Inspecting a bridge underside, a 60 m wind-turbine blade or kilometres of overhead line traditionally means rope access, lifts, helicopters or foot patrols: slow, costly and risky. Drones capture dense imagery from every angle; computer vision then screens thousands of images per asset, flags cracks, corrosion, damaged insulators, leading-edge erosion or vegetation encroachment, and places each finding on a 3D model or asset tree. Engineers verify and decide — the AI shrinks the review workload and makes inspections comparable over time.
Problem & value
- Today: visual inspections by rope-access technicians, climbers, snooper trucks or crewed aircraft. They are periodic, expensive, hazardous (work at height, live conductors) and the results are subjective and hard to compare year over year. After a drone flight, the new bottleneck is manual image review: one turbine or one tower can yield hundreds of images.
- AI value: less work at height and fewer shutdown hours, objective and repeatable defect classification, a defect history per component, and condition-based rather than calendar-based maintenance. Savings claims from vendors vary widely by asset type and are only meaningful against your own baseline cost per inspection.
- Regulatory driver: bridge owners (e.g. DIN 1076 in Germany) must prove regular condition assessment; drones supplement, not replace, required hands-on inspection.
Variants
- Wind turbines: blade surface (leading-edge erosion, cracks, lightning damage, delamination via thermography), tower and nacelle; in-house drone teams, service providers, or tethered/crawler robots for close-up inspection. Offshore adds vessel-launched flights.
- Power lines and substations: tower/pylon corrosion, insulator and fitting damage, conductor damage, vegetation clearance via LiDAR, thermal hot spots in substations; BVLOS flights along corridors.
- Bridges and civil structures: crack mapping on concrete, corrosion, bearings and joints, often under the deck (GNSS-denied, so visual or LiDAR-SLAM navigation); also dams, chimneys, towers, facades.
- Other: pipelines, solar farms (thermal faults), railway catenary, ports.
- Private use: negligible beyond hobby roof checks under open-category rules.
How it works
- mission planning (asset model, waypoints, overlap, RTK)
- autonomous or piloted flight (RGB + thermal + optional LiDAR), geotagged
- upload / edge pre-screening (blur, coverage checks on site)
- photogrammetry / 3D reconstruction (orthomosaic, mesh, point cloud) + image-to-component registration
- defect detection & segmentation (cracks, corrosion, erosion, damaged insulators, hot spots)
- severity classification per component + change detection vs. previous campaign
- engineer review (confirm / reject / re-grade)
- report generation
- work orders in EAM/CMMS (SAP PM, IBM Maximo, Infor) + digital asset twin
- Detection: object detectors (YOLO-class, RT-DETR) for discrete parts and defects; segmentation models for cracks and corrosion areas (pixel-accurate width/area); anomaly detection for rare defects with few labels.
- 3D registration: each image is matched to the exact blade section, tower member or bridge element so findings are located and trackable — this is as important as the detector itself.
- VLMs: report drafting and semantic search over findings, not the sole safety-relevant classifier.
- Human-in-the-loop: a qualified inspector confirms every finding above a severity threshold and signs the report; low-confidence images are routed to review rather than silently dropped.
Data preparation
Data are high-resolution RGB and thermal images, video, LiDAR and flight logs with GNSS pose, plus the asset register and earlier inspection reports. They sit on drone SD cards, pilots' laptops, file shares and in reports (often PDF or paper). Typical gaps: images without asset ID or position, inconsistent folder structures, scattered historical reports, no common defect classification, few examples per defect type, and huge data volumes.
Preparation steps: (1) agree a defect catalogue and severity scale with engineers (PREP-019); (2) link every flight and image to an asset ID from the register; (3) set up a structured, backed-up storage with metadata (PREP-010, DATA-007); (4) digitise and extract old inspection reports (PREP-006, DOC-014); (5) label defects with experts, using model-in-the-loop to speed up (DATA-023); (6) blur persons and number plates (DOC-018 principle) before wider use.
Tools: CVAT, Label Studio, drone vendor software, a simple asset-keyed object store. Rough share of total effort: 40-60 percent.
