Attention Mechanism-based CNN for Surface Damage Detection of Wind Turbine Blades

Wind turbine blades undertake the task of energy conversion. Wind turbine blade damages detection can provide guidance to maintain wind turbine. In an effort to address the low accuracy of current blade damage detection, this paper proposed an attention mechanism-based CNN the classification of common damage to wind turbine blades. This framework uses morphological transformations to enlarge the dataset and adjusts contrast and chromaticity to enhance damage features. The preprocessed blade images are sent to CNN for feature extraction, and then the feature maps are re-calibrated by applying the proposed mechanism to build up the score of the damage feature in the classifier. The performance of the proposed framework for detecting blades damage was confirmed by manually collected images from commercial wind farms, with a multi-classification accuracy of 92.85%. And this paper analyzes the role of the attention mechanism in damage detection by visualizing the attention mechanism using grad-CAM.