Artificial Intelligence in Construction Engineering and Management

Lecture Notes in Civil Engineering
Limao Zhang
Yue Pan
Xianguo Wu
Mirosław J. Skibniewski
Artificial
Intelligence in
Construction
Engineering and
Management
Volume 163
Series Editors
Marco di Prisco, Politecnico di Milano, Milano, Italy
Sheng-Hong Chen, School of Water Resources and Hydropower Engineering,
Wuhan University, Wuhan, China
Ioannis Vayas, Institute of Steel Structures, National Technical University of
Athens, Athens, Greece
Sanjay Kumar Shukla, School of Engineering, Edith Cowan University, Joondalup,
WA, Australia
Anuj Sharma, Iowa State University, Ames, IA, USA
Nagesh Kumar, Department of Civil Engineering, Indian Institute of Science
Bangalore, Bengaluru, Karnataka, India
Chien Ming Wang, School of Civil Engineering, The University of Queensland,
Brisbane, QLD, Australia
Lecture Notes in Civil Engineering (LNCE) publishes the latest developments in
Civil Engineering - quickly, informally and in top quality. Though original
research reported in proceedings and post-proceedings represents the core of
LNCE, edited volumes of exceptionally high quality and interest may also be
considered for publication. Volumes published in LNCE embrace all aspects and
subfields of, as well as new challenges in, Civil Engineering. Topics in the series
include:
• Construction and Structural Mechanics
• Building Materials
• Concrete, Steel and Timber Structures
• Geotechnical Engineering
• Earthquake Engineering
• Coastal Engineering
• Ocean and Offshore Engineering; Ships and Floating Structures
• Hydraulics, Hydrology and Water Resources Engineering
• Environmental Engineering and Sustainability
• Structural Health and Monitoring
• Surveying and Geographical Information Systems
• Indoor Environments
• Transportation and Traffic
• Risk Analysis
• Safety and Security
To submit a proposal or request further information, please contact the appropriate

Springer Editor:
– Pierpaolo Riva at [email protected] (Europe and Americas);
– Swati Meherishi at [email protected] (Asia - except China,
and Australia, New Zealand);
– Wayne Hu at [email protected] (China).
All books in the series now indexed by Scopus and EI Compendex database!
More information about this series at http://www.springer.com/series/15087

Limao Zhang · Yue Pan · Xianguo Wu ·
Mirosław J. Skibniewski
Artificial Intelligence
in Construction Engineering
and Management
Limao Zhang Yue Pan
School of Civil and Environmental School of Civil and Environmental
Engineering Engineering
Nanyang Technological University Nanyang Technological University
Singapore, Singapore Singapore, Singapore
Xianguo Wu Mirosław J. Skibniewski

School of Civil and Hydraulic Engineering Department of Civil and Environmental
Huazhong University of Science Engineering
and Technology University of Maryland
Wuhan, Hubei, China College Park, MD, USA
ISSN 2366-2557 ISSN 2366-2565 (electronic)

ISBN 978-981-16-2841-2 ISBN 978-981-16-2842-9 (eBook)
https://doi.org/10.1007/978-981-16-2842-9
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore
Pte Ltd. 2021
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
The construction industry worldwide has been a late bloomer to adopting digital
technology, where construction projects are predominantly managed with a heavy
reliance on the knowledge and experience of construction professionals. Artifi-
cial intelligence (AI) works by combining large amounts of data with fast, itera-
tive processing and intelligent algorithms (e.g., neural networks, process mining,
and deep learning), allowing the computer to learn automatically from patterns or
features in the data. It provides a wide range of solutions to address many chal-
lenging construction problems, such as tunnel-induced damages, risk estimates, and
defect detection. A tremendous transformation has taken place in the past years with
the emerging applications of AI in the construction industry. This enables industrial
participants to operate projects more efficiently and safely, not only increasing the
automation and productivity in construction but also enhancing the competitiveness
globally.
This book highlights the latest technologies and applications of AI in the domain
of construction engineering and management (CEM). Various AI techniques have
been developed to make machines mimic human cognitive processes in terms of
learning, reasoning, and self-correcting. Relying on the important AI approaches,
we put emphasis on nine hot research topics of AI in CEM, including “Knowledge
Representation and Discovery” (second chapter), “Fuzzy Modeling and Reasoning”
(third chapter), “Time Series Prediction” (fourth chapter), “Information Fusion”
(fifth chapter), “Dynamic Bayesian Networks” (sixth chapter), “Process Mining”
(seventh chapter), “Agent-Based Simulation” (eighth chapter), “Expert Systems”
(ninth chapter), and “Computer Vision” (tenth chapter). Different AI approaches
have been proposed to achieve various kinds of functions which are beneficial to
CEM in terms of automation, safety, reliability, and digitalization. These applica-
tions serve as evidence to verify that AI techniques have created tremendous value
in revolutionizing the construction industry, leading to a more reliable, automated,
self-modifying, time-saving, and cost-effective process of CEM.
This book expects to (1) reveal complexities in conducting CEM activities and
unlock the potential of incorporating AI in supporting and improving CEM; (2)
display procedures and algorithms in a detailed manner on how to use AI to discover
knowledge and address CEM challenges; (3) present evidence-based practices to
v
vi Preface
highlight the benefits of AI techniques in providing intelligent solutions by learning

from multiple data resources in an automatic, efficient, and reliable manner; and (4)
improve the understanding of the current development and prominent future works
in AI-enabled CEM and promote the more widespread leverage of AI in CEM.
This book is an interesting reference for scientists, engineers, and graduate
students in various disciplines, including but not limited to civil engineering, artificial
intelligence, system science, mathematics, etc.
Singapore Limao Zhang

Singapore Yue Pan
College Park, USA Mirosław J. Skibniewski
Wuhan, China Xianguo Wu
February 2021
Acknowledgements It is our sincere desire to thank the anonymous peer reviewers for this work.
We further want to thank Dr. Leilei Chang, Mr. Kai Guo, Ms. Ying Wang, and Ms. Weiyi Chen,
who are postdoctoral research fellows or Ph.D. candidates at the School of Civil and Environ-
mental Engineering, Nanyang Technological University, Singapore, for their cooperative research
and assistances with gathering literature in some of the methods and theories considered.
Contents
Introduction to Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1 What Is AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 AI Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 AI Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 What Is CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Significance of CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Activities in CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Characteristics of CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 How AI Benefits CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4 Organization of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Knowledge Representation and Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Safety Leadership in Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Knowledge Representation, Learning, and Discovery . . . . . . . . . . . . . . . . . 19
4 SEM-Enabled Knowledge Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.1 Survey and Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Hypothesis on Causal Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Knowledge Learning and Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5 Knowledge Learning from Data and Validation . . . . . . . . . . . . . . . . . . . . . . 24
5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 Common Method Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3 SEM Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.4 Measurement Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.5 Structural Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6 Knowledge Discovery in Safety Leadership . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.1 Impacts of Path Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.2 Impacts of Stakeholder Participation . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7 Merits of SEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
8 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
vii
viii Contents
Fuzzy Modeling and Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2 Fuzzy Modeling and Reasoning Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3 A Holistic FCM Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1 FCM Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 FCM Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3 FCM Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 TBM Performance in Tunnel Construction . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1 TBM Failure Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2 TBM Failure Map Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5 FCM-Enabled TBM Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.1 Predictive RCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.2 Diagnostic RCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 Hybrid RCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6 Merits of FCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.1 FCM Interpretability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.2 FCM Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.3 FCM Capacities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Time Series Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2 Estimation of Tunnel-Induced Ground Settlement . . . . . . . . . . . . . . . . . . . . 68
2.1 Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.2 Analytical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.3 Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.4 Intelligent Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.5 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3 Time Series Prediction Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4 A Hybrid Time Series Prediction Approach . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1 Decomposing Original Data Using WPT . . . . . . . . . . . . . . . . . . . . . . . 76
4.2 Predicting Separate Time Series Using LSSVM . . . . . . . . . . . . . . . . . 78
4.3 Reconstructing Different Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4 Prediction Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5 A Realistic Tunnel Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2 Training Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Analysis of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Contents ix
Information Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2 Structural Health Assessment (SHA) in Tunnels . . . . . . . . . . . . . . . . . . . . . 97
3 Information Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4 A Hybrid Information Fusion Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.1 Probabilistic SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2 Improved D-S Evidence Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.3 Safety Risk Analysis and Assessment . . . . . . . . . . . . . . . . . . . . . . . . . 106
5 Information Fusion for SHA in Tunnels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1 Project Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 SVM Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3 Multi-classifier Information Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6 Merits Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.1 Uncertainty Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2 Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Dynamic Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
2 BN and DBN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
2.1 BN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
2.2 DBN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3 Dynamics in Tunnel-Induced Damages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
3.1 Tunnel-Induced Road Damage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
3.2 Control Standard for Road Damage . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4 DBN-Enabled Dynamic Risk Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.1 Risk/Hazard Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.2 DBN Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.3 DBN Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5 Dynamic Risk Analysis in Tunnel-Induced Damages . . . . . . . . . . . . . . . . . 137
5.1 Case Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.2 DBN Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.3 Predictive Analysis and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.4 Sensitivity Analysis and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.5 Diagnostic Analysis and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Process Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
2 Process Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3 Event Logs in BIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4 Process Mining Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.1 Process Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.2 Process Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
x Contents
4.3 Process Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

4.4 Social Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
5 Typical Applications in BIM Process Discovery . . . . . . . . . . . . . . . . . . . . . 160
5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.2 BIM Process Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.3 BIM Process Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.4 BIM Process Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.5 BIM Collaborative Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 167
6 Summarizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Agent-Based Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
2 Agent-Based Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
3 Pedestrian Evacuation Under Emergency . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
4 Simulation-Based Route Planning and Optimization . . . . . . . . . . . . . . . . . . 178
4.1 Evacuation Network Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
4.2 Influential Factors Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
4.3 Route Planning Strategy Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
4.4 Evacuation Efficiency Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5 Pedestrian Evacuation Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.1 Case Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.2 Simulation Model Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.3 Simulation Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6 Route Planning Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.1 Average Pedestrian Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.2 Average Evacuation Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
6.3 Average Evacuation Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.4 Average Evacuation Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7 Merits of Simulation and Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Expert Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
2 Expert Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
3 Relationships Between BIM and Construction Safety . . . . . . . . . . . . . . . . . 205
3.1 BIM for Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
3.2 BIM Data Reuse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
3.3 Construction Safety Risks in Tunnels . . . . . . . . . . . . . . . . . . . . . . . . . . 207
4 Knowledge Base Development for Construction Safety . . . . . . . . . . . . . . . 208
4.1 Knowledge Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
4.2 Knowledge Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
4.3 Knowledge Database Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
5 BIM-Based Risk Identification System (B-RIES) . . . . . . . . . . . . . . . . . . . . 215
5.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Contents xi
5.2 Engineering Parameters Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

5.3 Knowledge-Based Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
5.4 Risk Analysis and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6 System Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
6.1 Project Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
6.2 Safety Risk Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
6.3 Implementation Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
7 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Computer Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
2 Deep Learning and Computer Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
3 Computer Vision Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
3.1 Feature Pyramid Attention Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
3.2 Spatial Attention Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
3.3 Channel Attention Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
3.4 Learning Process and Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . 241
4 Computer Vision for Automated Crack Detection . . . . . . . . . . . . . . . . . . . . 243
4.1 Data Resource . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
4.2 Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
4.3 Analysis of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
4.4 Comparison with Existing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5 Merits of the Proposed Computer Vision Approach . . . . . . . . . . . . . . . . . . . 251
6 Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Conclusions and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
1 Main Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Introduction to Artificial Intelligence
1 What Is AI
The term Artificial Intelligence (AI) is a branch of computer science to make

computers perform human-like tasks, and thus, computers can appropriately sense
and learn inputs for perception, knowledge representation, reasoning, problem-
solving, and planning. Various types of innovative AI technologies are designed
to imitate the cognitive abilities of human beings, which can, therefore, deal with
more complicated and ill-defined problems in an intentional, intelligent, and adap-
tive manner. Typically, AI can be regarded as a conjunction of machine learning
and data analytics. That is to say, machine learning algorithms, a subset of AI, learn
sufficiently robust data from multiple sources and then act on the insights of data to
make smart decisions adaptively [16]. The investment in AI is experiencing rapid
growth, in which machine learning particularly accounts for a major proportion.
According to a report from Accenture company, AI is already altering every walk of
life, which heralds dramatic potential to boost labor efficiency by 40% and double
annual economic growth rates in 2035 [11]. For the purpose of making AI live up
to expectations, more and more companies are actively investing in various AI tech-
nologies, which put AI into a sharper focus and extend its application scope [2]. As
AI talent continues to mature and be applied, it is believed that the topic of AI can
become the next digital frontier to drive the increase of automation and intelligence.
1.1 AI Development
AI is not a completely new concept. Its development has gone through three major
stages [3]. The early golden age of AI research is during 1956–1974. The name
“artificial intelligence” made its advent at the 1956 Dartmouth Conference. Since
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 1
L. Zhang et al., Artificial Intelligence in Construction Engineering and Management,
Lecture Notes in Civil Engineering 163,
https://doi.org/10.1007/978-981-16-2842-9_1
2 Introduction to Artificial Intelligence
then, researchers have emphasized adopting algorithms to solve mathematical prob-

lems and geometrical theorems. However, the very limited data volume and computer
memory at that time caused difficulties to create intelligence in machines. The second
wave occurred from 1980 to 1990 along with the emergence of expert systems. It
is known that expert systems can logically mimic human behavior and knowledge
to realize decision making at a high degree of expertise, which is the predecessor of
current machine learning and deep learning. However, building an expert system is
not a straightforward task, which requires a lot of time and cost to prepare a reli-
able knowledge base by extracting sufficient rules from human experts. Besides,
this system is only suitable for narrow problems and its results are even mistakable.
Since the late 1990s, AI is booming again with the help of exponential growth of
data, computing power and storage, and more advanced algorithms.
The current AI has moved to a remarkable level, leading to dramatic breakthroughs
in categories like machine learning, deep learning, computer vision, natural language
processing, virtual assistants, robotic process automation, and others. As a result, AI
can power many aspects of society by offering new opportunities for intelligent
agents, automated data analytics, and improved decision making in diverse domains,
including business, robotics, transportation, healthcare, education, manufacturing,
and others. This book mainly presents and discusses how AI is transforming the
construction industry for delivering smarter construction management.
1.2 AI Techniques
Many AI techniques have been developed to make machines mimic human cogni-
tive processes in terms of learning, reasoning, and self-correcting. In other words,
with the help of AI techniques, machines can perform tasks under a certain degree
of intelligence. There are many ways to classify various developed AI techniques.
Herein, we category them into five major groups, as illustrated in Fig. 1.
Fig. 1 An overview of common AI techniques

1 What Is AI 3
Expert systems: An expert system is a straightforward and understandable method

towards intelligent decision making, which contains possible expert knowledge and
reasoning to address complicated problems. The earliest expert system was devel-
oped in 1965, which achieved the first success of AI. Three essential components are
included in an expert system, namely the knowledge base, user interface, and infer-
ence engine. More specifically, the knowledge base is a database of information,
notions, and rules prepared by different experts in a specific domain. The IF-THEN-
ELSE rule is widely used to organize the collected knowledge. More high-quality
knowledge will improve the system’s performance. The user interface offers the
interaction between non-expert users and the expert system to present questions and
input data to start the decision-making process in the inference engine. Finally, the
inference engine acts as a brain of the system to efficiently deduct reliable solutions
based on factual and heuristic knowledge stored in the base.
Fuzzy logic: As a proposal of fuzzy set theory, fuzzy logic was put forward in
the early 1960s to handle uncertainty and ambiguity in human language. It offers
flexible decision making as similar to human reasoning. In other words, it can deal
with input data in vagueness, uncertainty, impreciseness, and incompleteness through
converting them into computer understandable forms, and then make responses based
on a set of fuzzy rules. The fuzzy logic architecture contains the following parts. The
fuzzifier transforms the crisp number into fuzzy sets using membership functions.
Then, the fuzzy inference is responsible for the decision process, which conducts the
computational procedure to identify the match between the fuzzy inputs and rules.
In particular, rules stem from the rule-based expert knowledge, which can be flexibly
added and removed from the system. They help to determine the relationship between
fuzzy inputs and outputs. Lastly, the defuzzifier converts the fuzzy output back to
a crisp number. Since the mathematical concept of fuzzy logic is relatively simple,
this method is easy to implement and understand. Moreover, fuzzy logic has been
commonly incorporated into the expert system to design the fuzzy expert system [5].
Statistical model: Statistical models employ mathematical equations to inference the
relationship between variables, which often implement statistical approaches under
certain assumptions to make an inference, like hypothesis testing, confidence inter-
vals, and others. For instance, the Bayesian network, the hidden Markov model, and
the autoregressive integrated moving average (ARIMA) model are three typical statis-
tical models. To be specific, the Bayesian network, also known as belief networks,
is a kind of probabilistic graphical model to capture casual relations in variables.
Its focal points lie in the probability representation to measure the uncertainty. In
addition, a method of statistical inference named Bayesian inference can be realized
in the forward and backward directions based on the Bayes’ theorem, in order to
update the network when more evidence is fed in. The hidden Markov model is a
statistical Markov model based on the Markov chain, which can consider both the
observed and hidden events to calculate the probability distribution over a sequence.
The ARIMA model is specialized in time series forecasting. It explains the lagged
observation of time series to predict the future value, which is expressed by the form
of regression.
Machine learning: Machine learning is a great step of AI to enable machines to

have powerful learning abilities from data. Various algorithms are programmed to
access data and learn by themselves from these large volumes of available data with
no human involvement. That is to say, machine learning aims to teach machines how
to discover patterns hidden in large data and realize data-driven predictions on future
tasks. Supervise and unsupervised learning are the two main categories belonging
to machine learning. The significant difference between these two methods is that
supervised learning requires labeled data to create future predictions, while unsu-
pervised learning tries to reveal hidden structures from unlabeled data. As machine
learning evolves, deep learning and reinforcement learning have been developed at
a higher level to be a new trend. Deep learning is inspired by the neural networks of
human brains, which is made up of multiple processing layers to process information,
represent features, and gain knowledge. It is skilled in detecting intricate structures
in high-dimensional data, which has dramatically improved the state-of-art models
in natural language processing, computer vision, and others. As for reinforcement
learning, it interacts with the environment to generate trial and error, resulting in suit-
able actions and decisions sequentially to maximize reward. In other words, it faces
a game-like situation, where the next input depends on the output of the previous
input.
Process mining: Process mining is still a young discipline to bridge the gap between
process management and data science, which incorporates the process aspects into
machine learning and data mining. It mainly explores event log data with the intent
of monitoring, diagnosing, analyzing, and improving the actual process. In general,
three basic tasks are carried out by process mining. It begins with the automated
discovery of process models from event logs with the help of proper algorithms.
Afterward, conformance checking pairs the process model with the event log to check
compliance, detect deviations, and predict bottlenecks, in order to provide guidance
to optimize the complicated process. Besides, model enhancement can implement
more data analysis methods to provide insights into other aspects of the operation
rather than the process perspective. In short, process mining takes full advantage
of event logs to create full transparency on processes, which can inform strategic
decisions to raise the process efficiency and reliability.
2 What Is CEM
To ensure the resiliency and success of the construction industry, construction engi-
neering and management (CEM) highlights the significant contribution of theoretical
and technical knowledge in addressing key aspects of the infrastructure construction
project. CEM can be conceived as an executive process of the physical work on
the construction site [10]. Architectures, engineers, constructors, contractors, and
others work together to provide professional services in constructing buildings for
residential, commercial, or industrial applications and solving on-site problems.
2 What Is CEM 5
2.1 Significance of CEM
Since the assembly process is made up of several order events, there are complicated
patterns of individual time requirements along with restrictive sequential and coop-
erative relationships among the project’s many segments embedding in such process
[13]. The common interest of CEM chiefly arises from the following two sides.
For one thing, construction, as a large sector of the economy, plays a prominent
role in driving economic growth and long-term national development [4]. According
to a survey from McKinsey Global Institute in 2017, the global construction industry
makes up around 13% of the world’s Gross Domestic Product (GDP) and this number
is projected to rise to 15% in 2020. Meanwhile, construction projects create a broad
range of job opportunities for 7% of the world’s working population. Despite its
economic importance, an obvious issue is poor labor productivity during the construc-
tion procedure, negatively leading to the waste of manpower, material resources,
and financial resources. Since construction activities contribute a lot to our society
economically, it makes the most sense to take proper construction management to
improve product performance. If the construction productivity is enhanced by as
much as 50% to 60% or higher, it is estimated to bring an additional $1.6 trillion into
the industry’s value each year and further boost the global GDP.
For another, construction with inherent complexity is one of the most dangerous
industries worldwide, which is greatly susceptible to a variety of unpredictable
factors, such as participants in different roles, the changeable environment in large
uncertainty, and others [12, 15]. It is likely for workers to be exposed to a great threat
to health, and thus, they may suffer from mechanical injuries, degenerative disor-
ders, chemical hazards, psychosocial stress, and others [7]. Therefore, the construc-
tion industry tends to cause a small scale of fatal accidents with higher frequency
than other domains, which is even responsible for 30–40% of fatalities worldwide
[21]. In the Chinese construction market, the number of construction accidents and
fatalities remains stubbornly high [17, 18]. Numerous researches have revealed that
safety issues are tied up with management, emphasizing the necessity of construction
management for safety control and accident prevention [8]. Through identifying,
evaluating, and reacting to the potential risk in dynamic construction work at an
early stage, it is expected to eliminate safety hazards and then achieve a sustainable
reduction in fatalities in the construction industry.
2.2 Activities in CEM
CEM is consisting of a series of physical activities for constructing infrastructures,

residential and commercial buildings, which can be defined as a completed process
including designing, scheduling, budgeting, building itself. All sorts of construction
projects require appropriate management over the lifetime of a project, in order to
guide the project towards success under the control of time, cost, scope, quality,
and collaboration. Project managers are responsible for project management, who
coordinate closely with other participants to draw up plans, schema, timelines, costs,
personnel arrangements, and others. They monitor the entire progress of work and
concentrate on all aspects of the project (i.e., labor, capital, time, equipment, material,
risk), in order to give back corresponding instructions to lower the possibility of
delays, budget overruns, high risks, and great conflicts. In short, the main activities
in CEM can be categorized into three major phases of construction as follows.
(1) Planning
Prior to the start of physical construction, the conception of the project is proposed
by clients. Then, its feasibility is analyzed comprehensively to clear the project
objectives. For meeting such a specific goal, detailed plans for the project develop-
ment are created in terms of resources, schedules, budgets, and dependencies, which
can potentially help in reducing cost, duration, and irrational process in the prac-
tical project. For instance, the schematic design can be drawn to fully describe the
building systems. Scheduling can chronologically distribute multiple activities, and
then assign dates, workers, and resources to a certain activity. It aims to fit the plan
to a reasonable time scale and workflow. Cost estimation is the process of predicting
the needed funds and resources for conducting a project within a defined scope. That
is to say, the planning phase incorporates activities mainly for forming project plans,
drawing design, and allocating labors, sources, and time. These activities assist in
rationally streamlining the construction process. They can serve as a reference to
monitor the actual process and guide it to be finished punctually within the estimated
budget.
(2) Construction
This is a phase of executing the physical construction, and thus, the planning made
at the previous phase is expected to pay off. For construction workers, they are
the manual labor to perform the on-site construction, who conduct a range of
basic tasks in a general sequence, including layout marking, excavation, foundation
work, column casting, wall construction, lintel, roofing, plastering, fixing of doors,
windows, electrical and plumbing works, tiles laying, painting, and others. Some
skilled workers are also needed to operate sophisticated machine tools. Regarding
project managers, they take charge of the on-site monitoring and assessment. That is
to say, managers oversee the actual construction process in terms of scope, budget,
and schedule, and then compare the observations to the formed planning. If an incon-
sistency is detected, the corresponding action can be enacted to bring the process
back into conformance or adjust plans to cope with changes. Moreover, managers
are also expected to identify the exposure of risk and the associated impact on the
performance, schedule, or budget of projects. Furthermore, both the qualitative and
quantitative risk analysis can be carried out for quality control, which provides a better
understanding of the risk factors and their effects. As a result, timely responses can
be created to proactively address the potential issues, which can eventually minimize
the risk before it occurs.
2 What Is CEM 7
(3) Operation and Maintenance (O&M)

When construction is completed, the project will enter a new phase named O&M.
It is known that O&M takes most of the time within the lifecycle, leading to a large
amount of cost accounting for around 60% of the total project budget [6]. The goal
of O&M is to operate and maintain a constructed facility to meet the anticipated
functions over its lifecycle, which is crucial to ensure the safety and comfort of
users. Operation means the provision of day-to-day services to operate and control
the facility in an efficient, economical, and reliable manner. Maintenance aims to
minimize the possibility of system failure, which can be conducted from two aspects.
To be more specific, preventive maintenance creates plans and schedules to adjust and
guide the ongoing operation on a regular basis. It is useful in detecting the potential
risks, and thus, corresponding actions can be timely taken prior to unexpected events.
In contrast, corrective maintenance is implemented after problems have occurred. It
strives to repair the problematic parts to get them back on the normal status as quickly
as possible. Moreover, recent attention in O&M has focused on sustainability. The
execution of O&M must obey some energy regulations and standards to make the
facility run long-termly, safely, and energy-savingly. As a result, proper O&M is
proven to benefit the facility management a lot, such as to shorten response time to
issues, reduce unnecessary repairs, drive down costs, maximize the asset value, and
mitigate potential risk, which can eventually prolong the lifespan of the facility and
improve the users’ satisfaction.
2.3 Characteristics of CEM
Since a construction project is unique, temporary, and progressive in nature for

producing the desired objectives, CEM can be considered as a process to handle
a series of interrelated tasks over a fixed period within certain limitations. The key
characteristics of CEM are outlined from the following five points.
(1) Uniqueness
Due to the differences in client requirements, project size, conditions, influences, and
constraints, the construction projects differ from one another to enhance the difficulty
of project management. That is to say, we cannot simply replicate the scheduling,
design scheme, budget, and logistics of an existing project to a new one. Instead, it
only allows us to learn from previous projects with similar features to make decisions
on the current project. It should be noted that the construction project is very technical
and characteristically unique, and thus, a highly customizable solution is deemed as a
necessity to ensure the reliability and efficiency of the project. In addition, individuals
with differing roles, including designers, engineers, suppliers, contractors, managers,
and other service providers are temporarily organized in a project. It means that each
project is carried out by a unique team, and each team has its characteristics regarding
participants’ skills, knowledge, experience, communication, and collaboration.
(2) Labor-intensive
Typically, a huge amount of manual labor should be required in a construction project,
who can offer great quantities of physical effort in charge of the project imple-
mentation, timely completion, and quality assurance. In general, the proportion of
labor costs can take up 30–60% of the total project budget, indicating that inten-
sive labor is a critical component in construction. These flexible workforces can be
accommodated to the changeable projects dynamically based on their initiative and
problem-solving ability. It has been found that 40 million workers were employed in
the construction sector in the year 2007, among whom 27% was skilled labor, 55%
represented the unskilled labor, and the rest denoted the support staff [7]. Currently,
there is fast growth in the demand for skilled and semi-skilled workers with more
skills and expertise, since they play an important role in improving the performance
and productivity of construction projects. In other words, this kind of labor force can
provide useful information and techniques to drive the project forward more easily,
efficiently, and safely.
(3) Dynamics
Although a project has been clearly defined from the beginning to the end, the actual
execution of the project is unable to always stick to the plan. That is to say, inevitable
changes or adjustments occur dynamically throughout the whole lifecycle of the
project, which are caused by various reasons. For instance, the project alternatives
need to be reformulated due to some human factors, including clients’ dissatisfaction,
designers’ and engineers’ mistakes, financial problems from contractors, and others.
Moreover, unforeseen conditions resulting from the undesired delays, bad weather,
complex geologic environment, additional demands of labor, equipment, materials,
and others are the major causes of changes. Besides, when additions or reductions are
applied to the project scope, the scheduling and budget should be updated accordingly
to adapt to the changing circumstances. It is known that managing changes properly
is the guarantee of success in a construction project, and thus, project managers need
to flexibly identify changes, cope with dynamics, and perform effective controls.
(4) Complexity
The complexity of construction projects can arise from two aspects, namely the task
and participants. For one thing, construction tasks are heavy and diverse, which can
possibly meet conflicting scheduling or performance problems. Therefore, it is not
a straightforward job to formulate proper scheduling to align interconnected tasks,
resources, and teams. To make the operation process run smoothly, a variety of
factors, such as safety, security, environment, weather, workers, and others, should
be taken into full consideration. The strict time limit can even negatively boost the
pressure in executing the planned tasks. When some unexpected things happen in the
construction process, there is no enough time to consider solutions [20]. Additionally,
although the advanced technologies and new materials are adopted for more sustain-
able development, they can possibly contribute to a higher degree of complexity. For
another, groups of workers in different roles and varying proficiencies are commonly
2 What Is CEM 9
involved, who can work together, communicate, and share information with each
other. This kind of multi-disciplinary collaboration will result in intricate interactions
in individuals and tasks. Besides, they tend to have conflicting business interests, and
thus, negotiations are needed to pursue the common goal.
(5) Uncertainty
Uncertainties are unknown before they actually occur, which can be regarded as
threats to unfortunately raise risk in project failure. Notably, a high level of uncer-
tainty is inherent in complicated construction projects, which is closely related to
various factors [19]. For instance, before the site construction, scheduling and cost
need to be estimated reasonably under great uncertainty. The improper estimation
will impede the progress of the project. As for the architecture design, some ques-
tions remain to be answered, such as whether the design can pass audits, whether
clients are satisfied with the design, and others. Moreover, a great deal of unavoid-
able uncertainty exists in the construction phase, such as the ground conditions,
soil-structure interaction, weather conditions, building material properties, design
changes, the reliability of suppliers, and others. If these uncertainties can be detected
and measured at an early stage, the project management can be optimized to act as
a reliable tool in mitigating potential risk in the construction project.
3 How AI Benefits CEM
Although AI stands out as transformational technology to potentially bring about

unprecedented changes in our work and life, its applications in CEM with the nature
of complexity, uncertainty, and dynamics are still in its infancy. In the immediate
future, the construction industry is projected to increase more focus and investment in
AI to catch up with the fast pace of automation and digitalization [9]. With the advent
of the building information modeling (BIM) and wireless sensor network (WSN),
huge amounts of data can be collected from the worksite every day along with some
features of “big data”, which yield great values to CEM. AI techniques can, therefore,
be deployed in a range of ways to improve the performance of construction projects.
In other words, AI has proven to take advantage of such rich data sources to train
suitable models and then deliver on its promises, such as prediction, optimization,
decision making, and others. The substantial benefits of AI in CEM can be outlined
as follows.
• Automation. AI drives the process of project management more technically
automatable and objective. Conventional construction management is heavily
based on experts’ knowledge, experience, and subjective judgment, which is
more prone to bias and confounding. To overcome these distinct disadvantages,
AI-based solutions are taken into account to pursue smart control. For instance,
machine learning algorithms can be employed to intelligently handle the mass
of accumulated data and discover hidden knowledge. The insights gained from
advanced analytics help managers to better understand the construction project

and rapidly spot the project concerns in a data-driven manner. These advanced
algorithms have also been integrated into project management software, aiming
to automate the complete procedure for data analysis and decision making. As for
on-site construction monitoring, drones and sensors are utilized to automatically
record data and take images/videos for scanning, monitoring, and mapping the
construction status, environment, and progress, which offer a more comprehen-
sive picture of the site through various project stages without human interaction.
That is to say, evidence obtained by such techniques can replace the traditional
manual observation which tends to be time-consuming, tedious, and error-prone.
• Risk Mitigation. AI can monitor, recognize, evaluate, and predict potential risks
in terms of safety, quality, efficiency, and cost across teams and work areas. In
the risk and uncertainty analysis, AI tools based on probabilistic models, fuzzy
theory, machine learning, neural networks, and others equip computers with the
capability to efficiently learn data collected from the construction site. Thus,
computers can explore much more risk-related data than humans could ever review
from both the qualitative and quantitative perspectives. In consequence, these
identified risks provide assistive and predictive insights on critical issues, and
thus, project managers can quickly prioritize possible risks and concentrate on
them to take proactive actions instead of reactions for risk mitigation, such as to
streamline operations on the job site, adjust staff arrangement, and keep projects
on time and budget. In other words, AI presents valuable opportunities to realize
early troubleshooting to prevent undesirable failures and accidents in the complex
project management [1]. Additionally, robots can take charge of unsafe activities,
in order to minimize the number of humans working in dangerous environments.
• High Efficiency. AI allows for optimizing the construction process, in order to
make the project run smoothly and finish on time and reliability. For instance,
process mining is a new AI-enabled approach to generate valuable insights into
the complicated construction, such as to track key workflows, predict deviations,
detect invisible bottlenecks, extract collaboration patterns, and others. The discov-
ered knowledge is critical to project success, which can guide the optimization
of the construction execution process. It is expected to avoid unnecessary steps,
reworks and conflicts, potential delays, and poor cooperation. In turn, tactic deci-
sions can be made for trouble-shooting at an early stage, driving the improvement
of operational efficiency. It is also effective in preventing costly correction at the
remaining stage. Moreover, AI-powered robots and autonomous systems have
been directly adopted on construction site. Since they can potentially replace
human workers to take over the repetitive and routine construction tasks, such as
bricklaying, welding, tiling, and others, they can be an ideal solution to address
labor shortages. Besides, they are able to work continuously without taking a
break at almost the same rate and quality, indicating that the proper use of smart
machinery will ensure efficiency, productivity, and even profitability.
• Digitalization. It should be noted that BIM provides a pool of information
concerning planning, construction, operation, and management, which has gone
far more than the 3D modeling to support various digital solutions. In other words,
3 How AI Benefits CEM 11
BIM has been considered as a digital backbone of future construction, which can
be reasonably integrated into AI to facilitate the digitalization of construction
projects. For BIM, it can provide a platform for not only collecting large data
about all aspects of the project, but also sharing, exchanging, and analyzing data
in real-time to realize in-time communication and collaboration among various
participants. Thus, all members have immediate access to the latest documents
and information, who can cooperatively examine the shared rich data sources. For
the AI techniques, they can deeply explore these massive amounts of data from
BIM to automate and improve the construction process. Based on the discovered
knowledge, strategic decisions can be informed for streamlining the complicated
workflow, shortening operation time, cutting costs, reducing risk, optimizing staff
arrangement, and others. In addition, the integration of BIM and AI can move
the paper-based work towards online management. For one thing, it can deliver
the most efficient and effective information to keep continuous updating of the
ongoing project. For another, it can take advantage of BIM data to make real-
time analysis, and thus, immediate reactions can be performed without delays to
prevent unexpected events.
• Computer Vision. Computer vision techniques have gradually taken the place
of the laborious and unreliable visual inspection in civil infrastructure condi-
tion assessment. Current advances in computer vision techniques lie in the deep
learning methods based on neural networks to automatically process, analyze,
and understand data in images or video through end-to-end learning, which
can conduct various tasks, like image classification, object detection, semantic
segmentation, and others. Towards the goal of intelligent management in the
construction project, computer vision plays an important role in two perspec-
tives, which are inspection and monitoring applications [14]. To be more specific,
inspection applications perform automated damage detection, structural compo-
nent recognition, unsafe behavior, and condition identification. Monitoring appli-
cations are non-contact methods to capture a quantitative understanding of the
infrastructure status, such as to estimate strain, displacement, and cracks’ length
and width. To sum up, the vision-based methods in CEM are comparatively cost-
effective, simple, efficient, and accurate, which are useful in robustly translating
image data into actional information for the purpose of structural health evaluation
and construction safety assurance.
4 Organization of the Book
This chapter briefly introduces the concept, development, and techniques of AI and
reviews activities, characteristics, and significance of CEM. The substantial benefits
of AI in CEM are summarized.
Chapter “Knowledge Representation and Discovery” describes the knowledge
representation and discovery and the relevant application in CEM. Many types
of knowledge representation and discovery methods are reviewed. One promising
method, Structural Equation Model (SEM), is chosen to discover casual relationships

underlying safety leadership in construction. A model consisting of 5 latent variables
and 26 observed variables is established to learn from data and reveal relationships
between different stakeholders’ leaderships and construction safety performance.
The hypotheses are testified, and the overall model fit is assessed. Impacts of path
coefficients, stakeholder participation, and construction dynamics are investigated.
Chapter “Fuzzy Modeling and Reasoning” describes fuzzy modeling and
reasoning and its application in CEM. Many types of fuzzy modeling and reasoning
methods are reviewed. One outstanding method, Fuzzy Cognitive Maps (FCM), is
selected to assess Tunnel Boring Machine (TBM) performance and perform Root
Cause Analysis (RCA) in case of unsatisfactory performance. FCM is able to capture
and utilize construction experience and knowledge from domain experts, and a cause-
effect model consisting of nine concepts is established for simulating the TBM perfor-
mance. A tunnel case in the Wuhan metro system in China is used to demonstrate
the applicability of the developed approach. It is verified to be capable of modeling
the dynamics of system behaviors over time and performing many kinds of what-if
scenario analysis, including predictive, diagnostic, and hybrid RCA. It turns out to be
a more competitive solution that deals with uncertainty, dynamics, and interactions
in the approximate reasoning process.
Chapter “Time Series Prediction” describes the time series analysis technique and
its application in CEM. Several common time series analysis methods are reviewed.
A novel hybrid approach that integrates Wavelet Packet Transformation (WPT) and
Least Squares Support Vector Machines (LSSVM) is presented to enhance the accu-
racy and reliability in time series analysis and prediction. It is used to estimate the
magnitude of the tunnel-induced ground settlement over time. The original time-
domain signals (i.e., the measured settlements over a time period) are decomposed
into a series of sequences using WPT, while LSSVM models are then built to predict
the target sequences within high- and low-frequency regions. A realistic tunnel case is
utilized to demonstrate the feasibility and applicability of the proposed WPT-LSSVM
approach. Results indicate the proposed approach can achieve a higher accuracy and
reliability than the traditional time series analysis approaches.
Chapter “Information Fusion” describes the information fusion technique and its
application in CEM. Several common information fusion methods are reviewed. A
novel multi-classifier information fusion approach that integrates the probabilistic
support vector machine (SVM) and the improved Dempster-Shafer (D-S) evidence
theory is proposed. It is implemented to fuse classification results provided by the
base classifiers at the decision level for an overall risk evaluation. The proposed
approach is utilized to perform structural health assessment in operational tunnels.
Results indicate that the proposed approach exhibits more outstanding classifica-
tion performance and can efficiently fuse multi-sensory information with ubiquitous
uncertainties, conflicts, and biases.
Chapter “Dynamic Bayesian Networks” describes network modeling and
analytics in CEM using AI. An excellent method, dynamic Bayesian network (DBN),
is highlighted, as DBN is able to model construction dynamics and update the
system status given new observations. Differences between the Bayesian network
4 Organization of the Book 13
(BN) and DBN are reviewed. A DBN model consisting of various geological,
design, and mechanical variables is developed to reveal dynamics in tunnel-induced
damages over time. Results testify that the developed model is capable of performing
feed-forward, concurrent, and back-forward control respectively on a quantitative
basis.
Chapter “Process Mining” describes process mining and its application in CEM.
Process mining is regarded as an emerging data mining method, which leverages
various algorithms and mathematical models for the visualization and investiga-
tion of event logs and assists in discovering, evaluating, and improving the existing
process. Process mining is applied to deeply mining large volumes of event logs in
BIM project execution, aiming to improve the understanding and efficiency of the
operation process in the BIM-enabled project. A framework of process mining on
BIM event logs with step-by-step procedures is presented, where in-depth knowl-
edge discovery is conducted in terms of the process discovery, conformance checking,
social network analysis, and decision tree classification.
Chapter “Agent-Based Simulation” describes the agent-based simulation for
improved decision support in CEM. Agent-based simulation provides a near-real
solution to perform various what-if experimental scenario analysis, given some exper-
iments are impossible or hard to be conducted in reality, such as pedestrian evacu-
ation under an emergency. Several common methods in the agent-based simulation
are reviewed in this chapter. The promising social force model which can reflect the
anisotropic characteristics and dynamic features of pedestrian interaction is selected
to simulate and optimize pedestrian evacuation behaviors in a subway station. Several
route planning strategies are modeled in the simulation environment and validated
by actual observations. The developed approach can provide valuable theoretic and
practical insights into a deep understanding of route planning strategies during the
pedestrian evacuation, and thus, the improvement of safety and economic objectives
can further be achieved in the design or re-design of metro evacuation systems.
Chapter “Expert Systems” describes the expert system technique and its applica-
tion in CEM. The expert system provides a powerful tool for knowledge integration
and autonomous inference mechanism, which can solve complex problems to signif-
icantly reduce dependence on domain experts. Considered as a digital representation
of physical and functional characteristics of a facility, BIM has the capability of
generating, storing, exchanging, and sharing the building information in an interop-
erable and reusable way, providing an ideal solution for risk management. A novel
approach integrating BIM and expert systems is proposed to perceive the safety
risk at an early stage of the construction. Both case-based reasoning and rule-based
reasoning techniques are combined in the inference model to improve the flexibility
and comprehensiveness of the system reasoning capacity.
Chapter “Computer Vision” describes computer vision and its application in CEM.
Deep learning techniques are driving advances in computer vision, and different types
of deep learning techniques are reviewed in this chapter. A novel computer vision
approach named a spatial-channel hierarchical network (SCHNet) is proposed to
support the automated and reliable concrete crack segmentation at the pixel level.
It not only considers the semantic interdependencies in spatial and channel dimen-
sions, but also adaptively integrates local features into their global dependencies.
Experimental results indicate that SCHNet can outperform state-of-the-art models.
It is robust to noises with a better generalization ability under various conditions,
including shadows, rough surfaces, and holes.
Chapter “Conclusions and Future Directions” summarizes the conclusions and
points out future directions. The AI’s functions of modeling and pattern detec-
tion, prediction, and optimization are highlighted, which create tremendous value
in revolutionizing the construction industry, leading to a more reliable, automated,
self-modifying, time-saving, and cost-effective process of CEM. Several emerging
applications of AI in CEM are put forward, which pave a more affordable and effec-
tive way to relieve the burden on manual labor and facilitate smart construction in
CEM.
References
1. Chang L, Zhang L, Fu C, Chen YW (2021) Transparent digital twin for output control using
belief rule base. IEEE Trans Cybern 1–15
2. Chui M (2017) Artificial intelligence the next digital frontier? McKinsey Comp Glob Inst
47(3):6
3. De Spiegeleire S, Maas M, Sweijs T (2017) Artificial intelligence and the future of defense:
strategic implications for small- and medium-sized force providers. The Hague Centre for
Strategic Studies
4. Giang DT, Pheng LS (2011) Role of construction in economic development: review of key
concepts in the past 40 years. Habitat Int 35:118–125
5. Hassanzad M, Orooji A, Valinejadi A, Velayati A (2017) A fuzzy rule-based expert system for
diagnosing cystic fibrosis. Electron Physician 9:5974
6. Hu Z-Z, Tian P-L, Li S-W, Zhang J-P (2018) BIM-based integrated delivery technologies
for intelligent MEP management in the operation and maintenance phase. Adv Eng Softw
115:1–16
7. Kulkarni G (2007) Construction industry: more needs to be done. Indian J Occup Environ Med
11:1–2
8. Li X, Yi W, Chi H-L, Wang X, Chan AP (2018) A critical review of virtual and augmented
reality (VR/AR) applications in construction safety. Autom Constr 86:150–162
9. Pan Y, Zhang L (2021) Roles of artificial intelligence in construction engineering and
management: a critical review and future trends. Autom Constr 122:103517
10. Pheng LS, Hou LS (2019) The economy and the construction industry. In: Construction quality
and the economy. Springer, pp 21–54
11. Purdy M, Daugherty P (2016) Why artificial intelligence is the future of growth. In: Remarks
at AI now: the social and economic implications of artificial intelligence technologies in the
near term, pp 1–72
12. Sacks R, Rozenfeld O, Rosenfeld Y (2009) Spatial and temporal exposure to safety hazards in
construction. J Constr Eng Manag 135:726–736
13. Sears SK, Clough RH, Sears GA (2008) Construction project management: a practical guide
to field construction management. Wiley
14. Spencer BF Jr, Hoskere V, Narazaki Y (2019) Advances in computer vision-based civil
infrastructure inspection and monitoring. Engineering 5:199–222
15. Waehrer GM, Dong XS, Miller T, Haile E, Men Y (2007) Costs of occupational injuries in
construction in the United States. Accid Anal Prev 39:1258–1266
References 15
16. West DM, Allen JR (2018) How artificial intelligence is transforming the world. Report, 24
Apr 2018
17. Wu C, Song X, Wang T, Fang D (2015) Core dimensions of the construction safety climate for
a standardized safety-climate measurement. J Constr Eng Manag 04015018
18. Zhang L, Wu X, Zhu H, AbouRizk SM (2017) Perceiving safety risk of buildings adjacent to
tunneling excavation: an information fusion approach. Autom Constr 73:88–101
19. Zhang L, Wu X, Zhu H, AbouRizk SM (2017) Performing global uncertainty and sensitivity
analysis from given data in tunnel construction. J Comput in Civ Eng 31(6):04017065
20. Zhang L, Wu X, Chen Q, Skibniewski MJ, Hsu SC (2014) Towards a safety management
approach for adjacent buildings in tunneling environments: case study in China. Build Environ
75:222–235
21. Zhou Z, Goh YM, Li Q (2015) Overview and analysis of safety management studies in the
construction industry. Saf Sci 72:337–350
Knowledge Representation
and Discovery
1 Introduction
Construction is one of the most dangerous industries worldwide, leading to a common

interest in improving construction safety performance due to humanitarian reasons
and rising costs of worker compensation. Numerous researchers suggested the contin-
uous unsafe conditions were mainly because of a misalignment of management
commitment and stakeholders’ actions, that is, lack of safety leadership [15]. Effec-
tive leadership plays an important role in ensuring the success of temporary construc-
tion onsite organizations that are facing a high degree of uncertainty. Strong safety
leadership should be the key to enhancing construction safety performance.
Safety leadership has emerged within the Occupational Health and Safety (OHS)
literature as a key construct in construction safety management [31]. Many industrial
participants/stakeholders are involved in safety issues in construction, and effective
collaboration of those stakeholders can contribute to ensuring construction safety
performance. Zuofa and Ocheing [32] pointed out that senior managers had played
important role in safety leadership by assisting in effective safety management. Other
important participators include the owner (O), the contractor (C), the supervisor (S),
and the designer (D). Romero et al. [22] summarized that the collective profes-
sionalization by multiple stakeholders was the key to strengthening the safety and
leadership in the construction industry. The relationship between those safety-related
internal and external factors within different stakeholders in construction projects has
not been fully investigated.
A long debate on the allocation of duty and responsibility for construction safety
management among main stakeholders exists. Wu et al. [27] concluded that the
owner’s greatest leverage was the leadership to influence safety perception, moti-
vation, and behavior of other stakeholders. Hadidi and Khater [9] emphasized the
significance of selecting contractors at the bidding stage to assure loss prevention
and better safety orientation during project implementation. Conchie et al. [5] high-
lighted the importance of supervisors’ safety leadership in promoting construction
18 Knowledge Representation and Discovery
safety. In general, a common agreement on which party should play a leading role
of leadership in construction safety has not yet been reached.
The main objective of this chapter is to answer the following questions: (i) How
to perceive interactions among multiple stakeholders during construction safety
management practices? (ii) Who plays a leading role in safety leadership in a
changing project environment? In this chapter, an Structural Equation Model (SEM)-
based causal model is proposed for revealing relationships among multiple stake-
holders. The causal relationships among multiple stakeholders in safety leadership
are hypothesized and tested using data gathered via a questionnaire survey. Impacts
of path coefficients on construction safety performance are explored. This research
attempts to provide a dynamic perspective to perceive the updating interactions
among different stakeholders. A better understanding of the knowledge for managing
construction safety in a changing project environment is provided.
2 Safety Leadership in Construction
Safety leadership is the process of interaction between leaders and followers in order
to achieve organizational safety goals as well as increase motivation and perfor-
mance. There are mainly two aspects involved to perform the safety leadership
analysis, including the individual and organizational aspects. From the individual
aspect, there are mainly five factors, including personality, emotional intelligence,
self-efficacy, motivation, and experience. From the organizational aspect, there are
mainly six factors, including role demands, situational constraints, organizational
support, subcultures, migrant labor, and subcontractors. Not all factors need to be
to conduct a thorough safety leadership analysis. The selection of specific factors
should be determined according to practical conditions of construction projects. In
addition, the weights or contributions of different factors are also an important issue
that should be deliberately chosen.
Given “safety” and “leadership” are abstract concepts in nature, it is logical to
assume that a large quantity of the gathered data should be in the form of qualitative
form, which are mostly gathered through questionnaires. Therefore, it is important
to adopt strict mathematical measures and techniques to conduct surveys, and more
importantly, analyze the data.
In general, safety leadership has drawn tremendous attention from both construc-
tion practitioners and researchers. Safety leadership is considered as a significant
antecedent of safety culture, worker safety behavior, and project success. The causal-
ities between safety leadership and other internal and external variables and the inter-
actions among different stakeholders have not been fully investigated and validated
with panel data. The underlying mechanisms by which leadership may influence
safety are not yet well depicted and understood, where complex interactions among
stakeholders and their impacts on safety performance should be considered and
explored holistically.
3 Knowledge Representation, Learning, and Discovery 19
3 Knowledge Representation, Learning, and Discovery
In the construction industry, knowledge is represented in various forms and can be

