| The pantograph-catenary system(PCS)is an important part of the traction power supply system for high-speed railway,and the dynamic characteristics and service status of PCS are directly related to the safety of high-speed railway operations.The modules and components of catenary are mostly exposed to open-air condition and work under dynamic excitations all the time,then long-term vibrations of catenary-pantograph system,extreme weather and environmental factors have very serious impacts on catenary.These factors would probably result in not only damaging consequences in a direct or indirect way,but also plenty of local anomalies,which can largely reduce catenary reliability and place them at serious risk of unavailability.Therefore,obtaining the operating status and health status of the PCS in time is a prerequisite for effective maintenance of the PCS,which is of great significance to ensure the safe and stable operation of high-speed railways.At present,the continuous status monitoring of PCS can be achieved through the state-of-the-art inspection systems—high speed railway power supply safety monitoring system(6C).However,the main problem currently faced is how to accurately find out a variety of anomalies from the monitoring data of high-speed railway power supply system.To this end,this thesis conducts research from multiple perspectives to realize the intelligent detection of abnormal state for PCS.First,this thesis introduces the research background and significance of intelligent inspection for high speed railway,and studies the current situation of PCS dynamic characteristics monitoring and catenary component defect detection based on computer vision.Inspired by the current research,the focus of this thesis is on the intelligent monitoring and inspection method for the PCS based on the deep learning.Secondly,aiming at monitoring the dynamic characteristics for PCS,the principle of computer vision measurement is studied to build a measurement model based on PCS monitoring system.The model can be used to measure dynamic geometric parameters of catenary.However,due to the complex background in the monitoring video,accurately extracting the contact points has become the key to the measurement process.Chapter 3 analyzes the image characteristics of the PCS monitoring video and proposes a detection method,which integrates deep convolutional neural networks and traditional image processing methods to deal with contact point detection in complex background.Experimental results show that the deep convolutional neural networks greatly improves the robustness of the detection algorithm and the ability of adapting to complex environments.In addition,the factors of the detection error are analyzed,and a compensation method of vehicle vibration is proposed to further improve the detection accuracy.The effectiveness of the compensation method is verified through experiments.Then,for the defect detection of catenary components,this thesis mainly studies the detection method based on visual recognition.Since the defect detection of catenary components usually faces the problem of sample imbalance,that is,the extreme shortage of defect samples,which brings challenges to the application of deep learning method.Chapter 4 researches the deep convolutional neural network framework based on unsupervised learning,and proposes a way to recognize defective component only by learning the distribution of normal component sample.This method combines a deep denoising autoencoder,a deep generative adversarial network,and a deep classifier.The denoising autoencoder is used to map the normal sample distribution into the latent space.The deep generative adversarial network generate pseudo-defect samples from the pitted latent space.And the deep classifier learns the strict boundaries of normal samples from normal and pseudo-defect samples to identify defective component.The proposed method is verified on the catenary insulator data set,and the test results show that the method has good defect recognition ability.Finally,considering the catenary components and defects often have fixed structural features,chapter 5 proposes a kind of structure-awareness method for detecting crack defects in components.This method also does not need to learn the defect samples,but manually design the defect criterion based on the structural features of the component in the image.The key of the structure-awareness method is to accurately extract structural features.To solve this problem,the thesis proposes a high-precision instance segmentation method based on dynamic convolutional neural network to extract structural features of components.This method avoids the auxiliary merging operation of the instance feature maps,thereby reducing the introduction of noise,which can concurrently accommodate instance segments with different sizes.Furthermore,in order to extract crack structure features,an adaptive crack structure perception algorithm based on Hessian matrix is proposed,which can accurately extract crack structure information under different illumination conditions.The experimental results show that the proposed method can effectively detect the crack defects of the component.In this thesis,deep learning is applied to the field of computer vision measurement and visual recognition.A deep learning based computer vision framework is established.It has important practical value for the study of the intelligent detection for PCS. |