Implementation paths
Which path for whom
- SMB — Path A: most small operators buy drone-inspection services or platforms with defect detection included, since flight operations and models are specialised.
- Enterprise — Path B: large asset portfolios justify an own pipeline on cloud APIs and models, integrated with the asset register and maintenance system.
- Public — Path C: public utilities and authorities benefit from sovereign, self-hosted processing for critical infrastructure data and procurement rules.
- Developer — Path C: open-source detection and photogrammetry tools allow full control, but need HW3-class GPU capacity.
Path A — Buy (recommended for most operators)
- Inspection platforms with AI analytics: Skydio (US; autonomous drones, 3D Scan), Percepto (IL; drone-in-a-box), Sterblue (FR, EU; power lines and wind), Voliro (CH; contact inspection), Scopito (DK, EU; annotation and reporting), Aerones (LV, EU; robotic blade work), Flyability (CH; confined spaces), DJI Enterprise hardware (note security debates on Chinese drones in critical infrastructure).
- Service model: many operators buy inspection as a service (flights, analysis, report).
- Choose A when: you have fewer than a few hundred assets, no in-house drone and ML team, or need accredited reports quickly. Check EU data hosting, export of raw images, open report formats and integration with your EAM.
Path B — Build on cloud APIs
- Storage and compute: EU-region storage and batch GPU inference (AWS Frankfurt, Azure, OVHcloud, STACKIT); Pix4D (CH) or DroneDeploy for orthomosaics.
- Models: managed training and hosting (Azure ML, Vertex AI, SageMaker) for custom detectors; multimodal APIs (Claude, Gemini, GPT, Mistral) for report drafting and zero-shot "second opinion" triage with structured outputs.
- Integration: EAM/CMMS APIs, GIS (QGIS/PostGIS), review UI such as CVAT.
- Choose B when: you have your own flight operations and want custom defect taxonomies tied to your maintenance system.
Path C — Self-host (open source)
- Detection/segmentation: Ultralytics YOLO (AGPL, check licence), RT-DETR, MMDetection, Detectron2, SAM 2 (Segment Anything 2) for assisted labelling and crack segmentation; Anomalib for unsupervised anomaly detection.
- Photogrammetry/3D: OpenDroneMap/WebODM, COLMAP, PDAL, Open3D.
- Edge: NVIDIA Jetson Orin on the drone or dock for real-time coverage checks.
- Hardware: HW3 (one 48 GB GPU) for training and batch inference on a fleet; HW1 (Jetson) at the edge; photogrammetry benefits from a strong GPU workstation.
Path D — Custom training
Usually needed: public datasets rarely match your coatings, insulators or concrete. Fine-tune a pretrained detector on hundreds to thousands of labelled examples per defect class (rare defects: synthetic data, anomaly detection), labelled by inspectors against your defect catalogue (DIN 1076 codes, OEM or insurer scale); hold out a campaign from another site and season.
State of the art (2026)
- Autonomy: drone-in-a-box and BVLOS corridor flights are moving from pilots to routine ops; autonomous obstacle avoidance and 3D-scan flight planning reduce pilot skill needs.
- Autonomy and models: drone-in-a-box and BVLOS corridor flights are moving from pilots to routine use; SAM-class segmentation speeds labelling; VLMs draft text and enable search but are not yet reliable for quantitative severity grading. Contact drones add ultrasonic thickness data (VIS-053).
- Limits and failure modes: small hairline cracks (sub-millimetre) need very low ground sample distance and good light; reflections, shadows, wet surfaces, dirt and paint marks cause false positives; class imbalance means rare critical defects are the least well detected; models trained on one turbine type or climate generalise poorly; GNSS multipath near steel structures and under bridges degrades georeferencing; drone imagery cannot see internal damage (corrosion under coating, delamination inside blades, prestressing steel). Treat AI output as screening.
Shared resources & combinations
- Same stack: VIS-004, VIS-043, VIS-058, VIS-040, VIS-001 (detectors, labelling). VIS-021 and VIS-061 show where to fly; VIS-053 adds subsurface evidence. Downstream: PRED-046, ANOM-001, VIS-060 (shared edge hardware), ROBO-001 (crawlers).
