Automatic Optical Surface Inspection of Wind Turbine Rotor Blades using Convolutional Neural Networks

Abstract

The operation of wind turbines includes the regular surface inspection of their rotor blades. This leads to considerable downtimes and expenses due to the manual inspection process. A possible solution is the automation of this process by using drones or robots. In this article, we present a key component for such an approach by automating the visual surface inspection with convolutional neural networks (CNN). We provide insights into CNN model selection based on available hardware and training data. We further show that all CNN models reach over 96 % median classification accuracy with the best model, ResNet50, reaching 97.4 %.