learned and discovered in many ways. However, if not clearly defined or handled,
a misunderstanding may occur that can cause unnecessary conflict or even great
economic loss. Therefore, it is very important to conduct knowledge representation,
learning, and discovery-related studies in construction.
• Knowledge representation deals with multiple types of information in various
forms, e.g., qualitative and quantitative, linguistic and numerical, complete and
incomplete, etc. With proper knowledge representation, knowledge in various
forms can be transformed into unified inputs that can be well handled by the
respective model.
• Knowledge learning denotes using machine learning or artificial intelligence
techniques to construct models that can bridge the complex relationships between
the inputs and the system output in various forms. Normally, a larger portion of
data is classified as the training dataset to train the model while the rest is used as
the testing dataset to validate the model performance.
• Knowledge discovery is complex in nature, as various internal causal relation-
ships are involved, and certain types of the hypothesis are used to build the complex
model. The key issue in knowledge discovery is to quantify the hidden causal link
between various variables and possible system outputs, which serves as great
support for making further strategic decisions.
Among the three components, knowledge representation is the premise for
conducting all analysis, knowledge learning is the tool for constructing the hidden
relationship and knowledge discovery is the final goal. In recent years, two statis-
tical approaches, including SEM and exploratory factor analysis (EFA), have been
proposed to empirically reveal and test inter-relationships in the hypothetical model.
EFA is capable of identifying the structure among the questionnaire items, where
no prior knowledge of factors or patterns of measured variables exists. Compared
with EFA, SEM is not only able to perceive complex structural relationships among
variables, but also estimate all coefficients in the model simultaneously. Actually,
SEM is a statistical methodology that takes a confirmatory approach to the anal-
ysis of structural theory. It represents causal processes that generate observations on
multiple variables, which can handle complex relationships among variables, where
some variables can be hypothetical or unobserved (latent variables). The significance
and strength of a particular relationship in the context of the complete model can be
assessed. In addition, the hypothesized model can be tested statistically in simulta-
neous analysis of the entire system of variables to determine the extent to which it
is consistent with the data.
Several research studies have reported the successful use of SEM in knowledge
representation, learning, and discovery. For instance, Al-Refaie [1] examined the
effects of organizational safety management and workgroup level factors on safety
self-efficacy, safety awareness, and safety behavior in Jordanian companies using
SEM. Wu et al. [29] proposed a systematic SEM-based approach to the evaluation

of prospective safety performance on construction sites, with causal relationships
and interactions between enablers and goals taken into account. Generally, SEM has
proven to be an effective tool in testing and investigating inter-relationships among
hypothesized models. SEM can also provide a statistical test for all variables in the
complete hypothetical model simultaneously, which ensures the consistency of the
data. As a result, a promising knowledge representation and discovery method is
selected to reveal casual relationships underlying safety leadership in construction.
4 SEM-Enabled Knowledge Discovery
4.1 Survey and Participants
In recent years, the ramifications of construction accidents are growing with a trend
toward more large-scale and more complex construction projects, especially in devel-
oping countries, such as China. The rapid development of huge and complex infras-
tructure projects, such as urban transit systems and high-speed railways, increase
occupational fatalities in construction. Despite the existence of the development of
occupational health and safety provisions, the effects have been far from satisfactory
[28]. In this chapter, the Chinese construction industry is studied as an example to
investigate the relationship of safety leadership among different stakeholders, aiming
to explore effective strategies to enhance their safety leadership and management.
The owner, contractor, supervisor, and designer parties are the most important
stakeholders related to construction projects. Their safety management levels, profes-
sional cooperation, and coordination have great impacts on ensuring the satisfactory
performance of safe project delivery. Herein, it should be noted that the supervisor
party refers to the third-party supervisor who represents the owner to supervise the
contractor. Several causes in some typical accident cases in the Chinese construc-
tion market are revealed in terms of different stakeholders, as shown in Table 1. A
thorough understanding of accident causes and respective responsibilities of each
participant can provide practical guidance for exploring relevant influential factors
that affect construction safety performance on sites.
4.2 Hypothesis on Causal Relationships
To clarify relationships among different stakeholders in construction plays a vital role

in model development and safety leadership analysis. Design-Bid-Build (short for D-
B-B or DBB), is a project delivery method where the agency or owner contracts with
separate entities for the design and construction of a project. DBB is the traditional
and prominent project delivery mode, where the construction work is awarded to a
4 SEM-Enabled Knowledge Discovery 21
Table 1 Several causes in typical accident cases in the Chinese construction market
Participants Causes Construction accidents in different provinces, China
Hubeia Yunnanb Hunanc Henand Jiangxie
√ √
Owner (O) The bidding system is not
implemented strictly
√ √
Construction activities are
undertaken without a work
safety license
√ √ √
Contract terms on construction
performance are not stated
clearly
√ √ √
Contractor (C) Undertaking projects beyond
the certificate scope
√ √
Undertaking projects under the
name of other construction
enterprises
√ √ √
The illegal practice of
sub-contracting or
contract-transferring
√
Technical workers do not have
the required certificates
√
There is no valid construction
scheme on site
√ √
Protective staff provided to
workers are imperfect
√ √
Supervisor (S) The chief supervisor has a
long-term absence during
construction
√ √ √ √
Field supervisors are
unqualified, leading to
inadequate supervision
√ √ √
Examination of construction
organizational design is not
taken
√
Designer (D) Unreasonable design changes
cause misunderstanding and
arguments
√ √
Project design is not consistent
with safety standards
Note a Hubei [11], b SAWS [23], c News [19], d MOHURD [18], e Daily [6]
contractor according to a competitive bidding system. The DBB mode is extensively

encouraged to be applied in the public sector due to its transparency, equity, and
efficiency. Indeed, in the DBB delivery mode, a debate on the allocation of duty
and responsibility in safety leadership within the four main participants (namely the
owner, supervisor, contractor, and designer) exists. To address those issues, SEM is
used to verify causal relationships and interactions among different stakeholders.
Owner
Project headquarters
Supervisor
Contractor
Designer
Sub-contractors ... Suppliers
Contractual Supervisory
relationship relationship
Fig. 1 Contractual and supervisory relationships in the DBB mode in the construction industry
A useful hypothesis enables predictions and analysis by applying deductive

reasoning, and reasonable hypotheses are very crucial for establishing an SEM.
Through the analysis of typical accidents (see Table 1), it is obvious that the four
main participants are all directly related to construction safety. At the same time,
it is known that the allocation of duty and responsibility in construction projects
varies a lot, given the construction management mode is different. The contractual
and supervisory relationships between the involved participants in the DBB mode, as
shown in Fig. 1, can be used to reveal the interaction among four major stakeholders.
Those relationships can clearly define the respective rights and responsibilities of
a specific participant, providing references for proposing relevant hypotheses for a
casual model.
To build a theoretical model to analyze the impact of participants’ behaviors on
construction safety leadership, the following hypotheses are developed based on
literature and theoretical analysis.
• Hypothesis H1 : The safety management performance of the owner positively
affects that of the supervisor;
affects that of the contractor;
affects that of the designer;
• Hypothesis H4 : The safety management performance of the supervisor positively
• Hypothesis H5 : The safety management performance of the design positively
• Hypothesis H6a : The safety management performance of the owner positively
affects construction safety performance on site;
4 SEM-Enabled Knowledge Discovery 23
• Hypothesis H6b : The safety management performance of the supervisor positively

• Hypothesis H6c : The safety management performance of the contractor positively
• Hypothesis H6d : The safety management performance of the designer positively
affects construction safety performance on site.
4.3 Knowledge Learning and Discovery
There are two types of models (i.e., measurement model and structural model) incor-
porated in an SEM model. Specifically, the measurement model addresses the reli-
ability and validity of the measurement of the latent variables, while the structural
model reveals the relationships between latent variables and describes the amount
of the explained and unexplained variances. An example of an SEM model is shown
in Fig. 2, where Fig. 2a represents a measurement model and Fig. 2b represents a
structural model. The measurement model could be defined by Eq. (1), while the
structural model can be defined by Eq. (2).
In the measurement model, there is a latent variable which cannot be directly
observed and can be revealed by certain observable variables. As shown in Eq. (1)
and Fig. 2a, the observable variables (i.e., Ai1 , Ai2 , Ai3 , and Ai4 ) are used to illustrate
the latent variable (i.e., Ai ) as a demonstration with some degrees of error. The
latent variables that are demonstrated in the measurement model would be used for
further computation in the structural model. The structural model aims to illustrate
relationships among latent variables by assessing the variance. As shown in Eq. (1)
and Fig. 2b, the target latent variable T is illustrated by the related latent variables
A1 , A2 , A3 , and A4 with their path coefficients λ1 , λ2 , λ3 , and λ4 , respectively, as
well as the residual error R1 [10].
Ai j = ϒ j Ai + E j (1)

n
T = λi Ai + R1 (2)
i=1
Fig. 2 A sample SEM model consisting of a a measurement model and b a structural model
where Ai j refers to exogenous observable variables, T and Ai are endogenous latent

variables, ϒ j and λi represent path coefficients in the structural and measurement
models, respectively, E j represents the measurement error, and R1 represents the
residual error. In Eqs. (1) and (2), exogenous observable variables (i.e., Ai j ) and the
casual structures serve as inputs, while endogenous latent variables (i.e., Ai and T)
and path coefficients (i.e., ϒ j and λi ) serve as outputs.
According to the actual construction accident analysis and the theoretical anal-
ysis in the literature, a list of 26 items are identified, which are covered in the design
of the questionnaire. Descriptions of those 26 items affecting construction safety
performance on sites are presented in Table 2. It should be noted that 24 items
(O1 , …, O6 , C1 , …, C6 , S1 , …, S6 , D1 , …, D6 ) serve as inputs that exert influence
over the performance of construction safety, while 2 items (CS1 and CS2 ) serve as
outputs that measure directly the construction safety performance. Particularly, CS1
(Internal safety performance) measures the internal dimension of the safety perfor-
mance through the self-evaluation (within the stakeholder), while CS2 (Collabora-
tive safety performance) measures the external dimension of the safety performance
through the improvement in collaboration and coordination among different stake-
holders. Most items are evaluated by the 5-point Likert-type scale ranging from 1
(completely disagree) to 5 (completely agree), in order to elicit respondents’ percep-
tions or opinions on different items. Taking the first factor O1 (Safety training and
incentive mechanism) as an example, the question on this factor is stated as “Do you
agree that safety training and incentive mechanisms are performing efficiently?” A
respondent then chooses a number from 1 (completely disagree) to 5 (completely
agree) based upon his or her working experience and understanding.
SEM represents causal processes that generate observations on multiple variables.
It can handle complex relationships among variables, where some variables can be
hypothetical or unobserved (latent variables). The hypothesized model can be tested
statistically in simultaneous analysis of the entire system of variables to determine the
extent to which it is consistent with the data. SEM is used to explore interactions and
leading roles in managing construction safety from survey data. The causal relation-
ships between different stakeholders’ leadership and construction safety performance
are tested by using survey data.
5 Knowledge Learning from Data and Validation
5.1 Data Collection
In this research, almost 1000 questionnaires are distributed to workers in the construc-
tion industry by a variety of survey tools, including emails, mails, websites, and
hand-out types. All the surveyed respondents are front-line workers, including project
managers, on-site engineers, designers, supervisors, foremen, and others in construc-
tion projects, where the involved projects are mainly located in the south, central,
5 Knowledge Learning from Data and Validation 25
Table 2 Descriptions of 26 items affecting construction safety performance on sites

Stakeholders Items Descriptions
Owner (O) O1 Safety training and incentive mechanism (Safety training and
incentive mechanisms are performing efficiently)
O2 Safety committee (Safety committee is set up to demonstrate
commitment towards safety on site)
O3 Risk insurance (Risk insurance plans are arranged for the construction
project according to the industry practice)
O4 Permission achievement (All the required permissions are obtained
from relevant government departments via legal means)
O5 Bidding system (The bidding system is implemented strictly to ensure
openness, fairness, and justness in the bidding process)
O6 Contract management (Reasonable constraints are set in contracts to
manage the construction performance)
Contractor C1 Legal certificates (Undertaking the project within the business scope of
(C) its construction certificates)
C2 Safety incentive mechanism (The safety incentive mechanism is well
established to ensure that the organization management and
decision-making processes are professional, practical, accurate, and
efficient)
C3 Sub-contractor management (Procedures to select and evaluate
subcontractors should be developed effectively to ensure the quality and
professionalism of sub-contractors)
C4 Qualified employees (Riggers, electrician, and frame workers should
receive formal occupational training and relevant certificates)
C5 Professional construction schemes (Professional construction schemes
and severe regulations should be developed to reduce unsafe behaviors
of human beings to an acceptable level)
C6 Safety check (The reliability of construction resources should be
checked, tested and inspected on a routine basis)
Supervisory S1 Neutral third party (The construction supervision tasks should be
(S) carried out in an objective and fair manner in accordance with the
entrustment by the owner)
S2 Safety evaluation (Examination of organization design and
construction scheme should be performed strictly to check the safety,
feasibility, and efficacy of the project)
S3 Chief supervisor (The chief supervisor should have comprehensive
quality and leadership skills as to fulfill his duties on working sites)
S4 Field supervisors (A practical on-site inspection plan is executed
effectively to ensure the safety condition and project quality of the
ongoing construction)
S5 Change order (The documented procedures for review and
implementation of changing changes should be executed strictly to
ensure the configuration conformity of the project)
(continued)
Table 2 (continued)
Stakeholders Items Descriptions
S6 Safety inspection (Safety inspection should be conducted on a routine
basis to reduce the hidden hazards)
Designer (D) D1 Compliance in design (Construction design should be in accordance
with the quality and safety standards in the construction industry)
D2 Design change management (Changing inquiries in the design should
be well handled, with the construction quality, progress and investment
taken into account)
D3 On-site service (Technical consulting services should be provided to
relevant departments on construction sites)
D4 Dispute resolution (Procedures for dispute resolution should be stated,
and designers should discuss with responsible departments to solve
disputes in case of incomplete, vague or contradictory requirement)
D5 Safety design (Safety issues should be considered in an efficient way in
the project design)
D6 Scientific innovation (Reasonable guidance should be developed when
new structures, materials or techniques are adopted in the project)
Construction CS1 Internal safety performance (Each stakeholder is well aware of
safety (CS) accident prevention to ensure construction safety)
CS2 Collaborative safety performance (Collaboration among stakeholders
for construction safety management is improved)
and northeast China. To have a global perspective of the safety performance across
multiple stakeholders, all the respondents are asked to indicate their role as an owner,
contractor, supervisor, or designer and answer all the stated questions in the survey.
For instance, a supervisor could answer the questions related to the owner on his or
her perception of the project involvement.
The collected data from the valid 464 respondents act as a foundational resource
for the model establishment and analysis in subsequent sections. According to the
results of statistical analysis among 464 valid respondents, it is clear that 37% of the
respondents are between the age of 30 and 39, 43% have 5–10 years of working expe-
rience in the construction industry, and 52% have bachelor degrees. As a result, the
quality of the questionnaire data is ensured, as the majority of the valid respondents
have good knowledge of actual construction practices.
The fact that initial variables have a strong correlation with each other is a neces-
sary pre-condition during the factor analysis. Accordingly, the Kaiser–Meyer–Olkin
(KMO) test and Bartlett test of Sphericity are conducted. Specifically, the KMO
test measures the adequacy of a sample in terms of the distribution of values for
the execution of factor analysis. Bartlett’s test of Sphericity determines whether the
correlation matrix is an identity matrix. Generally, the minimum desired value of
the KMO coefficient and the Bartlett test coefficient are is 0.8 and 0.01, respectively
[12].
Table 3 Statistics of valid respondents within different stakeholders in the survey

Stakeholder Owner Contractor Supervisor Designer Total
Number of questionnaires responded 127 154 127 99 507
Number of valid respondents 116 142 114 92 464
Ratio of valid respondents (%) 91.34 92.21 89.76 92.93 91.52
The KMO measurement value of the collected data is calculated to be 0.922.

The values of Sig. (Significant probability) of the KMO and Bartlett tests are both
calculated to be 0.000, which are less than the significance level of 0.01. As a result,
the collected data is confirmed to be relevant and suitable for the SEM-based analysis
in the next sections (Table 3).
5.2 Common Method Bias
When a self-reported questionnaire is used to measure variables, there is a possibility

that Common Method Bias (CMB) attenuates the theoretical significance of the find-
ings. CMB is a measurement error that threatens the validity of a conclusion drawn
from statistical results when self-report questionnaires are used to collect opinions
from the participants [26]. It represents variance that is attributable to the measure-
ment rather than to the constructs that the measurement represents. Several analytical
methods have been proposed to estimate CMB, including Harman’s Single Factor,
Common Latent Factor, and Common Marker Variable. Among those methods,
Harman’s Single Factor is the simplest and most widely-used method in the literature
[21] to test if CMB is of concern to the study, in which all items (measuring latent
variables) are loaded into one common factor. If the total variance for a single factor
is less than 50%, it suggests that CMB does not affect the data and the results. Other-
wise, the quality of the data collected from the self-reported survey is not acceptable
for performing further analysis.
Following the guide of “Introduction to SPSS” regarding the calculation of CMB
(Page 34) [17], the Harman’s single factor score of the sample of the collected 464
records is calculated to be 31.883%, which is much lower than 50%, as shown in
Table 4. The 31.883% variance explained by a single factor shows that CMB is not
a major concern in this study, providing preliminary evidence that the measurement
model is robust to CMB.
5.3 SEM Development
Regarding the required sample size for the construction of an SEM, some of the
principles should be followed up (i) Thompson [25] indicated that the sample size
Table 4 Results of the total variance explained in the collected sample of 464 records
Component Total variance explained
Initial eigenvalues Extraction sums of squared loadings
Total % of variance Cumulative % Total % of variance Cumulative %
1 8.29 31.883 31.883 8.29 31.883 31.883
2 3.185 12.249 44.132
3 2.263 8.703 52.835
4 1.674 6.437 59.272
5 1.366 5.253 64.525
6 1.246 4.792 69.317
7 0.729 2.803 72.12
8 0.662 2.547 74.667
9 0.643 2.474 77.141
10 0.575 2.21 79.351
11 0.534 2.055 81.406
12 0.504 1.94 83.346
13 0.484 1.863 85.209
14 0.434 1.67 86.879
15 0.395 1.521 88.4
16 0.372 1.43 89.83
17 0.348 1.339 91.17
18 0.327 1.259 92.428
19 0.315 1.213 93.641
20 0.304 1.169 94.81
21 0.277 1.064 95.874
22 0.248 0.955 96.829
23 0.237 0.912 97.741
24 0.219 0.841 98.583
25 0.196 0.755 99.338
26 0.172 0.662 100
should be at least 10–15 times as much as that of the observed variables; (ii) Byrne
[4] noted that the maximum likelihood (that was the default method used in AMOS)
required the samples to be more than 400 as to better estimate the proposed model;
(iii) Tabachnick and Fidell [24] suggested that the sample number should be at least
200 to construct SEM to make empirical research on it, with the sample number
ranging from 200 to 500. For instance, Fernandez-Muniz et al. [7], Martinez-Corcoles
et al. [16], and Al-Refaie [1] conducted relevant investigations using SEM, and their
sample sizes were 455, 495, and 324, respectively. In this research, there are 26
Table 5 Results of the

Measurement Average variance Composite
internal consistency test in
model extracted (AVE) reliability (CR)
measurement models
Owner 0.53 0.87
Supervisor 0.54 0.87
Constructor 0.64 0.91
Designer 0.67 0.92
observed variables (see Table 6), and thus, the sample with a size of 464 can meet all
the above requirements.
Based on the proposed hypotheses and data collected from 464 valid respon-
dents, the AMOS 21.0 software packages are used to test the hypotheses and gain
insights into a better understanding of the relationships between the above vari-
ables. To examine the impacts of different variables among different stakeholders
on construction safety leadership, many arrows representing causal relationships
between variables are established. Finally, the SEM for investigating the impact of
different stakeholders on construction safety performance can be established.
5.4 Measurement Model Evaluation
Two measures, Average Variance Extracted (AVE) and Composite Reliability (CR),
are used to evaluate the internal consistency in measurement models. Specifically,
AVE estimates the amount of variance captured by a construct’s measure that is
relative to random measurement, while CR is analogous to coefficient α. Bagozzi
and Yi [2] indicated that estimates of AVE great than 0.50 and CR greater than 0.60
could support the internal consistency. The results of the internal consistency test in
measurement models in the established SEM are presented in Table 5. The values
of AVEs range from 0.53 to 0.67, while those of CRs range from 0.87 to 0.92. As a
result, all constructs exhibit good internal consistency in the proposed SEM model.
5.5 Structural Model Evaluation
(1) Hypothesis testing
Hypothesis testing is a primary task in the SEM-based applications. The results

of regression weights in the established model for construction safety leadership
are summarized in Table 6. Figure 3 illustrates the fitted SEM for investigating the
impacts of multiple stakeholders on construction safety performance. The hypothesis
is considered to be acceptable for model establishment if and only if the value of p,
that is the significance level, is less than 0.05. In this study, the results as shown in
Table 6 Results of regression weights in the established SEM

Relationships Regression weights Test results
Estimate Standard Critical p
error ratio
H1 owner → Supervisor 0.201 0.034 5.912 *** Supported
H2 owner → Contractor 0.179 0.058 3.086 0.002 Supported
H3 owner → Designer 0.228 0.056 4.051 *** Supported
H4 supervisor → Contractor 0.181 0.059 3.068 0.002 Supported
H5 designer → Contractor 0.102 0.021 4.857 *** Supported
H6A owner → Construction 0.209 0.094 2.223 0.027 Supported
safety
H6B supervisor → Construction 0.181 0.031 5.839 *** Supported
safety
H6C contractor → Construction 0.328 0.108 3.037 0.002 Supported
safety
H6D designer → Construction 0.280 0.102 2.745 0.016 Supported
safety
Note ***Indicates that p < 0.001
Fig. 3 Fitted SEM for modeling construction safety leadership

Table 6 indicate that all the hypotheses as mentioned above are confirmed. In other
words, the established model as shown in Fig. 3 passes all the hypotheses proposed
earlier in this research.
(2) Goodness-of-fit verification
The assessment of the overall model fit is a critical issue in an SEM, and each
parameter in the hypothetical model should be successfully estimated. The SEM
model should pass through the following three types of assessment indices, namely
absolute fit indices, incremental fit indices, and parsimonious fit indices. There are
several specific indicators in each type of assessment indices. A critical review of
the description and evaluation criteria of those assessment indicators can be seen in
Xiong et al. [30]. Table 7 summarizes the results of the goodness-of-fit verification
in the developed construction safety model (see Fig. 3).
Taking the absolute fit indices as examples, the values of RMR (Root Mean Square
Residual), RMSEA (Root Mean Square Error of Approximation), GFI (Goodness-
of-Fit Index), and AGFI (Adjusted Goodness-of-Fit Index) are 0.048, 0.052, 0.941,
and 0.912, respectively, which are all in line with the recommended value. Among
all the indicators, only the value of RFI (that is 0.784) is below the recommended
value of 0.9. In conclusion, the model is considered to be acceptable, as most of the
goodness-of-fit indices are fairly good.
Table 7 Results of the goodness-of-fit verification in the established SEM

Type of fit indices Indicators Evaluation criteria Value Test results
√
Absolute fit indices χ2 test <0.05, good fit 0.000 < 0.05
√
RMR <0.05, good fit 0.048
√
RMSEA <0.08, not bad fit; <0.05, 0.052
good fit
√
GFI >0.90, good fit 0.941
√
AGFI >0.90, good fit 0.912
√
Incremental fit indices NFI >0.90, good fit 0.908
RFI >0.90, good fit 0.784 ×
√
IFI >0.90, good fit 0.902
√
TLI >0.90, good fit 0.926
√
CFI >0.90, good fit 0.907
√
Parsimonious fit indices PGFI >0.50, good fit 0.763
√
PNFI >0.50, good fit 0.721
√
PCFI >0.50, good fit 0.729
Note AGFI = adjusted goodness-of-fit index; CFI = comparative fit index; GFI = goodness-of-fit
index; IFI = incremental fit index; NFI = normed fit index; PCFI = parsimony comparative fit index;
PGFI = parsimony goodness-of-fit index; PNFI = parsimony normed-fit index; RFI = relative fit
index; RMR = root mean square residual; RMSEA = root mean square error of approximation;
TLI = Tucker-Lewis index; χ2 test = Chi-square test
6 Knowledge Discovery in Safety Leadership
In the above sections, the SEM model for revealing the stakeholder’s relationships
within construction safety leadership is built and verified. Thus, it is rational and
reliable to apply the established model to many kinds of analysis. In this research,
to explore the cause-effect interaction within different stakeholders, the impacts of
path coefficient, stakeholder participation, and construction process on construction
safety performance are investigated.
6.1 Impacts of Path Coefficients
Path coefficients are standardized versions of linear regression weights which can
be used in examining possible causal linkages between statistical variables in SEM.
Path coefficients analysis estimates the effects of variables in a causal system and
has been used to decompose the influences of one variable on another into direct,
indirect, and total effects [8]. In the established SEM (see Fig. 3), relevance and
mediation effects among latent variables (that represent different stakeholders) are
revealed. To verify effect significance, Table 8 illustrates direct, indirect, and total
effects among different stakeholders through the use of bootstrapping in AMOS.
Results are analyzed in detail as follows:
(1) Direct effects. Direct effects measure the strength of the association between
one variable and another variable that is directly collected. As shown in Table 8,
the contractor party displays a most significant direct effect (with a value of
0.3300), followed by the designer party with a direct effect of 0.2800, the owner
party (with a direct effect of 0.2100), and the supervisor party with a direct
effect of 0.1800. It is easy to understand that the contractor party is the direct
stakeholder that is responsible for project safety leadership in construction
projects. The designer and owner parties contribute a lot to project safety issues
Table 8 Results of direct, indirect, and total effects on causal relationships between different
stakeholders and construction safety performance
Relationships Direct effects Indirect effects Total effects Ranking
Owner → Construction 0.2100 0.1563 0.3663 1
safety
Contractor → Construction 0.3300 – 0.3300 2
safety
Supervisor → Construction 0.1800 0.0420 0.2220 4
safety
Designer → Construction 0.2800 0.0330 0.3130 3
safety
6 Knowledge Discovery in Safety Leadership 33
in terms of safety design and strict bidding, respectively. As for the supervisory
party, its direct contribution to the project safety is low.
(2) Indirect effects. Indirect effects measure the strength of the association of one
variable with another variable that is mediated through other variables in the
model. As shown clearly in Table 8, the owner party has the highest indirect
effect (with a value of 0.1563), followed by the supervisor party with an indi-
rect effect of 0.0420 and the designer party with an indirect effect of 0.0330.
No indirect path exists between the contractor party and the construction safety
performance. In total, there are 6 indirect paths leading to construction safety
performance. Taking the owner party as an example, Fig. 4 illustrates the distri-
bution of indirect paths between the owner party and construction safety perfor-
mance. The ranking list of indirect effects displays a great change, compared
to direct effects. This is because the causal influence mechanism varies greatly
within direct and indirect effects. In general, the contractor party leads a promi-
nent role in the direct influence on construction safety performance, while the
owner party leads a prominent role in the indirect influence.
(3) Total effects. Total effects sum up both direct and indirect effects on the target
variable in the model. As shown clearly in Table 8, the owner party has the
highest total effect (with a value of 0.3663), followed by the contractor party
(with a total effect of 0.3300), the designer party (with a total effect of 0.3130),
and the supervisor party (with a total effect of 0.2220). That is to say, the
owner party plays a significant role in construction safety leadership, with
both direct and indirect effects taken into account. This is mainly because
the current Chinese construction industry is a buyer’s market, and the owner
party with a powerful position has the capacity of exerting a big influence on
project performance in terms of both direct and indirect effects. As a result, the
owner party plays a leading role in ensuring the safety performance of a real
project, and more attention should be given to safety issues, such as contractor
bidding, design checking, and contract supervision and management for the
owner party.
0.20
Supervisor 0.18 Contractor
0.33
0.18 Contractor 0.33
Construction
Owner 0.20 Supervisor 0.18
safety
0.23 Designer 0.28
0.23
0.33
Designer 0.10 Contractor
Fig. 4 Strength of indirect paths between the owner party and construction safety performance
6.2 Impacts of Stakeholder Participation
Effective participation of all stakeholders is essential for the delivery of a safe project
throughout the life cycle. In SEM, the observed variables are seen as functions of
the latent variable, whereby changes in the latent variable are reflected or mani-
fested in the observed indicators. A high factor loading on an observed variable
implies that the latent variables are better manifested in the observed variable. To
identify the better reflective variables that impact the safety leadership performance
of each stakeholder, the four measurement models as shown in Fig. 3 are used to
reveal factor loadings of each variable in different stakeholders. The regression coef-
ficients between observed variables and a specific stakeholder are then extracted. The
observed variables that manifest well the latent variables (stakeholders) are picked
up and analyzed as follows.
(1) Owner party. The owner plays a leading role in the project-licensing, approving,
bidding, construction preparation, and fulfillment of the project target. The
ranking list regarding regression coefficients between observed variables (γOi ,
i = 1, 2, …, 6) and the owner’s safety leadership performance is presented
as followed, O4 (γO4 = 0.86) > O5 (γO5 = 0.79) > O6 (γO6 = 0.72) > O3
(γO3 = 0.69) > O2 (γO2 = 0.68) > O1 (γO1 = 0.61). Obviously, O4 (Permis-
sion achievement) and O5 (Bidding system) are the most remarkable observed
variables that reflect the safety leadership performance of the owner. In other
words, the owner’s involvement, including both participation and leadership
in construction safety management, is mainly manifested through the permis-
sion achievement mechanism (O4 ) and the bidding system (O5 ) based on the
perceptions of all the participants in construction. Thus, those issues should
be handled properly in order to maintain the role of the owner’s safety leader-
ship in a project. To be specific, construction activities should be carried out
when the owner has received all the required permissions from government
departments via legal means (O4 ), and the bidding system must be performed
strictly to ensure the openness, fairness, and justness of the contractor selection
process (O5 ). Results indicate that illegal operations in obtaining the required
licenses from the government and black-box operations in the bidding process
are both urgent problems that need to be solved in the current Chinese construc-
tion industry. Particularly, it is commonly accepted by all stakeholders that an
open, fair, and transparent market environment is a key matter to eliminate
subsequent safety hazards in construction projects.
(2) Contractor party. The contractor party is in charge of the entirety of a construc-
tion project in the DBB delivery mode. It needs to handle multi-objective
optimization problems to achieve ultimate goals for costs, schedules, quality,
and safety management standards. The ranking list regarding regression coef-
ficients between observed variables (γCi , i = 1, 2, …, 6) and latent variable (the
contractor’s safety leadership performance) is presented as follows: C2 (γC2 =
0. 86) > C4 (γC4 = 0.84) > C1 (γC1 = 0. 82) = C6 (γC6 = 0.82) > C3 (γC3 = 0.72)
= C5 (γC5 = 0.72). Such variables as safety incentive scheme (C2 ), qualified
6 Knowledge Discovery in Safety Leadership 35
employees (C4 ), legal certificates (C1 ), and safety check management (C6 )
exert remarkable impacts on the manifestation of contractor’s role in safety
leadership, as their regression coefficients are all greater than 0.80. Suprapto
et al. (2011) indicated that the effectiveness of incentive schemes played a
significant role in the owner-contractor working relationships. Ling and Kerh
[13] observed that contractors had higher incident rates with less qualified
employees. It is clear that there are a large number of management and harmo-
nization tasks in the working interface on construction sites, including orga-
nization, technology, economy, and contract issues, and thus, the contractor’s
safety management performance in leadership are excessive and complex.
(3) Supervisor party. The supervisor party is acting as a link or mediating role
between the owner and other parties, who is responsible for providing the owner
with professional management services, including project planning, resource
organizing, quality inspections, and supervisory. The ranking list regarding
regression coefficients between observed variables (γSi , i = 1, 2, …, 6) and
the supervisor’s safety leadership performance is presented as follows: S1 (γS1
= 0.83) > S3 (γS3 = 0.79) > S2 (γS2 = 0.75) > S4 (γS4 = 0.73) > S5 (γS5 =
0.66) > S6 (γS6 = 0.61). Obviously, the completion of construction supervision
tasks in an objective and fair manner (S1 ) is the most prominent observed vari-
able. In addition, the comprehensive quality and professional skills of the chief
supervisor (S3 ) and field supervisors (S4 ) should be also regarded as important
reflective factors. Blazsin and Guldenmund [3] concluded that supervisors’ fair
judgment was strongly and positively related to construction safety. Mahdavi
and Hastak [14] indicated that supervisors with professionally qualified skills
could largely help promote onsite workers’ understanding of task requirements,
operating procedures, safety compliance regulations, and reward and punish-
ment provisions. Thus, the supervisor party should pay close attention to the
development of satisfactory coordination strategies on sites, as well as excellent
supervisors.
(4) Designer party. The designer party is responsible for contributing ideas and
knowledge to create unique architectural solutions, especially for cost-effective
design and reasonable construction methods. The ranking list regarding regres-
sion coefficients between observed variables (γDi , i = 1, 2, …, 6) and the
designer’s safety leadership performance is presented as follows: D5 (γD5 =
0.87) > D4 (γD4 = 0.86) > D2 (γD2 = 0. 82) > D1 (γD1 = 0. 81) > D6 (γD6 = 0.78)
> D3 (γD3 = 0.77). Those factors, such as safety design (D5 ), dispute resolu-
tion (D4 ), design change management (D2 ), and compliance in design (D1 ),
are the most remarkable reflective variables indicating the designer’s role in
construction safety management. Globally, the concept of design for safety has
gained momentum as an intervention to improve worker safety. Olander [20]
indicated that the change order in design was a big threat to construction safety.
Thus, the design company should not only be responsible for the preliminary
and detailed design but also be necessary to offer timely and effective worksite
services and reject unfair requests from the owner in case of the violation of
the stipulations.
7 Merits of SEM
In this research, a total of 464 valid responses are collected and further used for the
model establishment, model verification, and model analysis. A systematic SEM-
based approach is proposed to explore causal-effect relationships on construction
safety performance from a stakeholder’s perspective. The advantages and capacities
of the developed approach are discussed as follows:
(1) Perceiving causal relationships. The causality is the relation between a set
of factors (causes) and a phenomenon. How to determine causal relation-
ships among relevant variables in an effective way, however, is a challenging
problem, as numerous complex processes and behaviors are involved in
construction safety leadership. Managing those complex processes/behaviors is
“tuned” successfully through experience, rather than through the application of
pure mathematics principles. In the developed approach, a causal SEM model
consisting of 26 variables is established to reveal relationships between behav-
iors of different stakeholders and construction safety performance according
to survey data. The established SEM model is considered acceptable, as most
of the verification indices are fairly good. In this way, the hypothetical model is
validated by using survey data, and it is rational and reliable to use the verified
SEM model to perform various sorts of analysis.
(2) Identifying leading roles. Many stakeholders are involved in the construc-
tion safety performance practice, and a large number of complex
processes/behaviors are involved in the safety management for different stake-
holders. The question on which factor plays a leading role in ensuring satisfac-
tory construction safety performance is unknown exactly. The developed SEM
approach enables to depict causal effects, including direct, indirect, and total
effects, between factors in terms of path coefficient analysis. From a global
perspective (stakeholders) or a local perspective (processes and behaviors of a
specific stakeholder), the factor with a high path coefficient in the fitted SEM
model can be identified as a leading factor in the construction safety perfor-
mance model. As shown in Table 8, the owner party has the highest total effect
(with a value of 0.3663), where the owner’s involvement in safety leadership is
mainly manifested through the permission achievement mechanism (O4 ) and
the bidding system (O5 ) based on the perceptions of all the participants in
construction. In this way, the leading roles and factors leading to construction
safety performance can be identified.
(3) Revealing construction dynamics. Causes and effects are typically related to
changes, events, or processes, which are the same in the dynamic construc-
tion process. The allocation of duty and responsibility for construction safety
management among different stakeholders may suffer a change as the construc-
tion process advances. The developed SEM-enabled approach can model the
construction safety performance at different phases, in order to reveal causal
relationships over time among different stakeholders. The direct, indirect, and
total effects on construction safety performance within different stakeholders
7 Merits of SEM 37
over time can be calculated and analyzed in a dynamic manner, based on the
fitted SEM model from the survey data. In this way, the dynamic nature of
construction activities can be taken into account to explore updates of the duty
and responsibility for a specific stakeholder over time.
8 Summaries
This chapter develops an SEM-based approach to exploring interactions among

different stakeholders. A causal model consisting of 26 variables is learned from
the survey data, aiming to reveal relationships between behaviors of different stake-
holders and construction safety performance. The built model has been verified
through the hypothesis test and the goodness-of-fit test. The impacts of path coef-
ficient and stakeholder participation on construction safety performance have been
investigated and analyzed in detail. The developed approach can be used as a deci-
sion tool to reveal the leading roles and the most contributing factors in managing
construction safety performance in a changing project environment.
In light of the survey data and SEM modeling and analysis in this research, some
interesting findings have been obtained. The owner party displays the highest total
effect (or impact) on managing construction safety performance, the contractor party
shows the most significant direct effect, and the owner party holds the largest indirect
effect on managing construction safety performance. The novelty of the developed
approach lies in its capabilities in (a) perceiving causal relationships among stake-
holders from survey data, (b) identifying leading roles in managing construction
safety performance, and (c) revealing the transfer of duty and responsibility within
different stakeholders as the construction advances. The research findings can assist
to allocate the duty on construction safety management among different stakeholders
and further facilitate the enhancement of safety performance in the construction
industry.
References
1. Al-Refaie A (2013) Factors affect companies’ safety performance in Jordan using structural
equation modeling. Saf Sci 57:169–178
2. Bagozzi RP, Yi Y (1988) On the evaluation of structural equation models. J Acad Mark Sci
16:74–94
3. Blazsin H, Guldenmund, F (2015) The social construction of safety: comparing three realities.
Safety Science 71:16–27
4. Byrne BM (2001) Structural equation modeling with Amos: basic concepts, applications and
programming. Lawrence Erlbaum Associates, New Jersey
5. Conchie SM, Moon S, Duncan M (2013) Supervisors’ engagement in safety leadership: factors
that help and hinder. Saf Sci 51:109–117
6. Daily P (2015) A collapse accident on construciton site in Jiangxi. http://news.cntv.cn/2015/
01/29/ARTI1422497971393589.shtml, 29 Jan 2015
7. Fernandez-Muniz B, Montes-Peon JM, Vazquez-Ordas CJ (2009) Relation between occupa-

tional safety management and firm performance. Saf Sci 47:980–991
8. Fox J (1980) Effect analysis in structural equation models extensions and simplified methods
of computation. Sociol Methods Res 9:3–28
9. Hadidi LA, Khater MA (2015) Loss prevention in turnaround maintenance projects by selecting
contractors based on safety criteria using the analytic hierarchy process (AHP). J Loss Prev
Process Ind 34:115–126
10. Henseler J (2017) Bridging design and behavioral research with variance-based structural
equation modeling. J Advert 46:178–192
11. Hubei AOWSI (2013) Investigation report in a catastrophe construction accident. http://www.
hubeisafety.gov.cn/content/content_a.asp?id=7269, 11 Apr 2013
12. Krenzer U (2002) SPSS survival manual. A step manual. A step by step guide to data analysis
using SPSS for windows. Homo J Compar Hum Biol 53:1–81
13. Ling FYY, Kerh SH (2004) Comparing the performance of design-build and design-bid-build
building projects in Singapore. Archit Sci Rev 47:163–175
14. Mahdavi A, Hastak M (2014) Quantitative analysis of bidding strategies: a hybrid agent based-
system dynamics approach. In: Construction research congress 2014: construction in a global
network, pp 1129–1138
15. Martin H, Lewis TM (2013) Pinpointing safety leadership factors for safe construction sites in
Trinidad and Tobago. J Constr Eng Manag 140:04013046
16. Martinez-Córcoles M, Schobel M, Gracia FJ, Tomas I, Peiró JM (2012) Linking empowering
leadership to safety participation in nuclear power plants: a structural equation model. J Saf
Res 43:215–221
17. Mat Roni S (2014) Introduction to SPSS. Edith Cowan University, SOAR Centre, Australia
18. MOHURD (2014) Investigation report on a major construction accident in Henan. http://www.
mohurd.gov.cn/zxydt/201412/t20141230_219958.html, 26 Dec 2014
19. News CC (2014) Accident report on a fatal accident in a bridge construction in Hunan. http://
www.anquan35.com/info/managementdetail-172652.html, 14 Apr 2014
20. Olander S (2007) Stakeholder impact analysis in construction project management. Constr
Manag Econ 25:277–287
21. Podsakoff PM, MacKenzie SB, Lee J-Y, Podsakoff NP (2003) Common method biases in
behavioral research: a critical review of the literature and recommended remedies. J Appl
Psychol 88:879–903
22. Romero Á, González MDLN, Segarra M, Villena BM, Rodríguez Á (2019) Mind the gap:
professionalization is the key to strengthening safety and leadership in the construction sector.
Int J Environ Res Public Health 16(11):2045
23. SAWS (2014) Investigation report on major collapse accident in Yunnan. http://www.lishui.
gov.cn/zwgk/zwxxgk/ajj/ywgz/xzxk_1/t20140122_952651.html, 20 Jan 2014
24. Tabachnick BG, Fidell LS (2007) Using multivariate statistics, 5th edn. Allyn and Bacon, MA
25. Thompson B (2000) Ten commandments of structural eqution modeling. In: Grimm LG,
Yarnold PR, pp 261–283
26. Tsai H-T, Bagozzi RP (2014) Contribution behavior in virtual communities: cognitive,
emotional, and social influences. MIS Q 38:143–163
27. Wu C, Fang D, Li N (2015) Roles of owners’ leadership in construction safety: the case of
high-speed railway construction projects in China. Int J Project Manage 33:1665–1679
28. Wu C, Wang F, Zou PX, Fang D (2016) How safety leadership works among owners, contractors
and subcontractors in construction projects. Int J Project Manage 34:789–805
29. Wu X, Liu Q, Zhang L, Skibniewski MJ, Wang Y (2015) Prospective safety performance
evaluation on construction sites. Accid Anal Prev 78:58–72
30. Xiong B, Skitmore M, Xia B (2015) A critical review of structural equation modeling
applications in construction research. Autom Constr 49(Part A):9–70
31. Zhang L, Chen H, Li H, Wu X, Skibniewski MJ (2018) Perceiving interactions and dynamics
of safety leadership in construction projects. Saf Sci 106:66–78
References 39
32. Zuofa T, Ocheing EG (2017) Senior managers and safety leadership role in offshore oil and
gas construction projects. Procedia Eng 196:1011–1017
Fuzzy Modeling and Reasoning
1 Introduction
In recent years, the construction of subway systems and underground utilities has
increased dramatically due to population pressure and a lack of surface space. The
tunnel boring machine (TBM) has been observed with widespread applications in
tunnel construction, which can be used for excavating tunnels for nearly all types of
rock and geological conditions. However, occurrences of shield cutter-head failures,
cost overrun, and even catastrophic failures of drive motors take place frequently in
projects [26]. For instance, during the construction of the Lesotho Highlands Water
Project, vibration and impact loads on cutters damaged the cutter head, leading to a
serious delay [8]. TBMs experienced high cutter wear, low penetration and advance
rates, groundwater inflow, and tunnel face instability in the Deep Tunnel Sewage
System project, Singapore [27]. During the excavation of the Chengdu metro lines
in China, an extremely high cutter wear in TBMs was reported due to blocked disk
cutters [7]. In general, a decrease in the performance of TBMs can directly affect the
productivity of tunnel projects, leading to unacceptable damages.
TBMs have the advantages of limiting the disturbance to the surrounding ground,
which significantly reduce the cost and failures in lining the tunnel segments. A timely
and appropriate response to changing underground conditions plays a very important
role in ensuring satisfactory performance during the TBM excavation. Actually, one
big challenge is the mismatch between geological variables (e.g., soil density) and
operational variable settings (e.g., the gross thrust of the TBM), which exists in
most of the TBM failures [25]. To precisely predict geological conditions at every
location during a tunnel construction is nearly impossible. Complex TBM systems are
characterized by high dimensions and involve many kinds of variables or subsystems
that are strongly interconnected and mutually dependent. Given complex geological
conditions underground, the operation of TBM is generally “tuned” through expert
experience, where no pure mathematic principles are available [14]. In this situation,
42 Fuzzy Modeling and Reasoning
capturing experts’ knowledge in an effective and effective manner provides a solution

to develop and improve complex system models.
Fuzzy computing techniques have been proposed to create a platform to capture
and integrate expert knowledge in an effective and efficient manner [3, 19]. Fuzzy
Cognitive Maps (FCM) is one such fuzzy computing modeling technique [15]. The
implementation of FCM for modeling complex systems is a promising area, since
FCM can integrate relevant knowledge from domain experts in an efficient manner. In
addition, FCM can simulate the reasoning process of human experts by using a more
abstract representation of the cause-effect mechanism, general control knowledge,
and adaptation heuristics, thereby enhancing the performance of the whole complex
system [22].
There are causes behind every problem. Localization and elimination of those
causes are of the utmost significance in order to prevent the problem or failure. Iden-
tifying the root cause and effect of an event becomes a very important issue in some
domains, such as security, mechanical, and manufacturing [24]. Root Cause Anal-
ysis (RCA) was initially proposed as an error analysis tool by the airline industry
to determine the ultimate cause of a collision or near miss and has since enjoyed
universal adoptions, including fault localization, causal reasoning, causal discovery,
and problem determination [24]. Kum and Sahin [9] used RCA to investigate the
marine accidents/incidents which were recorded by the Marine Accident Investiga-
tion Branch from 1993 to 2011 to reveal their causes. Donauer et al. [6] combined
knowledge of different disciplines and adapted a general framework of a knowledge
discovery database to identify root causes of nonconformities for quality control in
the manufacturing industry.
RCA is useful to identify the root cause to prevent problems from occurring and
recurring. Uncertainty is common in risk assessment and fault analysis. It is possible
that uncertainty will accumulate in the judgment process and finally deduce the
quality of the results. As RCA lacks the ability in quantifying the causal relationship,
it generally cannot deal with the disagreement when facing multiple experts. In
addition, RCA does not consider the confidence level of the identified cause for
each step [25]. Fuzzy thinking considers relationships emphasizing “degrees”, that
is, there are many middle states rather than the two ends “with” or “without”. As one
of the typical fuzzy methods, FCM can address the problems with uncertainty in the
causal inference and merge the various judgment from multiple sources.
In this chapter, a holistic FCM-enabled RCA approach to assessing TBM perfor-
mance is proposed with detailed step-by-step procedures. It incorporates three major
steps, including FCM design, computation, and application. A model consisting of
nine concepts is established for simulating the TBM failure map. A tunnel case in
China is used to verify the effectiveness and applicability of the proposed approach.
Numerous what-if experiments are conducted to explore the cause-effect interaction
of the TBM performance, including predictive, diagnostic, and hybrid RCA.
2 Fuzzy Modeling and Reasoning Methods 43
2 Fuzzy Modeling and Reasoning Methods
The core idea in fuzzy thinking is that “Borders are inexact”. Fuzzy modeling
and reasoning consider the degree of relationship and incorporates the updates for
dynamic process analysis. Fuzzy modeling is a fuzzy valued approach compared to
the two-valued data analysis and reasoning methods. The whole fuzzy set includes
more than two values representing the relationship ranging from weak to strong. In
fuzzy system modeling, the inference is achieved by approximate reasoning. Several
methods have been developed for conducting fuzzy modeling and reasoning, which
can be generally classified into the following four categories.
(1) Probability-Based Reasoning (PBR). PBR aims to combine the capacity of
the probability theory to handle uncertainty with the capacity of deductive
logic to exploit the structure of formal argument, which investigates a natural
extension of traditional logic truth tables, whereas the result is derived through
probabilistic expressions instead. The result is a more expressive formalism
with a broad range of possible application areas. Fault Tree Analysis (FTA) is
one of the most effective techniques for estimating the occurrence frequency of
hazardous events within PBR [5], which is highly recommended as a tool for
risk analysis and assessment in tunnel construction according to “Guidelines
for Tunnelling Risk Management” published by the International Tunneling
Association (ITA) in 2004.
(2) Rule-Based Reasoning (RBR). RBR is known as a reasoning technique with
powerful deductive inference capacity, which can be employed to deal with
complicated realistic problems, such as goal programming, scheduling, and
budgeting [10]. The basic form of a rule is in the form of “if <conditions>,
then <conclusion>”, where <conditions> represents the conditions of a rule and
<conclusion> represents its conclusion. The conditions of a rule are connected
between each other with logical connectives, such as AND, OR, NOT, and
others. When sufficient conditions of a rule are satisfied, the rule is acti-
vated and the conclusion is derived [2]. Fuzzy logic is one typical computing
approach within RBR, which provides a decision tool to solve engineering
problems under uncertainty because of its suitability for representing uncertain
information.
(3) Case-Based Reasoning (CBR). CBR is the process of solving new problems
based on the solutions of similar past problems, which attempts to seek an
approximatively similar case using an analogous reasoning technique, and then
make a corresponding adjustment for problem-solving. The reasoning process
is fast and efficient at the expense of in-depth analysis [10]. CBR has a self-
learning and self-improvement capability by adding new cases continually as
the system operates. When a new input case has to be dealt with, the case-based
system performs inference in four phases known as a CBR cycle: including (i)
retrieve, (ii) reuse, (iii) revise, and (iv) retain [1].
(4) Fuzzy Cognitive Map (FCM). FCM merges cognitive map theory used for
decision making in social science with fuzzy logic to improve causal reasoning.
It is graphically illustrated as a signed fuzzy graph with feedbacks, consisting

of nodes and weighted interconnections. FCM can be used to carry out many
kinds of what-if experiments. For instance, FCM was used to perform RCA
to identify root causes in various conditions [24]. Maitra and Banerjee [12]
used FCM to improve the production and maintenance of mining and plant
machinery. FCM provides an immediate understanding and identification of the
root causes of an event. It allows multiple experts to collaboratively contribute
to an FCM, which enhances the confidence in using FCM [18].
The above PBR, RBR, and CBR methods have several advantages in under-
standing, inferencing, and estimating the performance of complex systems. However,
those methods have major constraints in applications. For instance, Mentes and
Helvacioglu [13] indicated that the occurrence probability of basic events in tradi-
tional FTA was always regarded as a crisp value, whereas it was difficult or nearly
impossible to obtain exact values of probability due to the lack of sufficient data in
reality. Prentzas and Hatzilygeroudis [16] claimed that RBR was inefficient due to
knowledge acquisition bottleneck and brittleness of rules, while CBR was hard to
express general knowledge due to the inference efficiency problems and the provi-
sion of explanations. In addition, Liu et al. [11] indicated that RBR and CRB were
difficult to detect causes in case a lower system performance was observed.
In regard to tunnel construction, some situ observation techniques are used to
capture real data as inputs for system reasoning. Particularly, the performance of the
TBM system is dependent on many inter-related factors that are sometimes contra-
dictory and/or can only be defined by subjective evaluations. Compared to PBR,
RBR, and CBR methods, FCM requires much fewer efforts to make the knowledge
explicit by listing the concepts and constructing causal interconnections between
concepts. Given imprecision and uncertainty are underlying the data measurement
in reality, FCM can provide a solution for capturing such elusive factors based on
expert knowledge and analyzing dynamic behaviors of a complex system. Thus,
FCM is employed as an appropriate tool to analyze the complex TBM performance
in tunnel construction in this research.
3 A Holistic FCM Approach
FCM enables systematic causal propagation in terms of forward and backward

chaining. With the reasoning technique of FCM, a systematic decision approach for
simulating, analyzing, and assessing the complex system performance in construction
engineering projects is developed. Figure 1 illustrates the workflow of the developed
holistic FCM approach. Description of the incorporated three steps is elaborated as
follows.
3 A Holistic FCM Approach 45
Identify nodes (concepts)
Inputs Step 1. FCM

