Research On Blade Surface Damage Detection Of Wind Turbine Based On Computer Vision

Wind energy,as a renewable energy,has been widely used in the world.Due to the harsh environment and intense mechanical stresses,the blade surface of wind turbines may generate crack,oil pollution and sand hole,which will seriously affect the efficiency and safety of wind power generation.At present,the traditional method of blade surface damage detection mainly relies on manual operation,which has some shortcomings such as time-consuming,low efficieney and strong subjectivity.Therefore,in this thesis,based on the blade images collected by UAV,computer vision technology,including image processing,machine learning and deep learning,is used to study the detection algorithm of blade damage and the development of a health maintenance system.This paper consists of the following parts:(1)In this thesis,the damage characteristics and image processing theory of wind turbine blades are analyzed.The image processing algorithm based on Halcon is studied.The blade image is calibrated by camera,filtered by fast adaptive weighted median,enhanced by image and segmented by dynamic threshold to complete the processing and preliminary recognition of the blade damage image.(2)The features of HOG,Haar-like,LBP and classification algorithms such as CatBoost,XGBoost and AdaBoost are compared and analyzed.Using the original and extended LBP feature set,a multi-cascade classifier including decision tree and support vector machine based on AdaBoost is developed as LBP-ADA model.The LBP-ADA is formed to extract the features of blade damages and train the model.The original image is expanded by rotating,translating,scaling,adding noise points and changing the light and shade to form a machine learning image database.(3)Based on the deep learning theory,the small damages detection algorithm is studied.By introducing the structure of feature pyramid network(FPN),You Only Look Once(YOLO)algorithl is improved,and a new YSODA algorithm is proposed.The algorithm fuses shallow and deep feature maps of YOLO,and adds FPN structure to its network structure for advanced feature extraction.The experimental results show that the algorithm can effectively improve the detection accuracy of small damages on blades.(4)The accuracy,recall rate and weighted harmonic mean were used as evaluation indexes.The results of cross experiments with various methods showed that the LBP-ADA algorithm was effective in detecting blade damage.At the same time,YSODA algorithm is tested with accuracy and real-time as indicators,and compared with LBP-ADA and literature algorithm.It can be seen that YSODA algorithm proposed in this thesis has high accuracy and efficiency in micro-damage detection.Finally,a wind turbine blade health maintenance system is designed based on the developed detection models.The system is applied to the blade danage detection in the actual wind field,and the blade quality detection report including wind field information and damage information is automatically generated,which meets the operation and maintenance requirements of the actual wind fields.