Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Powering Real-Time Drone Inspection of Power Lines
Transcript
- Lucas: So there are about 5.5 million miles of transmission and distribution lines in the U.S. alone. Most of them get inspected maybe once a year — if that. And the standard method hasn't changed much: a crew in a helicopter flies the line, someone with binoculars looks for problems, and they write it down. Luna: That is still how it's done? I assumed we'd moved past helicopter and binoculars by now. Lucas: You'd think, right? But utilities are risk-averse, and the FAA has historically been very restrictive about flying drones beyond the pilot's line of sight. So most drone inspections were limited to short segments, like a substation or a single tower. That's changing fast. Luna: Because of the edge computing piece, I assume. Lucas: Exactly. The bottleneck wasn't just regulation — it was data. A drone flying a 50-mile transmission corridor can capture terabytes of thermal and visual imagery in a single flight. If you try to stream that to the cloud or even to a ground station, you run out of bandwidth before you've covered the first mile. Luna: So you process it on the drone itself. Lucas: That's the play. And the hardware has finally gotten small and power-efficient enough to make it practical. The most common setup I've seen uses an NVIDIA Jetson module — the Orin NX, usually — mounted on a larger drone like a DJI Matrice or a Freefly Astro. The drone carries a high-res thermal camera and a 20-megapixel visual camera, and the Jetson runs a custom convolutional neural network trained to detect specific anomalies. Luna: What kind of anomalies? Like a hot splice on a conductor? Lucas: That's a big one. A loose connection or corrosion creates a hot spot, and thermal cameras pick it up. The model can also spot vegetation that's grown too close to the line, broken insulators, bird nests on crossarms, even early signs of tower corrosion. The drone processes each frame in real time, and if it scores above a certain confidence threshold, it tags that GPS coordinate and sends a tiny alert — just a few kilobytes — back to the ground. Luna: So the drone doesn't need to stay connected. It can fly into a dead zone, do its analysis, and only transmit when it has something important. Lucas: Right. And that changes the economics dramatically. Duke Energy ran a pilot program in North Carolina last year that covered about 300 miles of transmission lines. They compared it to their usual helicopter inspections. The drone with on-board inference took about 40 hours of flight time spread over two weeks. The helicopter would have taken a similar amount of flight time but cost roughly 40 percent more per mile, plus you have to factor in crew logistics. Luna: Forty percent cost reduction is significant for a utility. But what about accuracy? Does the model miss things a human would catch? Lucas: That's the key question. In Duke's data, the model's detection rate for hot spots was about 92 percent — meaning it flagged 92 percent of the anomalies that a follow-up ground crew confirmed. Human spotters in helicopters typically hit around 85 percent, and that's with fatigue setting in after a few hours. The drone doesn't get tired. Luna: So the edge model actually outperforms the human baseline. That's not something you hear every day in industrial AI. Lucas: It helps that the detection task is very narrow. The neural network sees one thing: temperature differentials and geometric patterns. It's not trying to understand context. But there's a flip side — the false positive rate was around 8 percent, meaning the drone flagged things that turned out to be nothing. That's manageable because a ground crew can check those locations quickly, but it's not zero. Luna: Still, 8 percent false positives seems acceptable when you're talking about preventing a wildfire or a blackout. Lucas: Absolutely. And the model improves over time. Duke's team is using a continuous learning pipeline — every time a ground crew verifies a false positive, that image gets fed back into the training set. So the next iteration of the model should have a lower false positive rate. Luna: What about the regulatory side? You mentioned the FAA being restrictive. Has that changed? Lucas: It has. In late 2025, the FAA expanded its Beyond Visual Line of Sight — BVLOS — waiver program. Previously, you needed a specific waiver for each flight, and it was hard to get. Now, under the new rule, utilities can apply for a blanket BVLOS waiver for defined corridors, as long as the drone has detect and avoid capability and a remote pilot who can take over if needed. That's a game-changer. Luna: So now a utility can map out a hundred-mile corridor, get a single waiver, and run automated inspections on a regular cadence. Lucas: Exactly. And a few startups are building turnkey services around this. One called Neara — based in Australia but expanding into the U.S. — combines drone imagery with digital twin software. They fly a line once, build a 3D model of the entire corridor, and then use edge-based analysis on subsequent flights to detect changes over time. So you're not just looking for hot spots; you're tracking slow-moving problems like tower tilt or vegetation growth rates. Luna: That's a different class of problem. It's not just 'find the fire risk now' — it's 'predict when that tree branch will be within two feet of the line.' Lucas: Right. And that kind of predictive maintenance is where the real ROI shows up. Instead of sending a crew out to trim vegetation every three years on a fixed schedule, you can trim exactly the branches that need trimming, exactly when they need it. One utility in Texas estimated they could cut vegetation management costs by 60 percent using this approach. Luna: Sixty percent is enormous. Vegetation management is one of the biggest line items for transmission owners. Lucas: It is. And it's also the leading cause of power outages during storms. If you can predict which branches are likely to fail, you can prioritize them before the storm hits. Luna: So the edge computing piece here isn't just about speed — it's about enabling a fundamentally different operational model. You shift from reactive to predictive, from schedule-based to condition-based. Lucas: Exactly. And that's the pattern we see across industrial edge computing applications. The hardware — the Jetson modules, the thermal cameras, the drone platforms — is becoming a commodity. The differentiation is in the models and the operational workflow. Luna: One thing I'm curious about: does this scale to smaller utilities? Duke Energy has the budget and the engineering team to run a pilot like that. What about a rural electric cooperative with a few thousand miles of line and a two-person IT department? Lucas: That's the next frontier. Right now, the service model is the most viable path. Companies like Skydio and Airobotics are offering inspection as a service: they bring the