Leveraging AI to Improve Manufacturing

Welcome everybody. My name is Mike Unger and I'm one of the co-founders of the Industry 4.0 Club. The club's purpose is to advance the deployment of Industry 4.0 for manufacturers really around the country and around the world. My background includes 35 years of experience with Michelin North America in a variety of continuous improvement operations, finance, and HR roles. Today I'm a focal point business coach helping business leaders achieve their full potential. I'm joined today by my co-host Mike Yost. Mike, would you introduce yourself please?

Sure, thank you. My name is Mike Yost. I'm one of the co-founders of the Industry 4.0 Club and I work alongside Mike to help engage industry in conversations like we're going to have today with Ashish. to learn about modern technologies, advancing opportunities for people to improve their businesses and such. 35-plus-year veteran in industrial automation, industrial software space. It's a lot of my time focused nowadays on this work to help educate industry around the topics of Industry 4.0.

Thanks, Mike. So we have the privilege of having these conversations today from the SESAME broadcast booth in the Smart Manufacturing Zone at SME South Tech Conference. And we're, like Mike said, we're joined today by Ashish Khan. So welcome, Ashish.

Thank you.

A little bit about Ashish. He's the Vice President of Smart Manufacturing AI Business at Falconry. Falconry, at Falconry, Ashish helps to enable enterprises with AI governance by operationalizing trust in AI/ML systems, and deploying responsible AI. He's previously held CEO and CRO roles enabling enterprises to achieve sustainability and productivity goals using AI, IoT, and he holds a BS in electronics and electrical communication engineering from the Indian Institute of Technology and a master's degree in electrical engineering from Stony Brook University.

Thank you.

Quite impressive. Welcome, Ashish. So to start our discussion, can you share a little bit about the relationship between Falconry and SESAME?

Sure. So Falconry is a SESAME member, first of all, and last year SESAME and Department of Energy and ArcelorMittal, we delivered a project for ArcelorMittal's steel plant in Calvert, Alabama. And we were able to solve a particularly hard problem using AI for ArcelorMittal, which is called strip break classification. So using AI, we are able to deliver a solution that normally would be taking, let's say, 8 hours for a maintenance— plant maintenance engineer to find the root cause. And we are able to solve that with the power of AI within 15 minutes to half an hour.

Wow. Fantastic.

And SESAME this year has picked out, have highlighted Falconry as one of their successful project execution for the— I guess they awarded like 60 or 70 of those, so they handpicked us and we are presenting yesterday.

Fantastic, congratulations for that, that's really great. So you've been, you mentioned You've been with Factornery for a short time, but Factornery has been in manufacturing and AI for many, many years now. What kind of problems— you mentioned one that you helped solve— what kind of problems does Factornery help companies address?

Yeah, so first of all, we help these manufacturing companies to migrate to data-driven operations, and the problems we solve. Number one is reducing the unplanned downtime. ABB just last month published a report showing that it costs $125,000 on average per hour for unplanned downtime for these manufacturing companies. So using AI, we reduce this because we provide these alerts to maintenance engineers for a potential problem that is coming. The second thing, second problem we solve for them, as I mentioned, if there is a failure, an outage of some kind, either a stoppage, slowdown, these maintenance engineers would like to find out the root cause, the root cause analysis. It takes a lot of time. In a particular steel manufacturer, they actually— the plant was down for 2 days because they could not find the root cause to solve. And they thought that they have found it, but that was not the right one. So using AI, we shortened that period significantly. And the third problem we sometimes solve is There are, especially in process industry, there are thousands of parameters.

Yes.

And they continuously monitor, but they do not know for a particular quality or particular KPI what are the most important contributing factor. For example, in case of the ArcelorMittal, that example that I gave you, there were 1,350 parameters and we boiled down to 7 parameters that you need to track using AI. So number one, as I said, the downtime reduction, the unplanned downtime reduction. For us, for one company we are working with, they attributed to $4.5 million annual, $5 million annual savings based on that they attributed to Falconry AI alerts.

That's fantastic.

