Every turbine · every pass · from orbit
The blade failure rate nobody publishes.
We watch every wind turbine on Earth from satellite, and count the blades that go missing. Confirmed events over turbine-years, sliced by platform, vintage and climate.
3,800
Blade failures per year — the number everyone still cites
2015
The year that estimate was made, from one underwriter's claims book
0
Credible global detachment rates with a proper denominator
Fleet status
Every installed turbine we can find. One dot each.
466,683 turbines from OpenStreetMap, the most complete public register of individual machines. As satellite passes accumulate, each dot turns green for operating, orange for stopped, and red for a broken or missing blade or a crane on site. Switch to the satellite basemap and zoom in to see the machines themselves.
First monitored farm: He Dreiht availability →466,683 turbines · positions © OpenStreetMap contributors (ODbL), fetched 2026-09-06 · status classification begins with the first satellite passes
The problem
Everyone quotes an eleven-year-old number.
The figure the industry still repeats — around 3,800 blade failures a year, roughly 0.54% of some 700,000 blades then in service — is a 2015 estimate from a single underwriter's claims book plus press scraping. It lumps leading-edge erosion claims together with detachments.
Since then it has been laundered through advocacy databases, which is where most public incident lists actually originate. No one has counted the population, and no one has a denominator.
Insurers price on it. Lenders stress-test on it. OEMs are compared against it. It deserves a real measurement.
Why this is measurable
A missing blade is a much easier target than a stopped rotor.
Persistent
It stays gone until repair
A detachment is a geometric anomaly, not a transient state. Revisit interval stops being the binding constraint: a handful of clear looks is enough, not a lucky pass.
Unmistakable
Two shadow arms, not three
A damaged turbine is parked. At low sun angles a parked rotor casts three distinct shadow arms. Two arms is a signature you can read even when the blade itself is sub-pixel.
Cheap
Free imagery does most of the work
Sentinel-1 and Sentinel-2 pass every few days at no cost. That already oversamples an event that lasts weeks to months. Paid imagery is spent only on confirmation.
Blade presence is the core layer and the only one we rate on. Running-versus-stopped is shown as a second, lower-confidence layer: a parked rotor at pass time can be a fault, maintenance, curtailment or simply low wind, so it is a prompt to look closer, not an event.
The method
Nobody looks at images. We difference them.
Every pass, a small chip is cut around each turbine and reduced to two or three numbers. That gives a time series per turbine. A blade loss is a step change in that series.
01
Register
Ingest every turbine on Earth with coordinates, OEM, model, hub height and commissioning date. Dedupe. The map already exists; we assemble it.
02
Chip
Each satellite pass, cut a small window around every turbine. SAR and optical alike.
03
Reduce
Compute backscatter and azimuth-smear extent, optical band-offset fringing, shadow-arm count when sun angle allows.
04
Detect
Run step detection on each per-turbine series. Thresholds tuned across turbine sizes and terrain.
05
Confirm
Spend money only on flagged turbines: an archive pull or a tasked sub-metre scene. Cost scales with events, not fleet size.
06
Classify
Order the blade disappearance against lifting equipment. Crane first: planned exchange. Blade first, long lag: failure.
07
Log
Record first-absent and first-restored dates per event. The gap is downtime — the number nobody publishes.
08
Rate
Confirmed events over turbine-years, with confidence intervals, by platform, blade type, vintage and climate zone.
The signals
Four physical handles on a rotor from 700 km up.
Sentinel-2 · optical · free
Band-time offset
The MSI detectors capture the colour bands a few tens of milliseconds apart. A rotating blade sits in a different place in each band and shows colour fringing; a parked rotor produces clean, registered edges. Modelled maximum offset is around 45 m — roughly 4.5 pixels at 10 m, well above the noise floor.
Revisit 2–5 days · cloud-limited
Any high-res optical
Blade shadows
At low sun angles the shadow is far longer than the tower footprint. Rotating blades smear it; parked blades cast distinct arms. Arm count gives blade count, and arm geometry gives rotor azimuth for parked units.
Best at high latitude, winter sun
Sentinel-1 · SAR · free
Micro-Doppler smear
Rotating blades violate the stationary-target assumption in azimuth compression, so they defocus and throw ghost returns along the azimuth direction. Backscatter and smear extent change when a blade goes. Sees through cloud and dark.
Revisit 6–12 days · all-weather
Tasked · paid
Sub-metre confirmation
For flagged turbines only: an archive pull or a tasked sub-metre scene, or a 30–60 second satellite video clip. Unambiguous, and only bought when the free layers have already raised a flag.
Cost scales with events
Coverage
Offshore is a census. Onshore is a panel.
Offshore · complete population
Every offshore turbine, all the time.
A few thousand machines, all at known coordinates, all large, all against uniform water. Sentinel-1 alone gets most of the way — free, all-weather, no cloud problem. This is the v1 product and it is doable end to end.
- Full population, no sampling error
- SAR-first, so weather is not a gap
- Uniform background makes detection thresholds simple
Onshore · stratified panel
A designed sample, not brute force.
You cannot chip 400,000+ turbines at 3 m. A stratified panel of 10,000–20,000 turbines, selected by OEM platform, blade type, vintage and climate zone, gives a defensible rate with confidence intervals per platform — which is more useful than a global average anyway.
- Per-platform rates with confidence intervals
- Strata weighted back to the global fleet
- Panel refreshed as new platforms enter service
The validity threat
A planned blade swap looks identical to a detachment.
From orbit, a blade missing for a scheduled uptower exchange and a blade that has torn off are the same picture. The discriminator is temporal ordering against the lifting equipment.
Planned exchange
Crane or jack-up appears first. Blade disappears within the same window.
Unplanned failure
Blade disappears first. Long lag before equipment arrives — and that lag is the downtime measurement.
Denominators and ground truth
The map already exists. So does the calibration set.
Turbine registers, free and public
- OpenStreetMapNear-complete global turbine coordinates
- Global Energy Monitor wind trackerProjects, capacity, status, ownership
- USWTDBEvery US turbine with model, hub height, rotor diameter
- MarktstammdatenregisterGerman register with commissioning dates
- Danish Energy Agency master dataPer-turbine production by month
Commissioning dates make turbine-years computable. That is the denominator.
Documented events for calibration
- Vineyard Wind, 2024Offshore blade failure with public dates and imagery
- Siemens B53 campaignOnshore fleet-wide event, known scope and timing
- GE 1.6-100 spar cap campaignOnshore fleet-wide event, known scope and timing
Known events with known dates let us measure our own detection rate and false-alarm rate before publishing anyone else's.
What you get
Three numbers the industry has never had.
Rate
Detachment rate with intervals
Confirmed events per thousand turbine-years, by OEM platform, blade type, vintage and climate zone. Updated as passes accumulate.
Log
Event log with dates
Every confirmed event with first-absent and first-restored dates, planned-versus-unplanned classification, and the confirming imagery.
Downtime
Repair lag distribution
How long a failed turbine actually stands with a missing blade, offshore and onshore. Nobody publishes this today.
Early access
Offshore first.
Building now.
We are onboarding a small group of insurers, owners and OEMs to shape the first dataset. Tell us who you are and which fleets you care about.