Identify arcs
Design
Aggregate arc weights
y = tanh(x)
Determine threshold
functions
Prior knowledge or Step 2. FCM
given observations Computation
Model dynamics
Predictive RCA
Feedback Step 3. FCM

Diagnostic RCA
Analytics
Hybrid RCA
Fig. 1 The workflow of the developed holistic FCM approach
3.1 FCM Development
(1) Identification of concepts/nodes
FCMs are dynamic systems that have the topology of a directed fuzzy graph
consisting of nodes, arcs, permitting cycles, and feedbacks. Figure 2 illustrates a
simple FCM model. Nodes in the graph refer to the concepts that are used to describe
behaviors of the system, and they are connected by signed and weighted arcs repre-
senting causal relationships among concepts. It should be noted that all the values
in the graph are fuzzy, indicating that the concept nodes have a fuzzy nature. In the
FCM framework, each concept, defined by C i with an interval of [−1, +1], represents
an influential variable that has an impact on the system performance (e.g., events,
Fig. 2 A simple example of

FCM
w12
C1 C2
w13 w42
C3 w34 C4
actions, goals, values, or trends of the system). Relationships and interdependencies

among concepts are represented by arcs (i.e., interconnections) among concepts in
FCM models.
Concepts in an FCM (C i , C j , …, C n ) represent the state of a system. At time t,
the state of an FCM is defined by the vector At of the values of the concepts (At =
[C 1 , …, C n ] ∈ [−1, +1]). In regard to the complex construction engineering projects
(e.g., tunnel construction), the identification of FCM concepts or nodes involves the
collection of prior knowledge resources. The resources for the identification of FCM
concepts can be derived from three aspects, including (i) the accepted knowledge
from existing design and construction standards, (ii) the empirical knowledge from
statistical models in the literature, and (iii) the tacit knowledge from domain experts.
In addition, in order to represent or indicate system outcome, flow process, and
performance states of the identified FCM concepts (or nodes), the data sources can
be obtained from two aspects, including the hard data that is generated by electronic
sensors and soft data that is generated by humans. Both hard and soft data resources
that indicate inputs of an FCM model are used to capture and perceive the magnitude
of model output.
(2) Identification of casual relationships/arcs
An arc eij (∈ [−1, +1]) defines a causal flow C i → C j between the concepts.
Based on the graphical representation, it becomes clear which concept influences
other concepts, and what the degree of the influence is. FCM permits thoughts
and suggestions to be collected or aggregated in the construction of the graph
by adding or deleting an interconnection or a concept. According to the poten-
tial failure-consequence scenarios for the target event, explicit relationships and
interdependencies among the identified concepts or nodes can be revealed.
Given an FCM model with N concepts, a weight from the concept C i to the
concept C j (1 ≤ i, j ≤ N, i = j), defined by W ij , is the relative influence from
concept C i to concept C j . Fuzzy weights can be used to describe the degree of
the causal interconnection among concept nodes. There are three types of causal
interconnections between concepts for representing the degree of influence from one
concept to the other:
• W ij > 0: The arc from C i to C j indicates a positive causal relationship. An increase
in the value of C i leads to a corresponding increase in the value of C j , and similarly,
a decrease in the value of C i leads to a decrease in the value of C j .
• W ij < 0: The arc from C i to C j has a negative causal relationship. An increase in
the value of C i leads to a decrease in the value of C j , and vice versa.
• W ij = 0: This implies there is no relationship between C i and C j .
(3) Aggregation of weights from multiple experts
The causal interrelationships among concepts are declared using a linguistic vari-
able that takes a value in the universe U = [−1, +1]. It is known that the number
of linguistic intervals can largely affect the reliability of estimation. Specifically, a
narrow interval indicates that the statistic is precisely known, while a wide interval
μnvs μns μnm μnw μz μpw μpm μps μpvs

1.0
Membership
0.5
0
-1.0 -0.75 -0.50 -0.25 0 +0.25 +0.50 +0.75 +1.0
Influence
Fig. 3 Membership functions for the 9-point linguistic scale regarding the degree of influence
between concepts
indicates great uncertainty. In order to reach a high precision of the influence estima-
tion among concepts, it is better to conduct a group of narrow intervals, rather than
wide intervals. According to Dawes’ experiment, the use of the 5–9 point scale is
likely to produce slightly higher mean scores relative to the highest possible attain-
able score, and besides, a larger number of the intervals is usually impractical [4]. For
instance, Stylios and Groumpos [21] used a 9-point linguistic scale to show the fuzzy
set of the linguistic variable influence terms. Accordingly, in this chapter, the 9-point
linguistic scale is used as a basis for an expert to describe in detail the influence of
one concept on another based on his/her knowledge. Figure 3 illustrates the corre-
sponding membership functions for these 9 linguistic terms, including negatively
very strong (μnvs ), negatively strong (μns ), negatively moderate (μnm ), negatively
weak (μnw ), zero (μz ), positively weak (μpk ), positively moderate (μpm ), positively
strong (μps ), and positively very strong (μpvs ), respectively.
In order to exploit the knowledge and experience of experts in the description
and modeling of a complex system, it is assumed that there are M experts who
evaluate every interconnection and describe it with a fuzzy rule inferring a linguistic
weight. The credibility weight for every expert is adopted to increase the objectivity
of the FCM development. In the traditional fuzzy decision making approaches, the
weights of experts are assumed to be equal to each other or fixed without any changes
[17]. However, the level of expert confidence can reveal the reliability of the data
obtained from interviews with various individuals. In the construction practice, it is
generally known that the judgment ability of individuals tends to become increasingly
sophisticated and stable with the accrual of educational background and working
experience.
In this chapter, all the experts are equally credible initially and have the same
credibility weight at the beginning. When determining the weight of the intercon-
nection in FCM, at least M/3 of the experts have to fully agree with suggestions, thus,
an average weight of the interconnection is calculated using Eq. (1). Otherwise, the
credibility weight of an expert should be reduced by r% every time when there is a
wrong suggestion for an interconnection. Next, the suggested linguistic weights for
an interconnection may be aggregated. A defuzzification method, center of gravity
(COG), is used to produce an aggregated linguistic weight, and a numerical weight
for the interconnection is calculated.
M
k=1 bk × W k
ij
Wi j = (1)
M
where W ij refers to the aggregated fuzzy weight from C i to C j ; M refers to the
number of the participating experts in total; bk refers to the credibility weight of the
kth expert; W ij k refers to the fuzzy weight from C i to C j based on the judgment of
the kth expert.
3.2 FCM Computation
The causal interconnections of an FCM model can be expressed by an n × n matrix

W, which contains all the n2 rules or pathways in the causal web between n concepts
in the FCM. The state of the FCM model can be expressed by a 1 × n vector A. Then,
the dynamics of the FCM model are dictated by these matrices, where the model can
be updated through iteration with other concepts. Equation (2) drives this interaction
from the previous state to the next state [22]. After a set of iterations, the values of
the concepts will evolve until being stabilized at a fixed point or a cycle limit for
stable systems.
⎛ ⎞

n

Ai (t + 1) = f ⎝ Ai (t) + W ji × A j (t) ⎠ (2)
j=1, j=i
where t refers to the interaction index; Ai (t + 1) refers to the value of concept C i

at time t + 1; Ai (t) and Aj (t) refers to the value of concept C i and C j at time t,
respectively; W ji refers to the interconnection weight from concept C j to concept C i ;
f (•) refers to a threshold function.
Generally, there are two main kinds of threshold functions used in the FCM frame-
work. One is the unipolar sigmoid function, where λ > 0 determines the steepness of
the continuous function f (•) and squashes the content of the function in the interval
[0, 1], as shown in Eq. (3). Another is the hyperbolic tangent function which can be
used to transform the content of the function into the interval [−1, +1], as shown in
Eq. (4).
1
f (x) = (3)
1 + e−λx
e x − e−x
f (x) = tanh(x) = (4)
e x + e−x
where f (x) is the threshold function, and x is the value of the node/concept in the
FCM framework.
The selection of the threshold function depends on the method that is used to
describe the concepts. In this chapter, we attempt to assess the impacts of root causes
on the reliability of the target event in terms of “positive influence” and “negative
influence”. As a result, the sigmoid threshold function as shown in Eq. (4) is selected
to be applied later in a case study in this chapter.
3.3 FCM Analytics
Once constructed, FCM can be used to simulate the behavior of a system and perform
what-if analysis, due to its capability in fuzzy modeling and reasoning [15]. Several
variables may be involved in the occurrence of a failure, however, which variable
has a greater contribution is unknown exactly. The FCM model should consist of
at least one concept representing the target event, while some concepts representing
the influential inputs. RCA is able to diagnose the root cause and predict the future
outcome of an event. In order to identify the root cause and effect of an event in a
complex system, we focus on the predictive, diagnostic, and hybrid RCA using the
FCM inference technique.
(1) Predictive RCA: Predictive RCA aims to forecast the future outcome or effect
of an event when some evidence or information is available. The propagation
of evidence of a possible cause concept allows an update of the outcome of the
target event in the FCM framework, leading to the change in the outcome of
the target event over time.
(2) Diagnostic RCA: Diagnostic RCA aims to identify the possible root cause(s)
given the occurrence of a failure is observed. The propagation of the observed
information of the target event allows an update of the outcome of other possible
cause concepts in the network, leading to the change in the outcome of the cause
concepts over time.
(3) Hybrid RCA: Hybrid RCA is an integration of both predictive and diagnostic
RCA. The given evidence information of both the target event and possible
causes allows an update of the outcome of both the target event and possible
causes in the FCM framework, leading to the change in their outcomes over
time.
It should be noted that the predictive, diagnostic, and hybrid RCA can be
conducted by using the same equation that is Eq. (2). The difference lies in that
only the evidence of certain root causes is observed in the predictive RCA, only the
evidence of the target event is observed in diagnostic RCA, while the evidence of
both the target event and certain root causes can be observed in hybrid RCA.
4 TBM Performance in Tunnel Construction
In regard to underground excavation in soft soils, there are two main types of TBMs,
namely earth pressure balance (EPB) TBM and slurry TBM. Compared with the
slurry TBM, the EPB TBM is relatively inexpensive and does not need a large space
on working sites, and thus, it is commonly used with a broad range of applications,
especially in the construction of underground tunnels in urban areas. In this chapter,
we mainly focus on modeling, predicting, and diagnosing the EPB TBM performance
using FCM, where the updating dynamics of the TBM system over time are taken
into full consideration.
4.1 TBM Failure Mechanism
EPB TBM is a mechanized tunneling method where the spoil is admitted into TBM
via a screw conveyor arrangement. It allows the pressure at the face of the TBM to
remain balanced without the use of slurry. A screw conveyor transports the exca-
vated material from the base of the excavation chamber onto a belt conveyor. The
interaction between the screw conveyor and the TBM’s advance rate ensures that the
support pressure of the soil paste can be controlled precisely. Generally, TBM perfor-
mance is measured in terms of Advance Rate (C T ), which is the actual excavating
and supporting distance divided by the total time, including downtime for TBM
maintenance, machine breakdown, and tunnel failure. C T depends on various types
of factors ranging from machine design to geological features. Therefore, it may not
be possible to capture this complex relationship in an explicit mathematical formula.
In accordance with an extensive review of the literature, the most important param-
eters that affect the TBM performance can be divided into two major categories, as
follows:
• Geological variables: The construction of a tunnel project is very sensitive to
geological conditions due to complex tunnel-soil interactions. Some particular
geological features may increase the hazardous nature of the TBM advancing
process. Soil Cohesion (C 1 ), Friction Angle (C 2 ), Compression Modulus (C 3 ),
and Soil Density (C 4 ) are four variables frequently used to illustrate the geological
features.
• Operational variables: In the process of the TBM-driven tunneling excavation,
some pressure and speed sensors are installed on the top and middle of the cutter
head in advance. Engineers pay close attention to the measurement of those oper-
ational variables, in order to maintain the face stability of the excavation and
minimize the risk. Rational and appropriate Gross Thrust (C 5 ), Cuter Torque
(C 6 ), Earth Pressure (C 7 ), and Grouting Speed (C 8 ) are key to guarantee the
smooth excavation process. These variables can be further adjusted, depending
on geological conditions underground, in order to maintain the TBM’s satisfactory
performance under excavation.
4 TBM Performance in Tunnel Construction 51
Table 1 Definition of nine concepts involved in the TBM operation

Code Concept Description
C1 Soil Cohesion Component of shear strength of a soil
C2 Friction Angle The force that resists the relative motion of solid surfaces, fluid
layers, and material elements sliding against each other
C3 Compressive Modulus The capacity of a material or structure to withstand loads
tending to reduce size
C4 Soil Density Mass of many particles of the soil divided by the total volume
they occupy
C5 Gross Thrust Total thrust during the TBM operation
C6 Cutter Torque The torque of the cutter disk during the TBM operation
C7 Earth Pressure The pressure at the face of the TBM to remain a balanced
performance
C8 Grouting Speed Speed of the concrete grouting at the shield tail
CT TBM Advance Rate Actual excavating and supporting distance divided by the total
time in TBM operation
Finally, a total of nine concepts C i (i = 1, 2, …, 8, T ) are selected to model

the performance of the complex TBM system. Table 1 illustrates the definition of
these nine concepts. Due to the imprecise and uncertain nature of the concepts (or
variables) in the real world, the determination of the most appropriate values of those
concepts, particularly for operational variables, remains a problem. One significant
reason is that the scope of favorable or unfavorable geological/operational variables
varies greatly among different tunnel cases in different areas, leading to challenges
in determining the optimum values of specific operational variables.
To solve this problem, a 5-point linguistic scale provides a basis for an expert to
describe in detail the favorability of a specific concept in response to the TBM perfor-
mance based on his/her knowledge. Figure 4 illustrates corresponding membership
functions for these 5 terms: negatively very unfavorable (δnvu ), negatively unfavor-
able (δnu ), favorable (δn ), positively unfavorable (δpu ), and positively very unfavor-
able (δpvu ). For instance, if the concept C 1 has a value of 0, it means C 1 lies in a
very favorable range in response to the TBM performance. On the contrary, if the
concept C 1 has a value of −1 (or +1), it means C 1 lies in a negatively very unfa-
vorable (or positively very unfavorable) range in response to the TBM performance.
Given several experts are involved in the determination and measurement of concepts
in a specific project, the credibility weight for every expert is used to increase the
objectivity during the integration of expert knowledge from varying backgrounds, as
well as to enhance the confidence in using FCM. As mentioned above in the section
of fuzzy weight determination, both the well-known fuzzy logic method and the
defuzzification method COG are used to obtain a crisp value of a specific concept.
δnvu δnu δf δpu δpvu

1.0
Membership
0.5
0
-1.0 -0.5 0 +0.5 +1.0
Favor
Fig. 4 Membership functions for the 5-point linguistic scale regarding concept measurement
4.2 TBM Failure Map Modeling
To demonstrate the effectiveness and practicality of FCM for analyzing the TBM
performance, a tunnel project, the construction of Metro Line 7 in the Wuhan metro
system in China, is used as a case study to demonstrate FCM’s capability in addressing
engineering problems with uncertainty. The Wuhan metro system is an elevated and
underground urban metro system in the city of Wuhan, Hubei, China. Several urban
metro lines are developed to serve the city of Wuhan, as shown in Fig. 5a. The Metro
Line 7 is a new line with 19 stations for Phase One in the Wuhan metro system,
and the first tunnel cross section of that line, namely Dongfangmacheng–Changfeng
Section (DCS) as shown in Fig. 5b, is used as a case to analyze and assess root causes
in case of poor TBM performance in this chapter. DCS is a double-track tunnel with a
length of 1714 m on the right track and 1717 m on the left track. Two EPB TBMs with
a diameter of almost 6.2 m are used to excavate soils from the Dongfangmacheng
Station to the Changfeng Station at an average depth of 20 m below the ground
surface. The excavation of the right track started on August 22, 2014, and that of the
left track started on September 15, 2014. Figure 5c is a photograph of EPB TBM at
the construction of DCS.
The increase in the complexity and sophistication of large-scale logistics systems
requires the implementation of new intelligent strategies. Capturing the heuristic
knowledge using FCMs may provide the foundation for a new direction in modeling
complex logistics systems. The main requirements for sophisticated systems are
the possession of human-like knowledge and expertise in specific domains [20]. To
explore the complex interactions/logics in the TBM operation system, the knowledge
and experience of a group of experts who operate, monitor, and supervise the system
were obtained through questionnaires. A questionnaire was prepared and distributed
among ten experts who participated in the construction of DCS, including seven
domain experts with at least five years of working experience and three professors
4 TBM Performance in Tunnel Construction 53
(a) (b)
Dongfangmacheng
Station
Changfeng Station
(c)
Fig. 5 Metro lines in the Wuhan metro system: a map of the Wuhan metro system for 2017;
b location of DCS at the Metro Line 7; c EPB TBM at the construction of DCS
in the domain. The experts were asked to evaluate the influence/weight between the
concept C i and the concept C j (i, j = 1, 2, …, 8, T; i = j) using the 9-point linguistic
terms as shown in Fig. 3. In accordance with the methodology, as mentioned in
Sect. 3, the linguistic inputs from the experts were transformed into precise values
using the defuzzification method of COG.
Taking the determination of the influence weight from C 1 to C T as an example,
among the above ten experts, there are seven experts who have a judgment of nega-
tively weak (μnw ), two experts have a judgment of negatively moderate (μnm ), and
only one expert has a judgment of zero (μz ). Figure 6 shows an illustration of how each
individual expert’s input is entered, and multiple inputs are automatically aggregated
and defuzzified. The weight of W 1T is then determined to be −0.275. With knowl-
edge from all the participated experts integrated in an effective manner, Fig. 7 depicts
the graphical representation of the obtained FCM model for the TBM performance
at the construction of DCS.
5 FCM-Enabled TBM Performance Analysis
The operation of the TBM system plays a significant role in ensuring satisfactory
TBM performance in tunnel construction. It is, therefore, necessary to investigate
the cause-effect mechanism of the TBM operation and understand how the TBM
performance evolves quantitatively in case either geological or operational variables
μnw
1.0
Membership
7×
Experts
0.5
7.0
0
-1.0 -0.75 -0.50 -0.25 0 +0.25
Influence
6.0
μnm
1.0
5.0
Membership
2×
Membership
0.5 4.0
Experts
3.0
0
-1.0 -0.75 -0.50 -0.25 0 +0.25
Influence
2.0
μz
1.0 1.0
Membership
COG = -0.275
1×
Expert
0.5
0
-1.0 -0.75 -0.50 -0.25
Influence
0 +0.25
0
-1.0 -0.75 -0.50 -0.25 0 +0.25
Note: Assume all the experts have the
Influence same credibility weight of 1.0.
Fig. 6 Aggregation of the weight W 1T from ten experts
undergo a change. After the FCM model is built, various sorts of what-if experiments
can be conducted to explore the cause-effect interaction of the TBM operation system.
In this chapter, we focus on the predictive, diagnostic, and hybrid RCA of the TBM
performance in tunnel construction.
5.1 Predictive RCA
Predictive RCA aims to forecast future outcomes when the evidence of a vari-
able/concept is observed. In the FCM framework, the propagation of evidence of
a possible cause concept allows an update of the outcome of the target concept in
the network in light of the newly found evidence. As mentioned above, the values of
concepts correspond to real measurements that have been transformed in the interval
[−1, +1] using the 5-point linguistic scale (see Fig. 4). The concept has a value
of 0 (−1 or +1), indicating it lies in a very favorable (negatively very unfavorable,
or negatively very unfavorable) range in response to the TBM performance. Hence,
when the evidence is set on a specific cause concept in the FCM model as shown in
5 FCM-Enabled TBM Performance Analysis 55
5 1.0
1.0 17
-0.
5
0.70
C8 -0.125 C1
68
0.5
0.8
22
0.
1
8
1.0 C7 0.4 C2 1.0
5
0.8
27
64
-0.
5
52
5
37
-0.
-0.
-0.35
5
0.42
0.2
5
CT
0.315
-0.63
6
-0.
1
8
0.
0.56
2
-0.
.575
81
0.35
0
8
1.0 C6 C3 1.0
5
0.37
0.
0.
42
25
42
0.
5
-0.636
5
0.12
C5 -0.22
C4
5
1.0 1.0
Fig. 7 Graphical representation of the obtained FCM model for the TBM performance at the
construction of DCS
Fig. 7, the evidence will propagate to the other concepts in the network, which will
cause the outcome of the target event to change.
In this experiment, we attempt to investigate the impacts of variations of a specific
cause concept in the predictive RCA on the TBM performance, in order to identify
the direct cause-effect relationship between the cause concept C i (i = 1, 2, …, 8)
and the target concept C T on a quantitative basis. To be specific, in the initial stage,
all the concepts are set to lie at a favorable range (the value is set to be 0), except
one specific concept C i (i = 1, 2, …, 8). It is assumed that the concept C i is set to
lie in sequence at a negatively very unfavorable level (−1.0), negatively unfavorable
level (−0.5), positively unfavorable level (+0.5), and positively very unfavorable level
(+1.0), respectively, and then we monitor the outcome of system behavior of the TBM
performance (that is the concept C T ) in different scenarios. Using Eqs. (2) and (4),
the value of the concept C T will evolve and then stabilize at a fixed point after several
interactions. Figure 8 illustrates the impacts of variations of geological variables on
the TBM performance in the predictive RCA. Figure 9 illustrates the impacts of
variations of operational variables on the TBM performance in the predictive RCA.
Table 2 shows the fixed values of C T after a set of iterations in different scenarios.
Taking the concept C 1 as an example (see Fig. 8a and Table 2), the concept C T tends
to have a fixed value of +0.7861 (or +0.6660) when the concept C 1 has a value of
−1.0 (or −0.5), and the concept C T tends to have a fixed value of −0.7861 (or −
(a) (b)
(c) (d)
Fig. 8 Impacts of variations of geological variables on the TBM performance in the predictive
RCA: a the concept C 1 ; b the concept C 2 ; c the concept C 3 ; d the concept C 4
0.6660) when the concept C 1 has a value of +1.0 (or +0.5), respectively. It is clear
that a markedly negative correlation exists between C 1 and C T .
In general, the concepts C 1 , C 2 , C 3 , and C 4 all have markedly negative corre-
lations with the concept C T , while the concepts C 5 , C 6 , C 7 , and C 8 all perform
markedly positive correlations with the concept C T . With respect to the measure-
ment of this kind of correlation among different concepts, the concept C 4 displays
the strongest negative correlation with the concept C T , while the concept C 8 displays
the strongest positive correlation with the concept C T . That is to say, the concepts C 4
and C 8 are both very sensitive to exert impacts on the TBM performance, and thus,
they should be given significant attention in ensuring the satisfactory performance in
tunneling excavation. This is very consistent with what is observed in actual construc-
tion practice, indicating the established FCM model is validated to some extent. In
fact, to maintain the face stability of the excavation and minimize risk, engineers
need to adjust the operational parameters continuously during the TBM operation in
order to fit the surrounding geological conditions in a dynamic manner. The results
achieved from the predictive RCA can provide an increased understanding of how
the TBM performance evolves quantitatively in case either the geological or opera-
tional variables change, which can provide guidance for the reasonable adjustment
of operational variables in tunnel construction.
(a) (b)
(c) (d)
Fig. 9 Impacts of variations of operational variables on the TBM performance in the predictive
RCA: a the concept C 5 ; b the concept C 6 ; c the concept C 7 ; d the concept C 8
Table 2 Fixed values of C T after a set of iterations in different scenarios in the predictive RCA
Items The fixed value of C T in different scenarios
P(C T | C i = −1.0) P(C T | C i = −0.5) P(C T | C i = +0.5) P(C T | C i = +1.0)
C1 0.7861 0.6660 −0.6660 −0.7861
C2 0.8374 0.7199 −0.7199 −0.8374
C3 0.9137 0.8105 −0.8105 −0.9137
C4 0.9426 0.8511 −0.8511 −0.9426
C5 −0.8263 −0.7078 0.7078 0.8263
C6 −0.8990 −0.7915 0.7915 0.8990
C7 −0.8570 −0.7417 0.7417 0.8570
C8 −0.9480 −0.8595 0.8595 0.9480
5.2 Diagnostic RCA
Diagnostic RCA aims to detect possible root cause(s) when a failure of the target
event occurs. In the FCM framework, the propagation of the observed information
of the target event allows an update of the outcome of other variables or causes in the
network. As mentioned above, the values of the TBM performance correspond to real
measurements have been transformed into the interval [−1, +1] using the 5-point
linguistic scale (see Fig. 4). Hence, when a specific level of the TBM performance
(such as the negatively or positively unfavorable performance) is observed in the
target event (that is the concept C T ) in the FCM model, the observed evidence will
then propagate to the other cause concepts in the network, resulting in the change of
the causes affected accordingly.
In this experiment, we intend to identify the most likely root causes that affect
the reduction of the TBM performance in the diagnostic RCA. In the initial stage,
all the concepts are set to lie at a favorable range (the value is set to be 0), except
the target concept C T . It is assumed that the concept C T is set to lie at a negatively
very unfavorable level (C T = −1.0), a negatively unfavorable level (C T = −0.5),
a positively unfavorable level (C T = +0.5), and a positively very unfavorable level
(C T = +1.0) in turn, respectively, and then we monitor the outcome of behaviors
of other causes in different scenarios. Using Eqs. (2) and (4), the value of a specific
concept C i (i = 1, 2, …, 8) will evolve and then stabilize at a fixed point after several
interactions. The impacts of variations of the TBM performance on geological and
operational variables in the diagnostic RCA are shown in Fig. 10. Results indicate
that the values of geological variables C i (i = 1, 2, 3, 4) remain unchanged over
time since no arcs exist from the concept C T to the concept C i (i = 1, 2, 3, 4)
(a) (b)
(c) (d)
Fig. 10 Impacts of variations of the TBM performance on geological and operational variables in
the diagnostic RCA: a C T = −1.0; b C T = −0.5; c C T = +0.5; d C T = +1.0
Table 3 Fixed values of operational variables after a set of iterations in different scenarios in the
diagnostic RCA
Scenarios Fixed value of C i in different scenarios
C5 C6 C7 C8
P(C i | C T = −1.0) −0.8374 −0.9006 −0.8476 −0.9426
P(C i | C T = −0.5) −0.7199 −0.7936 −0.7311 −0.8511
P(C i | C T = +0.5) 0.7199 0.7936 0.7311 0.8511
P(C i | C T = +1.0) 0.8374 0.9006 0.8476 0.9426
in the FCM model (see Fig. 7). This is due to the fact that the variations of the
TBM performance cannot lead to the change of geological conditions itself. The
fixed values of operational variables after a set of iterations in different scenarios
are presented in Table 3. Taking the scenario where the target concept C T suffers a
great negative reduction regarding the TBM performance (that is C T = −1.0) as an
example (see Fig. 10a), the concept C 8 tends to have the largest negative fixed value
(that is −0.9426). That is to say, this variable should be deemed as the suspected
cause which should be checked, given that a great negative reduction of the TBM
performance is observed.
In general, as seen in Table 3, the final fixed value of each concept varies greatly
in different scenarios, indicating that the degree of the cause-effect relation varies
a lot among different concepts. More specifically, the TBM performance is very
sensitive to the change of operational conditions. The concepts C 5 , C 6 , C 7 , and C 8
show positive correlation effects in response to the variation of the TBM performance.
Engineers should give much more attention to the adjustment of operational variables,
such as C 5 , C 6 , C 7 , and C 8 , since the values of geological variables cannot be changed
at the construction stage. The values of operational parameters, however, can be
adjusted to go up in case the TBM performance negatively reduces, and go down in
case the TBM performance positively reduces. In conclusion, the results achieved
from the diagnostic RCA can provide more insights on what kind of variables cause
the reduction of the TBM performance, as well as the direction of adjustment of
relevant variables to ensure the satisfactory performance of the TBM operation.
5.3 Hybrid RCA
Hybrid RCA allows a combination of predictive and diagnostic inferences at the

same time. In the FCM framework, the possible root causes can be detected in case
a failure of the target event occurs by means of the diagnostic analysis, and then the
condition of the detected causes will be checked and tested, in order to make the
target event return to a normal state in a short time. The given evidence information
of both the target event and possible causes allows an update of the outcome of both
the target event and possible causes over time. The combination of predictive and
diagnostic inferences will cause either a cumulative or a cancellation effect on the

update of a particular state of other variables in the domain. The cumulative effect
will increase the state of the target event to a higher value, as compared with the
individual evidence obtained separately for the predictive RCA or diagnostic RCA.
Conversely, the cancellation effect will trim down the value of the target event.
As mentioned above in the diagnostic RCA, the concept C 8 should be deemed as
the suspected cause when the target concept C T suffers a great negative reduction
regarding the TBM performance (that is C T = −1.0). We can then adjust the value
of this operational variable to make the TBM performance go back to a normal state.
However, how to efficiently adjust the values of operational variables in response
to the variation of the TBM performance remains a challenging problem, and this
problem can be addressed by using the hybrid RCA.
In this experiment, we aim to determine the specific value that an operational
variable should be adjusted when the target concept C T (TBM performance) suffers
a negative or positive reduction. We take the negative reduction of the target concept
C T (that is C T = −1.0 or −0.5) as an example in this experiment, since the positive
reduction performs exactly opposite impacts on other concepts (see Table 3). To be
specific, in the initial stage, the concept C T is set to lie at a negatively very unfavorable
level (C T = −1.0) or a negatively unfavorable level (C T = −0.5), and all the other
concepts are set to lie at a favorable range (C i = 0, i = 1, 2, …, 7), expect the concept
C 8 . The value of C 8 is set to be 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, and
0.10 in turn, respectively, and then we monitor the outcome of system behavior of the
TBM performance (that is the concept C T ) in different scenarios. Using Eqs. (2) and
(4), the value of the concept C T will evolve along with a series of iterations. Figure 11
illustrates the evolution of TBM performance in the case of different initial values
of the concept C 8 , when the value of the concept C T is initially set to be −1.0 and
−0.5, respectively. Herein, the number of iterations is set to be ten, since the TBM
performance is expected to return to the normal state (that is C T = 0) in a very short
(a) (b)
Fig. 11 Evolution of TBM performance in the hybrid RCA in the case of different initial values of
the concept C 8 in different scenarios: a C T = −1.0; b C T = −0.5
time, and the system components of TBM, such as the shield cutter head, are very
likely to be damaged otherwise.
As shown in Fig. 11a, in order to return the TBM performance from a negatively
very unfavorable level (C T = −1.0) to a favorable level (C T = 0) in a short ten
iterations, it would be a reasonable and sensible decision if we adjust the value of
C 8 to be 0.07. Similarly, as shown in Fig. 11b, to return the TBM performance from
a negatively unfavorable level (C T = −0.5) to a favorable level (C T = 0) in a short
ten iterations, it would be a reasonable and sensible decision if we adjust the value
of C 8 to be 0.05. This deduction is consistent with the actual tunneling construction
practice, since the real-time adjustment of operational variables is very necessary for
ensuring the satisfactory performance of the TBM operation in response to changing
geological conditions. The magnitude of the adjustment of operational variables
should be greater when the TBM performance is observed with a larger reduction.
In general, the observed evidence of operational variables plays a significant role
in the time required when the TBM performance returns to a normal state. The
TBM performance varies greatly, depending on both the initial value of the TBM
performance (C T ) and the adjusted value of operational variables. More specifically,
the value of one operational variable should be increased to be a certain number
when the TBM performance suffers a negative reduction, and the value should be
decreased to be a certain number when the TBM performance suffers a positive
reduction, respectively. At the same time, the capacity of the hybrid RCA can be used
to determine the exact value that the one operational variable should be adjusted to
be, in order to return the TBM performance from a reduction to a normal state in a
very short time.
6 Merits of FCM
This chapter presents an FCM-enabled RCA approach to assessing the TBM perfor-
mance in tunnel construction. The developed approach is capable of performing
predictive, diagnostic, and hybrid RCA. A tunnel case in the Wuhan metro system is
used to demonstrate the effectiveness of the developed approach. More discussions
are conducted in terms of the FCM interpretability, validation, and capacities, as
presented as follows.
6.1 FCM Interpretability
Modeling is the process of building a useful representation of a system. Models have

a purpose, for example, to precisely simulate or predict the behavior of a system.
Interpretability (comprehensibility or understandability) is often recalled, especially
in contexts such as knowledge discovery from data. Most of the times, there is a clear
tradeoff between model complexity and model interpretability. As noted by Taber
[23], the greatest strength of the FCM approach is perhaps ease of use. FCM allows
multiple experts to collaboratively contribute to the construction of an FCM, which
enhances the confidence in its model development and further applications.
The FCM model considers expert feedback on the interrelationships between
concepts that impact the TBM performance. Through fuzzy weights in the FCM
framework, causal interconnections between concepts can provide an intuitive expla-
nation to end-users. In addition, one thing that needs to be pointed out is the measure-
ment of concepts in the developed model. Since the scope of favorable or unfavor-
able geological/operational variables varies greatly among different tunnel cases in
different areas, a challenge exists in determining the value of a specific concept
representing a corresponding geological or operational variable due to uncertainty.
To solve this problem, the values of all concepts have been initialized into a range
[−1, +1] using the 5-point linguistic scale in this chapter, where various kinds of
expert knowledge can be integrated in a unified platform.
For a particular project in a geological area, the magnitude of a specific concept
measurement (i.e., −1.0, −0.5, 0, +0.5, and +1.0) within the range [−1, +1] can
be determined according to the historical data in certain geological conditions, and
the values of concepts can then be calibrated by the Measurement-While-Drilling
(MWD) data in the real world. In this way, an immediate explanation can be obtained
regarding the measurement and adjustment of operational variables in real projects.
This is very convenient and simple for users who are not domain experts, but they
can accurately identify the root causes of an undesirable event or failure.
6.2 FCM Validation
In most cases, to compare model outputs with observations using independent data
which is not used in the model development can provide a solution to validate the
developed model. It would be convincing that the simulation results can be compared
to the historical data of actual TBM failures for validation purposes. However, the
historical data regarding actual TBM failures are quite limited during the construc-
tion of tunnel projects in the city of Wuhan in China. This is partly because some
maintenance and repair works have been conducted ahead of the occurrence of actual
failures due to the consideration of excessive costs. In regard to this issue, the domain
experts have knowledge for telling when TBM failures are likely to occur according
to their working experience, which provides the potential to validate the simulated
results to some extent.
To validate the effectiveness of the established simulation model (see Fig. 7), the
above ten experts who participated in the questionnaire investigation were invited
again to evaluate the effectiveness of model outcomes in various what-if experiments.
They all considered that the results of model prediction and diagnosis (see Fig. 11 and
Tables 2 and 3) were fairly consistent with the expectation in the actual construction
practice. To some extent, the established model for the TBM performance can be
validated in this way. In addition, in accordance with the predictive and diagnostic
6 Merits of FCM 63
analysis results using the established simulation model, some corresponding control
measures and reasonable adjustment of operational variables were carried out in time
to mitigate the risk during the tunneling excavation. Finally, the TBMs successfully
passed through the right track of DCS on February 2, 2015, and the left track on
March 2, 2015, respectively, without an accident or duration delay.
6.3 FCM Capacities
FCM is considered as a soft computing method within the domain of fuzzy modeling
and reasoning. As mentioned above in Sect. 2, there are many kinds of methods for
fuzzy modeling and reasoning, including the probability-based reasoning methods
(e.g., FTA), rule-based reasoning methods (e.g., RBR), and case-based reasoning
(CBR) methods. Comparing FCM with these traditional methods, what are the advan-
tages of the developed FCM approach? As shown in Sect. 5, the successful imple-
mentation of the developed FCM approach indicates that this approach is capable of
performing predictive analysis, diagnostic analysis, and hybrid analysis. In addition,
the developed FCM approach is verified to be capable of modeling the dynamics of
system behaviors over time (see details to Figs. 8, 9, 10 and 11). For instance, the
evolution of TBM performance over time in the hybrid RCA in the case of different
initial values of the concept C 8 in different conditions is presented in Fig. 11.
From a perspective of the analysis capacity, a qualitative comparison among many
soft computing methods in fuzzy modeling and reasoning, including FCM, FTA,
RBR, and CBR, is presented in Table 4. It is shown clearly that all those four methods
can perform predictive analysis. However, the traditional soft computing methods,
including FTA, RBR, and CBR, have challenges in performing diagnostic analysis,
hybrid analysis, and modeling dynamics. Only FTA is able to conduct the diagnostic
analysis. In summary, the developed FCM approach provides a more competitive
solution in fuzzy modeling and reasoning than the other traditional methods.
Table 4 A qualitative comparison of the analysis capacity among several soft computing methods
Methods Analysis capacities
Predictive analysis Diagnostic analysis Hybrid analysis Modeling dynamics
√ √ √ √
FCM
√ √
FTA – –
√
RBR – – –
√
CBR – – –
√
Note “ ” means the method has the capacity, and “–” means that the method does not have the
capacity
7 Summaries
The operation of the TBM system is very complex since various kinds of variables are
involved. To ensure the satisfactory performance of the TBM operation becomes a
challenging task. In this chapter, an FCM-based model consisting of nine concepts is
developed to analyze the TBM performance in tunnel construction, with construction
experience and knowledge from domain experts taken into account. The primary traits
of the holistic FCM method and the findings from the case study are summarized as
follows:
(1) From a predictive RCA perspective, a deep understanding of how the TBM
performance evolves quantitatively over time can be provided in case either
the geological or operational variables suffer a change. Through the case study,
we find that C 4 (Soil Density) displays the strongest negative correlation with
the concept C T (TBM Advance Rate); while C 8 (Grouting Speed) displays the
strongest positive correlation with C T . In other words, both C 4 and C 8 are
very sensitive to ensuring the satisfactory performance of the TBM operation,
which should be given significant attention in tunneling excavation.
(2) From a diagnostic perspective, more insights on what kind of variables cause
the reduction of the TBM performance can be detected. Through the case study,
we find that the TBM performance is very sensitive to the change of operational
conditions. The values of operational parameters, C 5 (Gross Thrust), C 6 (Cuter
Torque), C 7 (Earth Pressure), and C 8 (Grouting Speed) can be adjusted to go
up in case the TBM performance negatively reduces and go down in case the
TBM performance positively reduces.
(3) From a hybrid RCA perspective, we can identify the magnitude of the adjust-
ment scope of operational variables when the TBM operational performance
suffers a reduction. For instance, we can adjust the value of C 8 to be 0.07, in
order to return the TBM performance from a negatively very unfavorable level
(C T = −1.0) to a favorable level (C T = 0) in a short ten iterations. It would
be reasonable and sensible that we can adjust the value of C 8 to be 0.05, in
order to shortly return the TBM performance from a negatively unfavorable
level (C T = −0.5) to a favorable level (C T = 0).
The novelty of the proposed holistic FCM approach lies in that it is capable
of modeling dynamics of system behaviors over time and performing many kinds
of what-if scenario analysis, including predictive, diagnostic, and hybrid RCA. It
turns out to be a more competitive solution that deals with uncertainty, dynamics,
and interactions in the approximate reasoning process, compared to other traditional
methods (i.e., FTA, RBR, and CBR). The proposed approach can be used as a decision
support tool for ensuring the satisfactory performance of TBMs, and thus, it can help
increase the TBM productivity in a complex project environment.
References 65
References
1. Aamodt A, Plaza E (1994) Case-based reasoning: foundational issues, methodological

variations, and system approaches. AI Commun 7:39–59
3. Crestani F, Pasi G (2013) Soft computing in information retrieval: techniques and applications.
Physica
4. Dawes J (2008) Do data characteristics change according to the number of scale points used?
An experiment using 5 point, 7 point and 10 point scales. Int J Mark Res 50:1–19
5. Deshpande A (2011) Fuzzy fault tree analysis: revisited. Int J Syst Assur Eng Manag 2:3–13
6. Donauer M, Peças P, Azevedo A (2015) Identifying nonconformity root causes using applied
knowledge discovery. Robot Comput-Integr Manuf 33–44
7. Gao M, Ma H, Zhao J (2010) Influence of cobblestone geometrical property on equivalent
elastic modulus of cobblestone-soil matrix. In: Analysis of discontinuous deformation: new
developments and applications. Research Publishing Services, Singapore, pp 531–537
8. Klados G (1998) Experiences with hard rock shielded TBMs in special conditions. In: 5th
international symposium on tunnel, construction, pp 83–92
9. Kum S, Sahin B (2015) A root cause analysis for Arctic Marine accidents from 1993 to 2011.
Saf Sci 74:206–220
10. Kumar KA, Singh Y, Sanyal S (2009) Hybrid approach using case-based reasoning and rule-
based reasoning for domain independent clinical decision support in ICU. Expert Syst Appl
36:65–71
11. Liu H-C, Liu L, Liu N (2013) Risk evaluation approaches in failure mode and effects analysis:
a literature review. Expert Syst Appl 40:828–838
12. Maitra S, Banerjee D (2014) Application of fuzzy cognitive mapping for cognitive task analysis
in mechanised mines. IOSR J Mech Civ Eng (IOSR-JMCE) 11:20–28
13. Mentes A, Helvacioglu IH (2011) An application of fuzzy fault tree analysis for spread mooring
systems. Ocean Eng 38:285–294
14. Nelson PP (2016) TBM performance analysis with reference to rock properties. Compr Rock
Eng 4:261–291
15. Papageorgiou EI, Salmeron JL (2013) A review of fuzzy cognitive maps research during the
last decade. IEEE Trans Fuzzy Syst 21:66–79
16. Prentzas J, Hatzilygeroudis I (2007) Categorizing approaches combining rule-based and case-
based reasoning. Expert Syst 24:97–122
17. Sadeghi N, Fayek AR, Pedrycz W (2010) Fuzzy Monte Carlo simulation and risk assessment
in construction. Comput-Aided Civ Infrastruct Eng 25:238–252
18. Salmeron JL, Papageorgiou EI (2012) A fuzzy grey cognitive maps-based decision support
system for radiotherapy treatment planning. Knowl-Based Syst 30:151–160
19. Shetty N, Herbert MA, Shetty R, Shetty DS, Vijay G (2016) Soft computing techniques during
drilling of bi-directional carbon fiber reinforced composite. Appl Soft Comput 41:466–478
20. Stylios CD, Georgoulas G (2011) Modeling complex logistics systems using soft computing
methodology of fuzzy cognitive maps. In: 2011 IEEE conference on automation science and
engineering (CASE). IEEE, pp 72–77
21. Stylios CD, Groumpos PP (2000) Fuzzy cognitive maps in modeling supervisory control
systems. J Intell Fuzzy Syst 8:83–98
22. Stylios CD, Groumpos PP (2004) Modeling complex systems using fuzzy cognitive maps.
IEEE Trans Syst Man Cybern A Syst Hum 34:155–162
23. Taber R (1991) Knowledge processing with fuzzy cognitive maps. Expert Syst Appl 2:83–87
24. Wee YY, Cheah WP, Tan SC, Wee K (2015) A method for root cause analysis with a Bayesian
belief network and fuzzy cognitive map. Expert Syst Appl 42:468–487
25. Zhang L, Chettupuzha AJA, Chen H, Wu X, AbouRizk SM (2017) Fuzzy cognitive maps
enabled root cause analysis in complex projects. Appl Soft Comput 57:235–249
26. Zhang L, Wu X, Skibniewski MJ (2015) Simulation-based analysis of tunnel boring machine
performance in tunneling excavation. J Comput Civ Eng 30:04015073
27. Zhao J, Gong Q, Eisensten Z (2007) Tunnelling through a frequently changing and mixed
ground: a case history in Singapore. Tunn Undergr Space Technol 22:388–400
Time Series Prediction
1 Introduction
Over the past decades, the rapid development of big cities is raising the demands of
underground space utilization. One of the favorable options for urban development
is to build underground tunnels. Notably, a lot of tunnels are located at a low depth
in soil or soft rock zones under densely populated areas, and thus the excavation
works of shallow tunnels in the soft ground tend to result in both lateral and vertical
surfaces. There are two major characteristics of those shallow tunneling projects
including the lower depth and soft ground, which occasionally cause considerable
surface settlements. It should be highlighted the tunnel-induced ground settlement
poses a big threat to both surface and sub-surface facilities, which could act as a
critical factor in evaluating tunnel-induced damages. Therefore, a necessary step to
avoid excessive settlement is based upon the estimation, analysis, and controlling of
the tunnel-induced ground settlement accurately and timely, which has proven to be
a key procedure to ensure construction safety in tunnels [29].
Methods for estimating tunnel-induced ground settlement can be classified into
four main groups, including empirical [16], analytical [2], numerical methods [10],
and intelligent methods [20]. They have their respective merits in predicting tunnel-
induced ground settlement. However, some weaknesses still exist to cause the appli-
cation restrictions. For instance, Loganathan [13] claimed that the empirical methods
failed to give highly accurate results due to the limited adaptability in different ground
conditions. Chou and Bobet [4] indicated that analytical methods were prone to
underestimate the maximum soil deformations or overestimate the settlement trough,
since it took no account of time-dependent consolidation and creep ground loss
components. Łodygowski and Sumelka [12] mentioned that the numerical methods
suffered from low accuracy and validity, since simulation models could be sensitive
to boundary conditions (i.e., stress concentrations and local stresses) and difficult to
be validated. At the same time, most of the existing models are mainly concerned
68 Time Series Prediction
with the prediction of the maximum ground settlement without regard to the evolu-
tion of the ground settlement over time. The static results are incapable of estimating
tunnel-induced settlement in a dynamic and real-time manner [5].
At present, an increase in the number of urban tunneling projects can bring about
a large amount of project-related data, which is helpful in developing a more accurate
approach with improved prediction performance in the tunnel-induced daily ground
settlement. In this regard, a novel hybrid approach is proposed under the integration
of Packet Transform (WPT) and Least Square Support Vector Machine (LSSVM),
which combines the complementary advantages of both WPT and LSSVM. That
is to say, WPT with the process of time-frequency signal analysis is carried out to
handle the monitored tunnel-induced settlement sequences, while LSSVM can make
a reliable prediction by learning the high- and low-frequency information. A case
study is provided to estimate, test, and analyze the ground-induced settlement of a
real tunnel project in China in detail, aiming to demonstrate the feasibility of the
proposed approach and its potential in future implementation.
2 Estimation of Tunnel-Induced Ground Settlement
Excavation work of shallow tunnels in the soft ground is bound to generate soil move-
ments, and thus unwanted settlement will occur through developing above the tunnel
on the ground surface. A key problem is to protect existing structures and under-
ground utilities against damages by excessive ground movement. The understanding
of the influence degree and scope of tunneling excavation is particularly an impor-
tant task to address the problem. Typically, the following four methods, including
empirical, analytical, numerical, and intelligent methods, have been developed to
accurately estimate the tunnel-induced ground settlement.
2.1 Empirical Methods
The well-established empirical methods available to date are used primarily to esti-
mate ground settlements in soft soils. Martos [14] firstly proposed that a Gaussian
or normal distribution curve could well describe the shape of the daily settlement
trough. Later, Peck [16] analyzed the settlement data collected from several tunnels
and mines by fitting Gaussian curves, which could provide suggestions about the
equation of settlement, as shown in Eq. (1).
2 Estimation of Tunnel-Induced Ground Settlement 69
⎧
⎪
⎪ x2
⎪
⎪ Sv (x) = Sv,max exp − 2
⎪
⎪ 2i x
⎪
⎪
⎪
⎪
⎨ Sv,max = √ VS
2π i x (1)
⎪
⎪
⎪
⎪ VS = Vl π R 2
⎪
⎪
⎪
⎪
⎪
⎪ Z
⎩ ix = √
2π tg(45◦ − ϕ/2)
where S v (x) is the ground settlement at a transverse distance x from the tunnel center-
line, S v,max is the maximum transverse settlement at x = 0, x is the horizontal distance
between the location and the tunnel centerline in the transverse direction, and ix is
the horizontal distance from the tunnel centerline to the inflection point of the curve.
V S is the stratum loss caused by tunneling excavation, V l is the stratum loss ratio,
and ϕ is the friction angle. Based on the hypothesis of Peck, many scholars offered
further suggestions about the function of ix [15].
However, these empirical methods fail to generate highly accurate results due to
two important limitations as follows. One is that their applicability may be limited
due to different ground conditions and construction techniques, while the other is
that horizontal movements and subsurface settlements cannot be predicted using the
established empirical relationships [13]. Besides, Ding et al. [5] indicated that the
field has a limited number of settlement markers available, which could be another
reason for the unsatisfying estimation accuracy. Celestino et al. [3] claimed that the
real ground settlement was sometimes significantly greater than that obtained from
the approximately fitted curves.
2.2 Analytical Methods
Multiple attempts have been devoted to developing analytical solutions for estimating
tunneling-induced ground settlements. For example, Sagaseta [17] presented closed-
form solutions to obtain the strain field in an initially isotropic and homogeneous
incompressible soil due to near-surface ground loss from tunneling. Verruijt and
Booker [25] proposed an analytical solution for tunnels in homogeneous elastic
half-spaces using an approximate method suggested by Sagaseta [17] for the case
of ground loss. This solution is applicable for the incompressible case and arbitrary
values of Poisson’s ratio. The effect of tunnel lining deformation in the long term is
taken into account [13]. Equation (2) expresses the Verruijt and Booker’s closed-form
solution for estimating settlements.