Yeah, so just a question, 2 things maybe, just One about Falconry specifically, and then translate that more broadly into industry. You mentioned steel. So do you have a particular vertical focus from an application perspective? You have domain expertise in-house. And I guess my thinking is the amount of data that you need to do the work that you do. If you're talking about large steel processes, probably lots of data. Does it translate out of steel into other industries? A lot of the people that we deal with are coming from smaller companies. You know, maybe machine shops, that kind of stuff. Should they be thinking AI is an answer for them, or is there a correlation between how much data is created to run the analytics and such and, you know, the value of the—

Okay, excellent question. No, actually, you know, first of all, if you don't have a lot of data, that's all right for us because the particular AI technology we are using is we do not create the labels. And in other words, we do not need a lot of data on particular failure modes to study. So we use a generative AI technology where we actually look at the data and find what the normal behavior is, and we generate or construct what will be the behavior of the signal in the next window, next window, just like ChatGPT tries to find out what is the next word, the next word, and so on. So it's the same foundational technology in some sense. So we, we can start in small, medium businesses as well. As a matter of fact, those companies do not have data scientists, and therefore our particular technology is very appropriate because all they have to do is to ingest the data without having to any kind of cleaning or preparing or labeling the data, and we find these anomalies which are precursors to problems, and they can use these to identify what could be a potential problem coming and so on.

And then how about in that, you know, that work between you working, your company providing expertise, technical expertise, understanding the domain extent, And then data scientists within the user side, and then the process engineers, the process people. What's that typically look like? Because like I said, a lot of the people that we deal with probably don't have data scientists on staff. So you work with the process engineers, like what does it typically look like? We're learning something from the data we're getting, and now you have to go out and affect process changes or what have you. What's that typically?

Yeah, good question. We work very closely with the plant engineers, like the reliability engineers, the maintenance engineers, the process engineers, the quality engineers, and we help them like a— again, just like a— we call it AI copilot. Just like, you know, your smartphone, you do not know what microprocessor is used or what the LTE, what the 5G is.

Okay.

You're just using it. So we try to make it as simple as that, that hey, this is the signal that is— we are predicting that this signal is not behaving correctly. Now you have the knowledge, for example, a motor current spikes up, what you are supposed to do. So that they do. But of course, there is a, you know, we, every week we sit down with our customers, go through the dashboard, go to the heat map, and try to point out that, hey, you should be looking at these signals. And also, once we work with these engineers, we also pick up a lot of our understanding of what they're looking for and so on. So it's a somewhat collaborative effort. But we don't expect, at least from our side, we do not expect our users to know anything about AI or algorithm or data or anything. We try to hide the complexity as much as possible.

That's good. Heat maps and dashboards sound like things that are familiar to people in a plant environment, so it probably makes sense they can understand those things.

Yeah.

So we try to, you know, as you all know, the goal of technology is always to hide the complexity. And this best example is probably the old telephone that, you know, used to dial and pick up, right? And that shows how technology could be. There is enormous complexity, right, in the backend. But that's what we try to do with AI as well.

That's a good example. You mentioned in your answer earlier, one of the things we were talking about was generative AI. And I had 2 questions really. So should we be worried about generative AI? That's kind of one that's on the side, you know, for just our audience. But secondly, you touched on how you integrate that into your solutions. Can you elaborate more on that as well?

Yeah. So first of all, you know, when we uttered the word generative AI because of OpenAI's announcement in December '22 about the ChatGPT, everybody associates generative AI with these large language models. you know, the natural language processing type of models like a chatbot and so on.

Yes.

But, you know, generative AI is essentially different from what is another class of AI, which is we call discriminative AI, which is like, you know, the typically image recognition type, like this is a cat, this is— you train the AI, and this is not a cat, that type of thing. But generative AI can be used beyond just the LLMs, and that's what we are using. So you can think of this as opposed to the corpus of everything, every text that's ever been written, that is the body on which the AI is working. Here it is the time series data from the entire plant that we are working on to predict the, uh, or work on the generative side. So number one, on the, uh, in case of generative AI, there is always a concern that copyrights, because, you know, a lot of texts. That is why you will notice that many of the, uh, creative artists are suing OpenAI and etc. So there is one side to it, and the other side is the hallucination, etc. In our case, you know, there is, uh, all those issues are obviously not there. And we are— we have been very effective because, you know, the other thing is that many of these technologies are obviously dependent on very high-performance computing and so on. Fortunately, thanks to NVIDIA, the GPUs are available. They have brought down the cost of computing significantly.

Yes.

So our technology can work flawlessly now.

Super.

