Advances in High-speed Rail Technology

“Advances in High-speed Rail Technology” presents the latest and most cutting-
edge theories, techniques, and methodologies in the multidisciplinary field of high-
speed railways, focusing on advances and findings from China. This series includes
monographs, professional books and edited volumes from dedicated conferences
and workshops. All volumes are authored or edited by established experts in their
fields and undergo rigorous peer review, based on the editors’ preview and selection
and refereeing by independent experts. The intended audience includes researchers,
engineers, industrial practitioners, graduate students, and professionals. Topics of
interest in “Advances in High-speed Rail Technology” include, but are not limited
to: Infrastructure, Mobile Equipment, Communication & Signal, Traction Power
Supply, Operation Organization, etc.
Zhigang Liu · Wenqiang Liu · Junping Zhong

Deep Learning-Based
Detection of Catenary
Support Component Defect
and Fault in High-Speed
Railways
Zhigang Liu Wenqiang Liu
School of Electrical Engineering Department of Civil and Environmental
Southwest Jiaotong University Engineering
Chengdu, Sichuan, China Hong Kong Polytechnic University
Hong Kong, China
Junping Zhong
Department of Industrial and Systems
Engineering
Hong Kong Polytechnic University
Hong Kong, China

ISSN 2363-5010 ISSN 2363-5029 (electronic)

Advances in High-speed Rail Technology
ISBN 978-981-99-0952-0 ISBN 978-981-99-0953-7 (eBook)
https://doi.org/10.1007/978-981-99-0953-7

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2023
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

By the end of 2022, China’s high-speed railway operating mileage has exceeded
40,000 km and can circle the earth’s equator, which is the largest high-speed railway
operation network in the world at present. China’s high-speed railway development
has gradually transitioned from the construction stage to the operation and main-
tenance stage. With the gradual improvement of the requirements for the operation
quality of high-speed railways, higher requirements are put forward for the reli-
able operation of the equipment of the traction power supply system of high-speed
railways.
The catenary, also called overhead contact system (OCS), is the critical equipment
of traction power supply systems, mainly responsible for the power supply of multiple
power units (EMUs). And its reliability will directly affect the safe operation of
EMUs. Therefore, efficient detections and estimations of the catenary in high-speed
railways are significant and necessary for the whole railway’s maintenance to ensure
the catenary’s regular operation.
The catenary support device is the critical part of the catenary in high-speed rail-
ways and mainly includes 12 categories of support components, such as insulators,
isoelectric lines, screw pins, etc. The actual operation shows that under the influence
of complex factors such as external environment and train vibration, defects or faults
of support components perhaps occur, such as looseness, missing, fracture, and crack,
which can result in the decline of their structural reliability. They may lead to the
failure of the catenary system and endanger the operation safety of high-speed rail-
ways. Therefore, the detection of defects and faults in catenary support components
is vital for ensuring the safe situation of high-speed railways.
Due to the characteristics of the catenary support devices that have a complex
structure, many types of components, large-scale differences, fewer defect and fault
samples, and complex detection routes, it brings huge difficulty for the detection of
defects and faults of the catenary in high-speed railways. Aiming at these problems, a
series of solutions based on deep learning for detecting catenary support component
defects and faults are presented in detail in this book.
The book includes seven chapters. It begins with an overview of catenary detection
in high-speed railways in Chap. 1. The advance of deep learning is introduced in

v
vi Preface

detail in Chap. 2. Catenary support components and their characteristics in high-

speed railways are presented in Chap. 3. For the preprocessing of catenary support
components’ images, the corresponding methods and technologies are described in
Chap. 4. The two key parts for the detection of catenary support component defects
and faults, namely component positioning and defect and fault detection, which are
also the main contents of this book, are discussed in Chaps. 5 and 6, respectively. In
addition, Chap. 7 induces the detection methods based on the 3D point cloud.
This book systematically shows the latest research outcomes of catenary detec-
tion technologies in high-speed railways, especially the detection of catenary support
component defects and faults. Some methods or algorithms have been adopted in
practical engineering. They can provide important references and help the researcher,
scholar, and engineer with pantograph and catenary technology in high-speed rail-
ways. Different from traditional catenary support component detection methods of
catenary support components based on image processing, some advanced methods
in the deep learning field including convolutional neural networks, reinforcement
learning, generative adversarial network, etc., are adopted and improved in this book.
In this book, we try to indicate an increasing interest in deep learning applications
in the detection of high-speed railway catenary for actual engineering applications.
We hope to provide some references and help for other worldwide researchers who
engage in maintaining high-speed railways.

