Research On Pollution Detection Technology Overhead Catenary Insulators On Complex Background Based On Computer Vision

As the railway medium and long-term planning and the implementation of the country for further development of the western region traffic facilities,construction and planning in the western region of the railway is increased,and as an important part of electrified railway traction power supply system of the catenary,play to the role of the electric locomotive power,The insulator of catenary is one of the most widely used parts of catenary,so that there is enough distance and insulation between the live conductor of catenary or between the live conductor and the earth,and the mechanical support and positioning of catenary,so the quality of the insulator is very important to the safe operation of the line.The insulators of catenary are exposed to the air and various environments outdoors for a long time.At this time,the power supply of electrified railways will have problems,resulting in more serious accidents.Therefore,the following studies are carried out in this paper:(1)The YOLOv3 network model was improved,after the sampling by spatial attention mechanism and channel attention mechanism combining mechanism of cascading double attention to fusion filtering characteristics,improve the ability of feature extraction,the introduction of gaussian function to the maximum inhibition method was improved,reducing the presence of keep out target miss rate,improve the accuracy of insulator detecting,Shorten the detection time of insulators,complete the identification of catenary insulators under complex background.(2)I optimize the semantic segmentation network structure by pruning,accelerate the semantic segmentation algorithm,and increase the local perception field of the convolution layer by reducing the resolution of the feature graph in the down-sampling part,so as to enhance feature extraction ability.Figure resolution in the portion of the sampling,reduction features,the segmentation results are expanding,connect multiple connected domain,increase the robustness of the algorithm,using samples of Gao Yuyi transposed convolution operation information,the low resolution characteristic figure resolution reduction process,within a certain range,adding more layers of transposed convolution divided semantic classification network structure,network structure into the convolution The pruned full convolutional semantic segmentation network was obtained to complete insulator segmentation,further reduce the detection range,and extract the insulator region.(3)The insulator surface sand,rust and other impurity detection,improved based on texture feature and possibilities mean clustering insulator contamination detection method,using the gray level co-occurrence matrix,respectively,to calculate the three types of filth texture feature vector of the space,through the method of clustering to soften points samples,established between sample and categories of uncertainty,Using principal component analysis algorithm for characteristic vector fusion and dimension reduction,improved possibilities mean clustering algorithm,relax the membership degree of restraint,makes the membership degree is no longer on a Shared or divided,through to the membership matrix and cluster center constantly iteration,until meet the conditions,in order to test the insulator possible each impurity,Improve the accuracy of insulator pollution detection.Through experiments,good results have been achieved in the pollution detection of insulators,and the sand and rust on the surface of insulators can be comprehensively detected,which proves that the method proposed in this paper is feasible.