How AI is transforming the modern factory? 

Can AI really spot risks faster than any sensor? How far are we from the autonomous supply chain?  In this episode of the AI at Scale podcast, Mike Labhart, North America Global Supply Chain Director for Smart Operations & Data Analytics of Schneider Electric, walks us through the most valuable AI use cases in manufacturing and what’s next in smart factories. 

Starting from safety, through quality, maintenance and energy, Mike reveals what scaled industrial AI truly looks like. He takes us inside smart Schneider Electric factories, recognized as lighthouses by World Economic Forum and ranked number 1 in the Gartner top 25 Supply Chain, to share the most exciting AI use cases on the shopfloor. 

What you will learn from this conversation : 

• How AI detects safety risks earlier than traditional sensors, from smoke to missing Personal Protection Equipment (PPE)
• How deep‑learning and image recognition improves quality and helps to identify ergonomics risks across dozens of sites
• How machine learning prevents costly downtime thanks to predictive maintenance
• Why clean, consistent data naming is key to scaling AI across factories
• What agentic AI means for the future of manufacturing
• Why autonomous supply chains are within reach now. 

Get a unique chance to hear directly from a seasoned smart factory leader on what it means to build AI that powers the next generation of factories. 

Listen to the AI at Scale podcast 

Listen to Mike Labhart: On the shopfloor of AI-driven factories episode. Subscribe to the show on the preferred streaming platform (Apple Podcasts, Spotify or listen to the episode on YouTube.)

Like what you hear?

• Visit our AI at Scale website to discover how do we transform energy management and industrial automation with artificial intelligence technologies to create more sustainable solutions.  
• Listen to the other episodes of the AI at Scale podcast or watch them on YouTube 

Transcript

Host:  

Welcome everyone. My name is Gosia Górska, and I’m the host of the AI at Scale podcast by Schneider Electric. Today I have a pleasure hosting Mike Labhart, who is the North America Global Supply Chain Director for smart operations and data analytics at Schneider Electric. Welcome, Mike. 

Guest:  

Gosia, it’s a pleasure to be here. Hello. 

Host:  

So, Mike has spent his time turning factories into smarter and more connected, more data-driven operations. From AI to industrial IoT, Mike has seen what it really takes to bring emerging technologies to the shop floor. Before that, he spent seven years at General Motors working on major automation projects in high-volume manufacturing. Today, we will talk about what it looks like to modernize operations at scale. How data becomes a performance engine. How new technologies move from proof of concept to everyday use and scale. And what are the most valuable AI use cases in manufacturing. So, I’m glad that we will have this conversation, Mike. 

Guest:  

That sounds exciting. 

Safety in Supply Chain 

Host:  

And I wanted to start with some kind of an overview, because global supply chain is such a vast, really large territory where we can have different kinds of AI applications. Even if we look at Schneider Electric, global supply chain is an important part of our operations and also the one that is really highly recognized worldwide. We are the number one global supply chain according to Gartner. We received many accolades from the World Economic Forum. And I wanted to hear from you, what are some of the examples that you observe in your daily operations of the AI use cases, maybe starting from the ones that are less obvious. I know that safety is a really important topic in global supply chain. So what kind of examples could you bring to us from safety? 

Guest:  

Sure. Safety is the number one priority at Schneider Electric. So anytime we can utilize the new technology available with artificial intelligence, particularly deep learning, we’ve tried to take advantage of that in our facilities. I think about the way that deep learning technology has affected some of our lighthouse facilities. As an example, in all 48 plants here in North America, we’re now using tools to monitor and record the actual employee performing the task and using artificial intelligence to evaluate the ergonomic risk for those tasks. So we’re able now to take images or videos of the person performing the task, the workstation, the reach, the dimensions, and artificial intelligence provides us an immediate, quick evaluation of where those risks occur for that particular task. We’ve also tackled some very difficult problems in our El Paso facility in El Paso, Texas regarding the use and adoption of PPE in our high-voltage test bay areas. This has often been a difficult task to ensure that employees were wearing proper shock protection. We’ve been able to use deep learning technology and deploy models at the edge on our AVEVA platforms to actually detect the presence of gloves on those employees and notify and warn when we are missing that critical shock protection. So that’s been very impactful for us. 

