Wind Turbine Blade Condition Detection And Analysis Based On Deep Learning

The blade is the main component of energy acquisition in wind turbines,and its health directly affects the operating efficiency and safety of the wind turbine.Although the fault detection technology of fan blades has been studied,there are some shortcomings,which have not been widely used,so that regular visual inspections are widely used in most wind farms.In this paper,a wind turbine blade condition detection and analysis method based on deep learning is proposed,and the fault detection and analysis of wind turbine blades based on deep learning is mainly studied.The specific research contents and results are as follows:Firstly,the improvement of model performance based on the basic structural characteristics of convolutional neural networks,including the under-fitting and over-fitting solutions,and the optimization of network operation speed,are studied.A detailed model optimization process is presented.Secondly,the research on wind turbine blade condition detection based on convolutional neural network is carried out.A dataset containing 25,773 wind turbine blade images and six different conditions was constructed by acquiring the wind turbine images by the UAVs.And three optimized convolutional neural network models are built,and the weights are compressed by applying the ADMM algorithm.Through training on the constructed dataset,the VGG-11 model was selected by comparing F1-Score.The classification accuracy reaches 79.3% in the testing set,and 20% compression rate is achieved by applying the ADMM algorithm.Finally,aiming at the problem of locating the wind turbine blade fault accurately,the research on the fault detection of the wind turbine blade based on the improved YOLOv3 is carried out.A wind turbine blade fault object detection dataset is constructed based on former dataset,due to the traditional target detection algorithm cannot accurately detect the inclined wind turbine fault objects,the rotated layer and the reduced layer are added to solve the inclined fault target detection problem;in addition,a multi-scale detection network is proposed to meet the purpose of detecting large targets.The validity of the improved algorithm is verified on VOC2007 dataset,and the mAP reaches 76.1% on self-labeled dataset,and the running speed reaches 31.2f/s,which realizes the real-time wind turbine fault detection.