Asset inspection: moving the model to the data

The shape of the problem

An operator with distributed assets — wind turbines, pipeline segments, substations — flies drone inspections and ships the footage to a central team who review it frame by frame.

Getting through the whole estate takes months. By the time an engineer reviews footage of a given asset, the footage is old. Defects that were early-stage when filmed are not early-stage when found, and a repair that would have been a repair becomes a replacement.

Attempts to automate this stall on the same rock: the sites have no meaningful bandwidth. An offshore platform cannot ship raw 4K inspection footage to a central cloud, and the compliance position on where that footage may be processed is usually not flexible.

Why the obvious fix does not work

The bottleneck gets framed as review capacity, and the plan is to hire more engineers.

It is a transport problem. The review work that matters — deciding what to do about a defect — is a small fraction of what the engineers spend their time on. The rest is looking at footage of assets that are fine.

Hiring scales the wrong half.

The design

Inference at the edge, on the site's own hardware

A quantised vision model runs on a single accelerator in the site's existing equipment cabin, processing footage where it lands. Frames with no candidate defect never leave the site — they are scored, logged and discarded on the retention schedule.

Only flagged frames go back to the centre, which is a volume the available link handles comfortably, and which satisfies the compliance position because the raw estate footage never moves.

A governed defect corpus at the centre

Every confirmed defect, every engineer disposition and every eventual maintenance outcome is written back into a governed corpus with lineage to the frame, the asset, the position and the flight. That corpus is what the next model version trains on.

This is the loop that makes the programme compound. The first model trains on public data and whatever historical annotations exist. Later versions train on the operator's own confirmed dispositions, and get materially better at the failure modes specific to its equipment and its weather.

Engineers on judgement, not triage

Engineers see a queue of candidate defects with the frame, the asset's history, the model's confidence and comparable prior dispositions already assembled. They decide. They do not search.

What to measure afterwards

• Full-estate inspection cycle time, against the pre-deployment baseline.
• Proportion of defects caught at an early stage — the category that separates a repair from a replacement.
• Unplanned downtime, year over year.
• Share of captured data that leaves the site, as a compliance measure rather than a performance one.
• False negative rate on a human-reviewed sample of discarded frames. This is the one teams forget, and it is the one that matters.

Where it runs

Edge inference on the operator's own hardware at each site. The governed corpus and the training estate sit in the operator's private region. The model weights are the operator's — which matters more than usual here, because the corpus those weights encode is an operating history nobody else has.