z1 z2 z 1 (kx 2 − z 22 ) z 2 (kx 2 − z 22 )
Uz = −ε R 2 2
+ 2 + δ R2 +
r1 r2 r14 r24

2ε R 2 (m + 1)z 2 mz(x 2 − z 22 )
+ +
m r22 r24
2
x − z 22 m 2zz 2 (3x 2 − z 22 )
− 2δ R h
2
+ (2)
r24 m+1 r26
where U z is the estimation of the vertical settlement, ε is the uniform radial ground
loss, δ is the long term ground deformation due to the ovalization of the tunnel lining,
R is the tunnel radius, H is the tunnel depth, and ν is the Poisson’s ratio of soil. z1 =
z − H, z2 = x + H, r 21 = x 2 + z21 , r 22 = x 2 + z22 , m = 1/(1 − 2 ν), and k = ν/(1 − ν).
In most cases, analytical solutions cannot consider the excavation process with
sufficient approximation. Many simplifications, such as a plain strain assumption,
elastic behavior, and/or soil isotropy, are adopted in those solutions. Nevertheless,
those solutions take no account of time-dependent consolidation and creep ground
loss components, it is likely to underestimate the maximum soil deformations or
overestimate the settlement trough [4]. Therefore, those methods were only suitable
for short-term undrained conditions, which could be restricted to cases with small
ground deformations, such as stiff clays and rocks.
2.3 Numerical Methods
The finite element method is effective in overcoming some limitations in empirical

and analytical methods, which indeed has been widely used in tunneling analysis. Due
to the three-dimensional nature of the excavation-structure interaction and non-linear
behavior of geomaterials, the tunnel-soil-ground interaction has been regarded as a
highly complicated process. To handle the complex problem, numerical simulation
methods have been adopted for construction safety analysis and control owning the
ability to consider all relevant factors, like ground heterogeneity, nonlinear behavior
of soils, tunnel-soil interaction, and construction methods. For instance, Liu et al.
[11] performed a parametric sensitivity analysis in FLAC3D to explore the interaction
between a surface building and the enlargement of a metro station. Azadi et al.
[1] conducted a numerical simulation model to study the settlement of a structure
adjacent to a tunnel, and then the obtained results are used in a neural network
for assessment of the tunnel-induced settlement effect. Zhang et al. [30] analyzed
the magnitude of tunnel-induced damage to adjacent historical buildings using the
simulation technique, aiming to investigate the impacts of tunneling excavation on
distributions of equivalent and principal stresses of the building in detail.
The numerical methods, however, are criticized for their accuracy and validity
[12]. Karakus and Fowell [9] indicated that it could be a hard task to set up a realistic
model that would be able to calculate tunneling-induced settlement profiles. Besides,
due to the little information about the soil properties available, simulation of an
excavation process is relatively difficult. Another issue to note is that results from
numerical models tend to be different from field measurements and case histories. In
particular, Karakus [8] presented that the settlement trough produced by numerical
analysis was always much shallower and wider than the measured data.
2.4 Intelligent Methods
Along with the rapid development of artificial intelligence (AI) technologies, intel-
ligent methods have been gaining increasing attention around the world due to their
powerful capacities and comprehensive adaptabilities. Those intelligent methods,
like artificial neural networks (ANNs) and support vector machines (SVMs), can be
helpful in solving complex real-world problems. Especially for ANN, it acts as a
typical class of intelligent methods to model nonlinear relationships between input
and output variables. Therefore, it is reasonable to apply this method to the analysis
of the ground settlement, which has become a popular topic.
In an existing work [20], the neural network has been constructed based on a
series of inputs, including tunnel depth, tunnel diameter, soil cohesion, volume loss,
and groundwater level, to effectively predict the output of the maximum surface
settlement. However, a great challenge of using ANN is to collect multiple variables
relevant to the tunnel-induced ground settlement [23]. Besides, how to properly
determine an optimal combination of controlling parameters in ANN models, such
as the number of hidden layers/nodes and learning rates, remains another question.
2.5 Comparison
Moving the existing research forward, the importance of accurately assessing tunnel-
induced deformations has been emphasized in recent years to tackle an increasing
number of urban tunnel projects. However, most of the existing methods for deter-
mining ground movements could be somewhat inadequate, over-simplistic, or overly
complicated [13]. It is increasingly necessary to develop a more accurate approach
to better predict tunnel-induced ground settlement. Although many existing models
have focused on the maximum ground settlement, they pay less attention to dynamic
and real-time predictions [26]. This kind of static results is insufficient to meet the
dynamic and real-time requirements regarding updates of the ground settlement over
time [5]. For a clear comparison, Table 1 lists the advantages and disadvantages of
different methods in estimating tunnel-induced ground settlements.
In pursuit of a reasonable solution, some assumptions have been made in tradi-
tional empirical, analytical, and numerical methods. Nevertheless, large amounts
of randomness and uncertainties are bound to exist in the assumptions and actual
construction practice due to the great level of dynamic and complexity in underground
conditions, which possibly lead to remarkable deviations between the predicted and
observed effects. To resolve the above-mentioned inefficiencies, the field monitoring
technique can be a useful solution, which implements the scientific method to check
Table 1 Advantages and disadvantages of methods for estimating tunnel-induced ground

settlements
Category Advantages Disadvantages References
Empirical methods (1) Simple empirical (1) Results are not highly [3, 5, 14, 16]
formulas that are easy to accurate
use (2) Limited applicability to
(2) Provide initial idea and different ground
estimation of settlement conditions and
(3) Wide applications in construction techniques
actual construction (3) Limited empirical
practices relationships to predict
horizontal movements
and subsurface
settlements
(4) Consider few variables
Analytical methods (1) Relatively simple (1) Many simplifications, [4, 13, 17, 25]
formulas that are easy to such as elastic behavior,
use and/or soil isotropy, are
(2) Cover both horizontal adopted in those
and vertical solutions
displacement (2) Only applicable to
(3) Relationship between short-term undrained
affecting variables can conditions
be understand
Numerical methods (1) Take tunnel-soil (1) Building numerical [8–10, 12]
interaction into account models are
(2) Capability of simulating time-consuming to
tunneling process create and verify
(3) Adaptation to various (2) Be criticized by the
ground conditions and accuracy and validity of
construction techniques the results
(3) Models are sensitive to
boundary conditions
(stress concentrations,
local stresses)
Intelligent methods (1) No need for prior (1) Require a large set of [19, 20, 23]
domain knowledge training data
(2) Consider multiple (2) Computation algorithms
influential variables are complex
(3) Adaptation to various (3) Collection of influential
ground conditions and variables and their
construction techniques values is less rigorous
interventions in the real world. As a result, outcomes can be observed in a natural

setting rather than in a simulation or assumed environment, meaning that the field
monitoring technique can bring a higher external validity. Meanwhile, the field moni-
toring data, as the outcome of the comprehensive effect by all the influential variables,
such as site condition (e.g., soil types) and load condition, can offer a deeper insight
into the future settlement in a direct way.
The intelligent methods have demonstrated their promising performance largely

due to their merits of simplicity and high accuracy. To date, studies on AI technolo-
gies are progressing continuously and rapidly in both theory and application. This
chapter intends to develop a novel hybrid approach under the integration of the field
monitoring technique named WPT and the AI algorithm termed LSSVM, and then a
time-series analysis and estimation are performed to enhance the accuracy and inter-
pretation regarding updates of the tunneling-induced ground settlement over time.
This approach relied on past naturally observed settlement data to predict the future
tunnel-induced settlement, which can be suitable for various types of geological
conditions.
3 Time Series Prediction Methods
During the process of tunnel excavation, the ground settlement for a specific cross-
section is measured continuously, and thus, the data points collected at regular inter-
vals constitute a time series sequence to track the excavation process over time.
Notably, there are some potential internal structures, such as autocorrelation, trend,
cyclicity, stationarity, and seasonality, embedded in a time series. The implementa-
tion of a time series analysis could act as an essential task to discover meaningful
statistics and characteristics from such serially correlated data with timestamps, By
making full use of changes that occurred in the past, it is possible to predict future
conditions reliably and inform tactical decisions at an early stage.
As computer and machine learning technologies develop, a lot of prediction
algorithms have been proposed over the past decade. As reviewed, the forecasting
approaches can be grouped into two categories, namely the linear methods and non-
linear methods. Early research work pays more attention to the smoothing technique,
error component model, and auto-regressive integrated moving average (ARIMA)
based on the assumption of linearity and stationarity to infer future trends. Due
to the simplicity of these models, they are more suitable for exploring short-term
predictions. In particular, ARIMA is one of the most widely-used models in time
series forecasting, which adopts a minimal number of parameters to ensure relatively
high prediction accuracy. However, some previous research claims that ARIMA is
improper to interpret and model nonlinear relations of many real-world problems,
leading to the increased errors in prediction problem about the nonlinear ground
settlement.
Compared with linear methods, the non-linear prediction methods specialize in
describing non-linear characteristics, aiming to achieve more accurate forecasting
performance. Algorithms belonging to the artificial intelligence field, like artificial
neural networks (ANN) and support vector machine (SVM), have received increasing
attention due to their powerful capacities and comprehensive adaptabilities. It has
been seen that ANN is adopted as an alternative approach for analyzing the ground
settlement, which has become increasingly popular [20, 23]. However, a difficult task
should be highlighted in the determination of network architecture, since the corre-

sponding direction or analytical method has not been available yet. Besides, ANNs
are likely to suffer another problem rooting in their relatively poor generalization
performance [26].
In regard to another intelligent method named support vector machine (SVM), it
is mainly responsible for classifying data and estimating function, which has been
successfully applied in engineering practices regarding tunnel construction, such as
geo-mechanical back analysis, displacement prediction, and other highly nonlinear
geotechnical areas. To be more specific, SVMs map input data into feature space
in a high dimension where it could be linearly separable. Compared with ANNs,
SVMs possess a better generalization ability, since they replace the empirical risk
minimization (ERM) with structural risk minimization (SRM).
To be more specific, the goal of ERM is to minimize the error on both the training
and testing data, and thus, ANNs rely greatly on the size of training data to achieve
a good estimation performance. On the contrary, SVMs can demonstrate great esti-
mation capacity in handling the problem with a small sample size, which can be less
vulnerable to the over-fitting problem [18]. Based on that, Suykens and Vandewalle
[24] further developed a modification of SVM termed least squares SVM (LSSVM),
which could create a set of linear equations instead of a quadratic programming
problem.
Some signal processing algorithms, like Fourier-based and wavelet-based trans-
form methods, have been proposed over the last decade for performing an interesting
task of monitoring and measurement. It is noted that traditional Fourier methods
are unsuitable to detect and classify signals with non-stationary characters, while
wavelets are able to approximate time-varying and non-stationary signals in a better
way than the Fourier methods. Moreover, a modification of Fourier transform named
the windowed Fourier transform (WFT) has been proposed to obtain a transform
with better localization properties. However, WFT is still criticized for its limitation
that WFT uses a fixed window size [7]. As for Wavelet transform (WT), it allows an
arbitrary function to be expressed as a series expansion, where each term is one of the
basic wavelets multiplied by its magnitude. One of its advantages over Fourier trans-
form is the perfect analysis of the resulting decomposed components with well-scaled
resolution.
Due to the time-frequency and multi-resolution property, WT has recently served
as a promising approach for the purposes of signal processing, fault diagnosis, risk
identification, and others. As stated by Sun and Chang [22], one drawback of WT
is the poor frequency resolution in the high-frequency region, which could raise
difficulties in discriminating signals containing close high-frequency components.
To relieve the issue, Wavelet packet transform (WPT), as an extension of WT, can
provide complete level-by-level decomposition, enabling the extraction of features
from signals that combine stationary and non-stationary characteristics with arbitrary
time-frequency resolution.
4 A Hybrid Time Series Prediction Approach 75
4 A Hybrid Time Series Prediction Approach
To ensure construction safety in tunneling projects, issues remaining in moni-

toring tunneling safety need to be solved. In this regard, China has released an
industry standard “Technical code for monitoring measurement of subway engi-
neering (DB11/490-2007)” in 2007. For studying the ground settlement induced
by tunneling excavation, it is necessary to monitor several items during the tunnel
construction and set up the monitoring section every 5–30 m. Herein, a hybrid time-
series analysis method that integrates WPT and LSSVM is introduced with step-
by-step procedures, aiming to improve the accuracy of the real-time prediction of
the tunnel-induced ground settlement. The workflow of the developed WPT-LSSVM
prediction approach has been provided in Fig. 1, incorporating the following four
stages.
Original
settlement data
Decomposing original data using WPT

Performance analysis
j
fj1 (t) fji (t) fj2 (t)
Feedback
LSSVM LSSVM LSSVM

j
fj1 (t)' fji (t)' fj2 (t)'
Reconstructing different time series using WPT
Predicted
settlement data
Fig. 1 Workflow of the developed WPT-LSSVM prediction approach

4.1 Decomposing Original Data Using WPT
As a generalization of wavelet decomposition, the wavelet packet method offers a

richer range of possibilities for signal analysis, allowing for a best-matched anal-
ysis of signals. In particular, it utilizes a recursion of filter-decimation operations to
provide a level-by-level transformation of a signal from the time domain into the
frequency domain. There are a set of linearly combined usual wavelet functions in
wavelet packets, which can bring in some promising properties, like orthonormality
and time-frequency localization from their corresponding wavelet functions. Equa-
tion [3] presents a function ψ ij,k (t) of a wavelet packet with three parameters i, j, and
k.
⎧ i
⎪ ψ j,k (t) = 2 j/2 ψ i (2 j t − k), i = 1, 2, . . . , 2 j
⎪
⎪
⎪
⎪ ∞
⎪ ψ 2i (t) = √2 h(k)ψ i (2t − k)
⎪
⎨
(3)
⎪
⎪
k=−∞
⎪
⎪ √ ∞

⎪
⎪
⎪
⎩ ψ 2i+1
(t) = 2 g(k)ψ i (2t − k)
k=−∞
where i, j, and k are the modulation, scale, and translation parameters of the wavelet
function ψ ij,k (t), respectively. The first wavelet is the so-called the mother wavelet
function, which is defined as ψ 1 (t) = ψ (t). h(k) and g(k) are the discrete filters
that are quadrature mirror filters associated with the scaling function and the mother
wavelet function.
The great challenge in wavelet analysis lies in the selection of the proper wavelet
function as well as the decomposition level of the signal. Several factors should be
taken into account, including (i) orthogonal or nonorthogonal, (ii) complex or real,
(iii) width, and (iv) shape. Ding et al. [5] presented a selection criterion to choose
the optimal wavelet function in the analysis of tunnel-induced ground settlement.
They developed a few wavelet functions under different decomposition levels and
conducted a unit root and normality tests to compare the performance of each alterna-
tive. Finally, it is concluded that decomposing with DB16 performed best. Therefore,
it is reasonable to employ DB16 in the case of this chapter.
In the wavelet framework, the decomposition process is executed by a recursive
filter-decimation operation. Figure 2 illustrates the full WT and WPT of a time-
domain signal f (t) up to the 3rd level of the decomposition. It can be observed that
the WT contains one high-frequency term from each level and one low-frequency
residual from the last level of the decomposition, while the WPT is composed of
complete decomposition at every level to make it achieve a higher resolution in the
high-frequency region. As a more detailed description of the WPT decomposition,
the recursive relations between the jth and the j + 1th level components are expressed
in Eqs. (4) and (5), respectively.
f31 (t)
1
f2 (t)
f11 (t) f32 (t)
2
f (t) f2 (t)
2
f1 (t)
(a)
H
H f31 (t)
1
f2 (t)
H f32 (t)
1
f1 (t) G
H
f33 (t)
f22 (t)
G f34 (t)
G
f (t) H
H f35 (t)
3
f2 (t)
f36 (t)
2 G
f1 (t) H
G f37 (t)
4
f2 (t)
G f38 (t)
G
(b)
Fig. 2 Three-level decomposition trees of a time-domain signal: a WT; b WPT
2i−1
f j+1 (t) = H f ji (t)
(4)
2i
f j+1 (t) = G f ji (t)
⎧ ∞
⎪
⎪
⎪
⎪ H {.} = h(k − 2t)
⎪
⎨
k=−∞
(5)
⎪
⎪
∞

⎪
⎪ = g(k − 2t)
⎪
⎩
G{.}
k=−∞
where H and G are filtering-decimation operators and are related to the discrete
filters h(k) and g(k), respectively. h(k) and g(k) are the quadrature mirror filter (QMF)
associated with the predefined scaling function and the mother wavelet function. See
more details to Gao and Yan [6].
From Fig. 2, it is clear that the calculation in WPT contains more steps than WT.
Although WPT could probably suffer from a higher computational complexity, it is
not an issue for the current computing system. Especially in the construction industry,
the amount of the monitored data is relatively limited in real construction practices,
and engineers have more concerns about the accuracy of prediction results, instead
of the calculation complexity. Each component in the WPT tree can be viewed as
the output of a filter that is tuned to a particular basis function. For this reason, the
whole tree can be taken as a filter bank. It can be seen from Fig. 2 that a three-level
resolution of the original signal will produce eight time series. After j levels of the
WPT decomposition, the original signal f (t) can be presented by a set of time series
f ij (t), (i = 1, 2, …, 2j ). The wavelet packet component signal f ij (t) can further be
expressed by a linear combination of wavelet packet functions ψ ij,k (t) in Eq. (6).
⎧ ∞
⎪
⎪
⎪ f i
(t) = cij,k ψ ij,k (t)
⎨ j
k=−∞ (6)
⎪
∞
⎪
⎪
⎩ cij,k = f (t)ψ ij,k (t)dt
−∞
where cij,k is the wavelet packet coefficient of the ith wavelet packet component
signal f ij (t), (i = 1, 2, …, 2j ). The wavelet packet functions are orthogonal, which
satisfies ψ mj,k (t)ψ nj,k (t) = 0 if m = n. See more details to Wu and Liu [27].
4.2 Predicting Separate Time Series Using LSSVM
In the last stage, the original data is decomposed into several time series, and then
prediction models can be built to estimate the target time series within high- and low-
frequency regions. For this purpose, the support vector machine (SVM), a popular
machine learning tool relying on statistical learning theory, is adopted to solve multi-
dimensional functions. The least-squares version of SVMs is called LSSVM, which
is a set of supervised learning methods to analyze data and detect patterns for classi-
fication and regression analysis. Let a given training set of LSSVM be formulated as
N data points {xk , yk }k=1
N
with input data x k ∈ RN and output yk ∈ r, where RN is the
N-dimensional vector space and r is the one-dimensional vector space. As given in
Eq. (7), the LSSVM approach formulates the regression as an optimization problem
in the primal weight space.
⎧
n
⎨ Minimi ze : 1 ωT ω + γ
⎪
e2
2 2 i=1 i (7)
⎪
⎩
Subject to : y(x) = ωT ϕ(xk ) + b + ek , k = 1, 2, . . . , N
where ek is the error variable, ω is the weight vector, and γ is the regularization
parameter.
In general, it is not a straightforward task to directly handle the weight vector ω
due to its high dimensionality. In order to simplify the computation, a dual formation
is constructed and a kernel function is adopted. Then, the solution of the regression
problem can be obtained according to Eq. (8).

N
y(x) = ak K (x, xk ) + b (8)
k=1
where K is the kernel function satisfying Mercer’s condition, x k is the training data,
and x represents the new input data. See more details to Xu [28].
Different from Lagrange multipliers, the kernel and its specific parameters
together with regularization parameter γ in LSSVM cannot be set from the opti-
mization problem. To determine and fine-tune these model parameters, some candi-
date methods, like Vapnik–Chervonenkis bounds, cross-validation, independent opti-
mization set, or Bayesian learning, can be carried out. In this chapter, the widely-used
radial basis function (RBF) is chosen as a kernel function, which has been formulated
in Eq. (9). To search optimized values of the two parameters including the regular-
ization parameter γ and the RBF kernel function parameter σ, genetic algorithm
(GA) is utilized as a search heuristic that mimics the process of natural selection in
the field of AI.

(x − xk )(x − xk )T
K (x, xk ) = exp − , k = 1, 2, . . . , N (9)
2σ 2
where σ is the width of RBF as a kernel function, and K(x, x k ) is the kernel function.
As shown in Fig. 1, the prediction task is performed separately in regard to each
specific decomposed sequence. Given the length of the decomposed sequence in the
last stage is T, the ith decomposed sequence can be defined as a vector {f ij (1), f ij
(2), …, f ij (T )} (i = 1, 2, …, 2j ). For prediction and evaluation purposes, the whole
sequence can be divided into two clusters, where one is for training and the other is
for testing. To deal with the limited observed data in the construction industry, the
rolling forecast technique is applied for the time series analysis in this chapter.
In each sample, the length of the input and output is defined to be N and 1, respec-
tively, which means that the output is decided based on the past N observations in a
time-series perspective. In this case of estimating tunnel-induced ground settlement,
the value of N is defined as 5. In other words, it is assumed that the current ground
settlement could be closely associated with the observed settlement information from
the past five days (Note: the tunnel-induced ground settlement data is monitored and
recorded on a daily basis). In other words, when the size of the training set is L, the
size of the testing set will be T-L-N. The samples in the training set and the testing
set have been given in Tables 2 and 3, respectively. The optimized values of γ and σ
are set by the GA-based search method in the training set. Subsequently, predictions
defined by f ij * (N + l), (l = 1,2, …, T − N), can be obtained, as shown in Tables 1
and 2.
Table 2 Samples of the training set and predictions in LSSVM

No. Input Output Prediction

1 f ji (1), f ji (2), . . . , f ji (N ) f ji (N + 1) f ji∗ (N + 1)

2 f ji (2), f ji (3), . . . , f ji (N + 1) f ji (N + 2) f ji∗ (N + 2)
… … … …

L f ji (L), f ji (L + 1), . . . , f ji (L + N − 1) f ji (L + N ) f ji∗ (L + N )
Table 3 Samples of the testing set and predictions in LSSVM

No. Input Output Prediction

1 f j (L + 1), f j (L + 2), . . . , f j (L + N )
i i i f j (L + N + 1)
i f ji∗ (L + N + 1)

2 f ji (L + 2), f ji (L + 3), . . . , f ji (L + N + 1) f ji (L + N + 2) f ji∗ (L + N + 2)
… … … …

T-L-N f ji (T − N ), f ji (T − N + 1), . . . , f ji (T − 1) f ji (T ) f ji∗ (T )
4.3 Reconstructing Different Time Series
This stage intends to obtain the predicted tunnel-induced ground settlement over time
through reconstructing the predicted sequences. In both the training set and testing
set, the predicted tunnel-induced ground settlement at the ith decomposed sequence
can be gained. Based on Eq. (10), all the decomposed sequences can be reconstructed
by following a level-by-level way, which is similar to the decomposition process in
the first stage [27]. Figure 3 shows an example of a three-level WPT reconstruction
for a predicted time-domain signal.
⎧ i∗
⎪
⎪ f (t) = f j+1
2i−1∗
(t) + f j+1
2i∗
(t)
⎨ j
, t = 1, 2, . . . , T
j
2 (10)
⎪ ∗
⎩ f (t) =
⎪ f ji∗ (t)
i=1
where f ji∗ (t) is the predicted ith decomposed sequence at the jth level, and f ∗ (t) is
the predicted sequence, with all the decomposed sequence reconstructed.
f31* (t)
f21* (t)
2*
f3 (t)
f11* (t)
3*
f3 (t)
f22* (t)
4*
f3 (t)
f * (t)
5*
f3 (t)
f23* (t)
6*
f3 (t)
f12* (t)
f37* (t)
f24* (t)
8*
f3 (t)
Fig. 3 Three-level WPT reconstruction of a predicted time-domain signal
4.4 Prediction Performance Evaluation
When the predicted results are obtained, prediction performance needs to be numer-
ically evaluated to reveal the accuracy of the developed prediction model, which can
rely on several standard evaluation measures presented in the literature [21]. This
chapter refers to two indicators, namely Mean Absolute Error (MAE) and Root Mean
Square Error (RMSE), to illustrate the correspondence between individual pairs of
model predictions and observations. To be more specific, MAE describes the average
magnitude of the error between the predicted and observed distributions. When the
model is performing well, the value of MAE will arrive at zero. Its value larger than
zero indicates that the predicted value is more or less than the observed value. As
for RMSE, it measures the standard deviation between the predicted and observed
distributions. RMSE serves to aggregate the magnitudes of the errors in predictions
for various times into a single measure of predictive power. Similarly, RMSE is zero
when the model precisely predicts the observed value. For its value greater than zero,
it means that the model predictions have certain deviations from observations.
In general, MAE reflects the mean prediction accuracy, while RMSE quantifies
the dispersion degree of the prediction performance. According to MAE and RMSE
provided in Eqs. (11) and (12), it is expected to effectively assess the prediction about
the time series of the tunnel-induced settlement.
1
T
M AE = f (t) − f ∗ (t)∗ × 100%, t = 1, 2, . . . , T (11)
T t=1

1 T
RMSE = ( f (t) − f ∗ (t)∗ )2 , t = 1, 2, . . . , T (12)
T t=1
where f (t) is the observed sequence of the tunnel-induced settlement, and f ∗ (t) is
the predicted sequence, with all the decomposed sequences reconstructed.
Based on predicted results, decision analysis can be performed to make better
decisions in terms of various aspects, such as techniques for structuring a decision,
uncertainty measurement, alternative assessment, option evaluation for arriving at
the course of actions. That is to say, effective analysis and management in decisions
contribute to improving project safety and performance. In this case, the tunnel-
induced ground settlement is accurately predicted on a daily basis, and then relevant
control measures can be proposed and conducted in time to deal with the risk situation
that the settlement is going to exceed the warning standard. Besides, the safety
warning standard herein is set as 2 mm/d regarding the daily ground settlement
according to the industry standard, “Technical code for monitoring measurement of
subway engineering (DB11/490-2007)”. Meanwhile, the previous stages can receive
feedback constantly to guide the model modification and updating.
5 A Realistic Tunnel Case
5.1 Data Collection
For demonstrating the practicability of the developed prediction approach, a case

study is performed in a Hongkong-Huiji (H-H) tunnel section in the Wuhan metro
system in China. To be more specific, the Wuhan metro system is an elevated and
underground urban metro system in the city of Wuhan, Hubei, China. The H-H tunnel
section, a twin-line tunnel with a length of almost 750 m, is located at Metro Line 3
in the Wuhan metro system. In each line, one earth pressure balanced (EPB) tunnel
boring machines (TBMs) with a diameter of almost 6.2 m are used to excavate soils
from the Hongkong Road Station to the Huiji 2nd Road Station at an average depth
of 15 m below the ground surface.
A photograph of one adopted EPB TBM in the H-H tunnel construction is shown
in Fig. 4. At the beginning of 2015, the excavation of the H-H tunnel section starts.
Along the H-H tunnel section, two monitoring points, DK26460 and DK26580 are
installed to collect data about the tunnel-induced ground settlement on a daily basis.
Figure 5 displays the evolution of the measured daily ground settlements in the period
of 13th April–11th June in 2015 at DK26460 and DK26580 during the H-H tunnel
construction. For each original signal of the measurement at DK26460 or DK26580,
there are 60 data points (T = 60) in total during the construction period.
5 A Realistic Tunnel Case 83
Fig. 4 Photograph of one

EPB TBM adopted during
the H-H tunnel construction
Fig. 5 Evolution of the measured daily ground settlements from 13th April to 11th June in 2015
at DK26460 and DK26580 during the H-H tunnel construction
5.2 Training Details
In this case, the measured daily ground settlements at DK26460 and DK26580 from
13th April to 11th June (see Fig. 5) are taken as two original time-domain signals to
implement the detailed computation process in the developed approach. According
to the WPT process, the DB16 wavelet is used to decompose each original signal f
(t) (t = 1, 2, …, 60) into two levels with four separate sequences f 12 (t), f 22 (t), f 32
(t) and f 42 (t) (t = 1, 2, …, 60). As a result, the decomposition sequences of original
daily ground settlements at DK26460 and DK26580 have been displayed in Fig. 6.
Regarding each sequence f i2 (t) (i = 1, 2, 3, 4; t = 1, 2, …, 60), the length of the input
(that is N as mentioned in Table 2) and output of each sample is set to be 5 and 1,
respectively, indicating that each sample is made up of 6 data points. Besides, when
the size of the training set (that is L as mentioned in Table 2) is defined to be 35, the
size of the testing set can be calculated to be T-L-N = 60-35-5 = 20. From the rolling
forecast technique, the obtained data at the first 40 days (from 13th April to 22th
May) in each time sequence constitute 35 samples which serve as the training set,
while the obtained data at the last 25 days (from 18th May to 11th June) constitute
20 samples which are used as the testing set. Then, the sequence in both the training
set and the testing set can be predicted based on LSSVM. Figure 7 illustrates the four
separate sequences predicted at DK26460 and DK26580, respectively, which can be
reconstructed into a comprehensive sequence according to Eq. (10).
Subsequently, it is necessary to evaluate the prediction performance of the devel-
oped WPT-LSSVM approach by analyzing the correspondence between individual
pairs of model predictions and observations. The scatterplot in Fig. 8 shows the obser-
vations and WPT-LSSVM enabled predictions for the tunnel-induced daily ground
settlement at DK26460 and DK26580, respectively. It can be seen that the observed
daily ground settlement is in general low when the predicted value is low. Similarly,
the observed daily ground settlement is high when the predicted value is high at
both DK26460 and DK26580. To better visualize the average agreement between
actual observations and model predictions, regression lines are fitted in both training
data and testing data. Figure 8 provides the fitted regression lines along with rele-
vant parameters. Special focus needs to be paid on the testing data, where the fitted
regression lines at DK26460 and DK26580 have a gradient of 0.9507 and 0.9630,
respectively. Since the gradients are both very close to 1, it validates the effectiveness
of the WPT-LSSVM model in accurately predicting the average magnitude of the
daily ground settlement.
For a more comprehensive assessment, the hybrid WPT-LSSVM model is
compared to the traditional LSSVM. In the same way, the originally measured ground
settlement sequence is divided into a training set with a size of 35 and a testing set
with a size of 20. LSSVM is applied to directly predict the ground settlement on a
daily basis with no procedure of decomposition and reconstruction. Also, the scatter-
plot is drawn in Fig. 9 to display observations and LSSVM enabled predictions for the
tunnel-induced daily ground settlement at DK26460 and DK26580, respectively. In
terms of the fitting results in the testing data from Fig. 9, the fitted regression lines at
f21 (t)
f22 (t)
f23 (t)
f24 (t)
12/04 17/04 22/04 27/04 02/05 07/05 12/05 17/05 22/05 27/05 01/06 06/06 11/06
Date
(a)
f21 (t)
f22 (t)
f23 (t)
f24 (t)
12/04 17/04 22/04 27/04 02/05 07/05 12/05 17/05 22/05 27/05 01/06 06/06 11/06
Date
(b)
Fig. 6 Decomposition sequences of original daily ground settlements from 13th April to 11th June:
a DK26460; b DK26580
DK26460 and DK26580 have a gradient of 0.5248 and 0.5077, respectively, which are
significantly lower than that with the developed WPT-LSSVM approach (see Fig. 8).
In other words, the developed WPT-LSSVM approach in this chapter can return
predictions in significantly higher accuracy than the traditional LSSVM approach.
To sum up, the prediction results strongly advocate the possibility of hybridizing the
f21 (t)
f22 (t)
f23 (t)
f24 (t)
12/04 17/04 22/04 27/04 02/05 07/05 12/05 17/05 22/05 27/05 01/06 06/06 11/06
Date
(a)
f21 (t)
f22 (t)
f23 (t)
f24 (t)
12/04 17/04 22/04 27/04 02/05 07/05 12/05 17/05 22/05 27/05 01/06 06/06 11/06
Date
(b)
Fig. 7 Predicted sequences of daily ground settlements using LSSVM: a DK26460; b DK26580
WPT and LSSVM algorithm to attain more preciseness and reliability in estimating
the tunnel-induced ground settlement.
3 3
Training data Training data

Testing data Testing data
2 2
Prediction=Observation Prediction=Observation
1 1
Predictions
Predictions
0 0
Slope=0.9761 Slope=0.9824
-1 Intercept=-0.0020 -1 Intercept=0.0087
R2=0.9867 R2=0.9881

-2 Slope=0.9507 -2 Slope=0.9630
Intercept=-0.0043 Intercept=0.0821
R2=0.9566 R2=0.9283
-3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Observations Observations
(a) (b)
Fig. 8 Scatterplot of observations and WPT-LSSVM enabled predictions for tunnel-induced daily
ground settlement in both training data and testing data: a DK26460; b DK26580
3 3

2 2
Prediction=Observation Prediction=Observation
1 1
Predictions
Predictions
0 0

-1 -1
R2=0.7825 R2=0.7693
-2
Testing data -2
Testing data
R2=0.5158 R2=0.4841
-3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Observations Observations
(a) (b)
Fig. 9 Scatterplot of observations and LSSVM enabled predictions for tunnel-induced daily ground
settlement in both training data and testing data: a DK26460; b DK26580
5.3 Analysis of Results
To further test the accuracy of the developed prediction approach, some traditional
prediction approaches are selected to perform the prediction for the same data sets.
Comparisons among different methods are made in terms of their prediction perfor-
mance, where the measured ground settlement provides a benchmark. Since one
typical class of intelligent methods named ANN has been widely applied to analyze
the tunnel-induced ground settlement in the last decade, two representatives of the
ANN-based approach, including Back Propagation (BP) ANN and Radial Basis
Function (RBF) ANN, are chosen for comparison purposes. As a result, Fig. 10
visualizes the Comparisons between the actual measured values and the predicted
Training data Testing data

Safety warning line
12/04
(a)
Training data Testing data

Safety warning line
12/04
(b)
Fig. 10 Comparisons between the actual measured values and the predicted values using different
methods concerning the tunnel-induced daily ground settlement: a DK26460; b DK26580
Table 4 Results of prediction efficiency indicators using different methods

Monitoring point BP RBF LSSVM WPT-LSSVM
MAE RMSE MAE RMSE MAE RMSE MAE RMSE
DK26460 0.3984 0.4996 0.3669 0.5035 0.3931 0.5248 0.1023 0.1476
DK26580 0.3201 0.4766 0.3640 0.4338 0.2068 0.3110 0.0989 0.1600
Average 0.3593 0.4881 0.3655 0.4687 0.3000 0.4179 0.1006 0.1538
values using different methods at DK26460 and DK26580, and Table 4 measures the
prediction performance. Those results are further analyzed and discussed as follows.
(1) From a perspective of the evolution of the ground settlement, all the prediction
methods, including BP, RBF, LSSVM, and WPT-LSSVM, are competent to
capture the rough trend of the tunnel-induced ground settlement evolution in the
real world. For DK26460 as shown in Fig. 10, in the testing set, the maximum
positive and negative daily ground settlement is predicted to appear on 3rd June
and 25th May, respectively. For DK26580 in the testing set, the positive and
negative maximum daily ground settlement is predicted to appear on almost
11th June and 3rd June, respectively. These predictions are quite consistent
with the actual observations. Either the maximum positive or negative ground
settlement at DK26460 appears earlier than that at DK26580. That is because
DK26460 is at the front of DK26580 along the H-H tunnel section, and the
TBM first passes through the cross-section at DK26460, and then DK26580
in a sequential order. At DK26460, it is found that the maximum positive and
negative settlement values almost exceed the safety warning line, and thus some
corresponding control measures are necessarily carried out in time to mitigate
the risk. Accordingly, the magnitude of tunnel-induced ground settlement at
DK26460 is expected to quickly go back to the normal level.
(2) From a comparison between ANN-based and SVM-based approaches, the
prediction performance of the SVM-based approach is proven to be greater
than the ANN-based approach in this case. For DK26460 and DK26580,
Table 4 reveals that the average values of MAE by using BP and RBF are
0.3593 and 0.3655, respectively, while the average values of MAE by using
LSSVM and WPT-LSSVM are 0.3000 and 0.1006, respectively. The average
values of RMSE by using BP and RBF are 0.4881 and 0.4687, respectively,
while the average values of RMSE by using LSSVM and WPT-LSSVM are
0.4179 and 0.1538, respectively. Clearly, the SVM-based approach results in
a lower value of MAE and RMSE, indicating that the SVM-based approach is
prone to generate higher accuracy in the prediction task than the ANN-based
approach. Besides, these two popular AI approaches (ANN and SVM) are
significantly different in theory and applications. More specifically, the devel-
opment of ANNs follows a heuristic path with applications and extensive exper-
imentation preceding theory, while the development of SVMs involves sound
theory, implementation, and experiments. ANNs can suffer from multiple local
minima, while SVMs training always finds a global minimum. A possible
reason for the outstanding performance of SVM over ANNs in practice is that
SVMs can tackle the biggest problem with ANNs, since ANNs use empirical
risk minimization, while SVMs use structural risk minimization. SVMs utilize
kernels to represent nonlinear relationships by a linear model, which can be
less likely to be overfitting [26]. Besides, two more advantages of SVMs can be
highlighted as the simple geometric interpretation and the flexibility to provide
sparse solutions [18]. In short, SVMs are suitable to reach relatively high clas-
sification accuracy even in the case with high-dimensional data and a small
set of training patterns. Therefore, this method can assist in the time series
analysis of tunnel-induced ground settlement, where the size of the training set
is always small, particularly for one specific tunnel section.
(3) From a comparison between LSSVM and WPT-LSSVM methods, the predic-
tion accuracy of the WPT-LSSVM method exceeds the LSSVM method. For
DK26460 and DK26580 listed in Table 4, the average values of MAE and RMSE
by using LSSVM are 0.3000 and 0.4179, respectively, while the average values
of MAE and RMSE by using WPT-LSSVM are 0.1006 and 0.1538, respectively.
Due to the smaller value of MAE and RMSE from the WPT-LSSVM method,
it can be concluded that the WPT-LSSVM method has a better estimation
performance with higher accuracy than the LSSVM method. Therefore, the
improvement of prediction accuracy is likely to root in the WPT process. As
for the original time sequence of the observed daily ground settlement in Fig. 5,
it is impossible to capture any patterns during this time-dependent settlement
process, since there are some unavoidable measurement errors. The measure-
ment error always presents a high frequency in a similar way to noise, while the
actual settlement presents a low frequency. Accordingly, we can then separate
the monitored data into the high- and low-frequency parts with the support of
WPT, resulting in four separate sequences f 12 (t), f 22 (t), f 32 (t), and f 42 (t) (t =
1, 2, …, 60) that come from the original signal f (t) (t = 1, 2, …, 60). Based on
the sequence within the lowest frequency (see f 12 (t) (t = 1, 2, …, 60) in Fig. 6,
we can capture a key pattern that the ground surface experiences an upheaval
first and then is followed by continuous settlements. This is due to the fact of
the improvement effect on the surrounding soil because of plastic flows as the
EPB TBM progresses [23]. In conclusion, the proposed WPT-LSSVM model
possesses the following two advantages in the task of predicting tunnel-induced
ground settlement based on the past ground settlement. For one thing, WPT-
LSSVM has a high capability in multi-resolution decomposition, which can
offer better localization properties in both spatial and frequency domains and
provide flexibility for the signal representation to achieve better classification
accuracy. For another, it can effectively avoid disturbance between strong and
weak signals during the training and learning of prediction models, and thus
the accuracy and reliability of prediction results can be further enhanced.
6 Summaries 91
6 Summaries
It has been highlighted that accurate prediction of tunnel-induced ground settle-

ment is an important prerequisite for ensuring construction safety during tunneling
projects. Most of the existing models focus on the maximum ground settlement,
but they are unable to realize dynamic and real-time predictions. In this chapter,
a hybrid prediction approach is developed to raise the estimation accuracy in esti-
mating tunnel-induced ground settlement on a daily basis. To verify the feasibility
and applicability of the developed approach, a case study is performed in a real-
istic tunnel case in the Wuhan metro system in China. Also, the developed approach
has been compared to existing methods in terms of prediction performance. Some
conclusions are summarized as follows.
(1) A systematic WPT-LSSVM approach is developed and described step-by-step
in this chapter. To begin with, WPT is conducted to decompose the original
time-domain signal into a series of sequences using WPT. Then, LSSVM
models are built to predict the target time series within high- and low-frequency
regions separately. Next, all the predicted sequences are reconstructed to
recover the predicted tunnel-induced ground settlement over time. Finally,
two indicators, MAE and RMSE, are calculated for performance measurement
through quantifying the correspondence between individual pairs of model
predictions and observations for performance analysis.
(2) The SVM-based approach displays higher prediction accuracy than the ANN-
based approach, resulting in a more reliable estimation of tunnel-induced
ground settlement. There are two significant advantages of SVMs to be empha-
sized: one is to have a simple geometric interpretation, while the other is to
give a sparse solution. To explain the superiority of SVMs over ANNs in
practice, a possible reason is that SVMs can address the biggest problem of
ANNs, since SVMs are less prone to over-fitting. Even under the situation of
high-dimensional data and a small set of training patterns, SVMs are sufficient
to reach a relatively high classification accuracy. In the case of studying one
specific tunnel section, it should be noted that the size of the training set in
the tunnel-induced ground settlement is always small. This makes SVMs an
ideal candidate for the time series analysis and estimation of the tunnel-induced
ground settlement with a small training set.
(3) The developed WPT-LSSVM approach has demonstrated a more promising
prediction performance than the traditional LSSVM approach, resulting in
predictions under higher accuracy. It is believed that the WPT process plays a
critical part in raising the prediction accuracy mainly due to its multi-resolution
decomposition capability. Therefore, WPT can completely decompose the
original time-domain signal into high- and low-frequency regions at every
level, which can effectively avoid disturbance between strong and weak signals
during the training and learning of SVMs. As expected, the accuracy of the
prediction results can be improved.
In the end, the novelty of this chapter can be summed up from two aspects: (a) the
state of knowledge by proposing a hybrid method that integrates signal processing
method named WPT and a popular machine learning algorithm called LSSVM,
aiming to significantly improve the prediction accuracy and reliability; (b) the state of
practice by precisely estimating the tunnel-induced ground settlement as the construc-
tion project progresses toward completion, which can help to dynamically detect the
possible risk and sound a warning ahead of time.
References
1. Azadi M, Pourakbar S, Kashfi A (2013) Assessment of optimum settlement of structure adjacent