Chengdu, China Zhigang Liu

Hong Kong, China Wenqiang Liu
Hong Kong, China Junping Zhong

Acknowledgments This book is partly supported by the National Natural Science Foundation of
China (51977182, U143420) and the Sichuan Province Youth Science and Technology Innovation
Team under Grant (2016TD0012). It is also a summary of the achievements of our team for many
years, supported by doctoral and master graduate research, such as the works of Ye Han, Changdong
Wu, Liyou Wang, Kai Liu, Yang Lv, Yuyang Li, Cheng Yang, Changjiang Yang, Changjiang Li,
Xusong Liu, Weiping Guo, Qiao Li, and Junwen Chen.
Contents

1 Overview of Catenary Detection of Electrified Railways . . . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Catenary Detection and Monitoring System . . . . . . . . . . . . . . . . . . . . 3
1.3 Catenary Component Detection Technologies . . . . . . . . . . . . . . . . . . . 5
1.4 Catenary Parameter Detection Technologies . . . . . . . . . . . . . . . . . . . . 9
1.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Advance of Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Image Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.1 Handcrafted Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.2 Deep CNN Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Deep Learning Frameworks in Computer Vision . . . . . . . . . . . . . . . . 26
2.3.1 Image Object Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.2 Image Object Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.3 Image Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Other Deep Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.1 Deep Autoencoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.2 Generative Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . 36
2.4.3 Deep Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3 Catenary Support Components and Their Characteristics
in High-Speed Railways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Catenary System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 Catenary Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

vii
viii Contents

4 Preprocessing of Catenary Support Components’ Images . . . . . . . . . . 55

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Catenary Image Denoise with LWBCTCS . . . . . . . . . . . . . . . . . . . . . . 55
4.2.1 Denoise Method LWBCTCS . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2.2 Catenary Image Denoising Experiment and Analysis . . . . . . 60
4.3 Catenary Image Enhancement with Deep Learning . . . . . . . . . . . . . . 64
4.3.1 Catenary Image Enhancement Based
on Zero-DCE Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.2 Catenary Image Enhancement Based on Retinex
Theory and Generative Adversarial Networks . . . . . . . . . . . . 72
4.3.3 Catenary Image Enhancement Based
on Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5 Positioning of Catenary Support Components . . . . . . . . . . . . . . . . . . . . . 95
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Simultaneous Positioning of CSCs with Variant Deep CNN
Object Detection Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2.1 CSCDNET—CSC Positioning with Multiple Scale
Feature Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2.2 CSCNET—CSC Positioning with Unsupervised
Coarse Image Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2.3 CSCSIN—CSCs Positioning with Position
Relationship Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.3 Positioning Refinement of CSCs with Deep
Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3.1 Horizontal Positioning Box Refinement with Deep
Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.2 Obliqued Positioning Box Refinement with Generative
Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.3.3 Segmentation Boundary Refinement
with the CascadePSP Network . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6 Detection of Catenary Support Component Defect and Fault . . . . . . . 163
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.2 Messenger Wire Base Defect Detection and Evaluation Using
the Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.2.1 Suspected Crack Region Extraction . . . . . . . . . . . . . . . . . . . . . 164
6.2.2 Crack Detection of Messenger Wire Bases . . . . . . . . . . . . . . . 172
6.2.3 Analysis of Experimental Results . . . . . . . . . . . . . . . . . . . . . . 175
6.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.3 Insulator Defect Detection and Evaluation Based
on Autoencoder Networks (AE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Contents ix

6.3.1 Insulator Piece Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

6.3.2 Insulator Defect Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.3.3 Insulator Defect Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.3.4 Analysis of Experimental Results . . . . . . . . . . . . . . . . . . . . . . 181
6.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.4 Isoelectric Line Defect Detection and Evaluation Based
on Generative Adversarial Networks (GANs) . . . . . . . . . . . . . . . . . . . 188
6.4.1 Generative Adversarial Representation and Fault
Diagnosis of Isoelectric_Lines . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.4.2 Analysis of Experimental Results . . . . . . . . . . . . . . . . . . . . . . 196
6.4.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7 Detection of the Parameters of Catenary Support Devices
Based on 3D Point Clouds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
7.2 Geometry Parameter Detection of Contact Wires
with RANSAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.2.1 Acquisition and Preprocessing of the Catenary 3D
Point Cloud Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
7.2.2 Detection and Extraction of Contact Wires . . . . . . . . . . . . . . . 209
7.2.3 Detection of the Catenary Conductor
Height and Stagger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.2.4 Analysis of Experimental Results . . . . . . . . . . . . . . . . . . . . . . 215
7.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.3 Structure Parameter Detection with 3D Deep Segmentation
Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
7.3.1 Cantilever Component Separation . . . . . . . . . . . . . . . . . . . . . . 221
7.3.2 Cantilever Structure Parameter Measurement . . . . . . . . . . . . . 224
7.3.3 Analysis of Experimental Results . . . . . . . . . . . . . . . . . . . . . . 228
7.3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Abbreviations