Risks in the production facilities 

Host:  

If we just highlight what it means maybe for those who are not familiar with the vocabulary. So PPE is the personal protection equipment, and ergonomics actually even for people who are outside of global supply chain, you may hear about ergonomics because this is the art and science of how you should actually behave in a safe manner and the most efficient way for your health, for your safety as well in the working conditions. So, even in the office we have some ergonomics rules like how to adjust your laptop, and the same applies for the production. So this is really super informative to know that we can use the deep learning technology to recognize how to best design the workplace for the employees. And outside of ergonomics and the personal protection, I also heard about one application and I was hoping that you can tell us a bit more about some of the other risks that are related to safety and may occur in production facilities such as fire. 

Guest:  

Absolutely. Traditional smoke detection with the use of a smoke alarm in a particular area requires that the smoke propagates to that alarm in order to set off the trigger. In a wide-open manufacturing space, that can often be difficult to detect smoke over a large area, and once the smoke actually propagates to a sensor, you can understand that considerable damage may have occurred and employee risk may have occurred. We’ve been able to utilize the deep learning technology in conjunction with our AVEVA platform to deploy at the edge a model that detects the image of smoke. So we can monitor a large area and we can detect and alarm for the image of smoke that exists. We have welding processes in our Monterrey 1 facility where we have dust collection equipment that poses a fire risk to the facility, and we’re able to detect that smoke pattern before it actually reaches a sensor or a detector in a wide-open area. So that’s been very beneficial and impactful. 

Host:  

Okay, so I understand that the most important part here is the sensor limitations, right? Because why couldn’t we just use sensors for smoke detection? As you mentioned, the facilities sometimes are so big and large that it would be quite difficult. So how big are they on average, if you can tell? 

Guest:  

As an example, the Monterrey 1 facility is roughly a million square feet. Facilities like Lexington, Kentucky or Lighthouse, it’s not uncommon to be half a million square feet in size. Often with a high bay area or a large open space, and the operations are not confined to small quarters or enclosures. 

Host:  

I see. So, it shows basically the challenge and the solution that we could find. That’s really great. So, we said safety first. Next comes quality 

Quality

Guest:  

Quality. This is probably where the deep learning technology has affected us or impacted us most in the quality room. We’ve been very adoptive of many off-the-shelf deep learning solutions for providing quality checks of our products. I believe in North America now in our operations, we have a little over 100 installations throughout our processes where we’re utilizing deep learning technology cameras to validate that quality has been achieved, that the correct operations, that the correct products are contained within the product. What that has given us over traditional machine vision, which is nothing new, we’ve used vision systems in operations for years, is the ability to rapidly deploy models that would have taken considerable development time with traditional machine vision. We actually have the capability now to learn what’s normal for a product and report anomalies to that product without actually having to define each one of those individual anomalies to check for. So, we have seen a considerable impact in the quality domain with the use of the deep learning technology. And now we’re branching out of that traditional use of monitoring merely the product and beginning to monitor the actual actions that the employees are taking in heavily manual operations to validate that the manual actions are properly taken to build quality into the product as a first priority, quality at the source. 

Host:  

Yes, I see. And something that also captured my attention in relation to quality and vision inspection in the manufacturing facilities is the fact that you often lack the data actually, because how many pictures of a failed package, for example, can you have, right? How many pictures of some issues that can happen? There is a big limitation here, and I heard that sometimes we are using synthetic data because basically it’s hard to get a decent amount of images so that the system can actually recognize all the flaws. So the solution that I understand you are mentioning is solving exactly this issue, that you don’t need to have so many images because it can detect the flaws even if you didn’t document them before. 

Guest:  

That’s correct. When you think about training deep learning models on a product to detect an anomaly, the environmental variation also, such as the lighting, slight color, product variations, oftentimes it’s difficult to train a model to be highly accurate without having that type of variation to train to. You mentioned the synthetic image generation. We’ve utilized synthetic image generation in our plants to generate thousands of models from a few product images with that exact variation so that we can train our models to a high level of accuracy instead of trying to collect that data over many, many months of runtime to make our model accurate. So that’s a very important point and it’s something that is also sort of a new concept that we’ve seen impact with. 

Predictive maintenance

Host:  

Yeah, that’s really fantastic. So let’s move on to the next topic, which is maintenance. Predictive maintenance, other types of maintenance, how can AI help in this area? 