urban tunnel by using neural network methods. Tunn Undergr Space Technol 37:1–9
2. Bobet A (2001) Analytical solutions for shallow tunnels in saturated ground. J Eng Mechan
127:1258–1266
3. Celestino T, Gomes R, Bortolucci AA (2000) Errors in ground distortions due to settlement
trough adjustment. Tunn Undergr Space Technol 15:97–100
4. Chou W-I, Bobet A (2002) Predictions of ground deformations in shallow tunnels in clay. Tunn
Undergr Space Technol 17:3–19
5. Ding L, Ma L, Luo H, Yu M, Wu X (2011) Wavelet analysis for tunneling-induced ground
settlement based on a stochastic model. Tunn Undergr Space Technol 26:619–628
6. Gao RX, Yan R (2011) Wavelet packet transform, wavelets. Springer, Berlin, pp 69–81
7. Gomes J, Velho L (2015) Windowed fourier transform, from fourier analysis to wavelets.
Springer, Berlin, pp 47–60
8. Karakus M (2007) Appraising the methods accounting for 3D tunnelling effects in 2D plane
strain FE analysis. Tunn Undergr Space Technol 22:47–56
9. Karakus M, Fowell R (2005) Back analysis for tunnelling induced ground movements and
stress redistribution. Tunn Undergr Space Technol 20:514–524
10. Kasper T, Meschke G (2004) A 3D finite element simulation model for TBM tunnelling in soft
ground. Int J Numer Anal Meth Geomech 28:1441–1460
11. Liu J, Qi T, Wu Z (2012) Analysis of ground movement due to metro station driven with
enlarging shield tunnels under building and its parameter sensitivity analysis. Tunn Undergr
Space Technol 28:287–296
12. Łodygowski T, Sumelka W (2006) Limitations in application of finite element method in
acoustic numerical simulation. J Theor Appl Mechan 44:849–865
13. Loganathan N (2011) An innovative method for assessing tunneling induced risks to adjacent
structures, PB2009 William Barclay Parsons Fellowship Monograph 25. Parsons Brinckerhoff
Inc., New York, pp 1–129
14. Martos F (1958) Concerning an approximate equation of the subsidence trough and
its time factors, International Strata Control Congress, Leipzig. Deutsche Akademie der
Wissenschaften zu Berlin, Section fur Bergbau, Berlin, pp 191–205
15. O’Reilly MP, New BM (1982) Settlements above tunnels in the United Kingdom-their magni-
tude and prediction, Tunnelling ’82. Papers Presented at the 3rd International Symposium. Inst
of Mining & Metallurgy, Brighton, Engl, pp 173–181
16. Peck RB (1969) Deep excavations and tunnelling in soft ground. In: Proceedings of 7th
International Conference on SMFE, pp 225–290
17. Sagaseta C (1987) Analysis of undraind soil deformation due to ground loss. Géotechnique
37:301–320
18. Samui P (2008) Support vector machine applied to settlement of shallow foundations on
cohesionless soils. Comput Geotech 35:419–427
References 93
19. Samui P, Sitharam T (2008) Least-square support vector machine applied to settlement of
shallow foundations on cohesionless soils. Int J Numer Anal Meth Geomech 32:2033–2043
20. Santos OJ, Celestino TB (2008) Artificial neural networks analysis of Sao Paulo subway tunnel
settlement data. Tunn Undergr Space Technol 23:481–491
21. Sun Y, Leng B, Guan W (2015) A novel wavelet-SVM short-time passenger flow prediction in
Beijing subway system. Neurocomputing 166:109–121
22. Sun Z, Chang C (2004) Statistical wavelet-based method for structural health monitoring. J
Struct Eng 130:1055–1062
23. Suwansawat S, Einstein HH (2006) Artificial neural networks for predicting the maximum
surface settlement caused by EPB shield tunneling. Tunn Undergr Space Technol 21:133–150
24. Suykens JA, Vandewalle J (1999) Least squares support vector machine classifiers. Neural
Process Lett 9:293–300
25. Verruijt A, Booker J (1996) Surface settlements due to deformation of a tunnel in an elastic
half plane. Géotechnique, pp 753–756
26. Wang F, Gou B, Qin Y (2013) Modeling tunneling-induced ground surface settlement
development using a wavelet smooth relevance vector machine. Comput Geotech 54:125–132
27. Wu J-D, Liu C-H (2009) An expert system for fault diagnosis in internal combustion engines
using wavelet packet transform and neural network. Expert Syst Appl 36:4278–4286
28. Xu Q (2013) Impact detection and location for a plate structure using least squares support
vector machines. Structural Health Monitoring, 1475921713495083
29. Zhang L, Wu X, Ji W, AbouRizk SM (2017) Intelligent approach to estimation of tunnel-
induced ground settlement using wavelet packet and support vector machines. J Comput Civil
Eng 31:04016053
30. Zhang L, Wu X, Skibniewski MJ, Fang W, Deng Q (2015) Conservation of historical buildings
in tunneling environments: Case study of Wuhan metro construction in China. Constr Build
Mater 82:310–322
Information Fusion
1 Introduction
Information fusion is a data-driven technique to combine data from multiple sources,

which can generate improved information in higher quality and accuracy for detec-
tion, inference, or characterization of an object than a single sensor alone [17]. The
techniques of information fusion have been extensively utilized in multisensory envi-
ronments with the following advantages. Firstly, the heterogeneous data collected
by multiple sensors are usually influenced by measurement errors, sensor accuracy,
environmental factors, inevitably leading to a high degree of uncertainty in practical
projects. Through aggregating multiple pieces of evidence, additional informative
and rich data can be generated to better tackle uncertain problems and reduce data
ambiguity. Secondly, it has been found that fusing observed information can reach a
higher accuracy of a classification decision. Moreover, information fusion supports a
wider spatial and temporal coverage. It helps decision-makers to better comprehend
the observed situation and achieve more convincing predictions.
Currently, machine learning begins to put into many types of industries to replace
traditional analytical approaches. It gives computers the ability to automatically learn
ever-increasing volumes of sensor data with no need for human interaction, which in
turn efficiently and accurately identify patterns and make predictions. Among various
machine learning models, support vector machine (SVM) has been maturely applied
in structural health assessment (SHA), which distinguishes different classes of data
with the maximum margins by creating hyperplanes. For instance, Chong et al. [4]
developed a nonlinear multiclass SVM-based approach to quantitatively estimate the
damage levels for a three-story smart building under ambient excitations, achieving
an accuracy of 98.96%. Ghiasi et al. [8] developed a least square SVM for structural
damage detection, aiming to improve the precision and calculation speed for classi-
fication, and then applied the algorithm to a four-story structure with 12 degrees of
freedom, which turned out to have a 98.14% accuracy. The merits of SVM can be
96 Information Fusion
outlined as follows [1]: (1) SVM has excellent prediction and generalization capabil-
ities; (2) SVM is global and unique, and thus, it can reach a global minimum to tackle
the overfitting problem; (3) SVM does not need too much concern about the dimen-
sionality of the input space; (4) Although SVM is a black-box model, the post-hoc
explainability techniques can be adopted to describe its inner structure mathemat-
ically. This research, therefore, intends to implement an SVM-based classification
for safety risk assessment using multi-sensory data.
When several SVM-based classifiers are deployed to review a set of data gathered
by sensors and predict the structural damage at different locations, it remains an
open question to perform a decision-level information fusion to realize an overall
risk assessment in the whole structure system. Specifically, data recorded by sensors
could be incorrect and incomplete with varying degrees of uncertainties, and thus,
evaluation based on a single SVM is hard to meet the demand of high prediction accu-
racy. To handle this issue, a typical data fusion tool named the Dempster-Shafer (D-S)
evidence theory is taken into account. It relies on certain combination rules to aggre-
gate all available evidence and model uncertainty, imprecision, and incompletion,
resulting in a degree of belief.
The D-S evidence analysis has been successfully applied to fault diagnosis in
complex systems. For instance, Bao et al. [2] carried out the D-S evidence anal-
ysis on a 20-bay rigid truss structure for structural damage detection, confirming
that the combination of damage detection results from different individual data sets
was better than individual results from each test data separately. Zhang et al. [28]
combined the D-S evidence theory with the cloud model and Monte Carlo simulation
to perceive the safety risk of buildings adjacent to the tunneling excavation, indicating
that the developed information fusion approach was able to synthesize multi-source
information for higher prediction accuracy and a more optimal decision making. Pan
et al. [16] proposed a hybrid risk analysis method based on interval-valued fuzzy
sets, the D-S evidence theory, and fuzzy Bayesian networks to fuse multiple expert
judgments in linguistic expressions, in order to efficiently evaluate the risk level in a
complex project environment under uncertainty. Additionally, it is indeed useful for
classification problems by treating individual classifiers differently according to the
evidence theory. It is noticeable that the Basic Probability Assignment (BPA) has
a great effect on the accuracy of the fusion results. However, there is no universal
standard for BPA calculation in fuzzy and complex conditions, which needs further
attention [28].
In this chapter, a hybrid approach that integrates SVM and the D-S evidence theory
is developed to facilitate risk analysis using multi-sensory data with uncertainties,
conflicts, and even bias. That is to say, the SVM classifier is built to detect the risk level
of a certain structure part based on a great deal of sensor data involving uncertainty
and incorrectness, and then the D-S evidence theory is utilized to combine predic-
tive results from different SVM models for achieving an overall risk assessment.
The main questions are: (i) How to train multi-class SVM with the most suitable
kernel functions and parameters? (ii) How to construct BPAs from the observed
evidence to fuse information from multi-classifiers at the decision level, leading to
1 Introduction 97
more accurate prediction results? (iii) How to verify the effectiveness of the devel-
oped approach under uncertainty? (iv) How to measure the global sensitivity of risk
factors contributing to system safety?
The proposed risk perception approach based on the hybrid information fusion
method is validated in a realistic operational tunnel case in China, which can reliably
evaluate the magnitude of the structural health risk. Results indicate the developed
SVM-DS approach is capable of fusing multi-classifier information effectively from
different SVM models and exhibiting a more outstanding classification performance
than the single SVM model under a high bias. Since the proposed reliable risk analysis
method can efficiently fuse multi-sensory information with ubiquitous uncertainties,
conflicts, and bias, it provides a comprehensive assessment for the structural health
status together with the most critical risk factors, and thus, proper remedial actions
can be taken at an early stage to mitigate the structural failure.
2 Structural Health Assessment (SHA) in Tunnels
With the acceleration of urbanization, more and more underground facilities have
been built for the greater use of underground space, which can relieve the stress of
land scarcity and cope with the increased demand for residence and transportations
in cities. As a major part of underground facilities to support human activities, the
tunnel system is experiencing a quick and mass extension, which brings a lot of
benefits, such as to alleviate both traffic congestions and environmental pollution
problems, to avoid the acquisition of costly land or property, and others [13]. For
example, metro networks in Singapore are planned to be doubled from the current
180–360 km by 2030. China is going to invest more than 1 trillion dollars to construct
90 metro lines in 44 cities with a total mileage of 4448 km from 2018 to 2030 [11].
It should be noted that structure diseases, like cracks, water leakage, settlement, and
others, are common to tunnels in the complex underground environment.
Unlike the surface infrastructures which have received enough regulatory scrutiny,
the prior knowledge about the underground geological conditions and environmental
parameters of underground structures is extremely limited, raising difficulties in the
early risk detection and perception. Since any kind of structural disease will incur
cascading structural failures in the whole tunnel system and eventually threaten the
normal operation of the tunnel, they should be evaluated carefully and timely. In this
regard, a network of smart sensors has gained traction in SHA. They are installed
in various structures of civil infrastructure to continuously collect data about the
structure and its environments, such as temperature, stress, delamination, strain,
vibration, and humidity [18], which can increase the diversity of the data resources.
Subsequently, these large amounts of information from several sources need to be
fused properly for forming a unified picture regarding the health status of the targeted
infrastructure, which can be defined as the multi-sensor data fusion. Hence, it raises
a critical issue that how to assess the structural health accurately and reliably using
T: Overall structural health risk
B1: Material deterioration B2: Seepage damage B3: Segment deformation B4: Exterior load risk
C9: Width of transverse joint C13: Differential settlement

C5: Deviation value of the
C1: Crack length and width C10: Width of longitudinal C14: Cumulative settlement
neutral pH
C2: Crack density joint C15: Water-earth pressure
C6: Leakage points number
C3: Peeling-off area C11: Transverse slab over vault
per 100ml
C4: Peeling-off depth staggering C16: Pressure difference of
C7: Leakage area per point
C12: Longitudinal slab reinforcement inside and
C8: Leakage water amount
staggering outside
Fig. 1 An SHA assessment in operational tunnels
data from multiple sensors, in order to offer a prospective insight into the structural
health status of the underground facilities.
In general, the structural health status of operational tunnels can be divided into
three levels, which are I (Safe), II (Low risk), and III (High risk). It means that when a
structure is evaluated as a level III, it tends to take a very high risk of structural failure.
Working under the complicated underground environment, the structural health of
operational tunnels is impacted by a number of factors. Commonly, risk factors,
which help in perceiving the overall safety risk of the monitoring sections in the
operational tunnel (T), are determined by actual probe data, domain expert judg-
ment, accumulated engineering, and Chinese technical specifications and manuals.
For a clearer framework, these factors can be grouped into four major structural
risk categories, namely material deterioration (B1), seepage damage (B2), segment
deformation (B3), and exterior load risk (B4). Figure 1 presents an SHA framework
consisting of 16 risk factors extracted from sensor data. A detailed description of
those factors is presented as follows.
(1) Material deterioration (B1): Cracks and peeling-off areas in the structure are
the two most visible forms of material deterioration on the structural surface.
They are prone to reduce structure durability and even increase the likelihood
of structural failure. Thus, the information about cracks and peeling-off areas
plays an important role in evaluating concrete structural performance. Relevant
factors, including the measured crack length and width (C1), crack density
(C2), peeling-off area (C3), and peeling-off depth (C4), are used to characterize
the degree of material deterioration precisely.
(2) Seepage damage (B2): The great amount of underground water can unavoid-
ably inflow and leak through the tunnel weak points, like cracks, joints, and
others, leading to structural erosion. In other words, seepage water has a long-
term negative influence on the structure. However, seepage hazards are easily
overlooked by engineers, since they are too tiny to be detected at the initial
stage. Basically, the monitored records, including PH deviation value of the
neutral (C5), leakage point number per 100 ml (C6), leakage area per point
(C7), and leakage water amount (C8) are considered to estimate the severity
of seepage damage, More specifically, C5 shows the chemical property of the
underground water. A poor PH value will inevitably accelerate rock damage,
2 Structural Health Assessment (SHA) in Tunnels 99
and then the saturated rock will give rise to structural diseases. C6 describes
the leakage distribution, and C7, C8 reflect the leakage volume.
(3) Segment deformation (B3): The structural health also depends on the quality
of joints. It is known that water leakage is more likely to occur in too large
joints, and thus, both the width of the transverse joint (C9) and longitudinal
joint (C10) should be concerned. Besides, the transverse (C11) and longitudinal
slab staggering (C12) are important, since the proper slab staggering can serve
to avoid the concentration of load in the concrete section.
(4) Exterior load risk (B4): The stability of the tunnel structure is very sensitive to
the exterior load. Notably, unbalanced pressure applied from the exterior load
tends to trigger considerable deformation and displacement in tunnel founda-
tion, valut, and reinforcement. The occurrence of structural failures can root
in the exterior load, which exceeds the allowable limit value. Generally, four
factors are refined out from the sensor data to measure the magnitude of the
exterior load risk, which are differential settlement (C13), cumulative settle-
ment (C14), water-earth pressure over valut (C15), and a pressure difference
of reinforcement inside and outside (C16).
In general, there are three levels of information fusion, namely the data level, feature
level, and decision level. At the data level, the information fusion process directly
utilizes raw data in form of signal refinement as inputs, such as structure vibration
velocity and acceleration. It can not only offer more accurate data than the individual
sources, but also provide a preliminary understanding of the observed situation. At
the feature level, characteristics of features are retrieved from the raw data, which are
then fused into a single concatenated feature vector. These extracted features can be
fed into machine learning models for further analysis. At the decision level, decisions
from different features are integrated under particular combination rules, resulting
in a more robust and accurate evaluation along with more specific inferences about
the structural health status.
To incorporate different kinds of risk information from multi-sensors into the
risk evaluation process, numerous soft computing methods using mathematical
models have been developed, such as the probabilistic approach [27], fuzzy logic
[7], interval theory [15] and analytic hierarchy process (AHP) methods [12]. For
example, Chileshe et al. [3] established the degree of risk by evaluating the likeli-
hood of risk occurrence in construction projects. Wang et al. [23] adopted the fuzzy
synthetic evaluation method to assess the structural health status of the tunnel, using
data from different types of sensors, where the predicted results were consistent
with the actual operational situation. Hyun et al. [10] performed a risk analysis by
adopting FTA and AHP, reaching a 90% predictability. Notably, the prerequisite of
these above-mentioned risk analysis methods is the high-quality data, but it is impos-
sible for actual projects to always implement precise data acquisition without errors
and missing values. In other words, those traditional soft computing methods are in
difficulties to address important aspects of randomness, noisy, incompleteness, and
uncertainties stemming from the complex dynamic circumstance, human factors, and
information conflicts [28, 29]. Unfortunately, these adverse factors tend to generate
remarkable deviations in decision-making.
Various machine learning algorithms have been developed to learn large data
from different sources for classification, regression, and clustering. It is worth noting
that artificial neural network (ANN) and support vector machine (SVM) are two
promising supervised learning algorithms in the machine learning domain, which
are proven useful in classification tasks [24]. To be specific, ANN can mimic human
learning processes by establishing relationships between process input and output,
which is inspired by the biological neural network with interconnected neurons [14].
However, it still presents some limitations, such as a black box, long computing time,
overfitting, dimensional disaster, and local minimum [9]. As a method in parallel with
ANN, SVM performs classification by constructing optimal hyperplanes in higher-
dimensional space to search for a global solution [22]. According to comparisons
between SVM and ANN, SVM can exhibit greater performance than ANN in some
cases. For example, Farooq et al. [6] made a comparative study to perform a structural
evaluation of damage detection and identification using a set of static strain data with
Gaussian noise by SVM and ANN, respectively. It was found that the prediction
accuracy of SVM was better than ANN under noise contamination. Dushyanth et al.
[5] concluded that SVM stood out due to a relatively small number of training data
and a dramatic reduction in computational time. Some variations of SVM have been
developed to fuse multi-sensor data in multi-classification problems [30].
Due to the effectiveness of SVM and the D-S evidence theory in information
fusion, we aim to propose a hybrid information fusion method based on these two
advanced soft computing methods. Besides, we intend to implement the proposed
approach in SHA where collected data is usually in a high degree of uncertainty and
impreciseness. Currently, the single or simple structures, like a beam, steel frame, and
others, are studied, and very few studies have concentrated on SHA for underground
facilities in a complicated and changeable environment. Moreover, the information
fusion with high conflicts has not gained much consideration in applications, which
tends to make the data fusion meaningless. Thus, a novel risk analysis approach inte-
grating the probabilistic SVM and an improved D-S evidence theory is developed in
this research, in order to provide insights into a better understanding of the structural
risk analysis and assessment subjected to uncertainty and conflicts.
4 A Hybrid Information Fusion Approach
Based on the SVM algorithm and the D-S evidence theory, a hybrid information
fusion approach is developed for achieving an overall structural safety risk assess-
ment. Figure 2 depicts the flowchart of the proposed hybrid approach for risk analysis.
To be specific, several SVM-based classifiers are built to learn data from different
4 A Hybrid Information Fusion Approach 101
Fig. 2 Flowchart of the

Start
proposed hybrid information
fusion approach
Sensor 1 Sensor 2 Sensor m
Probabilistic
Feature extraction SVM
Parameters and kernel

function selection
SVM 1 SVM 2 SVM n
BPA 1 BPA 2 BPA n
Conflict coefficient K calculation Improved D-S

evidence
theory
K<0.95?
Yes No
Dempter Weighted
fusion rule mean rule
Safety level identification
Monte Carol simulation for

robustness verification Safety risk
assessment
Global sensitivity analysis
End
sensors and output posterior probabilities, and then an improved D-S evidence theory
is utilized to refine and synthesize different classification results generated from
probabilistic SVM models. There are mainly three steps incorporated, as elaborated
below.
4.1 Probabilistic SVM
The core idea of SVM is to create the optimal line or hyperplane to separate data into
classes as far as possible, which can tackle both the linear and non-linear problems.
For the linear problem, it is observed that data is linearly separable. Linear SVM
can make linear classification by maximizing the distance (margin) between the
Table 1 Equations of
Kernel functions Equation
commonly used kernel
functions Linear kernel K(xi , x) = xi · x
Polynomial kernel K(xi , x) = (xi · x + 1)d
Gaussian kernel K(xi , x) = exp(−x − xi 2 /2σ2 )
Radial basis function kernel K(xi , x) = exp(−γx − xi 2 )
Sigmoid kernel K(xi , x) = tanh(αxi T x + c)
Note d is the degree of the polynomial kernel function, and γ is
the parameter of the RBF kernel function
hyperplane and the closest data point (support vector). Moreover, the kernel trick
makes SVM suitable for the non-linear problem, which relies on the appropriate
kernel function to map the original non-linear data in a low-dimensional space into a
higher dimension. Table 1 lists some commonly used kernel functions of SVM, which
define the inner product in the transformed space to avoid the expensive calculations
in the new dimensions [19]. Equation (1) gives the nonlinear decision function of the
hyperplane. It is known that the type of kernel function and its related parameters
greatly affect the classification accuracy of the SVM model. For example, large kernel
parameters are prone to cause overfitting. The penalty parameter C of the error term
controls the tolerance of the systematic outliers. A small value of C tends to ignore
some data near the decision boundary, while a large value of C can minimize the
training error as much as possible. However, if the value of C is too large, it is likely
to lose some generalization properties of the classifier and take a long calculation
period.
N

f (x) = sign αi yi K(xi , x) + b (1)
i=1
where αi is Lagrange multiplier, N is the size of the training data set, b is a threshold
parameter based on the training set, and K(xi , x) is the kernel function.
Regarding the linear SVM, it has shown many strengths, like the great generaliza-
tion capability, clear margins of separation between classes, effectiveness in handling
high-dimensional datasets. However, the model can be very sensitive to uncertain-
ties, such as noise, error-prone sensor data, and others, and thus, it is inefficient in
datasets with a high level of noise. Additionally, it does not take the confidence of the
prediction into account. To address the limitations of the linear SVM, a probabilistic
SVM is developed to give the classification result in a form of the posterior proba-
bility. Platt [21] put forward a method to map the output of SVM into the interval
[0, 1]. According to Eq. (2), a posterior probability can be estimated using a sigmoid
function.
1
P(y = 1|x) ≈ PAS BS (f(x)) = (2)
1+ e(AS f(x)+BS )
where x is the input data, and f(x) is the output of the standard SVM from Eq. (1). AS
and BS are parameters calculated by minimizing the negative log-likelihood function
as follows:

min − t(xi )log(P(xi )) + (1 − t(xi ))log(1 − P(xi )) (3)
i
N
+ +1
y
N+ +2 i
= +1
ti = 1
y
N− +2 i
= −1
i = 1, 2, . . . , l (4)
where P(xi ) = P(y = 1|x), N+ is the number of positive data (yi = +1), and N− is
the number of negative data (yi = −1).
4.2 Improved D-S Evidence Rule
A data fusion method is required to deal with multiple classification results that
are generated from different SVM models. It should be noted that the D-S evidence
theory specializes in tackling uncertainty and fusing multi-source information, which
quantifies the degree of belief in each hypothesis [25]. A frame of discernment is
defined as = {A1 , A2 , . . . , AN } to be a finite nonempty set of mutually exhaustive
and exclusive hypotheses, where N refers to the number of hypotheses. A power set
2 is the set of all subsets of , which includes itself and the empty set ∅. A basic
probability assignment (BPA) is a function m: 2 → [0,1] (where 2 is the power
set of ), which satisfies the following conditions:

m(∅) = 0
(5)
A∈2 m(A) = 1
where m(A) is the basic probability assignment of A for A ⊆ . More specifically, the
belief measurement m(A) between [0, 1] denotes the ratio of all related and available
evidence to support the claim that a certain element in belongs to the set A but not
to the subset of A. The degree of the belief in the interval [0, 1] is assigned to every
possible hypothesis under given evidence.
Based on the BPA, the belief function and plausibility function can be expressed
by Eqs. (6) and (7) to measure the confidence level. Figure 3 visualizes Bel and Pl
in a more comprehensible way. That is to say, Bel refers to the least probability to
hold the proposition A, while Pl is interpreted as the maximum uncertainty value of
A. The greater the uncertainty and impreciseness of the fused information, the larger
the distance between Bel(A) and Pl(A).
Fig. 3 Relationship between Bel ( A) Uncertainty Interval Bel ( A)

belief and plausibility
functions
0 1
Pl ( A)

Bel(A) = m(B) (6)
B⊆A

Pl(A) = m(B) (7)
B∩A=∅
where A, B ⊆ are the propositions in 2 .

When different BPAs {m1 , m2 , …, mn } are available, the Dempster’s combination
rule in Eqs. (8) and (9) can be calculated to aggregate multiple independent pieces
of evidence at one time based on the orthogonal sum of BPA [20]. Special emphasis
should be put on the conflict level K. When the value of K approaches 1, a very
high conflict will occur. Especially when K equals 1, the evidence will completely
conflict with each other, leading to the value of zero in the denominator 1- K in
Eq. (8). That is to say, the conventional Dempster’s rule is unable to handle high-
conflicting pieces of evidence, which tends to produce less informative combinations
and generate counter-intuitive results.
⎧
⎪
⎨
n
1
1−K
mj (Ak ) A = ∅
m(A) = m1 ⊕ m2 ⊕ · · · ⊕ mn = ∩N
k=1 Ak =A, A⊆
j=1 (8)
⎪
⎩ 0 A=∅

n
K= mj (Ak ) (9)
∩N
k=1 Ak =A, A⊆
j=1
where n is the number of pieces of evidence, j is the jth piece of evidence, N is the
number of hypotheses, k is the kth hypothesis, K is termed as the conflict coefficient
in the range of [0, 1] to represent the conflict level among multiple pieces of evidence,
and m1 ⊕ m2 ⊕ · · · ⊕ mn means the orthogonal sum of evidence.
To deal with high-conflict evidence, a hybrid combination rule is proposed based
on the Dempster’s rule and the weighted mean rule, which can aggregate pieces
of evidence directly at a time rather than a stepwise method [26]. The developed
weighted mean rule is given in Eqs. (10)–(12). More specifically, Eq. (10) measures
the similarity degree in different pieces of evidence by the Euclidean distance. Equa-
tion (11) calculates the weight of the ith piece of evidence. Equation (12) provides the
modified BPA m(A) to represent the belief measure of the target A in all hypotheses.
It is assumed that K > 0.95 (1–0.05 = 0.95) indicates an unacceptable high conflict,
since a 5% measurement error in actual construction is acceptable and a statistic

convention p > 0.05 usually defines a significant difference. Therefore, if K is larger
than 0.05, the weighted average rule is activated. Otherwise, the classical Dempster’s
rule in Eqs. (8) and (9) still takes effect.

j=n k=N 2
di = mi (Ak ) − mj (Ak ) (10)
j=1 k=1
di−1
wi = i=n −1
(11)
i=1 di

mi∗ (Ak ) = wi mi (Ak )
∗ (12)
mi () = 1 − N k=1 mi (Ak )
where n is the number of evidence, i (or j) is the ith (or the jth) piece of evidence, N is
the number of hypotheses, and k is the kth hypothesis. Specifically, mi () quantifies
the degree of uncertainty after the evidence fusion. The smaller mi () is, the more
reliable the fusion result is.
Herein, a toy problem is taken as an example to intuitively explain the
improved D-S evidence rule. There are two pieces of evidence, namely,
m1 (A) = {m1 (A1 ), m1 (A2 ), m1 (A3 ), m1 ()} = {0.98, 0.01, 0, 0.01}, and m2 (A) =
{m2 (A1 ), m2 (A2 ), m2 (A3 ), m2 ()} = {0, 0.02, 0.96, 0.02}. Their conflict coeffi-
cient K can be derived from Eq. (13). If we only refer to the classical Dempster’s
rule in Eqs. (8) and (9), we can calculate the BPA for A1 according to Eq. (14) and
obtain the mass function for A2 , A3, and in the same way. The combination result
turns out to be m1,2 (A) = {0.653, 0.02, 0.32, 0.007}. However, K = 0.97 is larger
than 0.95, which means there is a very high conflict between the given evidence. As
a solution, the weighted mean rule in Eqs. (10)–(12) should be triggered. Since only
two pieces of evidence exist, the distances d 1 and d 2 are the same, as expressed in
Eq. (15). Also. the weights w1 and w2 for the first and second pieces of evidence
are equal, as shown in Eq. (16). Equation (17) returns the new fusion result as
m1,2 (A) = {0.49, 0.015, 0.48, 0.015}, which gives less belief to A1 and A2 .
K = m1 (A1 ) · m2 (A2 ) + m1 (A1 ) · m2 (A3 ) + m1 (A2 ) · m2 (A1 )

+ m1 (A2 ) · m2 (A3 ) + m1 (A3 ) · m2 (A1 ) + m1 (A3 ) · m2 (A3 )
= 0.98 × 0.02 + 0.98 × 0.96 + 0.01 × 0 + 0.01 × 0.96
+ 0 × 0 + 0 × 0.02 = 0.97 (13)
1
m1 ⊕ m2 (A1 ) = [m1 (A1 ) · m2 (A1 ) + m1 (A1 ) · m2 () + m1 () · m2 (A1 )]
1−K
1
= (0.98 × 0 + 0.98 × 0.02 + 0.01 × 0) = 0.653
0.03
(14)

d1 = d2 = (0.98 − 0)2 + (0.01 − 0.02)2 + (0 − 0.96)2 + (0.01 − 0.02)2 = 1.372
(15)
1.372
w1 = w2 = = 0.5 (16)
1.372 + 1.372
m1,2 (A) = 0.5 × m1 (A) + 0.5 × m2 (A)

= {0.5 × 0.98 + 0.5 × 0, 0.5 × 0.01 + 0.5 × 0.02,
0.5 × 0 + 0.5 × 0.96, 0.5 × 0.01 + 0.5 × 0.02}
= {0.49, 0.015, 0.48, 0.015} (17)
4.3 Safety Risk Analysis and Assessment
In this step, when all information from different SVM models is fused at the decision
level, we can reach the final safety risk degree. Equation (18) presents a compound
decision-making rule to evaluate the belief of the fusion results, where the maximum
BPA determines the risk label for a certain sample. The confidence indicator m() is
utilized to measure the credibility of fusion results. That is, the smaller the value of
m() indicates the more credible the fusion result. Based on the perceived magnitude
of the safety risk, corresponding measurements can be taken in advance to prevent
potential structural failures.

m(Aw ) = max{m(Ai )}
(18)
m() < θ
where Ai (i = 1, 2, . . . , n) represents n safety levels, Aw denotes the fusion result, and

the belief distribution is expressed as m(A) = {m(A1 ), m(A2 ), . . . , m(An ), m()}.
In this research, the threshold is set to be θ = 0.1.
Sensitivity analysis is also an important aspect of risk analysis to examine how
results change with variables. Although there are a lot of related methods, most of
them mainly measure local sensitivities without the consideration of nonlinearity
and interactions among input factors. For this issue, the global sensitivity analysis
(GSA) is deployed to study the effect of an input factor on the performance of the
whole structural system. Consequently, the most critical factors can be determined,
which help to formulate appropriate remedial actions for risk mitigation.
The implementation of GSA is briefly described below [28]. Firstly, the defuzzi-
fication process for the overall risk level (T) in Eq. (19) is performed to explain the
magnitude of fuzzy sets by a defuzzied value. Then, the Monte Carlo (MC) simula-
tion approach is employed to simulate a series of input data {xi1 , xi2 , . . . , xiQ } by a
certain probability distribution. After Q times of simulation iterations, the relevant
outputs {t1 , t 2 , . . . , t Q } can be obtained. Finally, the Spearman’s rank correlation
coefficient expressed in Eq. (20) is calculated, which is not dependable on distribu-
tions with a linear relationship or similar shape. Since it quantifies the contribution
of each risk feature C i to the overall risk level, the most sensitive risk factors can be
determined. Accordingly, decision-makers can give more attention to these critical
factors for lowering the risk of structural failure.
⎧
⎨ t = 0.5 ϕ (Aj )−1 + j
ϕ (Aj )+1
i=n
⎩ ϕ Aj = i=j m(Ai ) (19)
i=j
i=1 m(Ai )
Q q q
q=1 (R xi − R(xi ))(R(t ) − R(t ))
q q
GSA(Ci ) = q
q 2 Q (20)
Q q 2
q=1 (R xi − R xi ) q=1 (R(t ) − R(t ))
q
where t is the defuzzified value for the belief distribution m(A) =

q
{m(A1 ), m(A2 ), . . . , m(An ), m()}; Q is the iteration number of simulations; R xi
q q q
is the rank of xi of simulated input data; R(xi ) is the mean value of R xi ; R(t q ) is
the rank of t q of all outputs after Q iterations; R(t q ) is the mean value of R(t q ).
5 Information Fusion for SHA in Tunnels
A realistic operational tunnel case in China is used to verify the effectiveness and
feasibility of the developed information approach. That is to say, multi-sensory infor-
mation with ubiquitous uncertainties, conflicts, and bias can be reasonably fused to
perceive structural risk level, leading to a better understanding of the structural health
status together with the most critical risk factors. Based on the in-depth analysis,
proper remedial actions can be taken at an early stage to reduce the potential risk in
the operational tunnel as much as possible.
5.1 Project Profile
The research object is an operational tunnel located at the Wuhan Metro System,
China. To be specific, the tunnel case (short for the C-H tunnel) is a double-spool
tunnel connecting the C-Station and H-Station in the Wuhan Metro Line 2. The whole
metro line begins operation in Dec. 2016. The C-H tunnel section has a length of
1174 m. The tunnel is mainly buried in the layer of muddy-silty clay with silt and
silty sand, and some sections pass through the artesian aquifer at 2.2–3.5 m below
the ground. The layout of the C-H tunnel case is provided in Fig. 4. Due to the under-
ground environment in the nature of complexity and uncertainty, various risks from
erosion, cracks, aging, and others are inevitably involved in the operational tunnel.
Therefore, the evaluation of its structural health status is not a straightforward task,
which could be considered as a complex multi-attribute decision problem. Figure 5
shows some examples of diseases in the operational tunnel.
With the help of sensor technologies and data acquisition platforms, various
sensors can be installed in the operational tunnel to continuously monitor the physical
behavior of the structure. In this case, we divide the whole tunnel into 35 monitoring
Fig. 4 Layout of the C-H tunnel case
Fig. 5 Examples of structural diseases: a water seepage; b diseases in the cross section of segments;
c cracks; d differential settlement
5 Information Fusion for SHA in Tunnels 109
sections (i.e., No. 1, No. 2, …, No. 35). Each section is 5 m in length along with
the tunnel central axis, which has several sensors with different functions to gather
a wide variety of data, including the displacement sensor, pressure sensor, and other
sensors. Subsequently, these large amounts of information from multiple sources
need to be fused properly for forming a unified picture regarding the health status
of the targeted tunnel, which can be defined as the multi-sensor data fusion. To be
more specific, four SVM models are performed directly on the features extracted
from the data collected by sensors. Classification results from SVM models need
to be further fused using the improved D-S evidence theory, aiming to achieve an
overall structural health status with high reliability and quality.
5.2 SVM Model Training
There are two major steps in SVM model training, which are described in detail
below:
(1) Choice of kernel function and parameters
The kernel function, kernel function parameters, and penalty parameter C exert a great
impact on the performance of SVM, which all contribute to finding out an optimal
hyperplane for separation in a high-dimensional feature space. Figure 6 visualizes
the scatter points about two factors C3 and C4 in the structure risk category B1 and
compares the classification results based on different kernel functions. From Fig. 6a,
b, it is observed that the linear separation is poor in classifying the two groups.
This is because there is an obvious overlapping in C3 and C4. In contrast, the RBF
kernel in Fig. 6c works well in separating points into three classes with the fewest
errors, which outperforms the linear kernel in Fig. 6a, b and the polynomial kernel
in Fig. 6d. In other words, the RBF kernel better fits the training data than others,
which is selected as the proper kernel function in this case.
Regarding the RBF-SVM model, two key parameters, namely the radius of RBF
gamma γ and the penalty parameter C, need to be fine-tuned by cross-validation.
Due to the limited input data, the sevenfold cross-validation is conducted here to
determine the best value of γ and C, aiming to establish an optimal SVM model in
pursuit of more reliable probability outputs. To be specific, the validation data set
(i.e., 35 sets of sample data (C1, C2, …, C16) from multiple sensors in 35 monitoring
sections) is randomly split into 7 mutually exclusive groups in equal size. For each
training, one of the 7 groups is taken out as a test set, while the remaining 6 groups
are put together as a training set. That is to say, all the data can be utilized for training
and testing, which can expectedly reduce the errors from bias. After 7 iterations of
training and validation, every group can be used as a test set once and as a training set
six times. In the end, the model performance is determined by the average accuracy
from the 7 iterations.
Pairs of (C, γ) in different values are tested in the RBF-SVM model and their
corresponding classification accuracy for four structural risk categories B1, B2, B3,
C3 C3
(a) (b)
C3 C3
(c) (d)
Fig. 6 SVM classification results for the classification of the feature B1 using factors C3 and C4
based on different kernel functions: a linear kernel; b linear SVM; c RBF kernel; d polynomial
kernel
and B4 are shown in Fig. 7. In Fig. 7b–d, it is obvious that the green point has the
highest accuracy than others, which denotes the value of C = 10, γ = 5. As the
value of γ in the green line increases, the accuracy value goes up first, reaches the
peak value at γ = 5, and then decreases gradually. It means that a too large curvature
of the decision boundary will decrease the classification accuracy. Although the best
classification performance in terms of the structural risk category B1 in Fig. 7a does
not originate from the pair of (C, γ) in (10, 5), C with a value of 10 can still reach
the highest accuracy under the condition that γ is equal to 0.5 or 1. Consequently,
the value of parameter C and γ is set as 10 and 5, respectively, in the RBF-SVM to
exhibit the robust classification performance.
(2) Calculation of posterior probability
Based on the constructed SVM model, the safety levels (i.e., I, II, and III) can be
assigned to the risk categories (i.e., B1, B2, B3, and B4). The confusion matrix in
Fig. 8 summarizes the number of true and false predictions, where a deeper color
refers to a larger number. It offers a clear insight into the performance of the SVM
classification model. From a qualitative point of view, the classification accuracy can
be calculated by the proportion of the sum of the value along the diagonal to the total
value, which works out to be 82.86% for B1, 94.29% for B2, 91.43% for B3, and
94.29% for B4, respectively. In other words, the overall accuracy for B1 is at least
8.57% lower than other categories, that is because the parameters C = 10, γ = 5 are
not the optimal ones in the SVM model of B1.
0.90 0.96
C=0.01 C=0.01
C=0.1 0.94 C=0.1
0.88 C=1
C=1
0.92
C=10 C=10
0.86 C=100 0.90 C=100
0.88
Accuracy
Accuracy
0.84
0.86
0.82
0.84
0.80 0.82
0.80
0.78
0.78
0.76 0.76
0 2 4 6 8 10 0 2 4 6 8 10
γ γ
(a) (b)
0.95
0.90
0.85 0.90
0.80
0.85
Accuracy
Accuracy
0.75
0.70
0.80
0.65 C=0.01 C=0.01
C=0.1 C=0.1
0.60 C=1 0.75 C=1
C=10 C=10
0.55 C=100 C=100
0.70
0 2 4 6 8 10 0 2 4 6 8 10
γ γ
(c) (d)
Fig. 7 SVM classification accuracy based on pairs of (C, γ) for four risk features: a B1; b B2;
c B3; d B4
To take full advantage of the SVM outputs, they can be mapped into the poste-
rior probabilities by a sigmoid function using Eq. (2). Table 2 takes the monitoring
sections No. 31, No. 32, No. 33, No. 34, and No. 35 as an example, which lists
their probability results corresponding to safety levels I, II, and III, respectively. As
a result, the predicted label is determined by the highest probability in bold font in
Table 2. Although the established four SVM models for different risk categories can
reach comparatively promising classification accuracy greater than 82.86%, they are
likely to return different predictions for the same monitoring section. For instance,
the predictions for risk categories B1, B2, and B4 in the monitoring section No. 31
all turn out to be a level II, while the risk category B3 is labeled as level III. What’s
worse, the monitoring section No. 35 suffers from a more significant inconsistency
in the classification predictions. Only predictions for B3 and B4 are the same (level
II), while the SVM models for B1 and B2 output contradictory results as level I and
level III, respectively.
It is revealed that SVM models tend to generate conflicting and imprecise predic-
tions for the same monitoring sector, which will inevitably cause confusion in the
18
15
12
0
Level I Level II Level III Level I Level II Level III
True label True label
(a) (b)
16
12
0
Level I Level II Level III Level I Level II Level III
True label True label
(c) (d)
Fig. 8 Confusion matrix for four risk factors: a B1; b B2; c B3; d B4
evaluation of the overall structure health status. Notably, the highest and second-
highest probability in an SVM model could be very close to unfortunately reduce
the credibility of the predicted result. The monitoring section No. 35 is taken as an
example. There is a very small discrepancy in the probability of a level I (0.581) and
level III (0.406) for risk category B1, implying that the conclusion drawn from such
results does not have a very high confidence level. The same problem also appears
in other monitoring sections. In this regard, when two or three risk levels have the
same proportion, the safety level can be defined by the combinations as (I, II), (II,
III), (I, III), and (I, II, III) and then transferred into a corresponding framework
= {I, II, III} with seven subsets except for the empty set for further investigation.
Table 2 Results of probabilistic SVM at five monitoring sections

Monitoring section Risk feature p(I) p(II) p(III) Predicted label
No. 31 B1 0.171 0.600 0.230 II
B2 0.078 0.671 0.252 II
B3 0.195 0.150 0.654 III
B4 0.170 0.617 0.212 II
No. 32 B1 0.609 0.142 0.249 I
B2 0.806 0.083 0.110 I
B3 0.270 0.105 0.624 III
B4 0.403 0.345 0.252 I
No. 33 B1 0.178 0.546 0.276 II
B2 0.052 0.235 0.713 III
B3 0.110 0.223 0.667 III
B4 0.110 0.223 0.667 III
No. 34 B1 0.170 0.552 0.278 II
B2 0.074 0.684 0.242 II
B3 0.218 0.637 0.144 II
B4 0.182 0.622 0.196 II
No. 35 B1 0.581 0.013 0.406 I
B2 0.214 0.181 0.604 III
B3 0.077 0.671 0.252 II
B4 0.170 0.617 0.212 II
5.3 Multi-classifier Information Fusion
According to the improved combination rule introduced in Sect. 3.2, the predicted
results from four SVM classifiers can be fused continually at the decision level
until all of them have been combined. It can generate more robust results than the
fusion at the raw data level and feature level, which is expected to enhance the
reliability and quality of the safety risk perception. Noticeably, BPA based on the
belief degree is critical in the information fusion process. Following the works from
[31], the outputs of SVMs in the form of posterior probabilities can be directly
converted into BPA. The frame of the discernment = {I, II, III} is defined to
refer to three risk levels. In consequence, the largest value among m (I), m (II),
and m (III) determines the label assigned to the targeted monitoring section, and
m() measures the uncertainty in the assessment result. The results of BPAs for all
monitoring sections after multi-classifier information fusion are outlined in Table 3,
which are analyzed as follows.
(1) The proposed data fusion approach can effectively reduce the uncertainty and
increase the precision in the overall structural health evaluation. Clearly, only
Table 3 Results of BPAs for all monitoring sections after multi-classifier information fusion
Monitoring section m(I) m(II) m(III) m(Θ) Predicted label True label
No. 1 0.0222 0.9398 0.0379 1.05 × 10−4 II II
No. 2 0.0207 0.9366 0.0426 1.21 × 10−4 II II
No. 3 0.0207 0.9366 0.0426 1.21 × 10−4 II II
No. 4 0.2735 0.6766 0.0500 4.85 × 10−6 II II
No. 5 0.0236 0.9407 0.0356 9.48 × 10−5 II II
No. 6 0.0200 0.9384 0.0415 1.31 × 10−4 II II
No. 7 0.0287 0.7895 0.1811 6.35 × 10−4 II II
No. 8 0.0596 0.9051 0.0353 4.47 × 10−5 II II
No. 9 0.0596 0.9051 0.0353 4.47 × 10−5 II II
No. 10 0.0192 0.9503 0.0305 9.27 × 10−5 II II
No. 11 0.2385 0.7242 0.0372 5.87 × 10−6 II II
No. 12 0.0192 0.9503 0.0305 9.27 × 10−5 II II
No. 13 0.0383 0.7640 0.1970 7.00 × 10−4 II II
No. 14 0.0099 0.9778 0.0122 3.91 × 10−5 II II
No. 15 0.0305 0.9240 0.0452 2.22 × 10−4 II II
No. 16 0.0305 0.9240 0.0452 2.22 × 10−4 II II
No. 17 0.0628 0.8099 0.1269 3.84 × 10−4 II II
No. 18 0.0104 0.9865 0.0030 1.45 × 10−5 II II
No. 19 0.0160 0.9628 0.0211 7.11 × 10−5 II II
No. 20 0.2368 0.7358 0.0273 1.74 × 10−6 II II
No. 21 0.0192 0.9617 0.0190 6.17 × 10−5 II II
No. 22 0.0178 0.9479 0.0342 1.05 × 10−4 II II
No. 23 0.0140 0.9202 0.0655 2.85 × 10−4 II II
No. 24 0.0126 0.9508 0.0364 1.25 × 10−4 II II
No. 25 0.0375 0.8418 0.1204 3.58 × 10−4 II II
No. 26 0.0971 0.7031 0.1994 3.46 × 10−4 II II
No. 27 0.1473 0.8015 0.0511 2.94 × 10−5 II II
No. 28 0.1109 0.8327 0.0564 3.64 × 10−5 II II
No. 29 0.0416 0.5083 0.4483 1.79 × 10−4 II II
No. 30 0.0100 0.9685 0.0214 7.96 × 10−5 II II
No. 31 0.0386 0.7636 0.1972 6.70 × 10−4 II II
No. 32 0.8416 0.0375 0.1205 3.60 × 10−4 I I
No. 33 0.0290 0.1884 0.7791 3.43 × 10−4 III III
No. 34 0.0207 0.9366 0.0426 1.21 × 10−4 II II
No. 35 0.3276 0.3144 0.3578 1.75 × 10−4 III II
the monitoring section No. 35 cannot be evaluated correctly, which means that
the classification accuracy based on SVM and the D-S evidence theory can
rise up to 97.14% (34/35). As for the confidence indicators m(), its value of
the 35 monitoring sections all gets close to zero (less than 1.0 × 10−4 ). That
is because the case has no combination (I, II, III) in Table 2.
(2) Data fusion has a powerful capability in dealing with the conflicting results
from four SVM models, and thus, it serves as a promising tool to fully synthe-
size multi-source information with high quality. The monitoring section No.
32 is taken as an example. In Table 2. there is no agreed result about the overall
safety level from the four SVM models (B1, B2, and B4 belong to the level I,
while B3 belongs to the level III). After the information fusion, Table 3 indi-
cates that the BPA value for safety level I (that is m (I)) is equal to 0.8416,
resulting in an ultimate prediction of the level I under a very high confidence
level.
(3) For all correct prediction results, No. 33 is rated at the high-risk level (level
III), which can be considered as the most dangerous section in this opera-
tional tunnel. This evaluation result consists of the facts, since the leakage
water amount recorded by sensors is 7.695 L/m2 /day above the limit value of
1 L/m2 /day. Therefore, more attention should be paid to the monitoring section
No. 33 for improving its safety performance. To mitigate risk at No. 33, it is
essential to take proper actions especially for controlling the seepage damage.
To be specific, some common methods, such as using the waterproof materials,
grouting to block the water, discharging water, and others, can be implemented
to restrict the amount of water leakage below the limit value.
6 Merits Analysis
The effectiveness of the proposed hybrid information approach based on the SVM
model and the improved D-S evidence theory is demonstrated from two aspects.
On the one hand, it can return a robust and comprehensive evaluation result, which
is useful in the structural safety analysis. On the other hand, it is superior to the
single SVM model in handling high-bias data, which is possible to achieve higher
classification accuracy. Discussions are presented as follows.
6.1 Uncertainty Modeling
The monitoring and measurement in structural health could involve a number of

inherent uncertainties, which mainly originate from measurement errors, compli-
cated environments, and others. Measurement errors are closely associated with the
characteristics of sensors, including quality, location, installation, and others, while
the environment-related uncertainties stem from the unpredictable geological under-

