Research On Transmission Line Recognition And Insulator Defect Detection Method Based On Deep Learning

The long-term exposure of transmission lines in the wild environment can easily lead to many hidden dangers such as foreign matter attachment,corrosion,broken strands,insulator damage,flashover and so on,which pose a great threat to the safety and stability of power transmission.Therefore,regular inspection of transmission lines and timely detection of faults in key components of transmission lines are of great significance to ensure the safety of power transmission.The traditional inspection method of transmission line is mainly based on manual inspection,which has low efficiency,high cost and high false detection rate.How to identify transmission lines efficiently,quickly and accurately in the complex and diverse backgrounds is the premise to realize the defect detection of transmission line components.In this paper,the aerial images of power transmission lines and the collected insulator defect images in the actual scene are taken as the research objects,and the deep learning method is used to realize the efficient identification of power lines and automatic detection of insulator defects.The specific work includes the following four aspect:(1)The advantages and disadvantages of domestic and foreign transmission line identification and insulator defect detection methods are analyzed and compared,and the deep learning-based semantic segmentation method and target detection method are used to identify the transmission line and insulator defect data set in this experiment.(2)UAV aerial photography is used to collect power line video data,and the video data is stored and processed in frames.Gaussian pyramid downsampling is used to reduce the image resolution,and Mosaic data enhancement is used for the power line image to expand the image.Three kinds of insulator defect datasets are collected,and the high-resolution images are cropped by overlapping sliding window segmentation method to crop out small target insulator images.Augmentor and Mosaic data enhancement methods are used to expand the insulator images,and the data sets of power lines and insulator defects are marked with the marking software to make them into VOC format.(3)According to the requirements of power line identification task,a power line identification model based on MobileNetV3-PSP codec network is constructed.The lightweight network MobileNetV3 is selected as the encoder backbone network,and shortcut links and deep stacking are drawn in the shallow layer of the encoder to enhance the ability of shallow feature reuse.The PSP module is added to the decoder to aggregate the semantic information of power lines in the global image and enhance the feature extraction capability.Use depthwise separable convolution instead of ordinary convolution to achieve network lightweight.The skip link structure is used to cascade the encoder and decoder to achieve multi-scale feature fusion,and transfer learning is used to speed up network training convergence.In order to understand the internal learning process of the network more clearly,visualize the feature layers of different depths of the network.Experiments show that the power line recognition model constructed in this paper can quickly and accurately identify power lines,with MAP and MIOU reaching 94.37% and 86.95%,respectively,and the recognition speed reaching 31 frames per second,which are better than other mainstream semantic segmentation networks.The idea of control variables is further used to compare the ablation experiments,which verifies the validity of the transmission line identification model.(4)The YOLOV4 network is improved to adapt to the task of insulator defect detection,and a variety of attention mechanism strategies are introduced into the YOLOV4 backbone feature extraction network CSPDarknet53.Through quantitative analysis,the CBAM3 strategy is finally used to improve the backbone network,and the feature areas that the network paid more attention to under different strategies are visually displayed by using the thermal map.The K-means algorithm is used to cluster the anchor box size suitable for the insulator features,which speeds up the convergence speed of network training.In PANet,the CSPlayer structure is used to replace the quintic convolution module,and the depthwise separable convolution is used to replace the cubic convolution module.The network parameters The amount is greatly reduced,and the model is lightweight.The SPP module is added to integrate the insulator defect features of multiple receptive fields to improve the detection accuracy.Increase the proportion of category loss in the total loss function to improve the multi-category classification accuracy.The use of flexible non-maximum suppression Soft-NMS instead of ordinary NMS solves the problem of missed detection of small target insulator defects with a high degree of coincidence.The improved YOLOV4 insulator defect detection network is compared by ablation experiments to verify the effectiveness of the improved network.Comparing it with the mainstream target detection network,the results show that the detection accuracy and speed of the improved YOLOV4 network are better than other networks,and the mAP and detection time are 92.26% and 19.82 ms respectively,which meet the requirements for accuracy and real-time performance in insulator defect detection tasks.