Guest:  

A lot of our plants, particularly in the engineer-to-order business and also facilities like the Lexington, Kentucky facility making make-to-stock product, are heavily vertically integrated. What I mean is that we have on-site fabrication and painting aside from assembly. And within that fabrication equipment, that painting equipment, we have a lot of critical equipment that is different than traditional assembly processes. A lot of rotating equipment that’s critical for the facility, a lot of pneumatic and hydraulic systems that are much more robust and much more industrial than the simple assembly process. Predictive maintenance has been crucial for us in terms of learning the patterns of data that describe normal for those critical pieces of equipment. Now we focused heavily in supply chain here in North America in our facilities with a very particular type of failure that occurs in rotating equipment: our pumps, our flywheels, our motors. We have partnered closely with Augury and we have implemented predictive maintenance using machine learning technology and vibration level detection. I think we’re now at nearly 130 assets in North America that are truly critical to the production process. We’ve also implemented using the AVEVA solution known as guided analytics. It provides a point-and-click machine learning model that deploys within the cloud and solves every 15 minutes based on the data that you point the solution to. And for non-rotating equipment like valves, temperatures, and times that describe the health of an asset, we’ve implemented over 50 models with AVEVA guided analytics for predicting equipment failure. So the machine learning portion of artificial intelligence has been very impactful for us for maintenance. We’re also now piloting what we call an advisor application for maintenance utilizing generative AI technology to glean from the records of repair services in recent history for the equipment, and be able to vectorize or match to the type of failure that the equipment has experienced to make recommendations to the maintenance employee on what to fix and how to repair. It also provides a large language model front end to the large collection of data and documentation that we have on the piece of equipment. So, we’re seeing some very exciting impacts to the maintenance world with artificial intelligence on our shop floor. 

Energy production

Host:  

So, this is showing really how AI works for maintenance. And last but not least, if we can double down on the energy. How the energy flows in the production systems. Industrial processes are usually energy-intensive. What are the best use cases that you observe in this area? 

Guest:  

We’ve centralized our smart operations program here in North America and also globally with standards on the adoption of energy management tools within our facilities. So, it’s a requirement for tools like Power Monitoring Expert, Resource Advisor, and Building Advisor to be deployed in our facilities, where the Power Monitoring Expert collects the data at the edge and provides the immediate visualizations for comparison of energy usage. But then lying in the app and analytic layer, the Building Advisor has the AI component that detects troublesome trends. Troublesome trends in temperatures, for example, we have documented cases here in the Lexington facility where dampers or controls may be operating slowly or sluggish in terms of mixing air for climate control within the building. And the Building Advisor has noted that troublesome area and alerted us to that fact such that we could repair before the cost accumulated with having to make up for that error with more heat from possibly boiler, oil, or natural gas type fired operations. So Building Advisor and Power Monitoring Expert are very critical to the smart operations strategy for North America. And the Resource Advisor, which our digital managers and regional managers use to gain insights on our energy use from facility to facility, and also as compared to weather patterns and year-over-year type comparisons for energy usage. So artificial intelligence has been very impactful in the energy sector in our facilities and it is a major part of our smart operations platform or requirements. 

Advice for companies 

Host:  

So energy definitely is one more area where we have very solid use cases and where we can clearly see the benefits. My next question is more about the difference between, how do you bring all these projects into life? All the examples that you have given us, these are examples of projects that are already running, and some of them are running already for a longer time. How would you recommend companies approach the challenge of the fact that you are running a lot of pilots, but then you are maybe struggling to put them into practice? How did you succeed with this? 

Guest:  

We faced that early on. Everyone faces it, and we still face it to some degree where we seem to be in what they call pilot purgatory, where you’re stuck. 

Host: 

Yes, exactly. 

Guest:  

Scaling and pilot success is really a mindset of usefulness. Let me be clear, starting with a problem solves a lot of problems. Starting with a problem that you actually are trying to solve with a technology. When you’re successful, you will find, and we often find and have experienced this ourselves, that scaling becomes automatic. When you solve a problem with a technology, everyone wants the technology. Now, early on in our programs, and there is some merit that can be given to disruptive type innovation, we may have focused too heavily solely on disruptive innovation. And trying technologies for the sake of technologies. We seem to have gained our biggest foothold when we explored, as an example, the vibration analysis for machine learning for vibration analysis and predictive maintenance, because these large pieces of critical equipment did shut our operations down and cause a massive impact to the operations. When we proved that we could solve that problem, we instantly gained engagement from others who wanted that technology. So, it’s very important to start with the problem. And clearly define your success criteria for solving the problem. Think about that endgame before you start a POC or start a technology. Otherwise, it’s very difficult to slip into that purgatory where you’re trying to figure out how the technology fits into your organization. We’ve made that mistake, and I’m not telling you that we don’t continue from time to time to make that mistake. But solve the problem, and you’ll gain the traction. 