Maybe talk just a second. Mike mentioned, I think, in your title or your role responsibility, AI governance. Could you maybe talk about what that encompasses, what you do? Are companies expecting a solution provider like yourselves to have somebody like you that'll talk governance with them?

Yeah, we just want to make sure that the— first of all, we use explainable AI. So beyond manufacturing, we also have applications in the This— our technology is used by US government, like the Department of Defense, US Navy, etc. They want to make sure that it's always explainable, that we did not come— we did not predict on the basis of something. And some of the elements do not apply to us. But in case of AI, you want to make sure there is no bias. There is no— there is— which is possible when you are using AI. in the context of, let's say, legal or hiring, etc., where a particular bias could be detrimental to the interests of a lot of people. So, but mainly the one is the model governance, that you know what kind of models you're using, and there is a provenance of all these models, and there is a— you want to make sure that You have a knowledge graph of all the models that you're using and so on and so forth, and you're maintaining properly, etc.

So for our audience, if just broadly speaking, can they trust AI? Should they, or should they be looking for something in a solution provider that would say, yes, it's safe to do this if there are these things Because I know a lot of people are hesitant and saying, you know, not going there yet, but then they run the risk of being left behind, right? So other people are cautious about it. So can you say, or yes, it's safe if there are these boundaries around it?

Yeah, absolutely. In this context that AI being used in the manufacturing plant, so we are using their data. What we are trying to do is to find the hidden things in your data, and data don't— so, and especially in this context that you are essentially— we are getting data from the PLCs, from the sensors, from the historian, from the SCADA, etc., etc. So those data that we look at and find the problems that are hidden that you cannot see without using AI, and this is what we are getting out of the data. So as long as there is, you know, it is embedded in the data, we don't see any problem at all. So it should be 100% trustworthy, and that's why, you know, every company is trying to use the power of AI. And, you know, when I come here, I see Smart Manufacturing Experience, Digital Solutions Showcase, So it's all about getting data and use this type of AI for the productivity improvement as well as the sustainability and the carbon footprint reduction, et cetera.

And the data sources you mentioned there, the PLCs and the SCADA systems and such, how about human information? Are you able to get information like manually input information? Does that play a role or is there not enough of that or?

Yes, absolutely. In some sense it does. For example, you know, we are looking for anomaly detection. Detection, but there is a reason why a particular section of a plant is— we are not getting any signals. And we get that in the form of a signal. We get it from the plant operation people that, hey, we have— this is a— we are going through a temporary shutdown or something like that. So we get that data. So yes, we receive that data. Now, there are a lot of data. You asked me about generative AI. That can be used for mining some of the data that they're recording, like the maintenance people are taking notes, right?

Yeah.

That what happened, what they did, and so on. So this is where the generative AI will play a big role because now instead of the corpus of the internet text, now you have the corporate data that you can work on and provide, you know, that chatbot-style help. that this is, uh, now that it is happening and I found out a similar thing happened, so therefore this could be applicable and so on.

You know what, um, today a lot of small and mid-sized companies are slow to start their Industry 4.0 journey, which includes they're slow to get involved with AI. You know, we, we hear that AI is probably not the first place a company should start. Can you talk about what needs to be in place in a in a small or mid-sized company before they start AI, an AI solution?

Yeah, so obviously, you know, AI needs data. So you should be getting the data from, you know, you should probably have some sort of data acquisition system in place. Otherwise, you know, we cannot start. But one of the mistakes a lot of times people make, you know, there's a perception that You must create a data lake of some kind, and we we say that no, whatever data you have, use that to score some wins early enough so that you build confidence in AI. Because AI also requires a cultural shift that you have to trust AI from and not AI not in just in manufacturing. You are using AI for you know efficient. You should treat AI like a co pilot that is helping you. in everything we do. Just like, you know, you and I are using ChatGPT for our emails, right? That are correcting— Don't tell anybody. So, so we, we, first of all, everybody is, you know, have a SCADA system generally, and they have added some sensors, they're getting some data. And we say that start with You know, whatever data you have, let's try to find anomalies from there, 'cause you don't have to invest a lot in creating all this. Therefore, you could be able to show results to your management that using AI, we have been able to reduce the unplanned downtime, or we have been able to reduce the, you know, the root cause analysis from 8 hours to half an hour, something like that. Once you show those, and then you build it up, And use that in an incremental fashion.

