A Recognition Software For HSR Catenary Parts Defects Based On Deep Learning

In recent years,the construction of high-speed railways in China has been continuously developing,and the mileage of high-speed railways has been increasing.This has also led to an increasing amount of analysis of railway contact network detection data within the jurisdiction of high-speed railways.Dependence on traditional manual data analysis methods has led to lengthy analysis work cycles,while the processing,analysis,and assessment of defects are constrained by factors such as personnel knowledge,experience,and work status.There is a risk of affecting the safe and stable operation of high-speed trains.This article focuses on the image data collected by the high-speed railway catenary suspension status detection and monitoring device(4C),and applies deep learning methods to the intelligent automatic defect detection process.The following work has been studied and completed:Firstly,in response to the detection of small targets and complex backgrounds in4 C image data,this paper proposes the use of an improved YOLOv5 algorithm for automatic recognition of defects in catenary construction.A multi-scale detection head is introduced to achieve multi-scale object detection,mainly by adding a new detection head to detect small targets while keeping the YOLOv5 s three-scale detection head unchanged,in order to optimize the recognition of catenary construction.Introducing the CBAM model to improve the Neck network,while paying attention to features in both spatial and channel dimensions,optimizing issues such as high noise in image data,small target scale,dense arrangement,and background interference,better extracting key features,clarifying the location of small targets,and preventing network speed and performance from being affected by improvements.The multi-scale fusion network based on Bi FPN is introduced to optimize the feature fusion method,and the deep and shallow features are fused to make the obtained feature information more accurate,comprehensive and rich.While retaining the advanced Semantic information,the detailed information is used to improve the detection ability of small targets.Secondly,based on the improved DCGAN model,data augmentation is carried out on the defect samples of overhead contact system parts to expand the number of defect samples.For the use scenario in this paper,the robustness and generalization performance of the model are improved by abandoning the symmetric form of discriminator and generator structure,using batch standardization for each layer of the generator,using Sigmaid activation function for the output layer of the generator,and using Re LU function for other layers.Finally,the model was trained using 470000 4C detection image data from a power supply section of a certain railway bureau as the dataset.Two criteria,classification and bounding box regression,were used to evaluate the performance of the algorithm.The performance of different algorithms was compared,and it was verified that the accuracy of the new model was higher than other YOLO series algorithms,indicating that it has more advantages in identifying faults in contact network components compared to other models.At the same time,the main program and subprogram flow diagram of the high-speed railway overhead contact line 4C defect detection software based on deep learning were designed,and the front and back programs and corresponding subprograms were written.Experimental analysis shows that the detection accuracy of this software meets engineering needs,and it has advantages such as stable operation,simple operation,and user-friendly interaction.