Drone & Aerial

Drone Transformer Inspections and AI: Lessons From NGED's 500

Team UAV captured imagery of 500 pole-mounted transformers for NGED's AI corrosion project. What the program shows about drone inspection data.

Ceptory Team8 min read
A drone pointing its coral camera cone at an electricity pylon with the inspection target bracketed

TL;DR

On October 21, 2026, Team UAV said it finished drone imagery of 500 pole-mounted transformers for National Grid Electricity Distribution, to train AI that spots corrosion. The lesson: consistent capture and linked metadata decide whether inspection AI works.

On October 21, 2026, UK drone services company Team UAV announced that it had finished a drone inspection program for National Grid Electricity Distribution (NGED), capturing high-resolution imagery of 500 pole-mounted transformers across the network. The imagery feeds an innovation project that asks a narrow, practical question: can AI and computer vision identify corrosion and other visible deterioration on those assets? This is a useful story for anyone running drone transformer inspection or any aerial inspection program, because the headline is not the AI model. It is the data collection that comes before it.

This post walks through what was announced, what it tells us about building inspection datasets, and what teams that already hold large libraries of drone video and photos can do now.

What exactly did Team UAV and NGED announce?

Team UAV announced a completed capture program of 500 pole-mounted transformers, funded as part of a project under the Network Innovation Allowance (NIA). According to DroneLife's report on the announcement, the project will "examine how drone imagery and computer vision can identify visible signs of deterioration, including corrosion."

The key facts from that report:

  • Scope: 500 pole-mounted transformers, planned from NGED's national asset list, across Cornwall, Wales and the Midlands.
  • Division of work: Team UAV developed and delivered the inspection program. Softwire, a fellow NGED supplier, is developing the computer vision models and a prototype web application that will assess transformer condition from the imagery.
  • Timeline: the companies expect the initial phase of the wider project to be completed by the end of 2026.
  • Who NGED is: it owns and operates the local electricity network across the Midlands, South West England and South Wales, and is separate from National Grid Electricity Transmission.

NGED's innovation engineer Favour Eboh described the aim in the same report: the project "will explore whether aerial imagery, AI, and computer vision can automate aspects of the condition assessment process, generating more consistent condition data to inform asset health reporting and investment decisions."

Note what is not in the announcement: no detection accuracy, no false-alarm rate, no claim that the models already work. The imagery is the input to a research phase. That honesty is worth copying in your own reporting.

Why does capture consistency matter more than the model?

Capture consistency matters because an AI model can only learn the difference between a corroded and a healthy transformer if the pictures differ in that respect and not in camera distance, angle or lighting. Team UAV's technical lead Mark DeGiovanni put it this way: "our focus was on developing a consistent and repeatable approach to drone inspections, ensuring the imagery we collected was suitable for the next stage of AI analysis."

Before inspections started, Team UAV ran technical trials to pick the equipment, camera positions and capture techniques. That is the step most inspection programs skip. The common failure pattern looks like this:

  1. Pilots fly the way they always have, each with their own habits.
  2. Months later someone wants to train or evaluate a model on the archive.
  3. The archive mixes resolutions, standoff distances and angles, so results are noisy and nobody trusts them.

A fixed capture protocol, written down before the first flight, prevents this. For a pole-mounted transformer it would typically specify which faces of the asset must be photographed, from what standoff distance, at what camera tilt, and what to do when vegetation blocks the view. Team UAV's report says that where "access restrictions, vegetation or missing infrastructure" prevented an inspection, the team identified suitable locations and alternative assets, which keeps the dataset balanced rather than silently shrinking.

They built a custom GIS database and delivery process. DroneLife reports that crews uploaded imagery from the field and the system linked each set of images to an individual transformer location together with survey information and metadata.

This is the second quiet lesson. An image of corrosion is only actionable if you know which asset it belongs to, when it was taken and under what conditions. Without that link, the best model in the world produces findings nobody can route to a maintenance crew. Whatever your sector, the minimum metadata worth capturing at the time of flight is:

FieldWhy it matters later
Asset ID and locationRoutes findings to a work order and enables repeat inspection of the same asset
Capture date and timeLets you compare condition over time and discount seasonal effects
Camera and settingsLets you separate real change from equipment change
Pilot or crew and flight IDSupports quality review and retraining on weak flights
Access notes (blocked, partial, alternate asset)Explains gaps instead of hiding them

Is this part of a wider shift toward AI in utility inspection?

Yes. Utilities in the UK are testing drones and AI through the same Ofgem funding route. Iberdrola reports that ScottishPower's distributor SP Electricity North West will start a 12-month pilot called VOLT (Visual Overhead Line Tracking) in Cumbria in November, using drones and AI analytics to speed up fault identification during adverse weather, funded through Ofgem's NIA. In Iberdrola's announcement the stated aims are better fault detection, field crew safety and operational efficiency, and less reliance on inspections on foot or by helicopter.

