to the system, if for example the AI, if we don't have these guardrails or interlocks within the system, if the AI output perhaps generates a signal that's out of bounds, it could have a cascading ripple effect across other parts of the system. So, that's why, again, having the guardrails monitoring the output of the AI to make sure that's within expected limits, because if it isn't, it could have an effect of other parts of the system which are dependent on decisions that the AI is making. 

And the other risk, I guess, is more of a governance and accountability risk. When you're using AI, you know, a lot of cases people are wondering potentially, why did the model behave this way? Why did the model go wrong? And who approved the update to the model, which is an effectiveness thing. So, the governance risk of identifying the cause of the problem, what caused that as a key thing as well?

TW: Then on the organizational capability side, do you think that organizations are aware of the changes they'll need to make to safely operate these AI edge systems at scale? Or are people learning on the fly what's going to be required from an organizational point of view?

AF: It depends on the organization. I think as AI has been rolled out and from, for example, proof of concepts and early adopters proven out what AI can actually do in terms of bringing off operational benefits, some companies are ahead of the game, are getting quite sophisticated in terms of understanding how they actually operate this at fleet scale successfully. The companies who are at the start of the start of that journey, what’s happened is best practice is starting to formalize, and there's ways of adopting this within your organization and following templates and best practice. 

One of the key things, I think, is that the key organizational shift will be the life cycle management of AI. That includes everything from how you develop the AI models initially and how you train them, how you version your data sets and version your models, how do you test the models, particularly potentially in a development environment or your lab environment, might be using simulators to start with, to start the plant, how you pilot the rollout. These are all things so you don't, you're not going to potentially just deploy the AI across systems until you've had maybe a pilot phase where you only deploy it on a subset of the system. 

There's an enormous challenge with respect to things like orchestration. How do you actually deploy your AI onto the nodes within the system where you want to run the AI? How do you monitor that AI? We talked about in terms of being able to make sure that it's running within the bounds. How do you update that AI over time? Because the models can drift, there's normally a process of using newer data sets to retrain the models to get improved behavior. How do you redeploy those models over time? And potentially, if things go wrong within the system, you need automated processes to roll back to predicted safe states. So the life cycle management of operations is becoming critical.

TW: You just provided me with a very big aha moment. The parallel between what current maintenance and reliability folks do with the storeroom and with what they'll be asked to do with life cycle management, there's a very strong parallel there. As long as those disciplines can wrap their heads around the idea that yes, AI is not something which will stay static, it's going to change, it's going to also be sunsetted and new ones will come in, just like you would a bunch of spare bearings or spare motors. At some point you have to go to the shelf and get them, get the newer, more effective model.

AF: That's exactly right, and I think with AI that the likelihood of rolling out updates in comparison to an existing DCS system is that the AI will probably be updated more frequently than the existing software infrastructure. So being able to do that efficiently and in a controlled process is actually, it's probably one of the most critical things. Of course, the AI is doing, there's a means to an end. It's making optimizations, it's making predictions, it's helping run your plant more efficiently, improving the yields of your processes. But without the lifecycle management and the automated orchestration and your rollback features, managing AI at scale is very difficult.

About the Podcast
Great Question: A Manufacturing Podcast offers news and information for the people who make, store and move things and those who manage and maintain the facilities where that work gets done. Manufacturers from chemical producers to automakers to machine shops can listen for critical insights into the technologies, economic conditions and best practices that can influence how to best run facilities to reach operational excellence.

Listen to another episode and subscribe on your favorite podcast app