Research On Detection Of Multi-type Surface Defects Of Wind Turbine Blades Based On Images Collected By UAV

As a non-polluting and renewable green energy,wind energy plays an active role in improving the national energy structure and optimizing the ecological environment.In recent years,wind power has developed rapidly,and being the core component of wind turbine,the safe operation of the blade is vital for ensuring stability of the wind turbines.However,the severe working environment and complex climate conditions inflict crack,breakage,oil pollution and other defects of different severity,which is unbeneficial to the safe operation and power generation efficiency of the wind turbines.At present,the defect detections on wind turbine blades(WTBs)are mostly manual operation,which have the disadvantage of high cost,low efficiency,strong subjectivity and high risk.Therefore,based on the images collected by unmanned aerial vehicle(UAV),the automatic and accurate detection of multi-type surface defects on WTBs applying computer vision algorithms including image processing and deep learning in this paper are studied as follows:(1)Based on the demand of actual project,the hardware and software system of surface defect detection on WTBs are designed and constructed.UAV is applied to collect the images of the WTBs in the actual wind field,and the collected data are screened,labelled and scaled,thus the WTBs image dataset applied in this paper is established.In addition,data augmentation methods such as translation,rotation,scaling and flipping are carried out on the images of training set.In particular,according to the characteristics of the dataset,the augmentation methods of Gaussian Blur and Multiply are employed to expand the training set which ensure the robustness and generalization of the trained model.(2)Based on the traditional image processing,the images in the dataset are processed by graying,image filtering,image segmentation,morphological processing and defect marking.Meanwhile,the iterative threshold segmentation and the open operation of morphological are combined to detect the surface defects on WTBs according to the characteristics of the dataset.The effectiveness of the above methods for the identification of surface defects on WTBs is verified by the experiments.The experimental results show that three types of surface defects on WTBs can be identified by the above algorithms separately.However,image processing is difficult to achieve the requirements of accurate multi-type classification,and the robustness of the algorithm remains to be improved.(3)Aiming at the actual demands of multi-type classification and high-accuracy detection,deep learning object detection algorithms are further applied in defect detection for a deep research.Cascade R-CNN based on Res Net-101 feature extraction network is employed as the basic detection model.The idea of transfer learning is used in training to obtain a faster convergence rate.Moreover,to alleviate the problem of false detection under a complex background,an improved bisecting k-means is presented during the test process.The experimental results show that various surface defects on WTBs can be automatically detected by Cascade R-CNN which has strong multi classification and recognition ability.In addition,it has a higher detection accuracy of 80.9% m AP than other advanced object detection algorithms.(4)In order to further improve the detection accuracy and robustness of the model,Deformable Convolution and Deformable Ro I Align,context information fusion,as well as PRe LU activation function are employed to improve the Cascade R-CNN after analyzing the characteristics of defects on WTBs.Consequently,a model named Contextual Aligned-Deformable Cascade R-CNN(CAD Cascade R-CNN)is proposed.It can be verified that each improved strategy can effectively improve the detection accuracy m AP of the model by abliation experiments,and the convergence rate can be promoted to prevent over-fitting.Compared with the detection results before and after improvement,it can be shown that the results of CAD Cascade R-CNN proposed in this paper can reach 92.1% of m AP,which is 11.2% higher than that of Cascade R-CNN,and has stronger robustness and regression accuracy with the increased evaluation threshold.Finally,a visual user interface is developed based on the trained model for applying in the field inspection.