The two projects differ in shape. The NGED work builds a labelled image dataset of static assets for later condition assessment. VOLT is oriented toward fast response after weather events. Both depend on the same underlying capability: turning a large volume of aerial imagery into answers someone can act on. For a step-by-step review workflow on exported flights, see how to review drone inspection footage faster. Public safety is seeing the same shift, as in the Evolver DFR drone-as-first-responder launch.

What does this mean for teams with drone video, not just photos?

Teams with drone video face the same data problem, with one more difficulty: footage is long, and the interesting moment may be four seconds of a thirty-minute flight. The DroneLife report describes still imagery, so we should be careful not to claim the NGED project involves video. But many inspection programs fly video as well as stills, and the capture-consistency lesson transfers directly.

Where video is involved, three questions become important:

  • Can you find the moment? If a reviewer remembers a corroded fitting on a pole near a particular junction, can they locate it without scrubbing through hours of flights? See our guide on natural language video search for enterprises.
  • Can you describe what is on screen and what is said? Pilots often narrate defects aloud or the camera captures asset tags and signage. Searching speech and on-screen text, not only visuals, recovers that context. Our post on multimodal video analysis for operational review explains the approach.
  • Can you automate the first pass? The point of the review step is to reduce the volume humans must watch, not to remove them. The case for this is laid out in beyond manual video review.

Where does Ceptory fit, and where does it not?

Ceptory is a video intelligence platform: you upload recorded video or add it by URL through the API, index it, and then search it in natural language and run multimodal analysis from the dashboard or REST API. It sells no drones or cameras. The drone monitoring feature page describes the workflow for aerial footage.

The honest limits, stated plainly:

  • Ceptory does not publish detection accuracy for corrosion or any specific defect. We have no figures to quote, and you should not assume a result for transformers without testing it on your own assets.
  • Ceptory is not a GIS or asset-management system. It would sit alongside the kind of GIS database Team UAV built, helping reviewers find and understand footage, not replacing the asset register.
  • Ceptory is not a trained, purpose-built corrosion model of the kind Softwire is developing for NGED. If you need a custom defect classifier, that is a separate modeling project.

What it does offer is a way to make a large drone archive searchable and reviewable without hardware changes. Pricing is public: $0.042 per minute indexed plus $0.0015 per minute infrastructure, search at $4 per 1,000 queries, and the first 30 minutes free, per the pricing page. You can estimate review effort with the free video review cost calculator and, for the API route, see the video intelligence API.

What should an inspection team do now?

Whether or not you plan to use AI this year, treat the NGED program as a template for preparing your data.

  1. Write a capture protocol before the next flight. Fix equipment, camera positions, standoff distance and handling of blocked assets. Trial it on a small set first, as Team UAV did.
  2. Attach asset IDs and metadata at upload, not later. Build or adopt a delivery process where crews link each image set or video to an asset record from the field.
  3. Record gaps explicitly. Log assets you could not inspect and why, and pick alternates so your sample stays representative.
  4. Make your existing archive findable. Index historic footage so reviewers can search by what appears on screen, what was said and any visible text, rather than by file name.
  5. Define what success means before you pilot. NGED's phrasing is modest: more consistent condition data to inform asset health reporting. Choose a measurable target, such as time to locate a past defect, and compare with a manual baseline.
  6. Keep humans in the loop. Use automated passes to prioritize, and have qualified engineers confirm any condition rating that drives investment.

What to watch next

The NGED project's initial phase is expected to complete by the end of 2026. The signals worth watching are whether Softwire and NGED publish any evaluation of how well the models identify corrosion, how consistent the labels were, and whether the prototype web application moves into operational use. Until results appear, treat the announcement as evidence of a well-run data collection exercise, not as proof that AI transformer inspection works at scale.

For teams with a backlog of aerial footage and no way to search it, the practical step is the same as it was last week: get the archive indexed and searchable. You can try Ceptory on your own drone video at Ceptory, with the first 30 minutes free.

Frequently asked questions

What did Team UAV complete for NGED?

Team UAV completed a drone inspection program covering 500 pole-mounted transformers across the NGED network, announced on October 21, 2026. The imagery supports an innovation project on whether AI and computer vision can spot corrosion and other visible deterioration.

Who is building the AI models for the NGED project?

Softwire, another NGED supplier, is developing the computer vision models and a prototype web application. Team UAV designed and delivered the drone inspection program. The companies expect the initial phase to finish by the end of 2026.

Why does consistent capture matter for inspection AI?

Models trained on images taken with different cameras, angles and distances learn those differences instead of the defects. Team UAV ran trials to fix equipment, camera positions and capture techniques so imagery from different sites works for AI analysis.

Can Ceptory detect transformer corrosion?

Ceptory does not publish accuracy figures for any defect type. It indexes drone video you upload or add by URL so you can search it in natural language and analyze scenes. Treat it as a review and retrieval layer, and test it on your own footage.

Is this drone footage or still photography?

The report describes high-resolution imagery, so it appears to be still photographs. The same capture discipline applies to video, which adds time-coded searchable evidence across a whole flight.

Ceptory

Search your own footage in plain language. The free tier indexes 60 minutes of video.