Research On Damage Detection Of Wind Turbine Blade Based On Computer Vision And Deep Learning

Wind turbine blades are the key components of wind turbines,and their safety and reliability are the prerequisites for the safe operation of wind turbines.Online early detection of surface damages on blades is critical for the safety of wind turbines,which could avoid catastrophic failures,minimize downtime,and enhance the reliability of the system.Monitoring the health status of blades is attracting more and more attention including on-site cameras and mobile cameras by drones and crawling robots.However,previous researches focus on the use of image processing and hand-crafted features,leading to problems such as poor positioning accuracy,inaccurate identification,and low detection efficiency.In this case,this paper explores a novel approach for multi-scale damage detection and multi-label damage identification through the combination of computer vision and deep learning methods and presents a hierarchical identification framework for wind turbine blades to deploy fast and efficient damage detection methods from image data.The research works of this paper are summarized as follows:Firstly,the causes and main types of wind turbine blade surface damage are deeply analyzed,statistical grey features and texture features are extracted from image samples.5 representative features are selected via person correlation analysis.Comparative analysis of representative features shows the damaged regions are darker,high contrast,and abundance of image textures compared to non-damaged regions.Most of the damaged regions can be identified based on representative features,which provides a reference for further design and extraction of damaged features.Secondly,a multi-scale damage detection algorithm based on Haar-like features and Ada Boost cascade classifier is established.The multi-scale sliding-window crop method is adopted to build the multi-scale damage dataset and train the Ada Boost cascade classifier.Multi-scale damage detection is realized through resizing the input image multiple times during the detection process.Test results show the recall rate of the multi-scale damage detection algorithm is superior compared to single-scale detection,excellent performance is shown on minor damage regions detection.The multi-scale damage dataset can effectively expand the image samples and improve the performance of the detector.Thirdly,a multi-label damage identification model based on a convolutional neural network(CNN)is presented.The VGGNet for Blade Damage Detection(VGG-BD)model,which is a modification based on VGG16,addressing the over-fitting problem of the traditional CNN model by replacing the fully connected layers with a global maxpooling layer,using dropout layers and LRe LU activation function.VGG-BD improves the training efficiency and robustness of the original model and prevents the model from overfitting during the training process.A multi-label damage dataset is established to train and validate the performance of VGG-BD.Experimental results show VGG-BD can identify 4 common types of blade damages and outperforms compared methods include SVM and VGG16 models.The sensitive analysis is conducted to validate the robustness of the proposed method under limited data conditions.Finally,a hierarchical identification framework for wind turbine blades is set up,which consists of a Haar-Ada Boost step for region proposal and the VGG-BD classifier for damage detection and fault diagnosis.Case studies are carried out on real data set collected from an eastern China wind farm.Results show that VGG-BD successfully reduces the fault detected regions in region proposals,which greatly improves the precision rate.The proposed framework can detect and identify the blade damages and outperforms other schemes include SVM and VGG16 models.The proposed scheme is faster than SVM and VGG16 methods.The proposed framework has the advantages of high detection accuracy and low computational cost and is suitable for massive deployment in wind farms.