Research On Foreign Object And Defect Detection Algorithm Of Overhead Catenary System Based On Deep Convolution Neural Network

China railway transportation regulations stipulate that the Overhead Catenary System(OCS)is an important equipment of electrified railway.The OCS of high-speed railway is a transmission line erected over the railway to provide electric energy to electric locomotives,which is mainly composed of positioning,support device and contact suspension.Due to the characteristics of long OCS line,numerous equipment on OCS and no standby,once foreign matters invade the OCS,such as bird’s nest or a part falls off,it will directly threaten the safe and stable operation of traction power supply system.Manual patrol inspection has high cost,low efficiency,unstable detection results affected by human factors,and labor intensity is too high.With the development of computer image technology,intelligent inspection image technology has become a new and efficient OCS state detection method.In recent years,the research of convolutional neural network has gradually deepened,which has brought new vitality to the field of computer vision.It has played a great role in security,the internet and other fields,and has become one of the most promising deep learning methods.In the field of OCS image detection,researchers in the railway industry mainly use traditional image algorithms and some direct and general deep learning algorithms for OCS image detection and recognition.OCS image detection technology is an important subject.Taking OCS bird’s nest and U-shaped hoop nut as examples,this paper studies OCS bird’s nest detection algorithm based on YOLOv3 and OCS U-shaped hoop nut detection algorithm based on SSD.The main work is as follows.Firstly,the digital image processing technology is used to preprocess the images of OCS bird’s nest and U-shaped hoop by graying,wavelet denoising,Wiener filtering and image enhancement,and then the detection models are designed for bird’s nest and U-shaped hoop respectively.Considering the characteristics of different sizes and unclear edges of the bird’s nest,firstly,a priori frame suitable for the size of the bird’s nest of the OCS is obtained by clustering,and the model is accelerated to fit the bird’s nest prediction frame.Then,the lightweight Efficient Net network is used as the feature extraction network to improve the feature extraction ability of the bird’s nest and shorten the detection time.Finally,after the feature extraction network,the spatial pyramid pooling module is added to fuse the global and local features to extract the multi-scale features of the bird’s nest.The experimental results show that compared with other deep learning models,the ESK-YOLOv3 model in this paper has higher detection accuracy and speed for OCS bird’s nest detection with complex background,and the map is 98.86%,which verifies the effectiveness and reliability of the algorithm in this paper.Considering that the target of U-shaped hoop nut is small,difficult to detect and the merger of detection speed and accuracy,firstly,Res Net network with stronger representation ability is used as the feature extraction network,and the attention mechanism module is added to the trunk of Res Net to improve the detection ability of the network for small targets,and then the nut small target data set is optimized by median filter and singular value decomposition,Finally,a priori frame scale suitable for the size of U-shaped hoop nut is designed,which can better cover the size distribution of U-shaped hoop nut of OCS,so that the model can be fully trained.The experimental results show that compared with other deep learning models,the SRC-SSD model in this paper has higher detection accuracy for the detection of OCS u-hoop nuts with complex background,and the map is 91.06%,which verifies the effectiveness and reliability of the algorithm in this paper.