ground conditions. Since the observed data is likely to suffer from unavoidable bias, it
is necessary to validate the robustness of the proposed hybrid method under different
bias levels. Therefore, various percentages of bias (i.e., 5, 10, 15, and 20%) are
added to the collected data from sensors as the preparation for the comparison test.
Commonly, the bias level of 5% is considered acceptable in measurement whereby
p ≤ 0.05. To describe and reduce the underlying uncertainty, the Monte Carlo simu-
lation approach is adapted to generate a large number of samples randomly according
to the distribution characteristic of the collected data. It is assumed that measured
data, in this case, follow the normal distribution, which is most commonly used
in applications. The simulation process will repeat 1000 times for each monitoring
section. The results are analyzed as follows.
(1) The BPA distribution is conducive to determine the structural safety level intu-
itively. Each iteration will recalculate BPAs and synthesize classification results
from four probabilistic SVMs to provide an overall evaluation of the structural
health level. The final health status of the specific monitoring section is deter-
mined by the safety level with the greatest value of BPA. Figure 9 illustrates
the results of BPA within different risk levels after 1000 iterations at the moni-
toring section No. 33 under different bias levels, where different colors and
shapes refer to different risk levels. When the input data is at the bias level of
5%, 10%, and 15%, Fig. 9a–c demonstrates that the most part of the highest
BPAs belongs to the level III. Thus, it is believed that section No. 33 can be
assigned the label as high risk, which is very consistent with the perception
result given in Table 3. Especially in Fig. 9a, most of the BPAs for the safety
level III are very close to 1 at all bias levels, which are separated apart from
other levels. However, when more biases are input into the measured data, the
BPA distributions for level II and level III are prone to be mixed together. As
shown in Fig. 9d, the model has difficulty in distinguishing the safety level
III with level II under the bias level of 20%. Besides, the value of confidence
indicator m () is substantially close to zero, indicating the low uncertainty
in safety level prediction. In short, it’s better to control the bias level within
15%, in order to ensure the accuracy and reliability of the perceived risk result.
Since the monitoring section No. 33 is rated at a high-risk condition, special
attention is required to minimize its possibility of structural failure at an early
stage.
(2) The proposed hybrid information fusion exhibits the merits of the robustness
to the bias. To have a better understanding of bias, Fig. 10 demonstrates the
frequency of the certain safety level, which is estimated under different bias
after 1000 iterations for four targeted monitoring sections (i.e., No. 31, No.
32, No. 33, and No. 34). It can be seen that the proposed approach is more
likely to misclassify the safety level with an increase of bias level. Especially
in Fig. 10b, c, the frequency of the misclassification is extraordinarily obvious
at a 15 and 20% bias. Although the model is possible to mistakenly predict the
risk condition throughout the 1000 iterations at a 20% bias, its error range is
6 Merits Analysis 117
Fig. 9 Results of BPAs after 1000 iterations at the monitoring section No. 33 under different bias
levels: (1) 5%; (2) 10%; (3) 15%; (4) 20%
acceptable. There are still 79.2% of predictions at No. 32 falling in the correct
label (that is a risk level I (Safe)) and 70.1% of predictions at No. 33 belonging
to a risk level III (High risk). As for the condition with data in a 5% bias,
the model can realize a nearly 100% classification accuracy, verifying that
bias limited less than 5% has little impact on the reliability of the prediction.
Therefore, the strong robustness of the hybrid information fusion method is
validated to some extent. Besides, it is worth noting that No. 34 is free of bias
since the prediction accuracy under various bias levels all reach up to 100%.
That is because the four SVM models of No. 34 give back consistent results,
namely the safety level II (Low risk).
(3) Global sensitivity analysis is conducted through the 1000 times iterations at
the monitoring section No. 33, since No. 33 is evaluated as the most dangerous
section (level III) and deserves special attention. Global sensitivity analysis
aims to identify the critical risk factors contributing more to the structural health
status. Accordingly, proper actions can be taken ahead of time for structural
failure prevention. The Spearman’s rank correlation coefficient from Eq. (20)
Risk level I Risk level II Risk level III Risk level I Risk level II Risk level III
1000 1000
800 800
Frequency
Frequency
600 600
400 400
200 200
0 0
5% 10% 15% 20% 5% 10% 15% 20%
Bias Bias
(a) (b)
Risk level I Risk level II Risk level III Risk level I Risk level II Risk level III
1000 1000
800 800
Frequency
Frequency
600 600
400 400
200 200
0 0
5% 10% 15% 20% 5% 10% 15% 20%
Bias Bias
(c) (d)
Fig. 10 Results of safety risk assessment after 1000 iterations under different bias levels at four
monitoring sections: (1) No. 31; (2) No. 32; (3) No. 33; (4) No. 34
measures the contribution of risk factors to the overall safety level. Figure 11
presents the global sensitivity analysis of 16 risk factors at the monitoring
section No. 33. It is clear that C7, C10, C13, and C15 turn out to be the top
four sensitive risk factors with the greatest positive global sensitivities, which
should be given more concern for reducing the risk magnitude at No. 33.
For C10, C13, and C15, their sensitivities gradually decline as the bias level
grows. When the bias increases to 20%, the most sensitive risk factors remain
unchanged, which verifies the robustness of the proposed method once again.
In contrast, C3, C4, and C11 perform the lowest sensitivity, which has little
influence on the structural system safety.
0.5
Bias 5%
Bias 10%
0.4 Bias 15%
Bias 20%
Global Sensitivity Measure
0.3
0.2
0.1
C3 C12 C16
0.0
C1 C2 C4 C5 C6 C7 C8 C9 C10 C11 C13 C14 C15
-0.1 Risk Factors
-0.2
Fig. 11 Global sensitivity analysis of 16 factors at the monitoring section No. 33
6.2 Comparative Study
It is known that the standard SVM model also owns the ability to estimate the overall
structural safety level. A comparison between the SVM model and the proposed
SVM-DS approach is presented below. To address unavoidable uncertainty, the
Monte Carol simulation is adopted, and thus, the calculation is repeated throughout
1000 times for the single SVM model.
(1) The single SVM model can effectively deal with input data at a very low bias
level, but it exhibits very poor classification performance at a high bias. Using
the standard SVM model, Fig. 12 describes the frequency distribution of the
total correct classification number among 35 monitoring sections at various bias
levels (i.e., 5%, 10%, 15%, and 20%). It is clear that nearly the 1000 iterations
can return correct prediction for 34 or even all the 35 monitoring sections
at 5% and 10% bias with a low standard deviation, representing a precise
and stable prediction performance. However, when the bias level grows up to
15% and 20%, respectively, the frequency distribution significantly changes.
The range of the correct number falls in the range [20, 35], which no longer
concentrates on the value of 34 or 35. In addition, the standard deviations at
15% and 20% bias increases to 4.09 and 5.83, respectively, which indicates
that the total correct number is spread out over a wide range, leading to great
fluctuations and unstable results. In Fig. 12c at a bias level of 15%, although the
largest frequency is in total correct number 35, the value of frequency is only
344. That is to say, only 34.4% of iterations can make the right classification
for all 35 monitoring sections. In Fig. 12d under the 20% bias, more than
1000 1000
Mean=35.00 Mean=34.90
Standard DeviaƟon=0.00 Standard DeviaƟon=0.36
800 800
Frequency
Frequency
600 600
400 400
200 200
0 0
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
Total number of correct monitoring points Total number of correct monitoring points
(a) (b)
350
Mean=28.92 Mean=22.26
500
300 Standard DeviaƟon=5.83 Standard DeviaƟon=4.09
250 400
Frequency
Frequency
200
300
150
200
100
100
50
0 0
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
Total number of correct monitoring points Total number of correct monitoring points
(c ) (d)
Fig. 12 Frequency distribution of the total correct classification number after 1000 iterations using
the standard SVM method under different bias levels: (1) 5% bias; (2) 10% bias; (3) 15% bias; (4)
20% bias
50% of iterations only classify 20 monitoring sections correctly. This reveals

that the classification accuracy of the single SVM model will experience a
sharp fall as the bias level increases. Moreover, the reduction in the mean
value of the total correct number in Fig. 12a–d also represents a remarkable
decrease in the classification accuracy, leading to imprecise and incredible
predictions. Overall, the single SVM is unable to ensure high accuracy and
reliability especially at a high bias larger than 15%.
(2) The hybrid SVM-DS model has better robustness than the single SVM model.
Figure 13 shows a comparison of the classification accuracy between the
proposed hybrid method and the single SVM intuitively. From the compar-
ison in Fig. 13, the misclassification percentage of the SVM-DS model under
a large bias is relatively lower than the single SVM, indicating that the hybrid
information fusion approach is capable of greatly improving the classification
accuracy and the robustness to bias. That is because the improved combination
method based on the D-S evidence theory and the weighted mean rule can
make full use of all the information, including conflicting evidence. Although
Fig. 13 Comparison of Single SVM

classification accuracy in the 1.0 SVM-DS
SVM-DS model and the
single SVM model
Classification Accuracy
0.8
0.6
0.4
0.2
0.0
5% 10% 15% 20%
Bias
the single SVM model owns 2.93% and 3.37% higher classification accuracy
than the hybrid method within a bias below 10%, its accuracy goes an apparent
decline when the bias is greater than 15%. More specifically, there is a sudden
drop in a 20% bias, reaching a classification accuracy of 63.59%. It implies a
poor classification performance of the single SVM model, since approximately
40% of data is classified incorrectly. Therefore, it can be concluded that the
single SVM model is very sensitive to bias. On the contrary, the proposed
SVM-DS model can still classify 87.99% of data correctly at a 20% bias,
whose classification accuracy only decreases 10% from a 5% bias to a 20%
bias. Since the underground environment of the operational tunnel is generally
filled with much complexity and uncertainty, it is more reasonable to imple-
ment the proposed SVM-DS model, which is proven to effective in tackling
high-bias information and then giving more accurate results of the safety risk
perception.
7 Summaries
The main objective of this research is to present a hybrid information fusion approach
with the combination of the probabilistic SVM and the improved D-S evidence theory.
Special emphasis should be put on the improved D-S evidence theory integrating the
Dempster’s rule and the weighted mean rule. The proposed risk analysis approach
serves as a decision tool to provide a reliable and comprehensive assessment of
structural safety level in the operational tunnel under uncertainty, which takes the
following three steps: (1) A total of 16 risk factors ranging from four risk categories
are identified and extracted from multiple sensors, which greatly affect the reliability
of the operational tunnel; (2) Four probabilistic SVMs are well trained by selecting the
appropriate kernel function and parameters; (3) The improved D-S evidence theory
makes the best of predictions from the four SVM models and fuse them at the decision
level. As a result, it is expected to achieve a more accurate safety assessment for all
the monitoring sections. Moreover, the Monte Carlo simulation is adopted to address
uncertainty for robustness validation and global sensitivity analysis. For verifying
its feasibility and effectiveness, it has been successfully applied in an operational
tunnel in the Wuhan metro system in China.
In view of the case, the hybrid SVM-DS method is helpful in fusing multi-source
information, which can considerably improve the accuracy and stability of the risk
level evaluation. The meaningful findings can be summarized as follow: (1) The
prediction results from the SVM-DS model are robustly reliable for the observed data
from the underground environment with a great deal of uncertainty and variability. It
can better synthesize highly conflicting classification results from four probabilistic
SVM models, reaching an overall accuracy of 97.14% for the targeted infrastructure.
(2) The SVM-DS model has a strong robustness to bias. Even when the bias level
grows to 20%, the proposed method can maintain the classification accuracy above
70% for each monitoring section. That is to say, SVM-DS has a superior bias tolerant
ability to minimize the adverse influences from bias, in order to ensure the reliability
and accuracy of the prediction results. (3) The SVM-DS model outperforms the single
SVM model in handling information with high conflicts and bias. Especially in the
condition of a 20% bias, the single SVM model only correctly classify 63.84% of
samples, which is 24.15% less than the accuracy of the SVM-DS model. In addition,
global sensitivity analysis is carried out to recognize the most critical risk factors on
structure risk throughout a series of repeated iterations. The identified most sensitive
factors can be the major concern for safety control measurements, and thus, managers
can tackle the potential risk in advance to prevent the occurrence of structural failures.
The novelty of the developed approach lies in two aspects. On one hand, it
contributes to the state of the knowledge by training data from various sensors by
SVM and synthesizing multi-sensory information using the improved D-S evidence
theory at the decision level to obtain a more precise and reliable assessment. On the
other hand, it contributes to the state of the practice by providing a promising tool for
evaluating the structural health and perceiving the potential risk of the operational
tunnel in the subsurface environment with great complexity and uncertainty, in order
to inform data-driven decisions for mitigating the risk of structural failure at an early
stage.
References
1. Arrieta AB, Díaz-Rodríguez N, Del Ser J, Bennetot A, Tabik S, Barbado A, García S, Gil-
López S, Molina D, Benjamins R (2019) Explainable artificial intelligence (XAI): concepts,
taxonomies, opportunities and challenges toward responsible AI. arXiv preprint arXiv:.10045
2. Bao Y, Li H, An Y, Ou J (2012) Dempster-Shafer evidence theory approach to structural damage
detection. Struct Health Monit 11:13–26
3. Chileshe N, Boadua Yirenkyi-Fianko A (2012) An evaluation of risk factors impacting
construction projects in Ghana. J Eng Des Technol 10:306–329
References 123
4. Chong JW, Kim Y, Chon KH (2014) Nonlinear multiclass support vector machine–based health
monitoring system for buildings employing magnetorheological dampers. J Intell Mater Syst
Struct 25:1456–1468
5. Dushyanth N, Suma M, Latte MV (2016) Detection and localization of damage using empirical
mode decomposition and multilevel support vector machine. Appl Phys A 122:250
6. Farooq M, Zheng H, Nagabhushana A, Roy S, Burkett S, Barkey M, Kotru S, Sazonov E (2012)
Damage detection and identification in smart structures using SVM and ANN, Smart Sensor
Phenomena, Technology, Networks, and Systems Integration 2012. International Society for
Optics and Photonics, p. 83461O
7. Feng SJ, Yao AM, Dong YF (2014) Fuzzy comprehensive evaluation of open-pit slope stability.
Adv Mater Res 962–965:1029–1033
8. Ghiasi R, Torkzadeh P, Noori M (2016) A machine-learning approach for structural damage
detection using least square support vector machine based on a new combinational kernel
function. Struct Health Monit 15:302–316
9. Huang Y (2009) Advances in artificial neural networks–methodological development and
application. Algorithms 2:973–1007
10. Hyun KC, Min S, Choi H, Park J, Lee IM (2015) Risk analysis using fault-tree analysis (FTA)
and analytic hierarchy process (AHP) applicable to shield TBM tunnels. Tunnel Undergr Sp
Technol Incorp Trenchless Technol Res 49:121–129
11. Li M, Yu H, Jin H, Liu P (2018) Methodologies of safety risk control for China’s metro
construction based on BIM. Saf Sci 110:418–426
12. Manca D, Brambilla S (2011) A methodology based on the analytic hierarchy process for the
quantitative assessment of emergency preparedness and response in road tunnels. Transp Policy
18:657–664
13. Pan Y, Ou S, Zhang L, Zhang W, Wu X, Li HJRE (2019) Modeling risks in dependent systems:
a Copula-Bayesian approach. Reliab Eng Syst Saf 188:416–431
14. Pan Y, Zhang L (2020) Data-driven estimation of building energy consumption with multi-
source heterogeneous data. Appl Energy 268:114965
15. Pan, Y., Zhang, L., Li, Z., Ding, L., 2019b. Improved fuzzy Bayesian network-based risk
analysis with interval-valued fuzzy sets and DS evidence theory. IEEE Transactions on Fuzzy
Systems.
16. Pan Y, Zhang L, Li Z, Ding L (2020) Improved fuzzy Bayesian network-based risk analysis
with interval-valued fuzzy sets and DS evidence theory. IEEE Trans Fuzzy Syst 28:2063–2077
17. Pan Y, Zhang L, Wu X, Skibniewski MJ (2020) Multi-classifier information fusion in risk
analysis. Inform Fus 60:121–136
18. Pan Y, Zhang L, Wu X, Zhang K, Skibniewski MJ (2019) Structural health monitoring and
assessment using wavelet packet energy spectrum. Saf Sci 120:652–665
19. Patle A, Chouhan DS (2013) SVM kernel functions for classification. In: Advances in
Technology and Engineering (ICATE), 2013 International Conference on. IEEE, pp 1–9
20. Peng Y, Lin J-R, Zhang J-P, Hu Z-Z (2017) A hybrid data mining approach on BIM-based
building operation and maintenance. Build Environ 126:483–495
21. Platt J (1999) Probabilistic outputs for support vector machines and comparisons to regularized
likelihood methods. Adv Large Margin Classif 10:61–74
22. Vapnik V (2013) The nature of statistical learning theory. Springer Science & Business Media
23. Wang B, Mo C, He C, Yan Q (2017) Fuzzy synthetic evaluation of the long-term health of
tunnel structures. Appl Sci 7:203
24. Worden K, Manson G (2007) The application of machine learning to structural health
monitoring. Philos Trans 365:515
25. Yuan K, Deng Y (2019) Conflict evidence management in fault diagnosis. Int J Mach Learn
Cybern 10:121–130
26. Zhang L, Ding L, Wu X, Skibniewski MJ (2017) An improved Dempster-Shafer approach to
construction safety risk perception. Knowl-Based Syst 132:30–46
27. Zhang L, Skibniewski MJ, Wu X, Chen Y, Deng Q (2014) A probabilistic approach for safety
risk analysis in metro construction. Saf Sci 63:8–17
28. Zhang L, Wu X, Zhu H, Abourizk SM (2016) Perceiving safety risk of buildings adjacent to
tunneling excavation: an information fusion approach. Autom Constr 73:88–101
29. Zhang W, Sun K, Lei CZ, Jr, Bill FS (2014b) Fuzzy analytic hierarchy process synthetic
evaluation models for the health monitoring of shield tunnels. Comput Aided Civil Infrast Eng
29:676–688
30. Zhao J, Xie X, Xu X, Sun S (2017) Multi-view learning overview: Recent progress and new
challenges. Inform Fusion 38:43–54
31. Zhou Q, Zhou H, Zhou Q, Yang F, Luo L, Li T (2015) Structural damage detection based on
posteriori probability support vector machine and Dempster-Shafer evidence theory. Appl Soft
Comput 36:368–374
Dynamic Bayesian Networks
1 Introduction
Underground transportation systems are in great demand in many large cities

all over the world. Tunnel construction has presented a powerful momentum for
rapid economic development worldwide. However, owing to various risk factors in
complex project environments, safety violations occur frequently in tunnel construc-
tion, leading to large problems on the surface transport operation. To avoid heavy
casualties and property losses caused by safety violations, innumerable studies have
introduced risk analysis into safety management and control.
Risk analysis can be divided into qualitative and quantitative categories [16]. The
former category includes Fault Tree Analysis (FTA), Comprehensive Fuzzy Evalua-
tion Method (CFEM), Safety Check List (SCL), and others, while the latter category
includes influence diagrams, Neural Network (NN), Support Vector Machine (SVM),
decision trees, and others. The above methods make a significant contribution to
safety management in complex engineering projects, however, they are confined to
static control management [1]. Khakzad et al. [9] claimed that FTA was unsuitable
for complex problems due to its limitation in explicitly representing the dependencies
of events, updating probabilities, and coping with uncertainties. When the associated
parameters, such as geological, design, and construction parameters are changed, the
aforementioned methods can not accurately illustrate the updating feature of system
dynamics as the construction progress evolves. Nor can professional supports or
suggestions be provided in real time as the parameters are updated.
In recent years, the Bayesian network (BN) has been proposed to model the
complexity in man-machine systems and is widely used in knowledge representa-
tion and reasoning [10]. However, in conventional BN-based analysis, it is a static
model representing a joint probability distribution at a fixed point or a time interval.
The dependencies among the random variables are not presented when constructing
the BN model, leading to possible deviations between the predicted results and obser-
vations in reality [17]. Instead, a dynamic Bayesian network (DBN) is an extension
126 Dynamic Bayesian Networks
of the ordinary BN, which allows the explicit modeling of changes subjected to time
series or sequences. DBN is capable of revealing complex multivariate time series,
where the relationships between multiple time series can be represented in the same
model. DBN is developed to unify and extend traditional linear state-space models
(e.g., Kalman filters), linear and normal forecasting models (e.g., ARMA), and simple
dependency models (e.g., hidden Markov models) into a general probabilistic repre-
sentation and inference mechanism for arbitrary nonlinear and non-normal time-
dependent domains [8]. That is to say, DBN can model the evolution of the proba-
bilistic dependencies within a random system. This chapter attempts to use DBN to
deal with the potential uncertainty and randomness underlying safety management
in tunnel construction.
A systemic decision approach based on DBN is developed, aiming to provide
guidelines for dynamic risk analysis of the tunnel-induced damage over time. The
proposed approach can be used to conduct various analysis tasks, including predic-
tive, sensitivity, and diagnostic analysis. Finally, a case study concerning the dynamic
risk analysis in a tunnel project in China is used to verify the applicability of the
proposed DBN-based approach.
2 BN and DBN
2.1 BN
Bayesian network (BN), a combination of graph theory and probability theory,

consists of a directed acyclic graph (DAG) and an associated joint probability distri-
bution (JPD). A BN model with N nodes can be represented as B < G, Θ > , where
G refers to a DAG with N nodes, and Θ refers to the JPD of the BN model. The
nodes {X 1 , …, X N } in the graph are labeled by related random variables. The direct
edges between nodes present association relationships among the variables. DAG
contains conditional independence assumptions, which can be revealed by means
of engineering models, expert judgment, or other known relations [17]. Figure 1
illustrates a simple BN model.
BN allows not only a forward (or predictive) analysis but also a backward (diag-
nostic) analysis, where the posterior probability of any set of variables can be
Fig. 1 A simple BN model

X2
X1 X3
Y
2 BN and DBN 127
computed. The relations represented by DAG allow the JPD to be specified locally by
the conditional probability distribution for each node. Assuming Pa(X i ) is the parent
node of X i in the DAG, the conditional probability distribution of X i is denoted by
P(X i |Pa (X i )). The JPD of P(X 1 , …, X N ) can then be written as Eq. (1).

P(X 1 , . . . , X N ) = P(X i |Pa(X i ) ) (1)
X i ∈{X 1 ,...,X N }
where X i is the ith node in the BN model, Pa(X i ) is the parent node of the ith node,
P(X 1 , …, X N ) is the joint probability of the node set {X 1 , …, X N }, and P(X i |Pa(X i ))
is the conditional probability of X i given Pa(X i ) has been observed.
2.2 DBN
Dynamic Bayesian network (DBN) extends the ordinary BN formalism by intro-

ducing relevant temporal dependencies that capture dynamic behaviors of domain
variables between representations of the static network at different time steps [13].
Thus, DBN is more appropriate for monitoring and predicting values of random
variables and is capable of representing the system state subjected to changes with
time. A DBN model is a way to extend BN to represent probability distributions over
semi-infinite collections of random variables {X 1 , X 2 , …, X N }.
Typically, we assume that each state only depends on the immediately preceding
state (i.e., the system is the first-order Markov), and two-time slices are considered
to model the system temporal evolution. A DBN model represents a discredited
Markov chain process, and this form is usually called 2TBN (two time-slice temporal
Bayesian networks). Accordingly, a DBN model is defined as a pair (B1 , B→ ), where
B1 is a BN model that defines the prior P(X t ), and B→ is a 2TBN which defines
P(X t |X t-1 ) using a transition probability table. Figure 2 illustrates a simple DBN
model. P(X t |X t-1 ) is given by Eq. (2).
X1[0] X1[t-1] X1[t] X1[0] X1[1] X1[2]
X2[0] X2[t-1] X2[t] X2[0] X2[1] X2[2]
X3[0] X3[t-1] X3[t] X3[0] X3[1] X3[2]
(a) (b) (c)
Fig. 2 A simple DBN model: a Prior BN; b 2TBN; c DBN


N

P( X t |X t−1 ) = P( X ti Pa(X ti )) (2)
i=1
where X ti refers to the ith node at time t (i = 1, 2, …, N), and Pa(X ti ) refers to the
parents of X ti in the directed acyclic graph.
The nodes in the first slice of the 2TBN do not have any parameters associated
with them, however, each node in the second slice of the 2TBN has associated a

N
conditional probability distribution for variables which define P( X ti Pa(X ti ))
i=1
for all t > 1 [6]. The distribution given by a 2TBN can be divided into two aspects,
namely, (i) the inter-slice distribution that models the probability of variables in X t
with parents at time t-1; (ii) the intra-slice distribution that models the probability of
variable in X t with parents in the same time slice.
Generally, we assume that the parameters of the conditional probability distribu-
tion are time-invariant, and the model is time-homogeneous. The parents of a node
Pa(X ti ) can either be in the same time slice or the previous time slice [6]. The arcs
between slices are from up to down, which represent causal flows of time. If there
i
is an arc from X t−1 to X ti , this node is called persistent. The arcs within a slice are
arbitrary, and one can put arcs in any way one wants to. That is to say, their relation-
ships are unconstrained, as long as the intra-slice-model is acyclic. The parameters
of the conditional probability distribution used by the proposed model are assumed
time-invariant. When we unroll the 2TBN until we have T time slices, the resulting
joint distribution probability can be defined by Eq. (3).

T
N

P(X 1:T ) = P( X ti Pa(X ti )) (3)
t=1 i=1
where X 1: T refers to the joint events of the observation for X 1 from time t = 1 to T.
3 Dynamics in Tunnel-Induced Damages
With the rapid development of urbanization in human communities, there is a rising

demand for exploring the underground space by tunneling. However, the tunneling
construction process will cause disturbance to the surrounding soil strata and rock
properties, resulting in different types of geo-hazards (e.g., groundwater leaking,
ground settlement, and soil collapse) [5]. These geo-hazards will lead to multiple
risks to the adjacent structures, and finally cause unrecoverable damage, which is
called tunnel-induced damage.
The tunneling excavation may induce damages to various kinds of nearby facili-
ties, such as buildings, roads, bridges, and pipelines [12, 22]. For example, the ground
movement induced by tunneling may lead to the settlement and lateral deformation
3 Dynamics in Tunnel-Induced Damages 129
of piles for adjacent bridges [23]. Several cases of collapses and sinkholes in road-
ways resulted from tunneling construction were reported [18]. Moreover, the ground
movement due to tunneling together with the aging corrosion issues may result in
deformation and fracture of pipelines, leading to the interruption of transportation of
water, oil, gas, and other supplies [22]. Most of the times, tunnels are built under the
urban trunk roads. Tunnel-induced road damage attracts more and more attention,
since the reliability and safety of the road system are directly affected. Typically, the
road damage displays obvious time-dependent characteristics. Thus, it is necessary
and meaningful to investigate the safety risk mechanism of the tunnel-induced road
damage, which can provide a reference for risk modeling and further investigation.
3.1 Tunnel-Induced Road Damage
The soil plays a mediating role in the tunnel-induced damage. The tunnel excavation
in soft soils inevitably leads to ground movement, greatly influencing the road surface
damage and further ground transportation safety. In accordance with the actual tunnel
construction practice, the tunnel-induced ground settlement displays a typical non-
stationary pattern along with time series [19]. Generally, the deformation tendency
following the excavation progress can be divided into three phases, including the
initial deformation, fast development, and deterioration phases. Figure 3 illustrates
the tunnel-induced ground settlement within three phases.
(1) Initial deformation phase. In this phase, both the accumulative ground settle-
ment and the settling rate begin to gradually grow from zero. Mostly, the ground
settlement takes place before the shield machine arrives at the excavation face.
The stress release of the surrounding soil is generally considered as the primary
cause of the ground settlement during this phase.
ground settlement
Accumulative
Initial deformation phase

Fast development phase
Deformation deterioration phase
Excavation face Tunneling shield machine
Excavation direction
Fig. 3 Patterns of the tunnel-induced ground settlement following the excavation progress
(2) Fast development phase. In this phase, the accumulative settlement is contin-
uously increasing at a high speed, while the subsidence rate starts to decrease
gradually. The ground settlement takes place when the shield machine passes
exactly through the excavation face from the shield head until the shield nail.
The generated settlement is mainly due to the increased shearing stress, which
is caused by the friction resistance between the shield machine body and the
surrounding soil.
(3) Deformation deterioration phase. In this phase, the settlement continues to
grow at a very low speed. The subsidence rate is gradually reduced to zero.
Normally, it is the secondary consolidation settlement of the soft soil that
contributes to the settlement. Generally, the maximum ground settlement is
observed at this phase.
3.2 Control Standard for Road Damage
The soil can be disturbed by tunneling activities of a shield machine, and the road
surface damage can be inflicted when the settlement deformation exceeds a certain
threshold. The impact of tunnel excavation on the road surface deformation and
damage can be presented in three aspects, namely cut-in ruts, dents, and carvings.
These aspects are very likely to appear on the road surface due to uneven surface
settlement, which could further affect the performance and durability of the road
pavement. According to statistics, more than 75% of existing roads are made up of
asphalt pavements [21]. Thus, the road surface damage assessment is mainly focused
on asphalt roads, and a case concerning the safety analysis of an asphalt pavement
induced by tunneling excavation is presented later.
To assess the road surface damage, the indicator, Pavement Condition Index (PCI),
is commonly used in the construction engineering practice. An industry standard
“Technical Specifications for maintenance of Highway Asphalt Pavement (JTJ073.2-
2001)” was issued in October 2001 in China, where PCI was calculated by Eq. (4).
Moreover, based upon the calculated results, the health condition of an existing road
surface can be rated into four different levels, namely Excellence (85–100), Good
(70–85), Fair (55–70), and Poor (0–55). The higher the value of PCI, the better the
condition of the road surface.
⎧
⎨ PC I = 100 − 15D R 0.412
⎩ D R = A1 × K 1 + A2 × K 2 + A3 × K 3 × 100%
(4)
A
where DR refers to the damage rate of the pavement condition; A1 , A2 , and A3 refer
to the damaged area of cut-in ruts, dents, and caving, respectively; A refers to the
total area of the investigated road surface; K 1 , K 2 , and K 3 refer to the conversion
coefficient representing the degree of damage due to cut-in ruts, dents, and caving,
respectively. The values (i.e., K 1 , K 2 , and K 3 ) can be obtained from the standard.
3 Dynamics in Tunnel-Induced Damages 131
However, challenges exist in determining the damaged areas, such as A1 , A2 , and

A3 , since the values of the damaged areas can only be measured when the damage
occurs. In other words, the current PCI-based approach is limited by post-accident
assessment, which cannot provide support for pre-accident control ahead of time.
It is known that the real-time monitoring of the settlement plays an important role
in guaranteeing the security in the tunnel excavation process. Once exceeding the
standard, early-warning signals can be released, and necessary timely actions can be
taken before an accident occurs. Thus, to solve the above problem, it is necessary
to investigate the relationship between PCI and the monitored settlement. Hu et al.
[7] and Yao et al. [21] defined the mathematical relationship between PCI and the
monitored settlement based upon the Peck formula [15], as shown in Eq. (5).
⎡ ⎡ 1/ 0.412 ⎤2 ⎤
100 − PC I 15 100 − K × A
⎢1 1
⎦ ⎥
Smax = S0 × exp⎣ ⎣ ⎦ (5)
8 0.4i K 2 + 2i K 3
2
where S max refers to the maximum ground settlement induced by tunnel excavation,
S 0 refers to the normal ground settlement in the free-field where no ground transport
is presented, and i refers to the width coefficient of the tunnel.
Following Eq. (5), a large number of monitoring records of prior cases are used
to conduct an intuitional representation of relations between PCI and S max , and the
fitted curve between the two is displayed in Fig. 4. The fitted curve is approximately
linear when S max falls in a certain range, and tends to be flattened when S max reaches
up to 38 mm. As for S max with a general distribution range of 0–90 mm, we can follow
up the suggested PCI division to separate S max into four different levels accordingly,
including I (Very Safe, 0–15 mm), II (Safe, 15–38 mm), III (Dangerous, 38–71 mm),
and IV (Very Dangerous, 71–90 mm). Specifically, the level I of S max is in response
Fitting Curve
Pavement Condition Index (PCI)
I
Smax vs PCI
SSE: 27.6
R-square: 0.9869
II Adjusted R-square: 0.9868
RMSE: 1.285
III
IV
Accumulative settlement Smax (mm)
Fig. 4 Fitted curve between PCI and S max in tunnel construction

to the excellent road condition (85–100), the level II of S max is in response to the
good road status (70–85), the level III of S max is in response to the fair road condition
(55–70), and the level IV of S max is in response to the poor road condition (0–55).
4 DBN-Enabled Dynamic Risk Analysis
Attempting to improve the effectiveness and accuracy of safety management in

complex project environments, a DBN-based dynamic risk analysis approach is
proposed to investigate the causal relationships among different factors. Figure 5
illustrates the overall flowchart of the proposed approach. There are mainly three
stages involved, as elaborated below.
4.1 Risk/Hazard Identification
Risk/hazard identification involves determining which risks (or risk factors) may
affect the project safety and documenting their characteristics. It is considered a diffi-
cult task for a complex system, particularly in tunnel construction. With the devel-
opment of the tunnel construction practice worldwide, large amounts of scattered
knowledge resources are accumulated, including explicit and tacit knowledge. For
the risk/hazard identification in tunnel construction, three procedures are explicitly
illustrated in this stage.
(1) Knowledge acquisition. Prior safety-based knowledge resources are acquired

for preliminary risk identification. The resources can be derived from three
Start
Predictive analysis
Knowledge acquisition Structure learning

Sensitivity analysis
Risk mechanism analysis Parameter learning

Diagnostic analysis
Relationship discovery Markov chain process

End
Risk/hazard identification DBN learning DBN analytics
Fig. 5 Flowchart of the proposed DBN-based dynamic risk analysis in complex projects
4 DBN-Enabled Dynamic Risk Analysis 133
aspects: namely, i) the accepted knowledge from existing design and construc-
tion standards, and ii) empirical knowledge from statistical models in the
literature, and iii) the tacit knowledge from domain experts.
(2) Risk mechanism analysis. The risk mechanism analysis for the target risk
event can be conducted by consulting metro design and construction standards
and technical manuals and retrieving literature. Relevant variables that repre-
sent or indicate system safety conditions, flow process, and performance states
can be determined.
(3) Relationship discovery. Based on the potential failure-consequence scenarios
in the risk mechanism analysis, explicit relationships and interdependencies
among the above variables can be revealed. Here, the range of each variable’s
observation data is defined. The state of each variable with a different attribution
is usually defined based on both historical data and expert estimation. The
failure modes, internal variables, exogenous factors, the explicit cause, and
effect relationships are determined.
4.2 DBN Learning
A DBN model is established to reveal the causal mechanisms and interdependencies

among system components. Specifically, the design of a DBN model involves deter-
mining the network structure, parameters, and the Markov chain process. There are
mainly three procedures incorporated in this stage, which are illustrated explicitly as
follows.
(1) Structure learning. Structure learning aims to figure out a proper DAG, and
confirm the association relationship between nodes. Every variable in the real-
world situation is represented by a Bayesian node. The nodes (variables) should
be first created based on the results of hazard identification in the last stage. The
network structure can be developed by creating the directed edges from nodes
(representing fault causes) to nodes (representing the consequences). A BN
model with N nodes may have N(N − 1)/2 directed edges at most. The inter-
vention of prior knowledge can reduce the blindness of the BN structure design.
As one of the most commonly used techniques for risk and reliability studies,
FTA is a logic diagram that displays the interrelationships between a critical
event and its causes in a system. A number of existing fault trees have been built
in construction engineering fields, and thus, the qualified fault trees regarding
tunnel construction safety can be employed to provide effective prior knowl-
edge for structure learning. Figure 6 illustrates the simplified procedure of
mapping Boolean gates of fault trees into BN’s directed edges. Meanwhile, the
over-fitting problem, which often occurs when depending entirely on training
data in the structure learning [2] can be avoided in this way.
(2) Parameter learning. Parameter learning aims to determine the conditional
probability distribution of each node under the established BN structure. The
conditional probability tables can be determined by learning the parameters
Fig. 6 Mapping Boolean R

gates of fault trees into R
directed edges of Bayesian
network X1 X2 X1 X2
(a) AND gate
R R
Bayesian
Qualified
network
fault trees X1 X2 X1 X2 structure
(b) OR gate
R R
X1 X1
(c) NOT gate
from the given data using some kinds of learning algorithms. Learning BN’s
parameters from complete data have received considerable attention in the
last decade. When modelers know the structure of BN and the data set has
no missing values, the maximum a posterior (MAP) estimation method is
usually used to learn BN’s parameters [20]. Taking the Dirichlet distribu-
tion as a prior distribution, the MAP method first counts the occurrence of
each value of a given variable to obtain the relative frequency. Then modelers
search for the best BN parameters using a practical search-and-score proce-
dure. Many search-and-score algorithms are available to evaluate the fitness of
the network, such as Greedy search algorithms (i.e., Cooper and Herskovits’
K2 algorithm [4]) and Markov Chain Monte Carlo (MCMC) method (i.e.,
Simulated Annealing and Minimum Description Length Principle). For more
detailed information, see Onisko et al. [14].
One of the practical learning problems is to deal with incomplete data. Compared
to the complete data, the occurrence of missing values leads to the analytical
intractability. It is tempting to simply delete incomplete records or fill in missing
values, because this leads to biased results or an over-confidence in the follow-
up analysis. To avoid such an ad hoc dilemma, some approximation methods can
be adapted to improve the increased uncertainty due to missing values, such as
Expectation-Maximization (EM), Data Augmentation (DA), and so on.
(3) Markov chain process. A discrete-time Markov chain is a mathematical
system that simulates transitions from one state to another in the real-world. It
is a random process usually characterized as memoryless, indicating that the
next state depends only on the current state and not the sequence of events that
precede it. Based on the Markov chain process as illustrated in Eq. (6). The
DBN model is a probability distribution function on the sequence of T hidden-
state variables X = {x 0 ,…, x T-1 } and the sequence of T observables Y = {y0 ,
…, yT-1 }that has the following factorization, which satisfies the requirements
for DBN that the state x t depends only on the state x t-1 .
−1
T −1
T
Pr ( X 1:T | Y1:T ) = Pr ( xt |xt−1 ) × Pr ( yt |xt ) × Pr (x0 ) (6)
t=1 t=0
where Pr(x t |x t-1 ) is the state transition probability density function, Pr(yt |x t ) is the
observation probability density function, and Pr(x 0 ) is the initial state distribution.
4.3 DBN Analytics
The DBN model can be used to conduct various types of analysis after it is estab-
lished. The most important use of DBN lies in updating probabilities in lights of
actual observations. It is, therefore, possible to predict the probability distribution
of potential safety risks and identify the most likely potential causes, given undesir-
able events are observed. In this stage, three sorts of analysis, including predictive,
sensitivity, and diagnostic analysis using the Bayesian inference capabilities, are
presented.
(1) Predictive analysis. The predictive analysis aims to predict the probability
distribution of a risk event (defined by R) under a combination of risk factors
(X 1 , X 2 ,…, X n ). The state of each risk factor, certain or uncertain, is treated as
evidence input into the DBN model. Compared to the traditional FTA, predic-
tive analysis in DBN models does not need to get minimal cut sets, leading
to the improvement of the computational efficiency. The probability distribu-
tion of R, represented by P(R = r), is given by Eq. (7). P(R = r) can serve
as an indicator to evaluate the risk of R, and decision-makers can take proper
preventions in advance if the risk magnitude of P(R = r) is very high.
P(R = r ) = P( R = r |X 1 = x1 , X 2 = x2 , . . . , X N = x N )
× P(X 1 = x1 , X 2 = x2 , . . . , X N = x N )

r = {r1 , r2 , . . . , r P }, xi = xi1 , xi2 , . . . , xiQ i , i = 1, 2, . . . , N (7)
where {r 1 ,r 2 , …, r P } is a range of P states for the risk event R; {x 1i ,x 2i ,…,x Qi

i } is a
range of Qi states for a single risk factor X i ; P(R = r|X 1 = x 1 , X 2 = x 2 , …, X N =
x N ) represents the conditional probability distribution of R; P(X 1 = x 1 , X 2 = x 2 , …,
X N = x N ) represents the joint probability distribution of X.
(2) Sensitivity analysis. The sensitivity represents how sensitive the model or
the system output is in response to minor changes in the input parameters.
Sensitivity analysis is useful in investigating the performance of each factor’s
contribution to the occurrence of the target event. The most natural way of
performing sensitivity analysis is to change the values of the input param-

eters and then monitor the impact of changes on the output probabilities.
A performance-based indicator, Sensitivity Performance Measure (SPM) is
proposed to measure the contribution of each risk factor X i to target event R.
Key risk factors can be identified to assist decision-makers in determining main
checkpoints in the construction phase. Under prior probabilities, the SPM of
each root node X i , represented by SPM (X i ), can be calculated by Eq. (8). In
light of actual observations of events (i.e., given evidence), for instance, X i is
q
observed to lie in a state of qi (X i = xi i ), and then SPM (X i ) can be calculated
by Eq. (9) in this situation.

Qi

1
j
P(R = r |X i = xi ) − P(R = r )
S P M(X i ) = , i = 1, 2, . . . , N (8)
Q i j=1 P(R = r )

1,...,qi −1,qi +1,...,Q i
P(R = r |X = x j ) − P(R = r |X = x qi )
1 i i i
S P M(X i ) = i

Qi − 1 q
P(R = r |X i = xi i )
j=1
i = 1, 2, . . . , N (9)
where r refers to the state of a risk-prone event R with P states; x i refers to the state
j
of risk factor X i with Qi states, and xi refers to the jth state of the risk factor X i .
Generally, SPM(X i ) can be used as an indicator to measure the degree of sensitivity
of the root node X i in case of the occurrence of a risk event. X i is more likely to
become the sensitive factor of the risk event R when SPM (X i ) is close to one, and
then should be given more attention.
(3) Diagnostic analysis. The diagnostic analysis aims to obtain the posterior prob-
ability distribution of each risk factor when an accident occurs. The suspected
causes can be detected in a very short period using a posterior probability
distribution, assisting in the real-time fault diagnosis. Compared to traditional
risk-based methods, such as FTA, CFEM, and NN, the feature of the diagnostic
analysis technique is matchless in the DBN reference [9]. The posterior prob-
ability distribution of risk factor X i , represented by P(X i = x i |R = r), can be
calculated by Eq. (10).
P(X i = xi ) × (R = r |X i = xi )
P( X i = xi |R = r ) = , i = 1, 2, . . . , N (10)
P(R = r )
where r refers to the state of a risk-prone event R with P states, and x i refers to the
state of risk factor X i with Qi states. Normally, X i is more likely to become the direct
cause of the occurrence of the risk event R when P(X i = x i |R = r) is close to one.
Based on the DBN-based risk analysis results, the critical risks and risk factors can
be determined, respectively. Then relevant safety control measures can be proposed
in advance for risk mitigation. Furthermore, safety monitoring and review can be
conducted to secure the rational status of the whole system or its system components.
In addition, the implementation of the control measures can provide feedback and
suggestions on the improvement and optimization of the established DBN model.
5 Dynamic Risk Analysis in Tunnel-Induced Damages
During the tunneling excavation process, the road surface damage is regarded as a
prominent example of tunnel-induced damages, which can significantly affect the
safe operation of the ground transport system. A case concerning dynamic risk anal-
ysis in the construction of the Wuhan Yangtze Metro Tunnel (WYMT) in China
is used to verify the applicability of the proposed approach. A DBN model is
developed to reveal the causal relationships among different influential variables
in tunnel-induced damages. Various types of what-if scenario analysis, including
predictive, sensitivity, and diagnostic analysis and control, are performed to analyze
the probabilistic risk in tunnel-induced damages in a dynamic manner.
5.1 Case Profile
Wuhan Yangtze Metro Tunnel (WYMT) is known as the first metro tunnel across
the Yangtze River in China, which acts as an important route connecting Wuchang
and Hankou districts in the city of Wuhan. It is a double-spool tunnel with a total
length of almost 6200 m. The tunnel construction started in April 2008, and a slurry
shield machine with a diameter of 6.52 m in cutter diameter was adopted to push
the tunnel from the Jiyu Bridge Station in Hankou to the Jianghan Road Station in
Wuchang. A specific monitoring section where the tunnel passed exactly under the
Wuchang Riverside Avenue (WRA) was taken as an example to present the detailed
DBN-based risk analysis.
Figure 7 presents the spatial relationship between WYMT and WRA. WRA was
built in 1927 and further upgraded several times. Due to the close distance to the
bank of the Yangtze River, WRA played a significant role in flood defenses and port
transport development in Wuhan. Therefore, no fault was allowed while the tunnel
was crossing under the existing road surface, leading to a great challenge for safety
management in the construction of WYMT.
5.2 DBN Model Development
Due to complexities in the underground conditions, the tunnel-induced road damage

is highly complicated with various temporal and spatial factors involved, such
Yangtze River Riverside

Avenue
Wuhan Yangtze Metro Tunnel
Wuchang Riverside Avenue
Legend
Tunnel axis
Tunnel outline
Monitoring point
Fig. 7 Spatial relationship between WYMT and WRA in the studied case
as geological, design, and mechanical variables. From a perspective of cause-

consequence analysis, the geological and design variables act as endogenous vari-
ables to the road surface damage, while the mechanical variables can be regarded
as exogenous variables. The description of the two types of influential variables,
affecting tunnel-induced road damages is elaborated as follows.
(1) Geological and design variables. Acting as the intermediary in tunnel-
environment interaction, the soil plays a crucial role in the tunnel-induced
damages. The tunnel excavation in soft soils inevitably leads to soil displace-
ment, which can subsequently affect the existing surface or subsurface struc-
tures [11]. The parameters, such as Soil Cohesion (X 1 ), Friction Angle
(X 2 ), Compression Modulus (X 3 ), and Soil Density (X 4 ), are four main vari-
ables frequently used to illustrate the geological features. Furthermore, the
tunnel related design variable also has a significant impact on disturbances
in surrounding environments. Cover Depth (X 5 ) is usually treated as the key
design parameter in finite element models [3], which can produce notable
impacts on the foundation deformation of surface facilities.
(2) Mechanical variables. In the process of shield-driven tunneling excavation,
some pressure and speed sensors are installed on the top and middle of the
cutter head in advance, and engineers pay close attention to the real-time
measurement of the monitored mechanical parameters, in order to maintain
the excavation face stability and limit settlements. At the construction stage,
the geological and design variables are determined and remain unchanged,
while the mechanical variables turn out to be the core factors which can be
adjusted depending on geological conditions in working sites. Rational and
appropriate Torque Ratio (X 6 ), Cut Slurry Pressure (X 7 ), Drilling Speed (X 8 ),
5 Dynamic Risk Analysis in Tunnel-Induced Damages 139
Thrust Ratio (X 9 ), and Grouting Speed (X 10 ) are considered as very important

technical factors. These factors can be further adjusted to guarantee the safety
of the ground transportation system in the tunneling excavation process.
Concerning the tunnel-induced road damage, prior expert knowledge acquired
from standards, technical reports, and expert experience can be used for risk mecha-
nism analysis. Also, some qualified fault trees regarding tunneling construction safety
are considered when investigating the relations between influential factors. Figure 8
illustrates the conceptual framework in tunnel-induced road damage, with various
influential factors and their relations taken into account. Then, with the support of the
developed early-warning web-based systems, a large number of measurement records
are applied for model learning. Table 1 illustrates the description of discrete states
of the above 10 influential variables. Next, following the conceptual model struc-
ture (see Fig. 8), the K2 algorithm, a well-known algorithm for Bayesian network
structure learning, is used for parameter learning to construct the static BN model.
Furthermore, a 2TBN model representation is used to reveal the system’s temporal
evolution over time. As mentioned in Sect. 3, the tunnel-induced ground settlement
displays a typical non-stationary pattern along with the excavation progress. Also,
expert estimation was considered to obtain the evolutional dynamics of the risk
event over time. Table 2 illustrates the transition probability table for the leaf node
R. Finally, a dynamic risk model for Road Damage Bayesian Network (RDBN) was
established, as shown in Fig. 9. Herein, the “self-link” for the leaf node R referred
to its dynamic behavior over time using a 2TBN representation.
X1: Soil Cohesion X6: Torque Ratio

X2: Friction Angle X7: Cut Slurry Pressure
X3: Compression Modulus X8: Drilling Speed
X4: Soil Density X9: Thrust Ratio
X5: Cover Depth X10: Grouting Speed
Geological and design

Mechanical variables
variables
R: Tunnel-induced road damage
Fig. 8 Conceptual framework in tunnel-induced road damage

Table 1 Description of root nodes and their states in RDBN

Root node Description States
Low (L) Normal (N) High (H)
X1 Soil cohesion <18 Mpa 18–20 Mpa >20 Mpa
X2 Friction angle <13° 13–15° >15°
X3 Compression modulus <2 Mpa 2–3 Mpa >3 Mpa
X4 Soil density <1.5 g/cm3 1.5–2 g/cm3 >2 g/cm3
X5 Cover depth <10 m 10–14 m >12 m
X6 Torque ratio <12 KN/mm 12–15 KN/mm >15 KN/mm
X7 Cut slurry pressure <0.2 Mpa 0.2–0.3 Mpa >0.3 Bar
X8 Drilling speed <30 mm/min 30–60 mm/min >60 mm/min
X9 Thrust ratio <650 KN/mm 650–1130 KN/mm >1130 KN/mm
X 10 Grouting speed <15 mm/min 15–30 mm/min >30 mm/min
Table 2 Transition probability table for the leaf node R over time
R(t − 1) R(t)
I II III IV
I 0.60 0.20 0.10 0.10
II 0.05 0.60 0.25 0.10
III 0.05 0.10 0.60 0.25
IV 0.05 0.05 0.10 0.80
Fig. 9 Established model of

