Deep learning models for damage type detection in wind turbines

This study presents deep learning models that are frequently used in the literature for the detection and classification of damage types in wind turbines and a new deep learning model (SatNET) that offers computational efficiency and rapid inference. Wind turbines, which are critical components of renewable energy systems, are sensitive to various damages (paint damage, erosion, serration, vortex, and vortex damage) that may endanger their operational efficiency and lifespan. The dataset consists of 1,794 high-resolution images taken under different weather conditions and angles, including damage and types. The images were increased by four times to 7,176 images using data augmentation techniques. Damage and types were detected using the developed SatNET deep learning model, 11 deep learning models, and the Faster Region-based Convulational Neural Network (R-CNN) object detection algorithm. Each of the models was evaluated with average sensitivity. Accordingly, SatNET achieved avarage precision (AP) values of 55.7% for paint damage, 76.7% for erosion, 95.2% for serration, 66.1% for vortex, and 27.3% for vortex damage. It demonstrated superior performance when compared to deep learning models frequently used in the literature, such as ResNet50 and VGG19. In addition, it has been shown that the model requires less computational cost than other models, with a memory requirement of 192 MB. The results show that SatNET’s computational efficiency and accuracy are competitive with other models. The model is suitable for systems with limited memory and computational capacity, which require real-time operation, and for systems with resource constraints. The results obtained can contribute to sustainability in renewable energy production by providing low-cost monitoring of damage and types in wind turbines.