Research And Implementation Of Wind Turbine Blade Damage Recognition For Mobile Devices Based On Deep Learning

With the depletion of fossil energy,countries around the world are increasing their efforts to develop and use clean energy represented by wind energy.The blade is the most important component of wind turbine(WT)to convert wind energy into kinetic energy.Its health is related to the efficiency of wind energy utilization and the safety of power generation process.For the detection of WT blade damage,the traditional artificial visual method has been gradually abandoned.In recent years,some wind farms have begun to use unmanned aerial vehicles(UAVs)to collect WT blades,and have also tried to use image processing methods in the blade damage detection stage.However,these methods require the images to be sent back to the maintenance company for analysis,which requires a lot of time.At the same time,the cost of manpower and material resources has not been reduced much.In this case,this paper proposes a method of detecting WT blade damage rapidly using mobile phones in wind farms.The damage identification tool is the lightweight deep learning model deployed on mobile phone.The main research works of this paper are as follows.Because the real-time detection of blade damage and the establishment of model set need a large number of WT blade images,the method of UAV automatic patrol-photography is proposed to collect the WT blade images.Based on the particularity of WT blades,a method for automatically collecting blade images by UAV is designed.From the beginning of positioning the blade to the final return,the UAV will cruise automatically according to the planned route,which can ensure that the blades of the entire WT can be photographed stably and comprehensively.In order to make the damage detection system run well on the mobile phone and detect WT blade damage efficiently,the model YOLOv4 for detecting the damage of blade is lightweight improved.Firstly,the backbone network of the YOLOv4 model is redesigned,and then the model is compressed using pruning and quantization methods.At this time,the volume of the model is reduced,and the computational complexity is also reduced.Then,in order to solve the problem that the detection accuracy of the model decreases after compression,the knowledge distillation method is adopted to improve the accuracy of the model.After the final lightweight R-YOLOv4 model is used for damage detection on the server,the detection frame rate is 174 fps,and the mAP value is 94.0%.Using tools such as Paddle Lite and Android Studio,the mobile end WT blade damage detection system based on the Android platform is designed and deployed.After the UAV collects the WT blades,the operator can detect the damage on the site without communicating with the equipment outside the wind farm.The results of damage detection can be displayed on the mobile phone screen,and the detection speed of each image is about 220ms,and the confidence scores are all above 85%.