Research On Arc Detection And Recognition Model Based On Improved MobileNet-SSD

The pantograph and contact network are important equipment to obtain electrical energy in high-speed trains,and the normal operation of the arch network system directly affects the quality of the train receiving current,so the arch network system arc detection is one of the means to ensure the normal operation of the train.For the current pantograph arc detection mostly through manual and contact arch network detection and other methods and there are low efficiency,slow detection speed and other problems,this paper based on deep learning theory proposed a use of target detection and identification of arch network arc detection method.In order to meet the needs of this paper target detection algorithm model training requires a large amount of data,for the total amount of data collected through high-definition cameras,the complex environment on the railroad route,the collection of images of varying quality,etc.,the construction of the arch network arc burning data set is proposed.The data set cleaning process is designed to screen out the images with inappropriate brightness,excessive covered area and high similarity.The image data enhancement library Imgaug and Open CV are invoked to enhance the data set,simulate the environmental and equipment factors in train operation,and improve the robustness of the data set.The image pantographs and arc-burning parts are labeled by Label Img software to obtain the arc-burning data set of the pantograph network.In order to improve the speed of pantograph combustion arc detection and solve the problem of real-time arc combustion detection,a stage of SSD target detection algorithm is selected.SSD algorithm model detection speed and high detection accuracy are applicable to the field of pantograph combustion arc detection.The model is loaded into the mobile embedded device carried by high-speed trains,which requires compression of the number of parameters of the model and reduction of computational effort.In this paper,we propose to replace the complex and large backbone network VGG-16 of the SSD model with the lightweight Mobile Net convolutional neural network.The input size is 300×300 pantograph arc data set images,and the average accuracy mean m AP of the Mobile Net-SSD model arc detection is 85.69% and the detection speed is 46.19 frames/s.In order to enhance the ability of Mobile Net-SSD algorithm model to detect smaller arc-burning targets as well as multiple arc-burning targets,the sensory field increase module and attention mechanism module are designed and added in the shallow Conv11 to enhance the ability of the model to extract and search feature information of small shallow targets and the ability of the model to learn feature information,respectively.Through the analysis of controlled experimental results,the improved Mobile Net-SSD algorithm model for arc-burning detection has an average accuracy mean m AP of 94.07% and a detection speed of41.85 frames/s,which substantially improves the detection accuracy of the model for small targets as well as multiple targets with a smaller reduction in detection speed.In order to test the ability of the model to identify arc-burning frames in the pantograph video,we first used manual screening of the video arc-burning frames and compared the detection results of the improved Mobile Net-SSD algorithm model with the actual results of the screening before and after the improvement.The improved Mobile Net-SSD algorithm model has the ability to detect arc-burning frames in pantograph video more accurately,and is suitable for arc-burning detection of pantographs in high-speed trains,which has high application value.