Research On Insulator Defect Detection Method Based On Convolutional Neural Network

As an important device for insulation and mecnanical support in transmission lines,insulators are exposed to complex natural environments all year round and are extremely prone to failures.In severe cases,they will affect the safe and stable operation of transmission lines.Therefore,during power inspections,the status of electrical insulators must be monitored.In order to improve the efficiency of power inspection and realize the intelligent detection of insulators in UAV aerial images,this paper introduces a deep learning method,based on convolutional neural networks to achieve target location detection and self-explosion defect detection of insulators.First,the Faster R-CNN is improved based on the complex background and different scales of aerial insulator images.ResNet-101 is selected as the feature extraction network to enhance the characterization ability of the network,and Sotf-NMS is used to process the samples to improve adj acent overlap.The detection accuracy of the insulator realizes the accurate detection of the target insulator;then,the single insulator in the insulator string has a small target,and the occurrence of self-detonation has little effect on the overall characteristics of the insulator.From the perspective of improving the network model’s ability to detect small targets Improve Faster R-CNN,select DenseNet,which has a higher utilization of extracted feature information,for feature extraction,fuse high-level features with strong semantic information and low-level high-resolution features,and input them to the RPN network layer by layer to generate different sizes Large and small candidate frames are used for further classification and regression.At the same time,ROI Align is used to unify the size of the ROI feature area,avoid errors caused by the offset of the candidate frame in the feature mapping process,and realize the accurate detection of the insulator self-explosion defect;finally,the insulator The target location network and the insulator defect detection network perform joint detection,which realizes the end-to-end output of insulator location and defect detection.Since there is currently no professional aerial insulator data set at home and abroad,this paper will use data enhancement to expand the relevant insulator data collected in this article,and use the open source tool LabelImg to manually label the data,and establish the insulator data set used in this article.It is used to train and detect the network model before and after the improvement.The experimental results show that the improved average accuracy of the insulator target location network reaches 92.1%,which is an increase of 11.9%compared to before the improvement.After the improvement of the insulator defect detection network,the average accuracy is increased by 28.5%to 93.2%.The two network models are j ointly detected.,The final detection accuracy rate reached 98.6%,and the final detection time reached 338ms by adjusting the detection candidate frame,which basically met the detection requirements,and verified the feasibility,effectiveness and robustness of the improved network model in this paper.