RSDBN in tunnel X3
construction X2 X4
X1 X5
Legend
B1
Root node
R 1
Intermediate node
B2
Leaf node
X6 X10
X7 X9
X8
5.3 Predictive Analysis and Control
Predictive control aims to use the predictive Bayesian inference technique to predict
the probability distribution of the occurrence of tunnel-induced road damage before
an accident occurs. Some reasonable control measures can be implemented ahead
of time for risk mitigation. In construction practice, many accidents occur when the
risk grows to be extremely high, and no measures can prevent the occurrence of the
accident.
According to the established RSDBN (see Fig. 9), we are able to calculate the
probability of a specific risk, as well as the tendency of the risk evolution over time.
We first enter the evidence information of risk factors (X i ) (i = 1, 2, …, 10) into
RSDBN within 18 time slices. Using the predictive analysis technique in the DBN
inference, the outcome of the risk levels of R is presented in Fig. 10. Typically, colors
are used to represent the degree of the risk. The color of Green, Blue, Orange, and
Red is in correspondence with the risk level of I, II, III, and IV, respectively. As seen
in Fig. 10, the Level I (Very Safe) tends to decrease rapidly at the beginning, while
the Level IV (Very Dangerous) tends to rise quickly. The tendency indicates that the
potential safety risk is growing fast in the initial stage. At the 9th slice, the potential
safety risk jumps up to the Level III (Dangerous), since the predicted probability of
the Level III is higher than that of other levels then. In this situation, early-warning
signals should be released to decision-makers, and the risk magnitude tends to be
smooth and steady afterward.
Furthermore, the interface displaying the risk evolution process as shown in Fig. 10
can be used as a dashboard to check and track the status of the potential safety risk.
0.4
I (Very safe)
II (Safe)
III (Dangerous)
IV (Very dangerous)
0.3
Predicted probabilities
0.2
0.1
0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Time slices
Fig. 10 Risk evolution tendency of tunnel-induced road damage over time

Figure 11 illustrates a dashboard for checking and tracking the status of the potential
safety risk in different scenarios. To be specific, the Scenario A represents the safety
status of R in the 1st time slice, where the predicted probability of the level of I is
calculated to be 0.296 and higher than that of the other three levels; the Scenario B
represents the safety status of R in the 4th time slice, where the potential risk suddenly
increases to the level of II; the Scenario C (or D) represents the safety status of R in
the 9th (or 11th) time slice, where the potential risk suddenly reaches up to the level
of III (or IV). As a result, the dynamics in the safety risk evolution over time can
be revealed, which can provide a visualized platform for decision makers to update
corresponding control measures in real-time.
1 0 1 0
0 1
0 1
0 1
0 1
90 90
0 1 0 1
(a) (b)
1 0 1 0
0 1
0 1
0 1
0 1
90 90
0 1 0 1
(c) (d)
Fig. 11 Dashboard for checking and tracking the status of the potential safety risk in different
scenarios: a Scenario A; b Scenario B; c Scenario C; d Scenario D
5.4 Sensitivity Analysis and Control
Sensitivity analysis and control aims to use the sensitivity analysis technique of the
Bayesian inference to identify critical and sensitive factors to the occurrence of risk-
prone events. As a matter of fact, the values of some geological and design variables
are fixed and can not change at the construction stage. Thus, engineers turn to give
much more attention to mechanical variables which have the flexibility to be adjusted
for ensuring construction safety. The most sensitive mechanical variables can be
determined as key checking points during the shield-driven tunneling excavation.
Under different excavation sections, the values of the input variables are varied
under given evidence. This results in the variation of the sensitivity of the involved
variables. The actual observations of influential variables X i (i = 1, 2, …, 10) are
entered into the model as the observed evidence. Given the risk magnitude of the
target event under different risk levels, Eq. (9) is used to calculate the performance
sensitivity of mechanical variables X i (i = 6, 7, …, 10). Figure 12 illustrates the
results of sensitivity analysis for mechanical parameters in different risk levels, with
the normalization process being adopted.
As shown in Fig. 12, under the risk level of I and II (indicating the low safety risk
level), X 6 (Torque Ratio) and X 7 (Cut Slurry Pressure) are identified as top sensitive
factors, since PSM (X 6 ) and PSM (X 7 ) are higher than that of other variables in such
conditions. Under risk level of III and IV (indicating the high safety risk level), X 8
(Drilling Speed), X 9 (Thrust Ratio), and X 10 (Grouting Speed) are identified as top
sensitive factors, since they are at the top of the ranking list regarding the performance
sensitivity. In general, the key checking points should be updated given the target
0.45
R=I
0.40 R = II
R = III
Sensitivity Performance Measure (SPM)
R = IV
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
X6 X7 X8 X9 X10
Fig. 12 Results of sensitivity analysis for mechanical parameters under different risk levels
event is observed at different risk levels. Accordingly, the major focus for safety
management strategies should be updated over time during the tunneling excavation
process.
5.5 Diagnostic Analysis and Control
Diagnostic control aims to use the diagnostic analysis technique in the DBN inference
to determine the most possible direct causes which have greater impacts on the
occurrence of risk-prone events. In case of an accident, the posterior probability of
each risk factor can be obtained, and the most suspected causes contributing to the
accident can be detected. More attention can then be paid to those suspected causes
in actual fault diagnosis.
Indeed, the values of some mechanical variables should be updated in accor-
dance with the changing geological conditions during the tunneling excavation
process. However, how to update these mechanical variables is not known exactly.
For instance, during the construction of WYMT, the tunnel-induced ground settle-
ment was observed to be 40 mm, indicating that the potential safety risk of R was
rated at a level of III (Dangerous). It is, therefore, necessary to reduce the risk and
perform the fault diagnosis. In this situation, the determined geological and design
variables were entered into the model as given evidence, and Eq. (10) was used to
obtain posterior probabilities of mechanical variables in the case of P(T = III) = 1.
Figure 13 illustrates comparisons of posterior probabilities at different states among
0.5 P(Xi=N|R=III)-P(Xi=L|R=III)
P(Xi=N|R=III)-P(Xi=H|R=III)
0.4
0.3
0.2
Posterior probabilities
0.1
0.0
X6 X7 X8 X9 X10
-0.1
-0.2
-0.3
-0.4
Suspected
Suspected causes
causes
-0.5
Fig. 13 Comparisons of posterior probabilities at different states among mechanical variables in

fault diagnosis
mechanical variables, where P(X i = N|R = III), P(X i = L|R = III), and P(X i = H|R
= III) referred to the posterior probability of the variable X i (i = 5, 6,.., 10) at the
state of Normal, Low and High, respectively.
When the result of P(X i = N|R = III) − P(X i = L|R = III) (or P(X i = N|R = III) −
P(X i = H|R = III)) is below zero, the variable X i can be identified as the suspected
variable, which is required to lower its value (or increase its value). Results as shown
in Fig. 13 indicate that X 8 (Drilling Speed) and X 9 (Thrust Ratio) were deemed as
the suspected variables which should be checked and updated in the case of P(T =
III) = 1. For this reason, the practical fault diagnosis and checking should then focus
on the reliability of these two factors and their reasonable parameter values until the
high risk was under control. That is to say, the posterior probabilities can provide
new evidential information for fault diagnosis in real time, and the values of relevant
variables can be updated in a dynamic manner. In this way, the suggestions on which
factor should be checked and how to adjust the suspected factor can be provided on
a quantitative basis.
6 Summaries
Underground tunnel transport has become a popular solution to relieve the pressure of
ground transportation systems worldwide, especially in developing countries such as
China. However, the tunnel-induced road damage has grown to be a public concern
due to its close relation to public safety in recent years. A DBN-based dynamic
risk analysis of the tunnel-induced road damage is developed in this paper. A case
study concerning the dynamic safety analysis of tunnel-induced road damage in the
construction of the Wuhan Yangtze Metro Tunnel in China is presented. Results
demonstrate the feasibility of the proposed approach.
In lights of the case study, some conclusions are drawn up. DBN provides a
powerful tool for knowledge representation and reasoning in a dynamic and complex
project environment. The proposed DBN-based approach can accurately illustrate
the dynamics of geological, design, and mechanical variables as the construction
progress evolves. This overcomes the deficiencies of traditional static fault anal-
ysis methods, such as FTA, CFEM, SCL, and influence diagrams. The proposed
DBN-based approach can perform predictive, sensitivity, and diagnostic analysis
and control on a quantitative basis. The critical and sensitive risks and risk factors
can be determined, and some corresponding safety control measures can then be
proposed in advance for risk reduction before an accident occurs. Moreover, the
suspected variables can be detected in real time in case an accident occurs, assisting
engineers in carrying out the fault diagnosis in a dynamic manner.
To summarize, the proposed DBN-based decision analysis approach can present
the causal relationships with time-dependent variables, as well as perform predic-
tive, sensitivity, and diagnosis analysis. The changes over nodes of DBN can reveal
the evolution in the construction progress, resulting in higher prediction accuracy.
The capability of DBN to model dynamics and uncertainty in the system enables
the proposed approach to be applied for decision support analysis in complex

infrastructure projects.
References
1. Alaeddini A, Dogan I (2011) Using Bayesian networks for root cause analysis in statistical
process control. Expert Syst Appl 38:11230–11243
2. Cawley GC, Talbot NLC (2007) Preventing over-fitting during model selection via bayesian
regularisation of the hyper-parameters. J Mach Learn Res 8:841–861
3. Cheng CY, Dasari GR, Chow YK, Leung CF (2007) Finite element analysis of tunnel-soil-pile
interaction using displacement controlled model. Tunn Undergr Space Technol 22:450–466
4. Cooper GF, Herskovits E (1992) A Bayesian method for the induction of probabilistic networks
from data. Mach Learn 9:309–347
5. Franco VH, Gitirana GdFN, de Assis AP (2019) Probabilistic assessment of tunneling-induced
building damage. Comput Geotech 113:
6. Hu J, Zhang L, Ma L, Liang W (2011) An integrated safety prognosis model for complex system
based on dynamic Bayesian network and ant colony algorithm. Expert Syst Appl 38:1431–1446
7. Hu Q-F, Huang H, Cheng L (2005) The risk evaluation on effect of shield tunnel construction
on urban asphalt pavement. Underground Space 2:263–267
8. Khakzad N (2019) Modeling wildfire spread in wildland-industrial interfaces using dynamic
Bayesian network. Reliab Eng Syst Saf 189:165–176
9. Khakzad N, Khan F, Amyotte P (2011) Safety analysis in process facilities: comparison of
fault tree and Bayesian network approaches. Reliab Eng Syst Saf 96:925–932
10. Langseth H, Portinale L (2007) Bayesian networks in reliability. Reliab Eng Syst Saf 92:92–108
11. Liao SM, Liu JH, Wang RL, Li ZM (2009) Shield tunneling and environment protection in
Shanghai soft ground. Tunn Undergr Space Technol 24:454–465
12. Liu W, Wu X, Zhang L, Zheng J, Teng J (2017) Global sensitivity analysis of tunnel-induced
building movements by a precise metamodel. J Comput Civil Eng 31:04017037
13. Montani S, Portinale L, Bobbio A, Codetta-Raiteri D (2008) RADYBAN: a tool for reliability
analysis of dynamic fault trees through conversion into dynamic Bayesian networks. Reliab
Eng Syst Saf 93:922–932
14. Onisko A, Druzdzel MJ, Wasyluk H (2001) Learning Bayesian network parameters from small
data sets: application of Noisy-OR gates. Int J Approximate Reasoning 27:165–182
15. Peck RB (1969) Deep excavations and tunnelling in soft ground. In: Proceedings of 7th
International Conference on SMFE, pp 225–290
16. Smith NJ, Merna T, Jobling P (2009) Managing risk: in construction projects. Wiley-Blackwell
17. Špačková O, Straub D (2013) Dynamic Bayesian network for probabilistic modeling of tunnel
excavation processes. Comput-Aided Civil Infrastruct Eng 28:1–21
18. Thomas T (2018) Perth TBMs on hold after sinkhole appears. Tunnel J
19. Wu X, Liu H, Zhang L, Skibniewski MJ, Deng Q, Teng J (2015) A dynamic Bayesian network
based approach to safety decision support in tunnel construction. Reliab Eng Syst Saf 134:157–
168
20. Yang J, Zhang X, Suo H, Lu L, Zhang J, Yan Y (2011) Maximum a posteriori linear regression
for language recognition. Exp Syst Appl 4287–4291
21. Yao G, Sun H, Wei Q (2008) Research on the damaging control criterion of blacktop due to
underground tunnel construction. J Beijing Jiaotong Univ 32:40–43
22. Zhang L, Wu X, Qin Y, Skibniewski MJ, Liu W (2016) Towards a Fuzzy Bayesian network based
approach for safety risk analysis of tunnel-induced pipeline damage. Risk Anal 36:278–301
23. Zhang Y, Zhang L, Wu X (2019) Hybrid BN approach to analyzing risk in tunnel-induced
bridge damage. J Perform Construct Facil 33:04019048
Process Mining
1 Introduction
BIM, as a new digital revolution for civil engineering, provides a collaborative plat-
form to facilitate information exchange and sharing among participants in different
roles for better decision-making [31]. The Architecture, Engineering, and Construc-
tion (AEC) industry has seen a rapid growth of BIM implementation. That is to say,
around 28% of projects use BIM in 2007 and the proportion increases to 71% in
2012 [3]. Notably, it is unreasonable to simply consider BIM as a three-dimensional
(3D) parametric model. Instead, BIM acts as a digital representation of building
characteristics, which allows users to gain quick and easy access to model informa-
tion. BIM can bring benefits more than the improved visualization due to the easy
retrieval of information, such as to enhance productivity, facilitate collaboration, and
reduce mistakes, conflicts, costs, and others. With the growing applications in BIM,
a huge amount of data can be generated and stored in the BIM platform to passively
track detailed activities in the modeling and construction process, leading to some
features of “big data” [16]. Therefore, more focus should be put on deeply mining
such large volumes of BIM data, in order to open a new way for better understanding
the project.
However, knowledge discovery from BIM data is not an easy task. In this regard,
the emerging data mining methodology called process mining can be implemented
as an optional solution to realize the data-driven decision support in terms of
work scheduling, task assignment, and productivity improvement for smart project
management. This can expectedly reduce the project’s risk of failure and improve
project performance.
It should be noted that the BIM event log associated with cases and events is
an important type of BIM data, which contains rich process-specific information
about project evolution [15]. They are shared and updated across the BIM platform
among users for information integration and interoperability. To logically organize
the event log data, a process is defined to consist of a collection of cases, where a case
148 Process Mining
is composed of ordered events with typical attributes, such as timestamp, activity,

actor, and others [23]. In other words, an event log is a collection of cases and events
with typical attributes. Regarding the practical value of event logs, they are initially
proposed to assist diagnosis and troubleshooting for software support staff, which
have recently been studied for other purposes.
BIM event logs are very similar to the web logs defined as a detailed and objec-
tive recording of all operations and events in an information system chronologically.
Currently, there are many pieces of research about the web log mining. They demon-
strate promising capabilities in identifying usage patterns and comprehending user
behaviors. They are useful in developing web intelligence, such as the web content
personalization, recommendation systems, and others. Similarly, BIM event logs also
deserve full exploration, where the valuable implicit information is extracted from
logs, such as compute-user interactions, productivity, and collaborative characteris-
tics. The captured information can serve as evidence to optimize the construction
project in a data-driven manner and enhance the likelihood of project success.
A successful project does not emerge by accident, but rather as a result of a
managed process that involves several participants from different disciplines and
backgrounds. It is found that poor management of the design process is the primary
cause of rework in construction projects. This problem is rooted in the lack of under-
standing of the complexities of the design process and how it influences design
productivity. Since event logs are a prime source of large data to document the
whole process, log mining can be an effective means to track and understand the
actual processes. To make the most of the event logs, an emerging and powerful AI
method, process mining, which specializes in retrieving hidden knowledge from
event logs for exploring, monitoring, and improving the real processes, can be
performed [13, 14, 29, 30].
The growing interest in process mining comes from two aspects. For one thing,
more and more event logs are generated automatically to provide sufficient and
detailed data about the historical process. For another, it can offer operational insights
into the complex process for continuous monitoring and improvement, which acts
as an ideal solution to address both efficiency and risk-related problems. That is
to say, the goal of process mining is to perform process-centric analysis, such as
process discovery, conformance checking, and model enhancement [25]. Process
discovery uses event logs as inputs to generate a process model without any prior
information. Conformance checking aims to judge the quality of the discovered
model by detecting the discrepancy between the reality recorded in event logs and
the modeled behavior. Process enhancement helps improve the process model based
on the actual process model in event logs, such as modifying the model to make it
more conform with reality. To deal with the large volumes of event logs systematically
generated by BIM, process mining provides an ideal solution which can narrow the
gap between data science and construction project management. It is expected to yield
tactic decisions to alleviate poor construction performance and facilitate cohesive
collaboration, leading to a high-efficient and high-quality construction process.
1 Introduction 149
In this chapter, several process mining techniques are conducted to uncover knowl-
edge from available BIM event logs in large volumes, aiming to offer a better under-
standing of the dynamics in the design and construction process. It is supposed to
create long-lasting positive impacts on the construction industry that has traditionally
suffered from chronic productivity problems. The main objective is to answer the
following questions: (1) How to automatically capture the behavior from event logs
and construct an understandable process model to describe this observed behavior;
(2) How to analyze the process model for checking compliance, detecting deviations,
predicting delays; (3) How to explore collaborative characteristics to reveal the flow
of work, information, and communication in the organization. Based on the results
obtained from process mining, project managers can easily identify the bottleneck,
collaboration characteristic, and performance preference relying on such valuable
results from process analysis. Insights can are provided to help managers arrange
and adjust workflows more efficiently and rationally, in order to strategically smooth
and speed up the operation process along with the increasing collaboration cohesion
in BIM projects.
2 Process Mining
As an emerging data mining method, process mining leverages various algorithms

and mathematical models for the visualization and investigation of event logs, which
assists in discovering, evaluating, and improving the existing process. It can be
regarded as a connection between event logs and the operational process, resulting
in objective, transparent, and fact-based insights from actual event logs. That is to
say, event logs can be learned to automatically create process models as a reflection
of the actual process and calculate process metrics. Then, a wide range of analytical
methods can be carried out in the obtained process model to detect possible problems
(i.e., inefficiencies, bottlenecks, and other weaknesses) and capture characteristics
of the organization in the process, which allow managers to quickly point at the key
parts of the process and inform data-driven decisions for strengthening operations
and accelerating the process. According to a recent survey, the top benefits of process
mining techniques are associated with objectivity, accuracy, speed, and transparency
[1]. To sum up, research topics with process mining can be grouped into three major
categories [5, 23], which are illustrated below:
(1) Process discovery
Process discovery relies on proper algorithms (i.e., α-algorithm, heuristic mining,
genetic process mining, region-based mining, and inductive mining, and others)
to automatically build process models based on the retrieved information from
event logs. The created model should meet the requirement of acceptable accuracy,
balancing recall, precision, and generalization along with minimal complexity [4].
Notably, the process model is able to deliver a high-quality model containing both
visible and invisible processes, which can answer the question on how the real-life
150 Process Mining
process is executed. That is to say, the process discovery aims to provide a fast, reli-
able, and cost-effective way for managers to identify large-scale processes with less
manual efforts, which serves as a basis for further analysis. Besides, animations of
the discovered process model can be produced to simulate the execution of events.
(2) Process conformance
Process conformance compares the modeled behavior in the process model and the
observed behavior in the event logs to check whether they are compliant with each
other. It mainly focuses on measuring the quality of the discovered process model
under four dominant dimensions, namely the fitness, precision, generalization, and
simplicity [26]. Regarding the application benefits, it is used for business alignment
and auditing, token replay, and footprints comparison, which assists in identifying
differences between a process model and the actual process in event logs. The detected
deviation can be explained as the undesired exception on the process model, which
does not conform to reality. It can serve as evidence to refine or improve the model.
(3) Process enhancement
Process enhancement is primarily responsible to extend or improve the existing data-
driven process model based on event logs for optimization potential. For instance,
when deviations are detected from the conformance checking, corresponding modi-
fications can be executed on the process model to make it replay most of the event
log. To enrich the process model described by the control flow, a more new perspec-
tive, such as the organizational, case, and time perspective can be added in by cross-
correlating the model with event logs. More specifically, these three new perspectives
concern the organizational structure, properties of cases, and timing and frequency
of events, respectively. Moreover, predictive approaches can also work together with
the process mining for various purposes, such as to cluster traces, identify the root
cause, and predict the activities. They can formulate suggestions to improve process
performance and mitigate risks of bottlenecks in advance.
Many software products for process mining are available to efficiently convert
event logs into process-related views and deliver insightful analytics. The following
software is commonly used for research and practice, including ProM, Disco (Flux-
icon), Celonis, ARIS Process Mining, Myivenio, and others. They aim to clearly
describe the step-by-step process by creating a visual map and then realize the func-
tions of diagnosis, checking, exploration, prediction, recommendation, and others.
At the list of tools, ProM and Disco, which are representative of the non-commercial
and commercial tools, respectively, attract more attention than other tools. ProM [27]
is the leading open-source framework to support process mining, which is easy to
extend. A variety of plug-ins (more than 1500) can be added to the ProM framework,
enabling users to implement the newest methods on their own data. But it could
be unfriendly for beginners. Disco [7], as the commercial process mining software,
is easy-to-use and fast for automated process discovery, process map animation,
deviation controlling, and performance optimization.
It should be noted that process mining is not merely a theoretical subject, which
has played important roles in a wide range of industries. Process mining has been
2 Process Mining 151
widely used in the healthcare industry which is characterized by great variability,

complexity, security, privacy, and multidisciplinary [23]. Process mining can make
the best use of data gathered in the hospital information systems to evaluate the clin-
ical pathways, track the medical protocols, check resource utilization, and others,
which helps clinicians to provide better services along with more promising predic-
tions and recommendations [21]. In addition, some other applications are reported in
the domains of the business, education, manufacturing, and logistics, which offer
higher flexibility on complex process management to uncover undesired behav-
iors, shorten the waiting and service time, promote collaboration, and others. Due
to the practicability of process mining, it is expected to extend it to the Architec-
ture/Engineering/Construction (AEC) and Facilities Management (FM) industry as
a powerful tool for knowledge discovery and decision making.
3 Event Logs in BIM
The starting point of process mining is the event logs, which map the complicated
process holistically from beginning to end. Typically, an event log is a list of cases,
where a case is a related set of ordered events identified by the unique ID, and thus,
it can be adopted to differentiate different patterns of activities. Some additional
information is added to the event, such as the timestamp to indicate the sequence of
events, the activities, and the resources to represent the participants or other assets
involved in the execution of an activity. Since the BIM event log also belongs to
the log data, it can be defined as a rich source of process-specific information to
capture flows of activities in chronological order, which contain several attributes,
including cases, activities, persons, and time [28]. Herein, the event logs from several
users over the course of working on multiple projects are taken as an example. They
can be generated passively by the Autodesk Revit to record all modeling activities
of designers in executing commands, which are then stored in the Program Files
directory under the Revit Product version folder called Journals [20]. To make good
use of the Revit logs in plain texts, it is necessary to extract several information items
about design activities from the event logs. Figure 1 displays a workflow of the data
extraction process from BIM design event logs, which contains three main steps,
namely data harvesting, data parsing, and data cleaning.
(1) Data Harvesting: In the BIM-based design project, a group of designers work
together on the BIM Revit platform. Each designer generates thousands of
access logs within Revit throughout the whole project. Several projects run
in parallel within a design firm, leading to the generation of a large number
of journal files. To streamline design logs across different projects from
numerous designers, a cloud-based database is developed to regularly harvest
the distributed log files onto the Server Centre via the firm’s Intranet.
(2) Data Parsing: BIM Revit logs are stored in a text format, which makes it chal-
lenging for data analysis. A journal file parser is developed to automatically
152 Process Mining
140901_atlcn0102_John_journal.0001.txt
1. Data
Journal 140901_atlcn0102_John_journal.0002.txt
Harvesting
files 140901_atlcn0102_Mike_journal.0003.txt
created 140901_atlcn0102_Mike_journal.0004.txt Server Logs
by Revit 140901_atlcn0102_Mary_journal.0005.txt 2. Data
140901_atlcn0102_Pete_journal.0006.txt Parsing
User ID
Open
Jrn.Directive "WindowSize" _
, "..\800654_UNT_Student_Union.rvt]", CSV Data File
"Floor Plan: Level 1" _
, 1580, 733 3. Data
Timestamp 'E 03-Oct-2014 14:51:18.036; 0:< Cleaning
Jrn.MouseMove 0 , 28 , 4
Design 'E 03-Oct-2014 14:51:18.036; 0:<
Command Jrn.Command "Ribbon" , "Create a wall ,
ID_OBJECTS_WALL" Cleaned Logs
Fig. 1 Workflow of the data extraction from BIM event logs
parse through a folder of journal files and read Revit logs [19]. Several infor-
mation items, such as user ID, project ID, timestamp, and design command,
are pulled out and stored in a Comma Separated Values (CSV) format.
(3) Data Cleaning: In the parsed CSV file, each line represents a design command
that is executed by a designer at a time point. A number of lines are involved
with noise, such as empty or meaningless commands, which creates a need
for removing noisy commands. For instance, in Lines 5 and 12 in Table 1,
Table 1 Example of records in the parsed CSV from BIM design event logs
Line User ID Timestamp Project ID Design command
1 John 4/1/2014 10:16:55 800,645 “Create a wall”
2 John 4/1/2014 16:26:51 801,320 “Pick lines”
3 John 4/2/2014 10:36:45 801,344 “Finish Sketch”
4 John 4/2/2014 11:12:25 801,344 “Create a wall”
5 John 4/2/2014 11:16:25 801,344 “Cancel the current operation”
6 Mike 4/8/2014 09:10:15 801,344 “Pick lines”
7 Mike 4/8/2014 10:01:05 800,645 “Split walls and lines”
8 Mike 4/9/2014 13:21:25 800,645 “Manage Links”
9 Mike 4/9/2014 14:21:11 801,320 “Move selected objects or their
copies”
10 Mary 4/9/2014 08:25:31 800,645 “Finish Sketch”
11 Mary 4/9/2014 12:28:13 801,320 “Create a wall”
12 Mary 4/9/2014 14:22:31 801,320 “Cancel the current operation”
3 Event Logs in BIM 153
the command “Cancel the current operation” shows that the designer presses
the “ESC” key while moving the mouse or in the middle of the execution of a
command, which has no specific meaning for the design and should be removed
from further analysis. Command lines are continuously evaluated, and those
containing valid commands without any null values or errors will be kept as
new data and are placed into the clean logs file.
4 Process Mining Framework
This section attempts to use tremendous volumes of event logs as a non-intrusive

data collection approach to capture human–computer and human–human interac-
tions. It opens a new way to delve into workflows, collaboration patterns, and perfor-
mance evaluation in project management. Therefore, the discovered knowledge from
event logs can guide managers to quickly identify root causes of poor performance
and immediately make corresponding responses for constant process improvement.
Figure 2 illustrates the overall framework of the proposed approach. It is clear that
the in-depth process analysis embedded in the event logs is conducted in terms of
the process model, social network, and decision tree.
Motivation Research Activities Knowledge Discovery
Worker1 217
217 217
Designer 1 Log File 1 211 2
165 212
5 212
Project 1 8 1
13
1
1
2 115
114
13
Worker4
12
13
2
8
44 Process model
Worker7 115
Designer 2 Log File 2 Intranet 115
13
Project .. 182
180
182
17
2
• Process discovery
Server Log Data
Designer .. Log File ..
Worker3
182
• Conformance checking
(b) Process discovery and analysis • Bottleneck analysis
Project M 51 3 6
50 11
Designer N Log File N 49

9
5
13
48 17
Social network
47 19
46 20
15
140901_atlcn0102_John_journal.0001.txt 45
4
21
22 8
Journal 140901_atlcn0102_John_journal.0002.txt 0
44
219 438 657 876 1095 1315 1534 1753 1972 2

23
• Macro-level analysis
43 27 10 24
files 140901_atlcn0102_Mike_journal.0003.txt 42 25
created 140901_atlcn0102_Mike_journal.0004.txt 41 26
by Revit 140901_atlcn0102_Mary_journal.0005.txt
40
39
16 14
29
28
• Micro level analysis

38 30
18
140901_atlcn0102_Pete_journal.0006.txt 37
36
35
12
34 33
32
31
• Correlation analysis
User ID
(c) Social Network Analysis
Jrn.Directive "WindowSize" _
, "..\800654_UNT_Student_Union.rvt]", Decision tree
"Floor Plan: Level 1" _
, 1580, 733
Timestamp 'E 03-Oct-2014 14:51:18.036; 0:<
• Performance
Jrn.MouseMove 0 , 28 , 4 prediction
Design 'E 03-Oct-2014 14:51:18.036; 0:<
Command Jrn.Command "Ribbon" , "Create a wall ,
ID_OBJECTS_WALL"
(a) Wealth of Design Log Files (d) Decision tree classification
Fig. 2 Framework of the process mining based on event logs

154 Process Mining
4.1 Process Discovery
The first task in process mining is process discovery, which extracts knowledge from
event logs to build rational process models. The discovered model demonstrates
the way the process actually works, which is taken as the basis for further process
analysis and optimization. One of the simple and fundamental algorithms is called
the α-algorithm, which can adopt relations, such as direct succession, causality,
parallel, and others to output the Petri net along with animations. Although the α-
algorithm is sufficient to catch the concurrency in the process, it is unable to deal
with problems arising from the unstructured process, including noise, incomplete
information, non-local dependencies, short loops, and others. In the meantime, the
α-algorithm tends to generate useless spaghetti-like models containing complete
processes with all details, which is hard to distinguish important and non-important
process information in noisy and less-structured logs. To overcome these limitations,
some advanced algorithms for process discovery have been developed. Herein, two
popular algorithms, namely fuzzy mining and inductive mining are introduced.
(1) Fuzzy mining. For model simplification and interpretability, fuzzy mining [6,
8] is proposed to display suitable abstractions or aggregations of the observed
process graphically using a map metaphor. It pays attention to the subsets of
the most significant behavior within the process, which is especially effective
in visualizing correlations of activities in a less-structured process. The funda-
mental idea of fuzzy mining lies in configuring two metrics named significance
and correlation. Significance representing the relative importance of behavior
can be easily quantified by frequency of events and routings, while correlation
estimates the closeness degree between two events. An efficient method for
simplifying the process model is to only preserve the high-level information.
Therefore, it is necessary to delete undesirable events and relations with both
low significance and correlation. As for the less significant but highly corre-
lated behavior, they need to be aggregated into clusters. From the map-like view
of abstract process models, primitive and cluster nodes are linked by edges in
different widths and colors representing relative significance and correlation
after conflict resolution and edge filtering. Besides, fuzzy mining has taken
effect in some practical applications, especially for interactively simplifying
models and investigating frequency and time duration [9, 10, 18]. However,
fuzzy mining is prone to suffer from unfitness and unsoundness due to its
deliberately imprecise model.
(2) Inductive mining: As an improvement over α-algorithm and fuzzy mining, a
promising technique termed inductive mining [11] is developed to tackle infre-
quent behavior and huge models. It belongs to a variation of the α-algorithm,
which can create the block-structured process in great fidelity, soundness, and
fitness. The implementation of inductive mining starts by recursively splitting
original event logs into sub logs. Then, each sub log is used to construct a
directly-follow graph which is defined by a set of activities by nodes and their
execution sequences by directed edges. As a result, the directly-follow graph
4 Process Mining Framework 155
is cut into smaller ones to partition the set of nodes into distinct subsets under
four kinds of cuts: exclusive choice, sequential composition, parallel compo-
sition, and the re-do loop. Overall, the splitting procedure will repeat until
every subset is only comprised of one node (activity). It has been proved that
inductive mining is flexible in creating a reasonable process model with no
duplicated activities, which can be fit and sound to the observed behavior in
the event log.
A process model serves as a concise description of the complicated process based
on the facts from event logs. It is typically in the graphical format to record the
workflow along with the related process knowledge, which helps people to better
understand the execution sequences and dependencies in a series of activities. To
represent the discovered process model, there are three common types, namely Petri
net, business process modeling notation (BPMN), and process tree. They can provide
different views of the discovered process in reality.
(1) Petri net: Petri net originally developed in the late 1960s is one of the oldest and
the most basic process modeling languages under the combination of mathe-
matical formalism and a graphical representation [17]. It is a bipartite graph
consisting of place, transition, arc, and token. To be specific, places and tran-
sitions are connected by a collection of directed arcs. Tokens flow through
the network, which are hold by places. The superiority of the Petri net lies
in its ability in demonstrating both the concurrency and asynchrony nature of
processes. Figure 3a shows a simple example of the Petri net model, where
circles represent the places, squares denote the transitions, and arcs indicate
various associations.
(2) BPMN. BPMN is a flow chart method widely used in business process manage-
ment. Two kinds of notations are incorporated to visually describe a detailed
sequence of business activities and information flows. One is a kind of activity
nodes on behalf of business events, which is comparable to a place in the
Petri net. The other is a kind of control nodes, where sequential and message
flows are used to indicate the flows and logic between activities. BPMN is
analogous to the Petri net to some extent, but it provides a more comprehen-
sive set of elements to express the flow behavior. That is to say, BPMN is a
high-level notation to offer an intuitive visual representation of the compli-
cated processes, which can be easier to understand even for people with no
professional knowledge. The BPMN in Fig. 3b has a similar meaning as the
Petri net in Fig. 3a.
(3) Process tree. Since the Petri net is likely to experience deadlocks and some
anomalies, the process tree can be another optional graph notation to ensure
the soundness of representations, which is especially beneficial to the inductive
process discovery. To be specific, a process tree is a block-structured workflow
net in a hierarchical structure, where the inner nodes refer to operators and the
leaves are labeled with activities. The Petri net in Fig. 3a is convertible to the
process tree in Fig. 3c.
156 Process Mining
XOR Split t2 XOR Join

AND Split
p2 p5 AND Join
t3
p1 t1 t6 p7
p3 t4 t5 p6
(a)
XOR-split Gateway
t2 XOR-join Gateway
AND-split Gateway AND-join gateway
t1 t3 t6
t4 t5
(b)
AND
XOR SEQ
t2 t3 t4 t5
(c)
Fig. 3 Examples of: a Petri nets; b BPMN; c process tree. Note AND means parallel composition,
XOR means exclusive choice, and SEQ means sequential composition
4.2 Process Diagnosis
Process diagnosis aims to check the conformance and determine the discrepancies
between the observed process in event logs and the discovered process model. It is
closely relevant to the quality evaluation of the discovered model. Herein, two major
quality dimensions called fitness and precision are used to check the model quality,
as presented below [2].
(1) Fitness. Fitness is the most dominant metric to measure the model’s competence
in replaying the event log. It can be defined by Eq. (1) to clarify how much of
the observed behavior can be demonstrated in the process model. Based on an
alignment-based calculation, the fitness can be obtained by counting deviations
in replaying events on the top of the model [24]. During the process of aligning
events to the process model, the cost should be given when events are skipped
or activities are inserted with no expectation. The better fitness is, the value of
Qf is closer to 1 where the cases in the model can well replay the log with fewer
problems. On the contrary, the fitness of 0 signifies that the process model is
unable to replay traces in the log. For example, a process model in the fitness
of 0.95 means that it fails to explain 5% of the observed behavior in the event
log. It should be also noted that a process model with a comparatively high
fitness may be overfitting.
f cost (L , M)
Qf = 1− (1)
move L (L) + |L| × move M (M)
where fcost(L,M) represents the total alignment cost for event log L and model
M. For example, if fcost(L, M) = 0, it means that the model M can perfectly
replay the log L. For the denominator, it refers to the maximal possible cost,
where moveL (L) is the cost of moving through logs rather than the model,
and moveM (M) is the cost only in the model. It is used to normalize the total
alignment cost.
(2) Precision: Precision in Eq. (2) is the proportion of actual behavior in logs to
the behavior allowed by the process model, indicating how well the simple
process model with minimal behavior can describe the event log [24]. It can
be observed that a poor precision approaching 0 can be caused by |enL (e)|
|enM (e)|, possibly leading to underfitting. This would imply that behaviors in
the process model are quite different from the event log. When almost all of
the behavior in the process model can be actually seen in the log, it returns
a high precision approaching the value of 1. That is to say, the lack of great
fitness or precision is likely to cause an oversimplified process model.
1 |en L (e)|
Qp = (2)
|E| e∈E |en M (e)|
where |enM (e)| represents the number of activities enabled in the model M,
|enL (e)| refers to the number of actual activities executed in event logs L under
the similar context, e ∈ E is events, and |E| is the number of events in logs L.
4.3 Process Prediction
As a tree-based supervised learning algorithm, the decision tree model provides an

intuitive way to focus more on the process instead of the result, which can easily
generate understandable decision-making rules for classifying objects. The decision
tree in the flowchart-like structure is adaptive to make predictions with great accuracy,
stability, and interpretability. For instance, if a person tries to decide whether he
should go out based on climate information, a decision tree diagram in Fig. 4 can be
built when information about this person’s past activities is learned by the algorithm.
It is drawn upside down, where the root is at the top and the internal node splits the
tree into branches.
Notably, the key feature of the decision tree lies in the splitting, which dives the
node into two or more sub-nodes to efficiently narrow the options among a lot of
158 Process Mining
Fig. 4 Example of a
decision tree Rainy ?
Windy ? Not Going out
Going out Not Going out
classes. The strategy of the node splitting heavily influences the predictive accuracy.
Herein, the widely used decision criterion named the Gini index is introduced, which
calculates the probability of a feature classified incorrectly when selected randomly.
As defined in Eq. (3), the Gini index can be derived from subtracting the sum of the
squared probability of each class from one, which is relatively easy to implement
even for large partitions. In particular, it is deployed in the classical CART algorithm
to create binary splits in a reasonable manner, which also performs well for nonlinear
datasets. The value of the Gini index is between 0 and 1, where 0 indicates the purity
of classification. Inversely, 1 means the data on the node is not pure at all, and thus,
several new nodes need to be generated from this node.
n
Gini = 1 − ( pi )2 (3)
i=1
where pi represents the probability of an element being classified for a distinct class.
4.4 Social Network Analysis
Since a group of people will involve in the construction project to finish the common
task together, their complicated collaboration can be reasonably mapped into a social
network for deeper analysis. A social network is a graphical model to intuitively
describe the interaction structure and characteristic, where nodes standing for people
who are connected by directed and weighted links. Besides, the network can be
mathematically expressed by a weighted sociogram (P, R, W) representing relations
among multiple performers, where P, R, W refer to performers, relations, and weights
of relations, respectively. In subsequence, social network analysis (SNA) can be
implemented on the built network to explore hidden knowledge about information
spreading and performers’ behavior from the following two levels:
(1) SNA at macro-level: SNA at the macro-level aims to analyze and measure
the characteristics of the entire social network established. Measurements at
the macro-level include the network size (S), density (D), average degree
(AD), mean distance (MD), diameter (ND), and clustering coefficient (CC),
as defined in Eqs. (4)–(8) [22]. Information can flow more easily in dense
networks compared to sparse networks, as there are multiple paths between
the nodes in the dense network.
E
D= (4)
N (N − 1)
E
AD = (5)
N

i= j d(n i , n j )
MD = (6)
N (N − 1)
N D = Maxni ,n j d(n i , n j ) (7)
N 2ei
CC = (8)
i=1 ki (ki − 1)
where E and N are the total number of ties and nodes in the network, respec-
tively, d(ni , nj ) is the shortest distance between the ith node and jth node, and
k i is the number of neighbors for the ith node.
(2) SNA at micro-level: SNA at the micro-level analysis aims to analyze and
measure characteristics of individuals at the established social network. It can
be conducted in two aspects. For one thing, the degree centrality C D can be used
to evaluate the position of a single node in the whole network. Equation (9)
defines the degree centrality for a certain node ni , which counts the number
of immediate links of an individual with others. A higher degree of centrality
indicates a higher importance of the node. For another thing, the Spearman
rank correlation can uncover the correlation between the degree centrality and
the production performance of individuals. It attempts to discover whether
individuals at the center of the interaction map are those most reachable for
others and assess the correlation between their centrality performance and
productivity. We will investigate whether the designers with high productivity
records in historical projects will gradually grow to be placed in core positions
in the working process.
j
1 ei, j
C D (n i ) = (9)
N −1
where j is the number of nodes linked to the node ni , and ei,j is the edge between
the ith node and the jth node.
160 Process Mining
5 Typical Applications in BIM Process Discovery
Some typical applications of process mining techniques are introduced in this section,
which can fully explore the BIM design event log from different perspectives. The
main objective is to facilitate a better understanding of the design process and to
enhance productivity in BIM-enabled collaborative practices.
5.1 Data Collection
A large dataset of event logs recording BIM Revit design applications throughout
several projects is collected from a major international design firm in the construction
industry. It is a Structured Query Language (SQL) database built in the design firm
for data sharing and storage. The design event logs are harvested and stored at the
Server Center on a weekly basis. The timeline of the design event logs collected is
from March 2013 to June 2014. The Journal File Parser is used to automatically read
the unstructured text files in order to capture the records of design commands. After
the data parsing process, 687,143 lines of design commands are generated and stored
in a CSV file, of which each line corresponds to the information of user ID, project
ID, timestamp, and the context of the command executed. By removing empty and
meaningless commands, 620,492 lines of valid commands are kept. The cleaned
dataset indicates that 51 designers are working together on 82 different projects.
Within the studied period, the production of designers varies significantly.
Figure 5a illustrates the scatter of the number of valid commands executed by different
90000 12
Y (Number of commands executed) 11.38 (max)
80000 11
Group I:
Y (Number of commands executed)
70000 High-producƟon
10 designers
60000 9.19
9
50000 Group II:
In (Y)
8 Median-producƟon
40000 In (Y) designers
Linear Fit of In (Y) 7.00
7
Equation y = a + b*x
30000 In (Y)
Plot
Weight No Weighting Group III:
6 Intercept 11.239 [0.128]
20000 Slope -0.119 [0.004] Low-producƟon
Residual Sum of Square 9.94769 designers
Pearson's r -0.96975
10000 5 R-Square(COD) 0.94042
Adj. R-Square 0.9392
4.80 (min)
0 4
0 4 8 12 16 20 24 28 32 36 40 44 48 52 0 4 8 12 16 20 24 28 32 36 40 44 48 52
Designer ID Designer ID
(a) (b)
Fig. 5 Classification of designers by accounting the number of valid commands executed: a scatter
of the number of commands executed by different designers; b three groups of designers (high-
production, median-production, and low- production designers). Note In Fig. 5b, the red dot line
is a linear fit of the scatter diagram, where the ordinate values are natural logarithms of ordinate
values (number of commands executed) as shown in Fig. 5a for purpose of a better classification
resolution
5 Typical Applications in BIM Process Discovery 161
designers, where all the 51 designers are represented by “#1”, “#2”, …, “#51”, respec-
tively. It is clear to see some designers, like “#1” ~ “#10”, contribute more than
20,000 valid commands, while some designers, like “#40” ~ “#51”, contribute less
than 2000 valid commands. To further investigate the correlation between network
characteristics and the production performance of designers, all the 51 designers are
classified into three separate groups for comparison study presented in later sections.
Figure 5b illustrates the results of the classified three groups of designers, namely
high-production, median-production, and low-production designers, where there are
17 members in each group.
5.2 BIM Process Discovery
Herein, we take a part of the prepared BIM design event log concerning seven workers
as an example. The targeted event log records how the tasks are transmitted among
the seven workers in detail, which is then fed into the inductive mining algorithm
for process discovery. To emphasize the critical relationships in the collaboration,
workers who are responsible for less than 2% of tasks are automatically removed
from the process. As a result, the process model can be visualized in three forms,
namely the Petri net, BPMN, and process tree. Workers 1, 3, 4, 7 are taken into
account in the discovered model, which work together to complete around 96.61%
of tasks. As shown in Fig. 6, all three process models can intuitively describe the
activities executed by workers from a “worker-centered” view, which are explained
below.
The Petri net in Fig. 6a is the basic process model, which has 10 places, 17 transi-
tions, and 34 arcs. More specifically, transitions in boxes represent workers to model
the possible process status. When tokens are input into the place, the corresponding
transition is activated. In this case, the important causal relationship between workers
is the XOR operation. Based on the XOR node in Fig. 6a, c and the XOR-split
gateway in Fig. 6b at the beginning of the process model, we can simply divide the
seven workers into three clusters. That is to say, the first cluster is made up of Workers
1 and 4, the second cluster contains Worker 7, and the remaining workers are in the
third cluster. It should be noted that the worker in a certain cluster is unable to take
over tasks executed by workers from other clusters.
5.3 BIM Process Diagnosis
Since the quality of the process model largely influences the reliability of analytical
results, it is an essential step to check how well the process model can reflect the
behavior captured in the event log. In this regard, the plug-in in the ProM tool for
a conformance checker is used to generate the Petri net annotated with diagnostic
information about the deviations between the log and the model, which is provided
162 Process Mining
Fig. 6 Discovered process models in a Petri net; b BPMN; c process tree

Fig. 7 Conformance checking in the process model discovered by inductive mining
in Fig. 7. To be more specific, the number in the box indicates how many tasks the
worker performs, and the number above the arc states how many times the process
traverses between participants. The red dash lines highlight the part in the process
model, which cannot conform with the event log. It is known that two types of
deviations exist in the process model [12]. The event, which appears on the event log
but does not reflect in the process, causes the log-move deviation. The event, which
is required by the model but does not occur in reality, is the model-move deviation.
It is clear to see that Fig. 7 only owns the log-move deviation, which is repre-
sented by the self-arc. The metrics of fitness and precision (expressed by Eqs. (1)
and (2), respectively) can be calculated to measure the discovered process model in
a numerical way. Table 2 outlines the evaluation results. Since no model-move devi-
ation appears in the model, the corresponding fitness reaches a high value of 1. The
log-move fitness is only 0.791, which is still acceptable for the model to demonstrate
a great re-discoverability property. In addition, all kinds of precision are greater than
0.85, indicating that there is no obvious underfitting in the model. Therefore, the
discovered process model from inductive mining is proven reasonable.
An effective method to ensure a smooth process is to shorten the unnecessary
waiting times as much as possible. Therefore, an important aspect of process analysis
can investigate the frequency and duration, in order to identify the key parts and
Table 2 Evaluation of the

Metric Value
process model discovered by
inductive mining Log-move fitness 0.791
Model-move fitness 1.000
Precision 0.868
Backwards precision 0.868
Balanced precision 0.868
164 Process Mining
potential bottlenecks within the process. The Disco Fluxicon tool based on fuzzy
mining is deployed for this goal. It can output two simplified process maps as shown
in Fig. 8, which are analyzed as follows.
Fig. 8 Simplified process map fuzzy mining focusing on: a absolute frequency; b median duration
(1) Absolute frequency view. The frequency in Fig. 8a states how often an activity
occurs. It is assumed that the higher frequency can indicate the greater impor-
tance of the worker or the process. Clearly, Workers 7, 1, and 3 are the three
most critical persons playing the central role in the construction project, who
are more active to finish about 36.34%, 33.98%, and 23. 70% of the total
tasks, respectively. As for the core process paths, the top two with the highest
frequency are from Worker 1 to Worker 1 (80), and from Worker 3 to Worker 1
(30). That is to say, the core paths are more likely to be executed by the critical
workers.
(2) Median duration view. Figure 8b concentrates on the performance metric
named median duration, which reflects the median time between the start and
the end of an activity. Since the deeper red color denotes a longer execution
time, it is observed that Worker 7 takes more service time than others. This is
reasonable since Worker 7 is the most key person in the collaboration. Besides,
the thicker and redder arrows highlight the potential bottlenecks, where the
longer waiting time is spent on task transmission between two workers. As
shown in Fig. 7b, the path from Worker 1 to Worker 6, from Worker 1 to
Worker 11, and from Worker 11 to Worker 1 are the three most problem-
atic sequences, which spend comparatively a longer time than others to cause
severe performance bottlenecks. Therefore, these identified bottlenecks can
guide the measurements for improving process efficiency. It is worth noting
that the execution process in paths going in or out Worker 1 tends to take a
longer time, which means the path related to Worker 1 is possible to exert great
impacts on the delay. Since Worker 1 has been regarded as a critical person,
it can be reasonably assumed that bottlenecks are likely to arise from the key
participants who are expected to conduct more tasks.
5.4 BIM Process Prediction
In this section, the decision tree model is used to predict the worker’s active time,
which can provide evidence to guide the manager in planning reasonable work
arrangements in a data-driven manner. All records about a certain worker named
“Tom” are extracted from the BIM design event log as a dataset for the case study.
The prepared dataset has in total 3240 records, from which two features called Time
(x1) and Weekday (x2) are extracted and one label termed is Work are prepared as
the foundation of prediction application. Figure 9 visualizes the distribution of two
extracted features. To be more specific, Time (x1) denotes one hour in a day, which
is given the range 0–23 on behalf of a one-hour time slot. For example, a value
of 9 indicates the time interval of 9:00–10:00. Weekday (x2) refers to Monday to
Sunday, which is represented by the index 1–7. Apart from features, each record also
owns one label termed isWork for the classification task. The label has two values:
1 means ‘worked’ and 0 means ‘didn’t work’. That is to say, the decision tree model
166 Process Mining
300
400
Count
Count
200
200
100
0 0
0 10 20 2 4 6 8
Time Weekday
Fig. 9 Histogram of the two features: a time; b weekday
will intelligently learn data about the two features, in order to make predictions about
whether the worker will work at a certain time.
To train and test the decision tree model, the prepared dataset needs to be split
into an 80%–20% ratio, which is a common practice for building and evaluating a
decision tree automatically. Herein, we take out a part of the decision tree diagram
as an example, which is provided in Fig. 10. Observably, the root node and the lead
nods all include criteria to be answered, which are typically extended to two branches
demonstrating the flow from the question to the certain answer. Regarding the root
node, it raises the question that whether the designer works before 7:30 am, which
needs to answer “True” or “False”. The “gini” indicates the Gini index at this node.
The term “value = [2490, 750]” states that when no further judgment is made, the
possibility of working before 7:30 am can be roughly calculated as 2490/3240. In
subsequence, the leaf node offers the “True” response, which continues to make
decisions on the new criteria “x1 ≤ 6.5”. As a result, the outcomes from the decision
tree are determined by the hierarchical relationships in a series of decisions. The
decision tree is easy to follow and comprehend, which provides a unique opportunity
in predicting the working time of a targeted worker.
Fig. 10 A part of the established decision tree