Data flows in the smart factory

Host:  

So, one thing is to define the problem. Next, is it the data that is used that is needed to actually succeed, and if you can tell us a bit more about how the data flows in the Smart factory? 

Guest:  

It’s a huge topic and not one that is easy to describe. It’s easy to describe a lot of the problems that you find when you begin trying to train models on data from different systems. We quickly realized that the biggest problem that we face is the contextualization of that data. What the data means in relation to other data from different systems. In Schneider Electric, as part of the smart operations program, we’ve adopted a very rigorous answer to the problem of what’s known as unique namespace, called the plant equipment hierarchy. The plant equipment hierarchy is a very detailed, rigorous naming structure for assets. Such that our assets and the resulting signals that we gain from the assets mean the same thing across our ERPs, across our MES systems, across the software systems that control and help us support production on the factory floors. When we defined that unique naming space and we stick to it, we now find that data flowing into our central or into our regional data warehouse from our ERPs and our MES systems, we find that it’s much easier to consume in terms of model training to infer certain outputs. But it is a problem that we faced, and a lot of companies face the fact that your data doesn’t mean the same thing from different places. It’s very important to adopt a rigorous naming structure, a unique namespace, for that data from those varying systems. 

Future 

Host: 

Yeah, so it sounds like a good starting point for everybody else who is thinking about implementing AI. Start with the problem, then define the data, and make sure that you are speaking the same language across different departments within global supply chain. And my last question is really about what excites you about the future. What technology do you think will be the next big thing for manufacturing processes? 

Guest:  

It changes so fast. Every day something else seems to disrupt. It’s often difficult to answer the question, and a lot of times the answer to the question changes within the next week or within the next month. But I think all eyes are on autonomous supply chain. The concept of agentic AI and self-controlling our processes for customer demand, scheduling, inventory levels, order execution on the production floor, logistics, and even customer satisfaction. That true autonomous supply chain, the agentic technology really seems to bring a level of excitement that maybe that’s now in reach. As an example, we’re taking a lot of the steps towards that autonomous supply chain with the adoption of digitizing the process of order execution on the shop floor, bringing customer demand directly from the ERPs automatically right to a screen on the production floor where the operators can execute the orders automatically and plan the workforce based on who’s going to be in the building for the given day and the volume that’s needed to produce directly to the supervisors so they can see if there is a problem in those workforce assignments to manage our queue levels. Those are the kind of steps that we’re taking towards a truly autonomous supply chain. As I mentioned, that agentic technology, the thoughts that we could automate that entire ecosystem across all the domains, that’s very exciting. And it looks like that’s now in reach. I’m very excited about the possibility of what the generative technologies and also the way that machines can relate to humans. How we’re going to be using humanoid technologies in our manufacturing processes. We spoke about ergonomics at the beginning, and it’s no great secret that it is a high priority to reduce that ergonomic risk. Imagine the day when we can take a humanoid style device and we can completely eliminate that ergonomic risk and focus on other tasks that are not as mundane for a human, and something that requires a higher level of intelligence, and reduce that safety threat. So that type of automation, what AI is enabling in the type of humanoid automation and the agentic piece is really what’s got me excited. I don’t think we really understand what’s available or what the potential is, and the potential will change within the next six months. 

Wrap up

Host: 

Yes, indeed. Thank you so much for this glimpse into the future and for sharing with us all the practical, real-life examples of AI in action. It’s been really a pleasure discussing with you, Mike. 

Guest:  

Gosia, I’ve had a lot of fun. Thank you very much. 

Host: 

Thank you so much. And for all of you who stayed with us, thank you so much for joining, and stay tuned for the next episodes that you can find on your favorite podcast platform or on YouTube. Thank you. 

AI at Scale Schneider Electric podcast series continues! 

The first Schneider Electric podcast dedicated only to artificial intelligence is available on all streaming platforms. The AI at Scale podcast invites AI practitioners and AI experts to share their experiences, insights, and AI success stories. Through casual conversations, the show provides answers to questions such as: How do I implement AI successfully and sustainably? How do I make a real impact with AI? The AI at Scale podcast features real AI solutions and innovations and offers a sneak peek into the future.