Enhancing Grid Reliability: How Buzz Solutions Uses Vision AI to Prevent Outages and Wildfires - Ep. 249

0:10Hello, and welcome to the NVIDIA AI Podcast. I'm your host, Noah Kravitz. Buzz Solutions is on a mission to enhance the safety and efficiency of the electric grid through innovative data analytics and machine learning. A member of NVIDIA's inception program for startups, Buzz's visual intelligence empowers utility companies to better monitor and manage their infrastructure. The result is improved reliability and reduced outages. Here to explain how Buzz does it and to talk about the impact AI can have on making our electric grids safer and more robust is Caitlin Albertoli, Bertolli, CEO and co-founder of BuzzSolutions.

0:46Caitlin, welcome, and thanks so much for joining the NVIDIA AI podcast. Thanks so much for having me, Noah. I'm excited to be here. So we're two Californians, and I mean, I think it's everywhere now, but certainly in our state, the electric grid has been a source of news for quite some time now. It's something that's on lots of people's minds. So I'm very excited to learn more about what Buzz is doing and the impact that AI can have going forward to really help all of us because we all use electricity, right? So maybe we can get into it with a little bit about your background and how you came to co-found BuzzSolutions.

1:19Absolutely. Thanks for the question. Happy to share. So we launched BuzzSolutions from a launchpad course at Stanford University in the spring of 2017. My background is actually not from the electric and power industry. Prior to Buzz, I was working in finance and I also ran a non-profit in the sustainable food world. Before that, I grew up in Southern California in a town that was close to the nuclear power plant, the San Onofre Power Plant. So that was really my first exposure to the power world, the power industry. I mean, it was actually shut down during the time that I was living in Southern California.

1:55So that was an interesting dynamic to grow up with and see. Yeah, did that leave an impression on you? I mean, did you think more about where power comes from than perhaps, you know, the average kid? Definitely. It definitely got me thinking more about our different generation sources and renewable energy more broadly. And so when I went into Stanford, I was certainly looking for more opportunities to learn about generation and power as a whole. That's great. So there's a lot of steps on the way, as you mentioned, with finance and sustainable food. But how did you wind up co-founding Buzz? How did Buzz come to be?

2:28Sure. So my co-founder, Vic, and I met in this Launchpad course that was in the School of Civil and Environmental Engineering. It was about building a startup in the School of Civil and Environmental Engineering focused on energy and sustainability. And there was something that called me to this class because I was always really entrepreneurial and interested in learning more about starting a business. But I guess the intersection of energy and sustainability, given my background, with entrepreneurship seems like a really unique opportunity. And so that was the course that we ultimately met. And when we were originally exploring ideas for what we wanted to, I should say, not really found a company at that time, but what we wanted to use for this class, we were looking at the wind turbine market.

3:15We were looking at all that was happening with optimization of placement of wind turbines and how you could use drones and technology to understand where it was best to place those wind turbines. But in those conversations with some of those, you know, those wind turbine companies and manufacturers, we very quickly were tipped off to the power space. You know, everyone said this is a huge market and there's definitely an opportunity in wind. But have you seen what's happening with inspections in the power space? in kind of an interesting time, because if you think about 2017, that was before a lot of the big wildfires, before a lot of the power outages, and before some of the major blackouts that, of course, we know about today and what has unfortunately transpired since that time.

3:58But we ended up taking that conversation and leveraged the Stanford Alumni Network to interview 35 investor-owned utilities. We were able to get conversations with 35 investor-owned utilities. 35, that's amazing, wow. Can you imagine that many? And they all had the same story. They were collecting a lot more data or they had plans to collect a lot more inspection data, whether it was due to improved technology that allowed them to do more inspections and also regulations that were driving more inspections. But they didn't have a way to manage or analyze that data. And so that was really kind of the genesis of ultimately how we launched Buzz.

