Wind turbines generate the clean energy that powers a more sustainable future but their blades are vulnerable to small defects that are extremely difficult to spot. Among the hardest to detect are barely-visible surface hairline cracks, which often look like nothing more than dirt streaks or faint scratches in drone-captured images. Missing these tiny defects can lead to severe, irreparable damage costing millions of dollars to fix.
Zeitview’s AI Research team set out to solve this problem. Their award-winning work introduces the most comprehensive dataset to date for barely-visible crack detection and presents a fully deployed, end-to-end AI pipeline that dramatically improves the accuracy, efficiency, and scalability of wind turbine inspections.
The paper introduces the Zeitview Crack Detection (ZVCD) dataset, built from real inspections across:
988 locations worldwide
195 blade models from 34 manufacturers
This dataset includes , many of which are nearly invisible to the human eye. Compared to existing public datasets, often small, simulated, or limited to obvious defects ZVCD provides the diversity and realism required to train reliable models for real-world inspections.
Rather than relying on heavy segmentation or object-detection systems, Zeitview uses a classification-based approach on image tiles. This design choice reflects constraints of real inspections:
Drone images are extremely high-resolution, requiring tiling anyway
Classification reduces false positives, a major issue in crack detection
Labels for classification are much cheaper to produce at large scale
The team trained and evaluated models such as ResNet-18, EfficientNet-B3, and MobileNetV3, emphasizing high precision to avoid missed cracks the most costly failure mode.
Zeitview’s deployed pipeline uses a multi-stage process:
Autonomous drone capture
Autonomous drones follow a planned path, capturing images to ensure full blade coverage.
Tile-based inference
Images are broken into 1024×1024 tiles, and the classifier flags tiles likely to contain cracks.
Localization via Grad-CAM
Since classification alone doesn’t show where the crack is, the system generates region proposals using Grad-CAM heatmaps and converts them into polygons after normalization, contour detection, and filtering.
Human-in-the-loop review
Analysts receive precise proposals, verify severity, and finalize inspection reports. This hybrid automation approach dramatically reduces manual review time while maintaining expert oversight where it matters most.
Zeitview’s system improves turbine maintenance in several critical ways:
Earlier crack detection → prevents catastrophic failures
Higher inspection consistency → fewer missed defects
Scalable to thousands of turbines → no added sensor hardware needed
Deployable on-drone hardware → enabling near-real-time assessment
By focusing on a high-impact but understudied class of defects, hairline cracks, the research directly supports fleet reliability and renewable energy sustainability.
To explore the full methodology, dataset, and deployment pipeline behind this award-winning work, read the complete paper here →https://arxiv.org/pdf/2407.07186