Research On Pantograph-catenary Arcing Detection Method Based On Image Deep Learning

As the key to the power transmission of high-speed trains,the pantograph-catenary system is the weakest link in the whole vehicle subsystem,and its health work is easily restricted by problems such as arcing.The non-contact arcing image detection technology has been widely promoted on the whole railway.Due to the complex and changeable operating background of high-speed trains and the similarity of arcing to other light sources,it is difficult for traditional image processing methods to obtain better detection results.The convolutional neural network under the deep learning framework has a qualitative improvement in image feature extraction and expression compared with traditional image processing methods.With the help of deep learning technology,the problem of how to accurately,efficiently and intelligently detect arcing under complex operating backgrounds will be effectively solved.The thesis focuses on the image deep learning-based arcing detection method,the details are as follows:First,designed the arcing region extraction method based on the overall offset of the pantograph head region,and improved the deep learning target detection model Faster RCNN from the perspective of image feature extraction and utilization,and by combining the Res Net50 and FPN to get multiscale feature maps,and the Soft-NMS method was used instead of the NMS method for proposal suppression,and the Ro I Align layer was used to replace the Ro I pooling layer to unify the feature map size.After the improved Faster RCNN model accurately located the pantograph head region,simultaneously offset the pantograph head location box from the four directions of up,down,left and right,and obtained the arcing region.Secondly,reset the number of fully connected layers and the number of nodes in each fully connected layer of the arcing image classification network VGG16,and added a BN layer and a dropout layer after each the fully connected layers and classification layers to prevent the network from overfitting.Then,used the improved Faster RCNN model to locate the arcing.Finally,a web application simulation detection platform based on the Flask framework was built,and the video test was carried out.The test results show that,in the location of the pantograph head region,the average accuracy of the improved Faster RCNN is 0.899,improved by 0.067 comparing with the original Faster RCNN model,the average accuracy when the Io U value set 0.85 is 0.972,improved by 0.039 comparing with the original Faster RCNN model,and the detection speed of a single image is 8.7 times improved.The improved VGG16 network achieves all correct classifications in the test dataset of 249 images of the arcing region.In arcing location,the average accuracy of the improved Faster RCNN model is 0.513,improved by0.156 comparing with the original Faster RCNN model,and the average accuracy when the intersection ratio set 0.85 is 0.377,also improved by 0.084 comparing with the original Faster RCNN model.Verified by the simulation testing platform,the test results show that the algorithms are logical and accurate,which provide idea and reference for engineering applications.