Research On The State Detection Algorithm Of Key Components Of High-speed Railway Catenary Based On Deep Learning

Catenary is an important part of high-speed railway traction power supply system,and its safety status detection is the top priority of railway safety.For this reason,my country has specially formulated the "General Technical Specifications for Railway Power Supply Safety Inspection and Monitoring System(6C System)",which is responsible for comprehensive inspection and monitoring of the traction power supply system of high-speed railways.Among them,the C4 catenary suspension status detection and monitoring device(4C system)is a subsystem of the6 C system,which is responsible for the high-precision imaging detection of the various components of the catenary suspension system,and guides the elimination of hidden dangers of catenary failure.At present,the 4C system has the function of initially identifying catenary components,but missed and misdetected frequently,so it is often used as an auxiliary means for catenary detection,and its detection accuracy needs to be further improved.Based on the characteristics of the parts in 4C image data that are often small,have a certain rotation angle,and the data samples are extremely unbalanced(abnormal samples are much smaller than normal samples),this paper divides the abnormal detection tasks of key parts of the high-speed railway catenary into Two parts of the study: the first part locates the area of the parts to be inspected;the second step performs part abnormality detection judgment on the part area located in the previous step.The specific research is mainly divided into the following parts:(1)In view of the complex and changeable background of railway 4C images,huge size,small proportion of target objects and most of them have rotation angles,this paper proposes an improved Faster R-CNN detection algorithm,which changes the basic backbone network VGG16 to more detection capabilities.Strong Res Net-50 network,while adding a multi-scale detection method to improve the detection accuracy of small objects,and finally according to the characteristics of the object’s rotation angle,add an angle parameter to the target detection frame,making the position recognition of the target object more accurate and improve Improve the detection accuracy.(2)Aiming at the problem that the proportion of the target object to be detected in the catenary is too small in the 4C image,this paper proposes to use the triple attention mechanism to significantly enhance the area of the target object on the image.In the process of perception,humans tend to selectively focus on a part of the area of interest in the given information and ignore other parts.This attention focus mechanism helps to improve the perceptual information while preserving its contextual information.Therefore,in this paper,the triple attention mechanism is added to the general target detection algorithm to enhance the object detection accuracy of the target detection algorithm,and solve the disadvantages of separate calculation of the traditional attention mechanism of channel attention and spatial attention mechanism,and improve Interaction of space dimension and channel dimension.(3)Aiming at the feature of few abnormal samples in 4C image data,this paper uses a semi-supervised method based on confrontation training to perform anomaly detection from the perspective of practical engineering.The specific method is to use the improved GANomaly network for adversarial training.In the training phase,only the normal sample data image is input,and the generator network generates the reconstructed image of the input data and the latent feature vector of the high-level space,so that the generator network model can learn normally The data distribution of the sample.When the input data is an abnormal sample image,the reconstructed image generated by the generator network and the latent feature vector of the high-level space will be significantly different from the reconstructed image and the latent feature vector of the high-level space generated when the normal sample is input.The device network sets an appropriate threshold to determine whether the input image is abnormal.In this paper,the network architecture of the original GANomaly algorithm is improved,and layer-jumping connections are added to improve the performance of the network.Based on the above research,the environment required by the algorithm is set up in the laboratory.All the railway 4C image data sets are used to verify the algorithm of this paper,and other advanced algorithms are used for comparison experiments.After a large amount of measured data image verification,the algorithm designed in this paper is reasonable,In line with practical engineering applications,can effectively detect abnormal parts of the railway catenary.The accuracy rate can basically reach the level of manual detection,which can effectively improve work efficiency.