- One GPU server (HW3) and one review/labelling tool can serve all of these.
Effort, cost & KPIs
- Path A: weeks to first campaign. Path B: 3-6 months. Path C/D: 6-12+ months incl. labelling and validation.
- Cost drivers: flight operations, labelling of rare defects, 3D registration, hardware, data volume.
- KPIs: recall for critical defects (the primary metric), precision/false-alarm rate, review time per asset, inspection cost per asset, hours of work at height avoided, time from flight to report, share of defects found early (before repair cost step-up), turbine downtime.
People & change
Involve management for budget and safety goals, the asset or maintenance lead as process owner, drone pilots and inspectors as end users, IT and IT security for data storage and remote access, legal for airspace and liability, and the DPO because flights can capture people and private property. Data owners provide the asset register and reports.
Works council: depends. Inspecting assets is not monitoring employees, but if staff such as pilots or maintenance crews are filmed, or results and flight logs are evaluated per person, §87(1) no. 6 BetrVG applies; inform the works council in any case about changed work processes (§90(1) no. 3). Public operators: staff council. A DPIA may be needed for systematic recording in public spaces or near housing; AVV with every cloud processor.
AI literacy: the AI Act (Art. 4) requires measures that support staff's AI literacy; train inspectors in reading detections, false alarms and limits. Communicate that engineers still decide on safety-relevant findings; the AI only prioritises images. Pilot on one asset type, such as one wind farm or bridge, with experienced inspectors, and keep expert sign-off. See GUIDE-005, GUIDE-008 and GUIDE-009.
Risks, regulation & compliance
- Drone law (EU): Regulation (EU) 2019/947 sets open/specific/certified categories; inspection usually needs the specific category (SORA risk assessment), BVLOS needs approval; national rules add geo-zones.
- EU AI Act: the inspection AI itself is typically minimal or "depends": it is high-risk if it functions as a safety component in the management and operation of critical infrastructure (Annex III, point 2, e.g. energy supply, road traffic, digital infrastructure) — assess case by case whether a failure of the AI endangers safety rather than merely supporting a maintenance decision. Drones themselves are regulated under drone law. No Art. 5 issue unless the pipeline is repurposed to identify people (see GDPR below).
- GDPR: flights may capture identifiable people and homes; minimise corridors, blur bystanders (VIS-030), set retention, run a DPIA for residential areas.
- Critical infrastructure security: imagery and 3D models are sensitive under NIS2/CER; use EU hosting and access control.
- Liability: the engineer signs the report; keep AI suggestion, human decision and raw images in an audit trail.
Getting started
- Pick one asset class (e.g. one wind farm or 20 bridges) and write down the defect catalogue and severity scale your engineers actually use today, with the baseline cost per inspection.
- Check legal basis: operator registration, authorisation category, geo-zones, GDPR/DPIA.
- Run a pilot campaign (own or service provider) with raw image export and ground-truth findings from human inspectors.
- Benchmark 2-3 vendors on your own images, measuring recall for critical defects first.
- Integrate findings with your EAM/CMMS and set an engineer-review workflow before scaling.
Sources
- EASA — Drones (UAS) regulations and categories — https://www.easa.europa.eu/en/domains/civil-drones
- Regulation (EU) 2019/947 on rules and procedures for the operation of unmanned aircraft — https://eur-lex.europa.eu/eli/reg_impl/2019/947/oj
- Regulation (EU) 2024/1689 (AI Act), Annex III point 2 and Art. 5 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Skydio — Inspect solutions — https://www.skydio.com
- Voliro — contact inspection drones — https://www.voliro.com
- Ultralytics YOLO documentation — https://docs.ultralytics.com
- Anomalib (Intel/OpenVINO) — https://github.com/openvinotoolkit/anomalib
- OpenDroneMap — https://www.opendronemap.org
- Directive (EU) 2022/2555 (NIS2) — https://eur-lex.europa.eu/eli/dir/2022/2555/oj