4:33Yeah. So, Caitlin, it's funny because I'm thinking about that kind of pre-pandemic time and going to NVIDIA GTC back, you know, probably around 2018, 2019, maybe 2017, and watching a demo of drones and kind of a, you know, is it a demo in the convention space? but a drone in a simulated warehouse environment flying around doing inspections and doing what's it called pick and pack and that kind of thing. So what you're talking about rings a bell for me, if in a different space. But how did you, you know, when you're talking about there's all this data, what do we do with it? And to me, that says AI, but this is, you know, seven, eight years later.

5:08So back then, did you see instantly, did you see all the data and think, oh, machine learning, AI, we can do something with this? Or kind of how did that leap to AI happen? It's a really good question because at that time for a lot of utilities, it was still just an idea or it was still just a plan that they were looking to put in place. You know, they were buying their first drones. They were training their first pilots. And on top of that, they were still doing and present day still are doing a ton with helicopters. So they were starting to actually collect images with helicopters at that time while also building drone programs.

5:41And so even in our first days of Buzz, a lot of our conversations with utilities revolved around, okay, we're starting to use drones. We're starting to collect data with drones. How do we scale this program? How do we collect data in the best way that's going to make AI the most successful, I guess you would say? We still have a lot of those conversations today, but data wasn't standardized in that space and it still isn't standardized. But if you can think in 2017 that a more traditional industry like utilities were they had five to seven year plans to scale these massive drone programs. I mean, it's it's lightning speed for some of these these more traditional industries like utilities.

6:20So it was more an idea at that time, but they had all of the right, I guess you would say, pieces to the puzzle to be able to build some of these programs. And so what's happened along the way? And I want to get into, you know, some of the use cases, some of the challenges and problems that you've solved and are helping energy companies solve. But I don't want to skip over the growth in between. Walk us through that. And if there are some of those, you know, use cases and things that kind of popped up and are still prevalent, definitely dig into those if you would. Sure. It was an interesting time back in the 2017-2018 timeframe.

6:54We're talking about these concepts like artificial intelligence, AI, and a lot of our user base at that time, they weren't really actively using software platforms. A lot of them were still manually drafting inspection reports. So when you're talking about going from manually drafting inspection reports to using AI... And these are the utility companies you're talking about. Correct. Yeah. These are the utility companies. So you're thinking of linemen. Digital transformation. Yeah. Exactly. Linemen, field technicians, engineers, a lot of them were still manually drafting inspection reports. And some of our linemen and field technicians that were using our solution, we're having to go through thousands of images, tens of thousands, hundreds of thousands of images manually before they're looking to use a tool like this.

7:36And so the reason I give you that background is in order to show up to a utility and ultimately bring value with a solution like ours, we had to plug into the utility workflow. And so from day one, when we were working with utilities to build this solution, to train our algorithms, to build our platform, we were really utility-centric. We wanted to make sure that the algorithms that we were training were very closely aligned with how a utility subject matter expert would have labeled this data manually. Is that or was that back then sort of standardized or was it that, you know, from company to company, even within a company, from sort of inspector to inspector, the format of those reports was just sort of freeform, I guess?

8:22I would say there's definitely a lot of subjectivity as it pertains to how linemen or field technicians would label the data and how they would classify a significant, as we call them, like anomaly or I should say a physical defect on a component is how someone could think of it more commonly. But it's if you can see it, the defect on the component or the anomaly, that could be a little subjective of how someone classifies it. But we see that there's a pretty standard set of labels or total classifications that a utility is looking for. And we worked with research organizations like EPRI, which is the Electric Power Research Institute, one of the leading electric research institutes, along with utilities to make sure that we got that nomenclature correct.

9:05And when you're talking about defects, anomalies, what kinds of things are, you know, are you looking for? Are they looking for? I'm not knowing much about it all. I'm just imagining, you know, there's a squirrel chewing through a line or there's a giant explosion. What kinds of things are being, you know, captured as anomalies. Actually, squirrels do have a big impact. So those are the ones I always hear stories about, you know, like, oh, what was the outage last night? Oh, is this squirrel at the substation again? Or, but is that really like they're that common one? Okay. It is. Actually, in substations in particular, a substation is where a utility changes the voltage.