So you don't need— what I think I hear you saying is a small to mid-sized company, if they have a problem where they think this might be an application, they don't need a lot of infrastructure to get started is what I hear you saying.

That is one of the things that we have tried, particularly at Falconry, that a very little investment in hiring the data scientists and so on So we really made it like an AI copilot rather than having them to build AI expertise inside the company.

Okay. And then also the, I say the infrastructure, so we talk about the people, but in terms of, obviously you said you've got to have sensors or a method of collecting the data from the PLCs, but beyond that?

No, generally, you know, if they have a historian, we take that data from the historian. There are a lot of, you know, we have a booth here for OPC, so OPC UA, helps to get the data from these historians into any kind of cloud-based data warehousing.

Mike started the conversation asking about the relationship between Falconry and Sesame, and Sesame's all about building that data foundation, disparate sources, common modeling, et cetera, which sets up nicely to do reporting analytics.

Yeah, actually, you know, this is so heartening to come to this kind of conferences and see that, you know, Currently, I feel like the United States of America, we are having a renaissance in manufacturing. The reshoring is happening. And good thing is that we are using all kinds of advanced technologies. You know, as I said in my booth, we are providing AI Copilot. You know, there are so many companies offering digital solutions. There is a Fanuc that is showing their robots. So the automation, AI, data-driven, you know, all these things are coalescing together. I hope that this, you know, we go through another renaissance of American manufacturing.

Yeah.

And one of the things that I'm fond of saying, and I said it in our last recording, but I would encourage people that are watching or engaging to take heart in the fact that there are a lot of people like yourselves that are doing sort of behind-the-scenes work to become an overnight success, right? Like, people are going to look and say, wow, this stuff is, is easy to use, it's trustworthy, etc. But that's because of all the work that's happening with all these types of folks that are putting in the effort. Um, and so that's, uh, I would just say to everybody that, um, there's a lot of this foundational work a lot of people are doing, um, a lot of experience, trials, bumps in the road, successes, learnings, et cetera, that are leading to a point where it's going to look like, oh wow, you know, we're just hearing about this, but you guys have been doing it for a long time.

Yeah, absolutely.

That's really neat. Well, as we get ready to close here, I have one last question for you really about, you've got a lot of experience putting AI into different environments and companies. Can you summarize maybe a couple of the key success factors For a successful AI implementation?

Yeah, so, you know, key success factors generally apply to a project. And so one first thing is that the outcome. So you must know what you are going to use it for and a very specific goal so that you don't treat it like a— in other words, you know, from the very beginning you should have the outcomes in mind that this is what I'm going to use it for.

Good tip.

And in the project area, I would also recommend that, you know, in both the people, the process, and the technology, all are very much aligned. So you need a little bit of a cultural change that you have to trust data, you have to be data-oriented. that this data is how we are going to leverage for our, um, uh, for the improvements we are looking for.

That's a good point. Those are 2 good points there. Mike, do you have any final thoughts before we wrap up?

Uh, yeah, I, I would say, you know, the, the talk about explainable AI, I think, is, um, uh, good to hear. A new term I haven't heard before. And I think probably the main thing would be the fact that folks have been doing this, folks like Falconry, folks like the people at SESAME, and that you are having success doing this with the goal of being that, you know, us as manufacturers don't have to know the nuts and bolts of what's going on under the hood, that we can trust the outcomes, use the outcomes to drive business, you know?

Right.

That's the main thing.

That's our goal, that just like I said, that technology should be hiding all the complexity and make it easy for— so this is why I use the term copilot, you know, just so that the maintenance people, the reliability people, the plant operations people feel very comfortable, just like a sergeant feels very comfortable using a robotic arm. So we, you know, we feel like doing that to a doctor and Let them use their subject matter expertise for what they are best at, right? So we are going to be just a tool.

It's a great example. It's a great story. Well, that's fantastic. Ashish, thank you. Mike, thank you very much.

Thank you.

It's been a great conversation. So if you like these kinds of conversations, please look for additional Industry 4.0 Club broadcasts. We speak often about women in manufacturing, The Future of Manufacturing, other fireside chats, and People, Culture, and Process on LinkedIn Live, YouTube, and also on Twitter. These events are advertised through our LinkedIn page as well as on our website. And please also join us at our next live event. We'll be at IME West in February of 2024. So again, thank you, Ashish. Thank you, Mike, and have a great day. Thank you.