The Research On A Detection System For Wind Turbine Blade Surface Defect S Based On UAV Image Acquisition And An Improved Depth Residual Network Classification Algorithm

With the increasing capacity and volume of wind turbine assembly machine,the traditional manual detection method is not only high cost,low efficiency,high risk,but also some damage is difficult to detect by naked eye,which leads to the detection accuracy is difficult to guarantee.With the rapid integration of computer technology into the industrial field,the methods to solve practical problems in the industrial field are more diverse and more intelligent.In this study,based on the rapid development of deep learning algorithm in recent years,combined with machine vision technology,a complete set of wind turbine blade damage detection system is established,which can realize the identification of blade damage and the division of damage level,which is of great significance to ensure the healthy operation of wind turbine and improve the efficiency of wind power generation.The main research work of this paper is as follows:(1)Data acquisition: DJI Spirit 4PRO quadrotor UAV was used in combination with DJI GS PRO of DJI ground station to plan routes for UAV and realize automatic data collection of fan blades.The data were collected from Wuchuan Guodian Wind Power Co.,Ltd.and Guyang Huachen Jiugongzhong Wind Power Plant.According to the different damage conditions,the video data and picture data of 30 wind turbines were collected.The video data were disassembled into picture data at a pace of 2pieces/second,and the total expanded data was 1751 pieces.(2)Data set creation: Since the input size of images is specified by deep residual networks ResNet,DenseNet,ShuffleNet,etc.,it is necessary to take pixel as the basic unit and cut the original full-size image to the input size required by each network.After cutting,a total of 501093 sub-images are obtained.For the damage classification task,10,000 fan blade charts were randomly selected from 22 sets to construct the training set,and 500 charts were randomly selected from each of the remaining 8 sets to construct 8test sets.Data enhancement was performed for the less number of damage types.For the injury classification task,the training set was constructed with 2976 injury pictures from the training set of the injury classification task,and 8 test sets were constructed with the injury pictures from each test set of the injury classification task.(3)Image processing: Because the traditional classification algorithm support vector machine(SVM)does not have the ability to automatically extract image features,it needs to do tedious pre-processing work on the data,mainly including graying,image enhancement,image filtering,image segmentation and feature extraction.(4)Classifier design and simulation experiment: In this paper,by reconstructing the tail structure of deep residual network,we design deep residual network classifiers ResNet101,DenseNet201,ShuffleNet,and propose improved deep residual network classifier.At the same time,in order to verify the effectiveness of the proposed method,the proposed deep residual network classifier is compared with SVM.The simulation results show that the proposed method achieves good results.For the damage classification task,Resnet101,Densenet201 and ShuffleNet reach 95.55%,96.23% and94.90%,which are all better than SVM’s 75.43%.For the damage grade classification task,Resnet101,Densenet201 and ShuffleNet reach 93.71%,95.60% and 92.74%,which are all better than SVM’s 72.08% accuracy.Improvement method 1,aiming at the imbalance problem of fan blade data set samples,Focal Loss Loss function in the target detection task was introduced.Damage classification task increased by 0.4%,0.3% and0.4% respectively,and damage grade classification task increased by 0.04%,0.14% and0.16% respectively.Improvement method 2,aiming at the problem of insufficient nonlinear capability of the model caused by single-layer full connection,the classification accuracy was improved by adding full connection layer,and the damage classification task was improved by 0.5%,0.92% and 1.8% respectively,and the damage grade classification task was improved by 0.09%,0.21% and 0.11%respectively.Improvement method 3,the deep residual network classification algorithm is combined with the traditional machine learning method SVM.The deep residual network feature extractor is constructed by removing the terminal layer of the deep residual network.After extracting the features,it is sent to SVM for training classification,which also improves the classification accuracy.The damage classification task increased by1.55%,2.32% and 1.95% respectively,and the damage grade classification task increased by 0.44%,0.6% and 1.21% respectively.After any combination of different improvement methods,it is measured that the three improvement methods have the highest accuracy when fused together,which are all higher than any improvement method and its pairwise combination.The damage classification task increased by 1.78%,2.47% and 2.07%respectively,and the damage grade classification task increased by 0.79%,0.77% and1.49% respectively.(5)Human computer interaction interface design: in order to facilitate the rapid deployment and application of the system and connect with the actual industrial site,a simple and easy to understand GUI human-computer interaction interface is designed to realize one key operation of data set selection,damage identification,damage classification and other functions,and realize the integration from input data to output test results.