9:41So it goes from highly energized lines to lower energized lines or vice versa. So the substation is the facility where that change takes place. A substation, a lot of our questions that we get from utilities are, hey, can you do things like birds, squirrels, snakes? Snakes is a huge one. Coyotes, mountain lions. So we actually get a lot of animal requests as well because animals, if you think about it, they're sitting on the infrastructure or they're chewing through something. It could have a significant impact on the equipment itself. and a person may not be out there to be able to see that, but they may start to get a sensor that's tripping and they may say, oh, there's something wrong here, but they may not know what caused it.

10:25And so having the ability to leverage visual data to say, oh, okay, there was a squirrel that came in contact with this or a snake can be really helpful for that utility to make a better decision on maintenance then. Right. Just to get a sense of the scale that we're talking about that you're dealing with, the Buzz is dealing with, what's the best way to, Like how much data, how many square miles, how many, you know, data points or images? How do you, what's the best way to express, you know, the scale of what these companies are doing and what Buzz is doing to help them? Sure. Some of our utility customers are collecting hundreds of thousands and millions of images a year.

11:00And that's just as a part of their standard inspection processes. Utilities have various different types of inspections. They could be doing what's called a flyby inspection, which is just looking for anything major glaring that's happening along the infrastructure. Or they could be getting down to a comprehensive inspection where they have a drone that's flying and taking multiple pictures to look for things like an insulator disc that's broken or a crack in a cross arm or a woodpecker hole in a cross arm, or even something as small as a cotter pin that could be missing or loose, you know, a bolt or something like that that's holding the insulators to the line.

11:36So there's all these different types of inspections that take place throughout the year. And a large utility may be collecting millions of images of their infrastructure that they're having to sift through on an annual basis. We see that utilities are doing data collection with helicopters, drones, fixed-wing aircrafts, ground-based data collection. Some are using satellites. So there's a variety of different ways that they're collecting this data. And typically, it'll serve a specific purpose for that utility. Got it. Okay. And then, so how does Buzz work? You have a couple of different products listed on your website, I would imagine.

12:11Well, I don't know if that's the full scope of it or not, but maybe walk us through some of the solutions that your customers, the utility companies are employing and kind of how it works from imagining data collection and analysis and action, but that may be entirely wrong. So how does it work? Sure. Great question. So at Buzz, we are here to modernize energy infrastructure with actionable intelligence. The central focus of what we do here is to help utilities take mass amounts of raw data and use AI, machine learning, computer vision, to turn it into actionable information or non-actionable information if it's not significant enough for utility to actually take action on it right away.

12:49And what that means for us is we're taking in all these different visual data sources. So as you'll see on our website, we have two different products that are listed. Our first one is more inspection-oriented. It's called Power AI. And so that's where we're taking in all of this inspection data that I've been talking about from the drones or helicopters, ground-based. It's really linear infrastructure data that we help the utility by managing it, meaning we'll ingest all that data. They can upload it to our platform. We'll help them map it so that they know exactly where all the images are on their corridor.

13:23And And then we'll analyze it with our machine learning computer vision algorithms. What was really important for us as we were training our algorithms and before we ultimately deployed was that we could show up to a utility with pre-trained algorithms, meaning on day one, we could start analyzing their data and start delivering immediate value. So our Power AI platform has pre-trained algorithms. We'll analyze all that data. And for us, we can analyze an image in a fraction of a second. So it's really quick, really fast. And then once those images are analyzed, we display them on a dashboard for a utility to be able to go through and they can actually edit, adjust labels, add, delete labels as they see fit.

14:05And then all of that gets pushed into an inspection report. We can deploy that in their workflow a variety of ways, whether it's directly into a work management system or to a geospatial information system, which is a GIS system for mapping and managing their inventory. We can deploy it in a variety of different ways for their workflow. So that's PowerAI. And then our second product is called PowerGuard. And that's the one I mentioned kind of as we do animal detections and all that type of thing. Right. But PowerGuard is a real-time or a near real-time alerting system that deploys on a fixed camera.