3D PointCNN Three-dimensional point convolutional neural network

A3C Asynchronous Advantage Actor-Critic
ADNET Action decision network
AE Autoencoder
ASIFT Affine Scale Invariant Feature Transform
BEGAN Boundary Equilibrium Generative Adversarial Networks
BLS-GSM Bayesian Least Squares Gaussian Scale Mixture
BS Brace_sleeve
BSS Brace_sleeve_screw
CascadePSP Cascaded pyramid scene parsing network
CEB Curve enhancement block
CGAN Conditional Generative Adversarial Networks
CNN Convolutional neural network
CRN Classification and regression network
CSC Catenary support component
CSCDNET Catenary Support Component Detection Network
CSCNET Catenary support component network
CSCs Catenary support components
CSCSIN Catenary support component structure inference network
CT Contourlet transform
CV Curvelet transform
CVAE Conditional Variational Autoencoder
DAE Denoising autoencoder
DBSCAN Density-based spatial clustering of applications with noise
DCNN Deep convolutional neural network
DFB Directional Filter Bank
DL Deep learning
DRL Deep reinforcement learning
EMU Electric multiple unit
FAB Feature attention block
Faster R-CNN Faster Region-based Convolutional Neural Network

xi
xii Abbreviations

FCN Full convolutional networks

FEN Feature extraction network
FPN Feature pyramid network
FPS Frame per second
GAN Generative adversarial networks
GIoU Generalized Intersection over Union
GRU Gated Recurrent Unit
GT Ground truth
HOG Histogram of Oriented Gradient
ICEN Illumination component enhancement Network
IDN Image decoupled network
IoU Intersection over union
IP_RANSAC Improved projection random sample consensus
LBP Local binary patterns
LEC Light-enhancement curve
LP Laplacian pyramid
LSCCRN Large-scale classification and regression network
LSCSCs Large-scale catenary support components
LWBCT Lifting wavelet-based contourlet transform
LWBCTCS Lifting wavelet-based contourlet transform with cycle shift
invariance
LWT Lifting wavelet transform
MABR Minimum area bounding rectangle
mAP Mean average precision
mBA Mean boundary accuracy
MDP Markov decision process
mIoU Mean intersection over union
MPN Mask prediction network
MSE Mean squared error
NMS Non-maximum suppression
PCA-SIFT Principal component analysis scale invariant feature transform
PGGAN Progressive growing generative adversarial networks
PixelRL Pixel reinforcement learning
PR curve Precision-recall curve
PSNR Peak signal noise ratio
PSPNET Pyramid scene parsing network
RANSAC Random sample consensus
RCCAEN Reconstruction and classification convolutional autoencoder
network
RCDN Reflection component denoising network
ResNet-101 Residual network with 101 layers
RL Reinforcement learning
RMC Reward map convolution
RoI Region of interest
RPN Region proposal network
Abbreviations xiii

SAE Sparse autoencoder

SB Sparse block
SC_LCCP Slope Constrained Locally Convex Connected Patches
SIFT Scale Invariant Feature Transform
SIN Structure inference network
SNR Signal noise ratio
SPCCD Structure parameters of catenary cantilever devices
SSCCRN Small-scale classification and regression network
SSCSCs Small-scale catenary support components
SSD Single-Shot MultiBox Detector
SSIM Structural similarity
SURF Speeded Up Robust Features
SVM Support Vector Machine
TOL-Framework Tighter Oriented Localization Framework
TPH-YOLOv5 Transformer Prediction Head You Only Look Once Version5
VAE Variational autoencoder
VGG16 A CNN with 16 layers (Developed by the Visual Geometry
Group)
WGAN Wasserstein generative adversarial networks
WGAN-GP Wasserstein generative adversarial networks with gradient
penalty
YOLACT You Only Look At CoefficienTs
YOLO You Only Look Once
Zero-DCE Zero-reference Deep Curve Estimation