drone, the edge processor, the trained model, and the reporting platform. The utility just provides access to the corridor and pays per mile inspected. That brings the cost down dramatically for smaller players. Luna: And the regulatory side — the BVLOS waiver — would be handled by the service provider. Lucas: Yes. A few of these providers already have blanket waivers that cover multiple states. So a cooperative in Kansas can essentially plug into an existing infrastructure without having to navigate the FAA themselves. Luna: That's a clever way to democratize the technology. Otherwise you'd see a widening gap between large investor-owned utilities and small co-ops. Lucas: Right. And the co-ops serve some of the most rural, hardest to inspect lines. They might benefit the most from drone-based inspection, but they have the least capacity to implement it internally. Luna: So the edge computing trend here is really enabling a services layer that wraps hardware and models into something a utility can just buy. Lucas: Exactly. And we're starting to see the same pattern in other infrastructure sectors — pipeline monitoring, cell tower inspection, even railway track inspection. The playbook is similar: drones plus on-board inference plus a service provider holding the waivers. Luna: Before we wrap, I want to circle back to something you said about the models improving over time. That continuous learning pipeline — is that happening on the edge, or is it back in the cloud? Lucas: The inference happens on the edge, but the training loop runs in the cloud. The drone captures the flagged images and sends those few kilobytes to a central server. Then a human reviews them, labels any false positives, and the model gets retrained. That updated model gets pushed back to the drone fleet during the next dock visit. Luna: So it's a hybrid architecture — edge for latency-critical inference, cloud for the training cycle. Lucas: Exactly. And that's a pretty good template for a lot of industrial edge AI applications. The edge handles the time-sensitive decision — 'is this a hot spot?' — and the cloud handles the slower, compute-heavy improvement loop. Luna: It reminds me of how the same architecture works in autonomous vehicles, just at a different scale. Lucas: Yeah, the parallel is strong. And like autonomous vehicles, the biggest challenge isn't the technology — it's proving reliability over millions of hours of operation. Utilities are conservative for a reason: if an inspection misses a fault and a line catches fire, the liability is enormous. Luna: So you need enough real-world data to convince regulators and insurers that the edge model is at least as safe as a human spotter. Lucas: Right. And that's where the Duke pilot is so valuable. They have thousands of miles' worth of comparative data now — drone vs. helicopter on the same lines. That data set is being shared with other utilities and with the FAA, and it's building the case for wider adoption. Luna: It's interesting that the slowest-moving part of the whole system is the regulatory and insurance side, not the silicon. Lucas: That's almost always the case with infrastructure technology. The hardware moves at Moore's Law, but the adoption moves at the speed of trust. Every new technology in the utility space goes through a decade-long proving period. We're maybe three or four years in with edge drone inspection. Luna: So the next five years should be interesting. I could see this becoming the default method for transmission line inspection by 2030. Lucas: I think that's a reasonable prediction. The cost curve is heading in the right direction, the regulatory barriers are falling, and the data on detection accuracy keeps getting better. It's one of those rare cases where the technology, the business model, and the regulation are all aligning at the same time. Luna: And the edge computing piece is really the enabler that makes it all work. Without that on-board inference, you're back to the bandwidth problem. Lucas: Exactly. The drone becomes a flying camera with no ability to decide what matters. And then you're still relying on humans to review hours of footage. The edge is what turns the drone from a data collector into a decision-maker. Luna: If this conversation gave you a new angle on how edge computing is evolving in practice, and you'd like to support the work that goes into researching and producing episodes like this, you can do so on a platform called Buy Me a Coffee. It's a simple way for listeners to contribute a few dollars to help us keep the show independent and ad-free. Lucas: And it genuinely makes a difference. The URL is buy me a coffee dot com slash fexingo — all one word, lowercase. If you've gotten value from the episodes, that's the place to let us know. No pressure, but it helps us spend more time on deep dives like this one. Luna: So back to the technology: one thing I'm watching is whether the same approach will extend to distribution lines — the lower-voltage lines that run through neighborhoods. They're harder to inspect because they're often under tree canopy, but some startups are working on smaller drones that can navigate that environment. Lucas: That's the next logical step. Transmission lines are the low-hanging fruit because they're in open corridors. Distribution is much more complex — you've got poles every hundred feet, transformers, streetlights, and a lot of occlusion. But the same edge computing principles apply: if you can process imagery on a small drone with a lightweight model, you can do the same kind of anomaly detection at the neighborhood level. Luna: And the potential impact is even bigger, because most outages happen on the distribution side. Lucas: Exactly. A transmission line failure might affect tens of thousands of customers, but a distribution failure affects a few hundred or a few thousand — and it happens much more frequently. If you can automate the inspection of distribution infrastructure, you could dramatically reduce outage times. Luna: So the edge computing story here is really about shifting from a model where you inspect everything on a fixed schedule to one where you inspect based on risk and condition. Lucas: And that's the deeper shift. Edge AI allows you to move from 'inspect everything equally' to 'inspect the things that need attention right now.' That's a more efficient allocation of resources, and it's only possible when you have real-time decision-making at the point of data collection. Luna: It's a good example of how edge computing isn't just a technical upgrade — it changes the operational logic of an entire industry. Lucas: And that's why it's worth paying attention to, even if you don't work in utilities. The same pattern is going to show up in water infrastructure, gas pipelines, cell towers, and rail. The playbook is being written right now, and it's being written on the edge. Luna: That's a good note to end on. Thanks for the conversation, Lucas. Lucas: Thanks, Luna. Talk to you next time.