Research On Self-explosion Detection Of Transmission Line Insulators Based On Deep Learning

Electric energy is an important basic energy for national economic development and national living standards,the safety of transmission lines and related devices is the premise of stable and continuous work of the power system,in order to ensure the reliability of power supply,operators regularly inspect transmission lines is an effective way to find faults.With the rapid development of China’s power industry,coupled with the complex terrain,vast territory,and large scale of power grid coverage,intelligent power grid construction and autonomous inspection system all put forward higher requirements for the accuracy and speed of line inspection and defect detection tasks.As the key device of support and insulation,the safety detection of transmission line insulators is an important research content to ensure the safety of transmission line operation,and the existing image processing algorithms can no longer meet the detection requirements.The development of deep learning theory and computer vision technology provides feasible ideas for the field of power system,and gradually becomes the mainstream research direction of line inspection.Therefore,the research on self-explosion insulator detection of transmission lines based on deep learning algorithm conforms to the development trend of power grid,and has certain practical application value.In this paper,the application of deep learning correlation algorithms in transmission line detection tasks is mainly studied,and targeted improvements are made according to the relevant characteristics of self-explosion insulators studied in this topic,so as to improve the performance of the detection model and complete the self-explosion identification.The main research contents are as follows:(1)Expansion of image samples for self-explosion insulators.Due to the particularity of the power industry,there is currently no high-quality public transmission line insulator dataset with diverse samples,based on the limited self-explosion images obtained,in order to ensure the rationality and effectiveness of model training,traditional image transformation technology and generative adversarial network related algorithms are adopted to enrich the data.In order to further expand the number and quality of images,this paper studies the related theories of DCGAN and CycleGAN,and introduces the two algorithms into the dataset production process,and uses DCGAN to achieve the expansion of quantity under the premise of ensuring the quality of self-explosion images.CycleGAN is used to exchange image style domains to improve the richness of datasets.Finally,the self-explosion insulator dataset required for this paper is obtained.(2)Mask R-CNN realizes the detection of insulator self-explosion.In view of the subjectivity of traditional image detection methods on self-explosion defect detection,in order to improve the accuracy and speed of the detection task,the two-stage semantic segmentation and detection network Mask R-CNN is used to realize the detection of small areas of insulator self-explosion,which can not only realize pixel-level positioning,but also directly segment the self-explosion area from the picture,but the complex background of the image and the small target scale make the network detection effect less than ideal.Therefore,this paper makes the following improvements to Mask R-CNN:(1)The attention mechanism CBAM is added to the feature extraction network to strengthen the effective extraction of small target features;(2)Improve the fusion strategy of feature fusion network for multi-scale features,add parallel fusion paths to fuse features of different sizes and scales,and design an overall feature fusion module to further improve the integrity of features in self-explosion areas;(3)Use GIoU for target similarity screening to avoid interference caused by high likelihood background on detection;(4)The mask branch loss is replaced with the Tversky loss,and the small number of suppressed samples causes overfitting to model training.The results show that the improved Mask R-CNN has significantly improved the detection accuracy of small-scale insulator self-explosion defects in complex background scenes.(3)Lightweight CenterNet to realize the detection of self-explosion insulators.CenterNet is designed based on the idea of anchor-free,the network can accurately detect small-scale targets,and alleviate the phenomenon of high time complexity of the model caused by a large number of anchor box screening,but due to the lack of feature fusion stage,the target information is lost to a certain extent.The details are as follows:(1)The feature extraction network is replaced with the lightweight structure GhostNet,which reduces the complexity of convolution operations;(2)In order to ensure the integrity of feature extraction in a small area of self-explosion,RFB is introduced to strengthen feature extraction for the feature layer obtained by the lightweight backbone network,so that the model pays attention to the feature information with higher correlation with the target;(3)A feature fusion network is constructed,and an effective feature layer with target multi-level information and global information is obtained for regression and classification,so as to improve the detection accuracy of lightweight networks.The results show that the lightweight design of the network only loses a small amount of accuracy,which realizes the effective reduction of the network scale,and the model design idea is feasible.The improved Mask R-CNN can achieve accurate detection of insulator self-explosion,the lightweight CenterNet not only performs well in detection accuracy,but also has certain advantages in model size and model operation.The insulator self-explosion detection technology based on deep learning studied in this paper has application value.