Wind Turbine Surface Damage Detection by Deep Learning Aided Drone Inspection Analysis

Applies deep learning to drone images of wind turbine blades to detect leading-edge erosion, vortex generator panels (including missing teeth) and lightning receptors. Faster R-CNN with several backbones was compared, and multi-scale pyramid augmentation with patching sharply improved accuracy. Trained on Danish turbine images, it also releases a public image set of one turbine. Showing model suggestions to human analysts shortened their per-image time.

Key findings

• Best mAP 81.1% (Inception-ResNet-V2 with pyramid, patching and regular augmentation) versus 27.5% with no augmentation
• Inference 1.36-2.11 s per high-resolution image versus about 200 s for a human, or about 131 s with model suggestions
• Human precision 83.2%; humans aided by suggestions 85.3%
• Public DTU dataset of 701 images of one turbine released

Limitations

• Surface damage only; internal damage such as delamination out of scope
• Small, imbalanced training data (e.g. 17 lightning-receptor annotations); lightning receptors 71.6% vs 98.7% for humans
• Main dataset private; public set covers a single turbine
• IoU threshold reported inconsistently in the paper