14:40And so for PowerGuard, we're doing things, same component monitoring that we're doing with PowerAI, but here we're doing it as an alerting system in a substation or in a facility where a utility needs more of those kind of eyes and continuous monitoring. So we call PowerGuard our continuous monitoring solution for security at a substation, whether that's an intruder person, car, animal, squirrel, exactly. Whether it's safety. So if someone's injured on the job or they're injured in their work, and then we do all sorts of component monitoring, whether that's overheating of components or physical degradation on those components that can be seen.

15:19And we send those alerts 24 seven to the relevant teams at the utility. So that's kind of our two different products. It's all vision based as well. Right. Of course. So in a, I don't know, like an emergency situation, a wildfire, a storm or something else that's causing, you know, whether it's a sudden unexpected blackout or if a utility company decides they need to roll out, you know, brownout blackouts. I can understand how PowerGuard obviously would alert, you know, there's something's gone wrong at this substation. But how do Buzz's solutions play into helping utility companies deal with, you know, these kind of real time emergency situations that are coming up?

15:58So for things like PowerGuard, we also do smoke detections. So we can detect smoke or overheating or sparking issues. Is that all visual? Yeah, so we're able to alert to any of those incidences that are happening, whether it's within the substation or at the utility facility or even in the surrounding area. So we can alert to any potential fire risk that's happening there. On the PowerAI side for inspections, we're identifying critical risk issues that could cause a wildfire before they start. It's preventative maintenance, basically. So we're identifying, for example, rusted components that could ultimately rust out and cause a line to drop down.

16:37We are able to do those types of detections on a transmission tower, a distribution pole. So a high voltage tower or a low voltage city line, we're able to do those detections. And then we also can do sparking detections as well. So if there's any sort of a physical component that could cause a sparking issue. We're able to detect that on the physical components. And then we do things like vegetation encroachment detections. So we can detect where vegetation is encroaching on a line. And where it's particularly valuable for utility is we can say, hey, there's vegetation encroaching, let's say on this distribution or a lower voltage city line.

17:14There's a vegetation encroaching on this distribution pole. And you also have five different anomalies or defects that are on this particular pole. This is a high-risk pole that you should definitely be maintaining. So we do those types of things to help with wildfire risk mitigation. And then on the storm side of things, utilities are starting to deploy more vision-based inspections like drone technologies post-storm to go out and fly corridors so they can understand what sorts of anomalies exist on the infrastructure before they re-energize that line. And by using Buzz, they're able to analyze and process through that data really quickly.

17:54So they understand exactly what they need to replace prior to re-energizing. So we're able to save mass amounts of time and help direct the field crews where to go to upgrade and restore that infrastructure post-storm. That's fantastic. I'm speaking with Caitlin Albertoli. Caitlin is CEO and co-founder of Buzz Solutions. And we've been talking about Caitlin's journey to starting Buzz. But Buzz's journey in their work using machine learning, using computer vision, other AI technologies to help electric grids, help utility companies deliver power safely, more effectively, do preventative maintenance, post-storm analysis and repairs, all these things that Caitlin's been talking about.

18:33And obviously, as you mentioned before, data kind of at the key of everything AI related. Let's talk about that a little bit. Maybe, Caitlin, you can talk about how you and your teams are training and deploying your AI models. You mentioned before how important it is to show up at a potential client's office with pre-trained algorithms kind of ready to go. So maybe talk a little bit more about your whole approach to using models. Definitely. We recognized that on day one, it was really important for us to show up with pre-trained algorithms that were tuned and trained specifically for utility infrastructure.

19:09We built our algorithms from scratch. We spent two years training our algorithms and collecting all of this data, highly proprietary data from utilities and worked with utilities to understand which anomalies were highest priority for them so that we could show up on day one with these algorithms that were already trained. And so we have dozens of different types of pre-trained detections today. We have data sets dating back over a decade, working with dozens and dozens of utilities across different geographies, both here in the U.S. and internationally, and also data sets that were collected from a variety of different sources.

