Early stage damage detection of wind turbine blades based on UAV images and deep learning

Abstract

In response to the shortcomings of existing image detection algorithms in the early damage detection of wind turbine blades, such as insufficient applicability and unsatisfactory detection results, this paper proposes an improved DINO (DETR with improved denoizing anchor boxes for end-to-end object detection) model for wind turbine blade damage detection called WTB-DINO. The improvement strategy of the DINO model is obtained by collecting and analyzing unmanned aerial vehicle (UAV) daily inspection image data in wind farms. First, the lightweight design of DINO's feature extraction backbone is implemented to meet the requirement of fast and effective video inspection by drones. Based on this, the Focus down-sampling and enhanced channel attention mechanism are incorporated into the model to enhance the feature extraction ability of the Backbone for damaged areas according to the characteristics of wind turbine blade images. Second, a parallel encoder structure is built, and a multi-head attention mechanism is used to model the relationship between samples for each type of damage with uneven distribution in the dataset to improve the feature modeling effect of the model for less-sample damage categories. Experimental results show that the WTB-DINO model achieves a detection precision and recall rate of up to 93.2% and 93.6% for wind turbine blade damage, respectively, while maintaining a high frame rate of 27 frames per second. Therefore, the proposed WTB-DINO model can accurately and in real-time classify and locate damaged areas in wind turbine blade images obtained by UAVs.