Research On Deep-learning-based Catenary Defect Detection Method For High Speed Railway

The pantograph-catenary system of the high-speed railway is the only way for the EMU to obtain electrical energy,and its operating state is directly related to the operational safety of the high-speed railway.With the continuous increase of train speed,the impact of pantographs on the catenary continues to intensify,which will lead to catenary components defects such as looseness and crack that seriously affect the safety of the high-speed railway.How to detect these kinds of catenary defects effectively is of great significance for the safety of high-speed railway operation.With the implementation of the 6C system,the maintenance of high-speed rail catenary has gradually evolved from traditional manual inspection to intelligent inspection.However,due to the complex structure of the catenary system,the intelligent defect detection still faces many challenges,such as complex environment,subtle defect characterization and high reliability requirements.Therefore,this thesis wil conducts research from multiple perspectives to improve the intelligence and reliability of the catenary inspection system.In the catenary defects detection process,localizing the catenary components from the complex background is a necessary condition for defect detection.Chapter 2 analyzes a cascade object detection frameworks based on deep convolutional neural networks,and performs key components detection experiments on the catenary image dataset.The proposed framework has achieved good performance.On this basis,this thesis focuses on the defect detection of catenary components.First,the defect detection of catenary components faces the problem of defective samples scarcity.How to train a classifier with good generalization performance using limited number of defective samples is an important issue for catenary components detection.Therfore,Chapter 3 uses a comprehensive transfer learning strategy to extract multiple kinds of image features through multiple pre-trained deep convolutional neural networks.On this basis,with limited number of training samples,the generalization ability of the classifier can be further improved by integrating multiple support vector machine classifiers trained using different kinds of image features.Finally,the validity of the proposed method is verified by experiments.Secondly,if defective samples are extremely scarce,the classifier-based defect detection method is no longer feasible.Chapter 4 converts defect detection into anomaly detection problems,and solves it through deep unsupervised learning method that does not require defective samples for training.The proposed method integrates a deep classifier and a deep denoising autoencoder into the same deep multi-tasking architecture.It not only does not require defective samples for training,but also can achieve the insulator segmentation and defect detection simultaneously.However,when the defect area is small and the gray value of the defect area is only slightly different from that of the normal area,the proposed method cannot detect the defect.Then,considering all catenary components have certain geometric characteristics and their defects can be evaluated by their geometric characteristics changes,Chapter 5 proposes to obtain components' geometric characteristics through image segmentation,and then use them to detect their defects.The proposed method cannot only overcome the problem of defective samples scarcity,but also provides the possibility to quantitatively measure defects.In this method,the accurate and reliable segmentation of components is the key to defect detection.This method uses a Bayesian fully convolutional network that fuses features of different levels of the backbone network to segment components,which cannot only accuratly segment the components,but also be able to evaluate the uncertainty of segmentation.The experimental results show that the defect detection method based on Bayesian fully convolutional segmentation network works well.Finally,Chapter 6 delves into the impact of data distribution shift on the reliability of deep learning models and the related solutions.Almost all the existing defect detection methods follow a fixed data distribution assumption: once be trained,the learning model no longer changes.However,in catenary defect detection,the data distribution is not fixed.The reason is that the visual inspection system not only needs to work on different railway lines,but also is inevitably affected by the movement of inspection vehicles and the changes in lighting intensity.Therefore,the learning of the detection model should not be a one-time event,but a continuous process that can adapt to the changes in data distribution.In view of this,we propose a defect detection framework based on adaptive deep neural networks to cope with data distribution shift,thereby ensuring detection reliability.This method defines an unreliability index that combines model uncertainty and prior knowledge,and designs an adaptive strategy based on the unreliability index to enable the segmentation network to adapt to the data distribution shift.The experimental results show that the method can achieve self-adaptation according to the change of data distribution,and ensure the reliability of decision-making.The thesis forms a technical framework of catenary defect detection based on deep learning,which provides a reference for further research on the intelligentization of catenary defect detection.