19:46As we talked about earlier, data is not standardized in the way it's collected in this space. And so the ability for us to train and tune our algorithms with high accuracy meant we had to work with a variety of different types of data sets. And so it took us a while to do so. But one thing I wanted to touch on is Buzz has also started to do a lot more with training our algorithms using synthetic data to be able to deploy a new algorithm with an anomaly that may not occur with high frequency. So if you think about some utility issues, they don't occur with high frequency. For example, smoking in a substation.

20:20You know, we obviously can't, you know, have a substation fire, light a substation on fire to train an algorithm. Right, right, right, okay. But we still had to train the algorithm to be able to deploy at a substation and say, hey, we can detect smoke if it occurs. And so we were able to use synthetic data to be able to train the AI for smoke in substations. We did the same thing for a lot of our animal detections as well, because capturing enough varied data of animal detections is more challenging. And I want to be really clear here too. The success of our solution is on a, I guess you'd say, the granularity of number of detections that we can provide was highly because of the amount of real data that we had.

20:59In order to be able to leverage synthetic data and use synthetic data successfully, we had to have a lot of real data in order to be able to use synthetic data more effectively. But there are certain instances where synthetic data can be quite valuable And those were just a couple examples of how we've been able to use that to deploy some of those newer algorithms that we have today for our specifically our power guard solution. Right. In working across the United States, but then also globally, are there great differences in the way that electricity is distributed in different parts of the world and the way that grids are set up?

21:34And I mean, my knowledge is very limited, but I know a little bit about burying power lines underground as opposed to putting them above ground. And so, you know, I'm thinking about it from the data standpoint, but then also just curiosity, like, is the physical infrastructure generally the same or does it vary greatly from place to place? We see it varies greatly from place to place, mostly because of when the infrastructure was built, you know, whether it's transmission, distribution. So what's the difference? Transmission is high voltage, high voltage infrastructure. You see those big steel towers more often than not that are traveling far distances.

22:09That's transmission. And then distribution are the lower voltage lines, like city lines. Most of the time you'll see them are those wooden poles that are throughout cities or surrounding areas. And so transmissions, much higher voltage, distribution is typically lower voltage. And so depending on locationally where you are, we'll see differences in how those towers look. For example, internationally, we've seen a lot more concrete structures than we have here in the U.S., whereas in the U.S., we see a lot more steel and wood. So the way that the structures look can be different internationally versus domestically.

22:44And then, you know, the components deployed on those structures can also look a little bit different depending on the region where you're located. Some infrastructure performs better in different terrain than others. And so that also factors into it. Kind of going back to the computing side of things, are there particular, and I don't know if it's components or types of disruptions that are harder to detect using, you know, computer vision and AI and everything that just have given you problems for whatever reasons? Definitely. So maybe I'll give three different examples. The first, more granular components, like a small cotter pin that's missing or loose.

23:20I mean, it's very hard for an algorithm to tell what's missing if you don't have a baseline data set of how it was deployed previously. So we find that loose or missing can be a little bit more difficult. So, you know, more granular components can be hard. We also find that, for example, for something like insulator detections, something as much as the time of day when the image was collected can play into it because shadows can have an impact. And so when we're starting to look at different types of damages on an insulator, the time of day that the image was collected and the shadows on that piece of equipment can definitely have an impact on the algorithms.

23:58So something like that can be harder. So that's the first thing is granular components or lighting time of day. The second is how data was collected. So depending on distance, angle, resolution, again, lighting, that all can play into the success of how an algorithm is trained as well. And then the third, the third piece that I will get into is going back to that anomalies that don't exist with high frequency. So whether it's a type of a squirrel, a type of a snake, or whether it's, you know, a rusting of a component in a territory where rust may not be prevalent, let's say. That can be harder for us to train and tune the algorithms with high degree of accuracy.

24:38And that's why we worked with the third piece or the third reason specifically is why we worked with so many utilities from varied regions with tons of different data sets. We were able to then get more access to anomalies that don't occur with high frequency. And maybe one more piece to add to And the third talk point here is utilities sometimes have tried to build in-house programs where they've tried to train and build algorithms entirely in-house. And they often come across the same challenge that, you know, there's not certain anomalies don't occur with high frequency. And so how do you train algorithms without a lot of data of that anomaly being present?