Research On Power Insulator Defect Detection And Multi-defect Classification Based On Deep Learning

Power insulators are an essential component of the power system,and their main function is to support the transmission lines and provide electrical isolation between the lines and supporting structures,ensuring the safety and stable operation of the power system.Due to climate change,mechanical damage,and other reasons,insulators often have defects,which can cause faults in the power transmission lines and threaten the safety and reliability of the power system.Therefore,regular inspection and maintenance of insulators are a crucial part of ensuring the stable operation of the power system.This thesis focuses on the collection of insulator images and applies deep learning methods to conduct in-depth research on insulator defect detection and defect classification.The main research content of this article is as follows:Aiming at the problem of insufficient samples of the current open-source insulator image data set,this thesis performs noise reduction on the collected insulator images,and uses superresolution convolution network to improve the clarity of the input image for the problem of insulator image distortion after noise reduction.In addition,this thesis uses the single sample and multi sample data set expansion method to establish the insulator data set needed for this study.Aiming at the problem of poor applicability of the original YOLOv5 model in insulator defect detection tasks,this thesis has made a series of improvements to the model.In this thesis,the path aggregation network of the neck structure of the original YOLOv5 model is replaced by a bidirectional feature pyramid network to improve the feature fusion ability of the network for different scale targets.The K-means++ algorithm is used to re-cluster the candidate box size to make it more suitable for insulator targets;EIOU Loss is used to replace the bounding box loss function of the original model,and Focal Loss is combined to improve the detection accuracy and convergence speed of the model.The experimental results show that the improved model has a detection accuracy of 93.2 % in the insulator defect detection task.In order to solve the small sample problem and improve the accuracy of insulator multidefect classification,this thesis uses the transfer learning method to pre-train the Res Net50 network on the Image Net dataset as the basic model of multi-defect classification.In addition,this thesis adds an attention mechanism to the model to make the model pay more attention to the characteristics of insulator defect regions,so as to improve the accuracy of multi-defect classification.In order to adapt to the number of categories of insulator data sets,this thesis improves the output structure of the fully connected layer in the original Res Net50 model.The experimental results show that the classification accuracy of the improved Res Net50 model is as high as 94.1 %.In order to achieve high-precision classification of insulator multi-defects under complex background conditions,this thesis constructs a cascade model of insulator detection and multidefect classification based on ’target detection + multi-defect classification’.The cascade model uses the improved YOLOv5 model to frame the insulator target detection,and then uses the cutting layer to cut the insulator image in the rectangular identification box and input the improved Res Net50 model to obtain the specific category of the defect.The experimental results show that the classification accuracy of the cascade model constructed in this thesis can reach more than 95 %,which can meet the needs of actual power inspection work and facilitate the maintenance of insulator defects by subsequent staff.