Table 3 Evaluation of the

Evaluation metric Value
predicted results based on the
decision tree Accuracy 0.828
Precision 0.977
Recall 0.795
F1-score 0.877
For the model validation, Table 3 presents the promising classification perfor-
mance of the developed decision tree. There are in total 2683 records to be classified
correctly, resulting in an overall accuracy of 0.828. For all records that are labeled
as “work”, 97.7% of them indicate the worker is active at that time. For all records
that truly represent “work”, 79.5% of them can be given the correct labels. F1-score
balances the precision and recall by considering both the false positives and false
negatives, which has an acceptable value of 0.877. That is to say, the decision tree is
proven to reliably predict the worker’s active time. It serves as convincing evidence
for managers to not only understand worker’s working performance, but also define
an appropriate staffing strategy for a project with the expectation of raising project
efficiency.
5.5 BIM Collaborative Network Analysis
For the social network analysis, the whole event log is taken into account, which
actually contains information about the executed commands by 51 designers from
the design firm. They cooperate to work on 82 projects and contribute 620,492
commands to the projects. Figure 11 displays an example of the collaboration among
four designers across multiple projects during the design phase. A social network
extracted from the whole event logs is displayed in Fig. 12. It is an undirected
network, where points represent all the designers involved, and the width of edges
#1 1 Project 1
2
2 (800645)
2
#2 1
Project 2
(801320)
1
Designers #3 2 Projects
1 2 Project 3
(801344)
#4 2
Collaboration relations
Fig. 11 An example of collaboration during the modeling process. Note The number in the arrows
indicate total valid commands a designer contributes to a project
168 Process Mining
High-production designers
Median-production designers
Low-production designers
Fig. 12 Map of the discovered social network among 51 designers in BIM-based collaborative
design
represents the weight of the relation. Table 4 summarizes the characteristics of the
network structure at the macro-level. The average degree (10.078) and the density
(0.202) indicate that the whole network of collaboration is finely intensive. The value
of the mean distance between nodes (1.984) points out that a worker can easily be
linked with another one in a few short steps. As for the network density and diameter,
Table 4 Characteristics of
Characteristics Value
the social network from the
event logs Number of nodes 51
Number of links 514
Density (D) 0.202
Average degree (AD) 10.078
Mean distance (MD) 1.984
Diameter (ND) 4.00
Clustering coefficient (CC) 0.771
11 51 3 6
50 11
3 9
49 13
51 19
21 5
48 17
42 47 4 25 5 1
47 19
1
46 20
50 8 7 15
46 27
48 14 45 21
49 20
18 4
22 8
44 23
34 2 6
0.00 0.07 0.14 0.21 0.28 0.35 0.42 0.49 0.56 0.63 24 0 219 438 657 876 1095 1315 1534 1753 1972 2
12 32 43 10 24
44 36 27
33 17
42 25
7
45 23
41 26
41 37
30 10
22 9 13 40 28
31 29
16 14
39 29
40 15 38
38 30
39 18
43 28
26 16 37 31
12
High-producƟon designers 36 32
35 34 33
35 Median-producƟon designers
Low-producƟon designers
(a) (b)
Fig. 13 Concentric map of centrality analysis among 51 designers by a degree centrality, and b link
centrality
20.2% potential links actually appear in the network, and the longest length of all
the shortest paths between node pairs is 4. The above-mentioned metrics all verify
that there is a cohesive collaboration hidden in the established network. Besides, the
clustering coefficient of the network (0.771) implies that the network can possibly be
divided into several clusters to facilitate better collaboration, and thus, connections
among designers in the same cluster are more intense.
The micro-level SNA concerns the degree centrality to investigate the embedded-
ness of each node and the position of the node in the whole network. Two measures
in this type are computed, including degree centrality by link and by weight. To
be more specific, degree centrality by link measures the width of the collabora-
tion between a designer and others, while degree centrality by weight measures the
depth of collaboration. Figure 13 illustrates the concentric map of centrality analysis
among 51 designers by degree centrality and link centrality. It can be seen that the
top three designers ranked in the degree centrality by the link are designers “#24”,
“#2”, and “#18”, while the top three designers ranked in the degree centrality by
weight are designers “#2”, “#7”, and “#4”. This indicates that these designers are
highly active and interact frequently with other designers in the network. These
designers are located in the center of the concentric map (as seen in Fig. 13) and
other designers would be likely to build up interactions with them. For instance,
designer “#2” is considerably active not only in the width but also the depth of
the collaborative network. In general, all these designers, including designers “#2”,
“#24”, “#7”, “#24”, and “#4”, can be regarded as “stars” in the network of BIM-based
collaborative design.
Moreover, the relationship between network characteristics and the production
performance of designers deserves further exploration. Correlation analysis can be
conducted between the measured node centralities and the recorded commands of the
170 Process Mining
Number of commands (Production) Linear Fit Number of commands (Production) Linear Fit without outliers
100000 100000
90000 90000
Y: Number of commands (Production)
Y: Number of commands (Production)

80000 80000 Outliers
70000 70000
Linear F it: Linear F it without outliers :
60000 60000
Y = 26.73 * X + 4405 Y = 24.82 * X + 2196
50000 Adj.R-Square = 0.3790 50000 Adj.R-Square = 0.7113
40000 40000
30000 30000
20000 20000
10000 10000
0 0
-200 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 -200 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400
X: Degree Centrality by weight X: Degree Centrality by weight

(a) (b)
Fig. 14 Linear fits between network characteristics (node degree centrality by weight) and produc-
tion performance of designers (number of commands executed): a with all data points; b with
outliers masked
involved designers within the organization. Considering the nature of the underlying
data, nonparametric Spearman correlation analysis is used to assess the relationships
between degree centrality and production performance of designers (represented
by the number of commands executed during the studied period). It turns out that
the correlation coefficients for the measured node degree centrality by link and by
weight are 0.253 (at a 10% significance level), and 0.805 (at a 1% significance
level), respectively. Thus, the calculated degree centrality by weight has the greatest
association with the production performance of designers, which can be described
by a linear regression test. Figure 14 illustrates the results of the linear fit. When all
data points are considered, an increase of one unit in the degree centrality by weight
generates an increase of 26.73 units in the production performance, associated with
an adjusted R-Square of 0.3790 (see Fig. 14a). When two outliers (marked in red
in Fig. 14b) are masked, the value of the adjusted R-Square is sharply increased
to 0.7113, indicating a strong interpretation for the regression fit. In other words,
the measured node degree centrality by weight is capable of explaining the greatest
percentage of variations (71.13%) in the production performance of designers. An
increase of one unit in the degree centrality by weight probably leads to an increase
of 24.82 units in the production performance of designers.
6 Summarizes
With the massive growth of BIM applications in the AEC/FM industry, large volumes
of computer-generated event logs are available to provide rich data resources. It is
worth noting that process mining provides a good opportunity to make good use
of these event logs. Various techniques from the process mining, including process
discovery, conformance checking, social network analysis, and decision tree, are
6 Summarizes 171
carried out to extract valuable information from event logs, in order to capture and
study the nature of the complicated workflow and collaboration. The analytical results
provide an objective approach to assess worker’s performance, which can assist
managers to make critical staffing decisions and optimize the workflow for enhancing
the likelihood of success of the BIM-enabled project.
In light of the application cases, the process mining has been successfully imple-
mented in a real BIM design event log to improve the understanding of the building
design development process. The benefits of the proposed approach can be summa-
rized as (1) Key processes hidden in the event logs can be efficiently mapped into the
process model based on the advanced process discovery algorithm, such as induc-
tive mining and fuzzy mining; (2) Conformance checking can not only verify the
correctness the discovered process model by metrics (i.e., fitness, precision), but also
identify the potential bottlenecks causing undesired delays for process optimization;
(c) Social network can be extracted from the event logs, which can be deeply analyzed
at different levels to reveal characteristics of collaboration within an organization; (d)
Decision tree as a supervised learning algorithm can intelligently predict a worker’s
work behavior during the design process, which provides managers some reference
to evaluate the worker’s performances and make appropriate work schedule planning.
From a broad perspective, process mining from event logs mainly provides two
kinds of merits. For one thing, hidden knowledge behind event logs can be revealed
and visualized, which can be used to monitor and analyze the project process in a
data-driven manner. For another, new insights can be provided into the execution
phase of the project to detect the potential deviations, delays, collaboration patterns,
and active time, and thus, project managers can reasonably adjust the process and
allocate workers to make the project perform better.
References
1. Ailenei I, Rozinat A, Eckert A, van der Aalst WM (2011) Definition and validation of process
mining use cases. In: International conference on business process management. Springer, pp
75–86
2. Buijs JC, Van Dongen BF, van Der Aalst WM (2012) On the role of fitness, precision, gener-
alization and simplicity in process discovery. In: OTM confederated international conferences
on the move to meaningful internet systems. Springer, pp 305–322
3. Construction M-H (2012) The business value of BIM in North America: multi-year trend
analysis and user ratings (2007–2012). Smart market report, pp 1–72. https://bimforum.
org/wp-content/uploads/2012/2012/MHC-Business-Value-of-BIM-in-North-America-2007-
2012-SMR.pdf
4. De Weerdt J, De Backer M, Vanthienen J, Baesens B (2012) A multi-dimensional quality
assessment of state-of-the-art process discovery algorithms using real-life event logs. Inf Syst
37:654–676
5. dos Santos Garcia C, Meincheim A, Junior ERF, Dallagassa MR, Sato DMV, Carvalho DR,
Santos EAP, Scalabrin EE (2019) Process mining techniques and applications–a systematic
mapping study. Expert Syst Appl 133:260–295
6. Günther CW (2009) Process mining in flexible environments
7. Günther CW, Rozinat A (2012) Disco: discover your processes. BPM (Demos) 940:40–44
172 Process Mining
8. Günther CW, Van Der Aalst WM (2007) Fuzzy mining–adaptive process simplification based
on multi-perspective metrics. In: International conference on business process management.
Springer, pp 328–343
9. Gurgen Erdogan T, Tarhan A (2018) A goal-driven evaluation method based on process mining
for healthcare processes. Appl Sci 8:894
10. Jaisook P, Premchaiswadi W (2015) Time performance analysis of medical treatment processes
by using disco. In: 2015 13th international conference on ICT and knowledge engineering (ICT
& Knowledge Engineering 2015). IEEE, pp 110–115
11. Leemans SJ, Fahland D, van der Aalst WM (2013) Discovering block-structured process models
from event logs-a constructive approach. In: International conference on applications and theory
of Petri nets and concurrency. Springer, pp 311–329
12. Leemans SJ, Fahland D, Van Der Aalst WM (2014) Process and deviation exploration with
inductive visual miner. BPM (Demos) 1295:8
13. Pan Y, Zhang L (2020a) BIM log mining: exploring design productivity characteristics. Autom
Constr 109:102997
14. Pan Y, Zhang L (2020b) BIM log mining: learning and predicting design commands. Autom
Constr 112:103107
15. Pan Y, Zhang L, Skibniewski MJ (2020) Clustering of designers based on building information
modeling event logs. Comput-Aided Civ Infrastruct Eng 35:701–718
16. Peng Y, Lin J-R, Zhang J-P, Hu Z-Z (2017) A hybrid data mining approach on BIM-based
building operation and maintenance. Build Environ 126:483–495
17. Petri C (1962) Kommunikation mit Automaten//Ph. D. thesis. Universitat Bonn, Schriften des
Instituts fur Instrumentelle Mathematik, Germany (in German)
18. Premchaiswadi W, Porouhan P (2015) Process modeling and bottleneck mining in online
peer-review systems. Springerplus 4:1–18
19. Revit A (2011) J File Parser. http://revitclinic.typepad.com/my_weblog/2011/11/journal-file-
parser.html. 18 Nov 2011
20. Revit A (2017) About J Files. https://knowledge.autodesk.com/support/revit-products/getting-
started/caas/CloudHelp/cloudhelp/2017/ENU/Revit-GetStarted/files/GUID-477C6854-2724-
4B5D-8B95-9657B636C48D-htm.html
21. Rovani M, Maggi FM, De Leoni M, Van Der Aalst WM (2015) Declarative process mining in
healthcare. Expert Syst Appl 42:9236–9251
22. Scott J (2017) Social network analysis. Sage
23. Van der Aalst W (2016) Data science in action. Process Mining. Springer, Heidelberg.
24. Van der Aalst W, Adriansyah A, van Dongen B (2012) Replaying history on process models
for conformance checking and performance analysis. Wiley Interdisc Rev Data Min Knowl
Disc 2:182–192
25. Van der Aalst WM (2011) Using process mining to bridge the gap between BI and BPM. IEEE
Comput 44:77–80
26. van der Aalst WM (2012) What makes a good process model? Softw Syst Model 11:557–569
27. Van Dongen BF, de Medeiros AKA, Verbeek H, Weijters A, van Der Aalst WM (2005) The
ProM framework: a new era in process mining tool support. In: International conference on
application and theory of petri nets. Springer, pp 444–454
28. Yarmohammadi S, Pourabolghasem R, Castro-Lacouture D (2017) Mining implicit 3D
modeling patterns from unstructured temporal BIM log text data. Autom Constr 81:17–24
29. Zhang L, Ashuri B (2018) BIM log mining: discovering social networks. Autom Constr 91:31–
43
30. Zhang L, Wen M, Ashuri B (2018) BIM log mining: measuring design productivity. J Comput
Civ Eng 32:04017071
31. Zhang L, Wu X, Ding L, Skibniewski MJ, Lu Y (2016) Bim-based risk identification system
in tunnel construction. J Civ Eng Manage 22:529–539
Agent-Based Simulation
1 Introduction
In recent years, with the continuous and rapid development of the urban economy,
a large number of people have migrated into cities, especially in developing and
thriving countries such as China and India, imposing a rigorous challenge on urban
transportation systems [12]. Due to the advantages of large transportation capacity,
low pollution, high speed, easy traffic, and low energy consumption, subways are
regarded as the modernized urban rail transportation system, playing a significant
role in transiting large passenger flows in large cities [28]. However, the predominant
disadvantages of the subway systems are densely distributed occupants and highly
restrained evacuation since a number of people usually reach durative summits in
density during the operating time, especially traffic peaks and holidays [21]. Due to
the heavily crowded population and the situation of underground space, the magni-
tude of injuries and casualties will be overwhelmingly disastrous in case of an emer-
gency [4]. In general, how to guide pedestrians to find out a more efficient route
for evacuation under emergency conditions attracts the public’s attention, and it is,
therefore, necessary to investigate a solution for optimizing evacuation route planning
strategies in metro stations.
The evacuation of pedestrians in metro stations is identified as a key issue in the
safety management of subway systems [29]. The challenges lie in the pedestrians
who are generally considered as the largest uncertainties since their motives and
actions are dynamic and independent [5]. Evacuation route planning is a strategy
that seeks to minimize the loss of life or harm to public security. In recent years,
the pedestrian evacuation has attracted more and more attention so far, where the
behavioral and psychological features of pedestrians and evacuation models have
been studied.
Existing pedestrian movement and behavior models mainly focus on pedestrian
evacuation simulation and route choice selection. Those models are most often
stochastic and oriented towards pedestrian dynamics and seldomly incorporate the
174 Agent-Based Simulation
interactions between pedestrians and the traffic [27]. It is easy to understand that one
pedestrian will move fast when the traffic is smooth while the pedestrian will move
slowly when the traffic is crowded. How to consider the interactions between pedes-
trians and the traffic to plan a more efficient evacuation route in heavily over-crowded
metro stations is a challenging task. Metro stations are generally situated under-
ground, where a large number of pedestrians get out from trains underground and
stream into escalators to upstairs in a constant manner, particularly in metro junction
hubs. If the pedestrians are not evacuated properly, especially in rush hours, newly
arrived passengers from the trains will increase the total volume of the pedestrians
stranded underground, which is very likely to result in stampede accidents.
The main research questions include: (i) What kind of factors should be considered
to come up with the best evacuation planning strategy during the pedestrian evacu-
ation in metro stations; (ii) How to validate the pedestrian movement and behavior
models that are stochastic and oriented towards pedestrian dynamics; (iii) How to
assess and improve the efficiency of the evacuation performance. To address those
research questions, a systematic multi-attribute decision approach with detailed step-
by-step procedures is developed for route choice planning. One metro station in
China is selected as a case to simulate the pedestrian evacuation process, and obser-
vations are compared with simulation results for model validation. The performance
efficiency of the proposed evacuation route planning strategy is assessed using the
proposed performance metrics.
For systems with a complex structure and over-numbered entities, e.g., people, vehi-
cles (cars/trains/planes), equipment, etc., it is difficult to directly observe and model
from the system level whereas the behaviors of the entities are relatively accessible
from the entity level [11]. For such complex systems, the agent-based simulation
(ABS) approaches have been proposed by considering the entities as agents [14] by
adopting the “bottom-to-top” perspective that focuses more on the dynamics and
procedures of the targeted system [15].
Normally, ABS is implemented in an iterative fashion [19]. First, the experts
determine the parameters based on their knowledge and understanding of the current
conditions. Next, the simulation process would go on for a while until it yields a
procedural result or stops at a pre-set pause point. Then, the experts would determine
whether to adjust their parameters to continue with the simulation or produce the
final output. ABS is especially fit for problems with dynamic procedural features
and interactive requirements, e.g., engineering construction with multiple phases,
combatant operations with varied scenarios, etc. Moreover, the results of ABS are
likely to be different in each round of simulation because the parameter configurations
are different. Therefore, the results of ABS are often given in a probability distribution
rather than a deterministic value [13].
2 Agent-Based Simulation 175
There are multiple advantages of ABS, including the ability to predict emergent
phenomena, natural description, flexibility, cost-effective, and time-saving. (1) Emer-
gent phenomena prediction [15]. ABS by nature can predict emergent phenomena
that cannot be modeled and predicted using analytical approaches because such
emergent phenomena cannot be reduced from the comprehensive system level to the
elementary and component level whereas ABS can make full advantage of designing
behaviors of single agents. (2) Natural description [13]. It is hard to directly design
a complex system from the system level because it is rather difficult to gather all of
the relevant information due to its scarcity. Comparatively, it is relatively easier to
obtain information on the entities or agents at the component level. (3) Flexibility
[18]. By designing agents rather than the whole system, more detailed information
can be used as inputs of ABS which provides much more flexibility to the modeling
process. (4) Cost-effective and time-saving [9]. Even if we can obtain information on
a complex system, it can be expected that the cost would be huge, and the resources
required would be enormous. ABS provides an effective and time-saving means by
exploring from the most elementary level that can be understood by stakeholders
from varied backgrounds. In this sense, adopting ABS would eventually save costs
in comparison with direct experiments.
It is admitted that there is a certain level of uncertainty in adopting ABS as the
final results are mostly given in terms of probability distributions [15]. This type of
uncertainty is almost unavoidable for systems in such a complex structure and with
over-numbered entities [22]. In fact, a deterministic result would almost be counter-
intuitive. By providing results in a probability distribution, along with the ability to
depict agent dynamics and other advantageous features, ABS has been successfully
applied in simulating and solving many problems with dynamic procedural charac-
teristics and interactive requirements from a wide range of theoretical and practical
backgrounds.
3 Pedestrian Evacuation Under Emergency
Pedestrians during evacuation exhibit complex and variable patterns of behavior, and
understanding these patterns is extremely important for improving evacuation proce-
dures and relevant regulations. The development of pedestrian dynamics for evacu-
ation is an interesting and ongoing research topic. To simulate the uncertainties and
dynamics of pedestrians’ behaviors, several famous models have been put forward
in the domain of evacuation, such as the cellular automaton model [1], discrete
event simulation [17], social force model [21], and others. The cellular automaton
is a discrete dynamic system with a grid-based motion decision, which consists of
a regular grid of cells, each in one of a finite number of states. Both the cellular
automation and discrete event simulation models are in a discrete state space, while
the social force model is in a continuous space and introduces the desired force to
describe the inner drive of pedestrians to escape, especially under stressful situations
[27].
Among the aforementioned models, the social force model is considered as one
of the widely accepted models for simulating individual movement behaviors [21],
especially in over-crowded situations. This is due to the fact that: (i) The social force
model is suitable to describe crowd dynamics since it can qualitatively reproduce
some self-organizing phenomena, such as lane formation and arching [6], (ii) The
social force model as a continuous space model can allow pedestrians to move contin-
uously within a pre-defined place by defining some forces while the discrete space
model, such as a cellular automaton, cannot allow pedestrians to move around in
an unrestricted manner [26], and (iii) the social force model is capable of importing
psychological and sociological factors in the real world by considering physical and
motivation forces [24]. Thus, the social force model is adopted in this chapter to
explore the crowd dynamics during the evacuation process in subway stations.
The conception of “social force” has been introduced to keep reasonable distances
among individuals. The social force model, proposed by Helbing et al. [8], is based
on the fluid dynamics equation, where the pedestrians were driven by different forces.
The original social force model is improved by [10], where an evolutionary optimiza-
tion algorithm is used to determine optimal parameter specifications for the social
force model based on suitable video recordings of interactive pedestrian motions. A
simple social force model is shown in Fig. 1, where the corresponding mathematical
expression of each pedestrian i can be represented by Eq. (1).
d vi (t)
mi = fi0 + fi j + fiw (1)
dt j (=i) w
vi Wall
f ij
vi0 ei0
Pedestrian i
Target
f iw fi 0
Pedestrian j
Wall
Fig. 1 A simple representative diagram of the social force model

3 Pedestrian Evacuation Under Emergency 177
where m i refers to the mass of pedestrian i; vi (t) refers to the actual walking velocity;
fi0 represents the diagram of the desired force; fi j represents the interaction force
between pedestrians i and j; fiw represents the interaction force between pedestrian
i and walls.
Generally, three kinds of forces are considered in the social force model, including
the desired force, the interaction force between pedestrians, and the interaction force
between a pedestrian and walls. The desired force reflects the pedestrian’s willingness
to achieve the desired velocity, and this force can be expressed by Eq. (2).
v (t)ei − vi (t)

0 0
fi0 = m i i (2)
τi
where vi0 is the desired speed with a certain direction ei0 , and τi is a constant of
adaptation time to adjust the pedestrian’s actual walking velocity.
The interaction force between pedestrians i and j can be expressed by Eq. (3),
which mainly consists of the psychological force, fisj , and the physical force, fi j .
p
Specifically, the psychological tendency between pedestrians i and j to stay away

from each other is expressed by the repulsive interaction force, which obtains its
maximum value when two pedestrians have minimal distance. The physical force
exerts on people when the distance between two pedestrian centers, d ij , is less than
the sum of the radii of these two pedestrians, ri j = ri + r j .
⎧
fi j = fisj + fi j
p
⎪
⎪
⎪
⎪
⎨ fi j = Ai exp (ri j − di j )/Bi ni j
⎪ s
fi j = kg(ri j − di j )
p
n i j + κg(ri j − di j )v tji ti j (3)
⎪
⎪
⎪
⎪ 0, i f x < 0
⎪
⎩ g(x) =
x, i f x ≥ 0
where r, A, B, k, κ, and τ are constant parameters.

The last force is the interaction force between pedestrian i and walls, and this
force can be expressed by Eq. (4). Herein, d iw is the distance between the center of
pedestrian i and the surface of walls.
fiw = Ai exp[(ri − diw )/Bi ]

n iw + kg(ri − diw )
n iw + κg(ri − diw )vwi
t
tiw (4)
where m, r, A, B, k, κ, and τ are constant parameters.

4 Simulation-Based Route Planning and Optimization
The social force model can reflect the anisotropic character and dynamic features of
pedestrian interaction, and thus, it provides a basis for simulating the pedestrian evac-
uation process. Aiming to improve the effectiveness of the route planning strategies
during the pedestrian evacuation, a systematic simulation-based decision approach
is developed to investigate how to simulate, assess, and improve the efficiency of the
evacuation route planning. The overall workflow is shown in Fig. 2. In the proposed
approach, the following four steps are adopted.
4.1 Evacuation Network Construction
The evacuation network is an abstract and geometric graph that reveals the spatial
distribution of all the possible evacuation paths. Basically, an evacuation network
can be denoted by a directed graph, G = (N, A, W ). Herein, N = {1, 2, …, n}
represents the set of evacuation nodes, such as escalators, turnstiles, and exits in a
metro station. Each evacuation node can be used to store the information, such as
evacuation speed, and pedestrian density at that specific area. A = {a1 , a2 , …, ak }
refers to the set of evacuation paths representing links between evacuation nodes.
Each evacuation path can be used to store information, such as the length and the
evacuation capacity of that specific path. Also, the time when a pedestrian enters and
leaves the specific evacuation path can be recorded. W = {w1 , w2 , …, wk } refers to
Evacuation nodes Evacuation paths/links

Evacuation Network Construction
Evacuation weights/impedence
Evacuation length (L) Evacuation time (T)

Influential Factors Identification
Pedestrian density (D)
Scenario I: Real situation of pedestrian evacuation process

Route Planning Strategy Design
Scenario II: Pedestrians choose the shortest evacuation route
Average evacuation length Average evacuation time

Evacuation Efficiency Assessment
Average pedestrian density Average evacuation capacity
Fig. 2 A systematic simulation-based decision approach for optimizing the route planning strategy
during the pedestrian evacuation
4 Simulation-Based Route Planning and Optimization 179
Fig. 3 A simple evacuation P

network
a1 a5
a4
S a2 E
a3 a6
O
the set of impedances (or weights) of paths in the graph. The value of path impedance
can be used to represent the uncertainty and complexity of a specific evacuation path.
Usually, the more uncertain (or complex) the evacuation path is, the higher the path
impedance is.
Figure 3 illustrates a simple evacuation network, where there are four evacuation
nodes and six evacuation paths in total. The nodes S and E refer to the start node
and end node, respectively. Besides, there are four combinations of paths during the
evacuation from the start node to the end node, namely a1 − a5 , a1 − a4 − a6 , a2 ,
and a3 − a6 .
4.2 Influential Factors Identification
Several factors, such as the evacuation length, evacuation time, risk perception of
the individual, and individual psychology, can affect individual behavioral decisions
in the evacuation process. The impacts of these factors on individual behaviors can
change over time. According to the experience-driven mindset, in emergency egress,
pedestrians are more likely to choose a route with the least evacuation time, since
this often coincides with the shortest path, as the shortest path is much easier to be
computed. Pedestrians often walk the shortest path if they do not want to think a lot
about their choice. However, pursuing either the shortest evacuation length or the
least evacuation time solely can easily result in the “faster-is-slower” phenomenon
[7] because of incomplete information.
The essence of the “faster is slower” phenomenon is a lack of complete informa-
tion. One pedestrian lacks a complete understanding of the crowded condition and the
required time for evacuation in real-time. As a result, s/he is unable to make a system
optimum decision. Accordingly, too many pedestrians choose the same evacuation
path based upon their incomplete information, and the congestion occurs due to high
pedestrian density in certain areas. That is to say, apart from the length of evacuation
route and the time of evacuation process, the density of pedestrian flow is another
significant factor that affects the efficiency of the evacuation process. The definition
and calculation of those three major factors are described in detail as follows.
(1) Length of evacuation route(L). This factor aims to record the total length of an
evacuation route from the start node to the end node. Basically, an evacuation
route may include several links between nodes in the evacuation network (see
Fig. 2), which can be a combination of some evacuation paths. Given the
importance and complexity of evacuation paths taken into account, the value
L at a specific evacuation route is represented by Eq. (5).

k
Lr = wi × ai × qi , r = 1, 2, ..., p (5)
i=1
where L r refers to the length of the rth evacuation route (r = 1, 2, …, p); wi

and ai refer to the impedance and length of the ith path/link (i = 1, 2, …,
k), respectively. The complexity of an evacuation link refers to the degree of
difficulty in identifying a reasonable path link for evacuation in the network.
The value of wi is defined by considering the complexity of the path link,
which should be higher when the layout of the path link is more complex; qi
is a Boolean function, and the value of qi is defined to be one when the link ai
is included in the rth evacuation route, and zero otherwise.
(2) Time of evacuation process(T). This factor aims to record the total duration
for a pedestrian to evacuate from the start node to the end node, which is a
combination of both the walking time and waiting time at the involved paths
between nodes. The value of T at the rth evacuation route (r = 1, 2, …, p) is
represented by Eq. (6).

k
Tr = (wki + wti ) × qi , r = 1, 2, ..., p (6)
i=1
where T r refers to the time of a pedestrian at the rth evacuation route (r = 1, 2,

…, p); wk i and wt i refer to the walking time and waiting time of the pedestrian
at the ith path, respectively, and their values can be obtained from pedestrians
who have just traversed the links in the simulation; qi is a Boolean function
with the same meaning mentioned above.
(3) Density of pedestrian flow(D). This factor aims to record the crowded condition
at an evacuation route from the start node to the end node. Basically, the
pedestrians are more likely to concentrate at or near the location of evacuation
nodes. Thus, pedestrian densities at some key evacuation nodes are used to
represent the crowded condition throughout the whole evacuation route. The
value of D at the rth evacuation route (r = 1, 2, …, p) is given by Eq. (7).

n
Dr = ρ j × ξ j , r = 1, 2, ..., p (7)
j=1
where Dr refers to the pedestrian density at the rth evacuation route (r = 1, 2,

…, p), ρ j refers to the pedestrian density at the jth evacuation node (j = 1, 2,
…, n), and its value can be monitored in a real-time manner in the simulation
model. ξ j is a Boolean function, and the value of ξ j is defined to be one when
the jth evacuation node is the key evacuation node at the rth evacuation route,
and zero otherwise.
4.3 Route Planning Strategy Design
Most models have been developed in the literature to simulate the evacuation process
according to the principles of the least evacuation time or the shortest distance, which
is easy to cause the “faster-is-slower” effect. In order to avoid that, pedestrians are
acknowledged that they should consider more than one factor in the choice of an
evacuation route, rather than rely on a sole factor. For instance, when too many
pedestrians choose the shortest or fastest evacuation route without considering the
crowded condition, the congestion is very likely to occur due to high pedestrian
density in certain areas. As a result, pedestrians cannot move forward and will waste
much more time. With many influential factors taken into account, a multi-attribute
decision approach should be used to discover the optimal route planning strategy for
guiding pedestrians.
In regard to the issue of the multi-attribute decision, there are generally two kinds
of methods, namely the comprehensive additive and multiplicative methods [25].
To be specific, the comprehensive additive method, also known as the weighted
arithmetic mean method, is suitable in situations where factors are basically inde-
pendent of each other. The comprehensive multiplicative method, by contrast, is
relatively more suitable in situations where the interaction among factors is high. In
this research, as mentioned above, each of the factors, such as L, T, and D, plays a
significant role in the decision of a reasonable evacuation route, and these factors are
highly related to each other. Thus, the comprehensive multiplicative method is used
to develop a multi-attribute decision strategy for evacuation route planning.
In the real world, to get rid of the “faster-is-slower” phenomenon, people might
give up a routine with the least evacuation time or the shortest distance, but choose a
routine with a lower density. In this research, two different route planning strategies
for pedestrian evacuation are proposed in the simulation environment. One is to
simulate the real-world pedestrian evacuation process for validation purposes, while
the other one is to improve the efficiency of the pedestrian evacuation. In the former
route planning strategy, defined by Scenario I, we assume that all passengers choose
an evacuation route which has both the shortest evacuation length and the least
evacuation time. The expression of the route planning strategy in Scenario I is given
by Eq. (8). Since this strategy is very consistent with the motivation of pedestrians
in the real world, the real situation of the pedestrian evacuation process can then be
simulated and validated via Scenario I.
⎛ ⎞
n

min Z r (t) = min ⎝ (L r (t))α × (Tr (t))β ⎠ (8)
j=1
where Z r (t) refers to the target value that colligates the condition of the influential
factors, such as the evacuation length and evacuation time, at the rth evacuation route
at time t; L r (t) and Tr (t) refer to the value of the evacuation length and the evacuation
time at the rth evacuation route at time t, respectively,
In the latter route planning strategy, defined by Scenario II, we assume that all
passengers choose an evacuation route which has the shortest evacuation length, the
least evacuation time, and the lowest pedestrian density. The expression of the route
planning strategy in Scenario II is given by Eq. (9). The factor of pedestrian density is
considered in the route planning strategy in Scenario II, to improve the performance
efficiency of the current pedestrian evacuation strategy specified in Scenario I.
⎛ ⎞

n

min Z r (t) = min ⎝ (L r (t))α × (Tr (t))β × (Dr (t))γ ⎠ (9)
j=1
where Z r (t) refers to the target value that colligates the condition of the influential
factors, such as the evacuation length and evacuation time, at the rth evacuation
route at time t; L r (t), Tr (t) and Dr (t) refer to the value of the evacuation length,
the evacuation time, and the pedestrian density at the rth evacuation route at time
t, respectively, and α, β, and γ refer to weights of the above factors, respectively.
For simplicity, we assume the above three factors, namely L r (t), Tr (t), and Dr (t)
share the same level of contributing to the emergency efficiency in this research, and
their values are defined to be 1.0 in Eqs. (8) and (9), respectively.
Route planning aims to identify an optimal path from a specified origin node to
other destination nodes for each pedestrian at time t. To search for the minimum
target value Z r (t) at time t, the most commonly used algorithms are the Dijkstra
algorithm [23], Floyd algorithm, and genetic algorithm. Chen et al. [2] indicated that
the Dijkstra algorithm was recognized as the best algorithm among them. Thus, the
Dijkstra algorithm is used to find the optimal evacuation route in this research. See
details on the Dijkstra algorithm to [20].
4.4 Evacuation Efficiency Assessment
A number of route planning strategies have been proposed, including the above two
strategies. It is, therefore, necessary to assess the performance efficiency of a route
planning strategy and determine which route planning strategy is more effective. In
this research, four key performance indicators (KPIs), namely average pedestrian
density, average evacuation length, average evacuation time, and average evacuation
capacity, are proposed to assess the performance of evacuation efficiency in different

route planning strategies. The definitions of these four KPIs are given as follows.
(1) Average pedestrian density(AveD). This indicator aims to assess the pedestrian
density at evacuation nodes on average during the evacuation. The value of
AveD can be calculated by Eq. (10). Generally, the lower the value of AveD is,
the better the evacuation efficiency is.

n
AveD = Aveρ j /n (10)
j=1

Ttotal
Aveρ j = ρ j (t)/Ttotal (11)

t=0
where Aveρ j (t) refers to the average pedestrian density at the jth evacuation
node (j = 1, 2, …, n), which can be defined by Eq. (11). ρ j (t) refers to the
pedestrian density at the jth evacuation node (j = 1, 2, …, n) at time t, and
T total refers to the duration of the whole evacuation process in total.
(2) Average evacuation length(AveL). This indicator aims to assess the distance
that a pedestrian walks during the evacuation on average. The value of AveL
can be calculated by Eq. (12). Generally, the lower the value of AveL is, the
better the evacuation efficiency is.
p
r =1(L r × Sr )
AveL = p (12)
r =1 Sr
where L r refers to the length of the rth evacuation route (r = 1, 2, …, p) [see

Eq. (5)], and S r refers to the number of pedestrians that evacuate along with
the rth evacuation route (r = 1, 2, …, p).
(3) Average evacuation time(AveT). This indicator aims to assess the time that a
pedestrian spends during the evacuation on average. The value of AveT can be
calculated by Eq. (13). Generally, the lower the value of AveT is, the better the
evacuation efficiency is.
p
r =1 (Tr × Sr )
AveT = p (13)
r =1 Sr
Herein, T r refers to the time a pedestrian spends at the rth evacuation route (r
= 1, 2, …, p) [see Eq. (6)].
(4) Average evacuation capacity(AveC). This indicator aims to assess how many
pedestrians can be evacuated within a certain period on average. The value of
AveC can be calculated by Eq. (14). Generally, the higher the value of AveC
is, the better the evacuation efficiency is.

p
AveC = (Sr /Ttotal ) (14)
r =1
where S r refers to the number of pedestrians that evacuate along with the rth
evacuation route (r = 1, 2, …, p), and T total refers to the duration of the whole
evacuation process in total.
5 Pedestrian Evacuation Simulation
In the past ten years, tunnel construction has presented a powerful momentum for
rapid economic development worldwide, especially in developing countries, like
China. To simulate and analyze pedestrian evacuation behaviors during rush hours,
one metro station located at the Wuhan metro system in China is used as a case for
modeling the pedestrian evacuation process.
5.1 Case Profile
The Wuhan metro system is an elevated and underground urban metro system in the
city of Wuhan, Hubei, China. It is a rapid transit system that serves to accommodate
increasing cross-borough traffic. By the end of the year 2020, there are nine metro
lines in operation, consisting of 240 metro stations with a total operation mileage
of 360 km. The Hongshan Square Station, located in the southeast of Wuhan, is
an interchange station of Line 2 and Line 4 in the Wuhan metro system. Figure 4
presents an internal 3D view of the Hongshan Square Station. There are three floors
below the ground in the Hongshan Square Station. Pedestrians get out of trains on the
second and/or third floor below the ground, go to the first floor (hall) using escalators,
and finally go to the ground through exits. Figure 5 illustrates the layout of the hall
of the Hongshan Square Station. In total, there are four escalators (i.e., Escalators 1,
2, 3, and 4), two turnstiles (i.e., Turnstiles A and E), and five exits (i.e., Exits B, C,
D, G, and F) at the hall of that station.
As shown in Fig. 5, Exit F is not visible. To make good use of exits for improving
pedestrian evacuation, the location of Exit F is made visible with the help of direc-
tional signs and optical clues. Pedestrians are persuaded that the Exit F locates in the
place of the relocated Exit F, and thus, the location of the physical Exit F is moved
to that of the relocated Exit F in the simulation model (see Fig. 5).
5 Pedestrian Evacuation Simulation 185
Fig. 4 3D view of the Hongshan Square Station in the Wuhan metro system
Escalator
Turnstile
Exit
0
W
00
C
Evacuation inflow
42
Unit mm
12
00
0
00
Ex 00
r3
Re
84
it F
to
loc
la
at
ca
ed
Es
r1
Ex
to
it F
la
ca
Es
B
it
Es
Ex
ca
PAID AREA
la
to
r2
A
ile
st
Escalator 4
rn
30000
Tu
Ex
it
C
Exit G
Turnstile E
0
00
55
160000
Exit D
Fig. 5 Layout of the hall of the Hongshan Square Station
5.2 Simulation Model Construct
One of the biggest challenges for the construction of the evacuation model lies in
the modeling of complex pedestrian dynamics. To address this problem, AnyLogic
provides a dynamic simulation tool with a pedestrian library for simulating pedestrian
F 1
a1
2
4 a9
G a2
a10 B
a4 3 a3 a5
Escalator A a6
Turnstile E a7
a8
Exit
Key node
D C
Fig. 6 Excavation network for the hall of the Hongshan Square Station
flows in a “physical” environment and allows the embedded models to simulate

pedestrian psychology. The pedestrian library is built based on the social force model
and provides a higher-level interface for the fast creation of pedestrian models in the
style of flowcharts, which is a widely accepted graphical representation of systems of
many important domains. Thus, the AnyLogic pedestrian library is used to simulate
the pedestrian evacuation process at the Hongshan Square Station during rush hours.
In this case, the first floor (hall) of the station should be the most crowded area
during rush hours since pedestrians, who get out of trains from the second and third
floors, join at the hall of the station. The built excavation network for the hall of
the Hongshan Square Station is illustrated in Fig. 6, where the escalators 1, 2, 3,
and 4 are considered as the start nodes, the turnstiles A and E are considered as
the intermediate nodes, and the exits B, C, D, F, and G are considered as the end
nodes. Both the intermediate and end nodes are regarded as the key nodes since the
pedestrian densities monitored at those key nodes play a significant role in the route
planning during the evacuation.
As shown in Fig. 6, there are ten path links, namely a1 , a2 ,…, a10 in the excavation
network at the hall of the Hongshan Square Station. Table 1 reveals the values of
path impedances and lengths of those ten path links. It should be noted that the
value of path impedance is defined by considering the complexity of the path link,
as mentioned above. In this case, a pedestrian at the start node (1, 2, 3, or 4) has
only one path link to access the intermediate node (A or E), while a pedestrian at
the intermediate node (A or E) has more than two path links to access to the end
node (B, C, D, F or G). In other words, a decision on choosing the path link from
the intermediate node to the end node is much more complex than that from the start
node to the intermediate node. As a result, the path impedances of the path links
from the start node to the intermediate node, which are a1 , a2 , a3, and a4 , are all
defined to be 0.5; while the path impedances of the path links from the intermediate
node to the end node, which are a5 , a6 , a7 , a8 , a9, and a10 , are all defined to be 1.0.
In total, there are twelve evacuation routes from the start node to the end node at
the hall of the Hongshan Square Station. Table 2 illustrates the values of the total
Table 1 Description of ten

No. Path link Path impedance Path length (m)
path links at the hall of the
Hongshan Square Station a1 1-A 0.5 30.0
a2 2-A 0.5 23.0
a3 3-A 0.5 13.0
a4 4-E 0.5 24.0
a5 A-B 1.0 27.0
a6 A-C 1.0 40.5
a7 A-D 1.0 52.5
a8 E-D 1.0 38.5
a9 E-F 1.0 48.0
a10 E-G 1.0 52.5
Table 2 Description of
No. Excavation route Total length (m) Key nodes
twelve excavation routes at
the hall of the metro station 1 1-A-B 42.0 A, B
2 2-A-B 38.5 A, B
3 3-A-B 33.5 A, B
4 1-A-C 55.5 A, C
5 2-A-C 52.0 A, C
6 3-A-C 47.0 A, C
7 1-A-D 67.5 A, D
8 2-A-D 64.0 A, D
9 3-A-D 59.0 A, D
10 4-E-D 50.5 E, D
11 4-E-F 60.0 E, F
12 4-E-G 64.5 E, G
lengths of all the evacuation routes. For instance, the 1st evacuation route “1-A-B”
is a combination of the path link “a1 ” (from Escalator 1 to Turnstile A) and the path
link “a4 ” (from Turnstile A to the Exit B). As shown in Table 1, these two path links
have a path impedance of 0.5 and 1.0, respectively. The total length of the evacuation
route “1-A-B” is then calculated to be 42.0 m using Eq. (5).
This model simulates the pedestrian evacuation process in a continuous manner.
Pedestrians successively get out of trains on the second or third floor below the
ground and go to the first floor (hall) using escalators. In this model, we assume
that pedestrians appear continuously at a flow rate of 30 p/min at the Escalator 1, 25
p/min at Escalators 2 and 3, and 50 p/min at Escalator 4. There are three passages
(or evacuation options) at each turnstile, and the width of each exit is set to be 1.8 m
in the simulation model. Figure 7 illustrates a flowchart for path selection in the
evacuation simulation for pedestrians located at Escalator 1. The bottom left box
pedSource1 pedGoTo pedSelectOutput pedGoTo1 pedSink1
pedGoTo2 pedSink2
Selection strategies:
Condition1: Z1<=min (Z4, Z7)
Condition2: Z4<=min (Z1, Z7) pedSink3
pedGoTo3
Condition3: Z7<=min (Z1, Z4)
Fig. 7 Flowchart for path selection in the evacuation simulation for pedestrians located at the
Escalator 1
in Fig. 7 shows the route selection strategies that are incorporated in the flowchart
in detail. For instance, a pedestrian at time t would select the 1st evacuation route,
rather than the 4th or the 7th evacuation route, once the calculated target value of Z 1
is less than that of Z 4 and Z 7 using Eq. (9) [or Eq. (8)].
5.3 Simulation Model Validation
The effectiveness of the simulation model is considered as an essential guarantee of

the plausibility in its applications. However, experience concerning model valida-
tion in evacuation conditions is lacking in the literature. Musolino and Vitetta [16]
indicated that one reason lied in an operational level since collecting field data was
burdensome, and it was not possible to monitor human behaviors in real evacua-
tions,while the other reason lied in a conceptual level since it was very expensive
to set up a real-world evacuation experiment. In many experiments, people are not
aware of whether it is a test or real. There may be (very likely is) a change in behavior
in imminent danger conditions. Indeed, there are no models that can exactly simu-
late the behaviors of the real world due to uncertainty, dynamics, and assumptions
[3], especially for human behaviors. To assess whether the model behaves sensibly
to changes and ensure the validation process is robust, we attempt to use the route
planning strategies that are specified in Scenario I for model validation since this
scenario can simulate the real-world situation of the pedestrian evacuation process.
In this scenario, all passengers choose an evacuation route with the shortest evac-
uation length and time [see Eq. (8)]. The key to model validation is tantamount to
comparing the simulation outputs with observations. Here, we intend to evaluate the
established evacuation model in terms of pedestrian densities due to its availability.
During the simulation process, the passenger densities located at key evacuation
nodes are monitored every 0.01 min in this case. After a six-minute simulation, a set
of 600 records regarding the pedestrian density is recorded and stored at every key
evacuation node. Equation (12) can then be used to obtain the average pedestrian
density at every key evacuation node. Taking the key evacuation node of A, for
example, Fig. 8 illustrates the statistics of the pedestrian density located at the node
of A in Scenario I, and the average pedestrian density is calculated to be 2.66 p/m2 .
At the same time, the pedestrian evacuation process at the key evacuation node of
A in the real world is monitored by using video cameras. Pedestrian densities at a
specific evacuation node are then recorded by drawing frames of the video at the same
position at different time within six minutes. Figure 9 illustrates some photographs
for monitoring the pedestrian density at the key evacuation node A in the real world
at different time. The observed average pedestrian density at the key evacuation node
A is calculated to be 2.69 p/m2 . With all the seven key evacuation nodes taken into
account, Fig. 10 illustrates the scatterplot of simulation results and observations in
Fig. 8 Statistics of the 130

pedestrian density at the key A_I
120
evacuation node A in A_I_Ave=2.66
110
Scenario I (unit: p/m2 )
100
90
80
Frequency
70
60
50
40
30
20
10
0
1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8
Pedestrian density
Fig. 9 Observed pedestrian density at the key evacuation node A in the real world at different time:
a T = 9: 00 am; b T = 9:01 am; c T = 9:02 am
3.0 Equation y = a + b*x

Adj. R-Square 0.99777
Value Standard Error
-- Intercept 0.01936 0.39269
2.8 -- Slope 0.98145 0.1862
2.6 A
2.4
D
Simulation
2.2 B
E
2.0
F
1.8
G Pedestrian density
1.6 C Linear of fit
1.4
1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
Observation
Fig. 10 Scatterplot of simulation and observation for the average pedestrian density at seven key
evacuation nodes (unit: p/m2 )
terms of the average pedestrian density. Results indicate that, in general, the observed
average pedestrian density is low when the simulated mean is low, and similarly, the
observed average pedestrian density is high when the simulated mean is high. The
fitted regression line has a gradient of 0.98 approaching 1, which indicates that the
established evacuation model can simulate the average magnitude of the pedestrian
density in a metro station well. In other words, the central tendency of the model
simulation results is not significantly different from that of the observations.
6 Route Planning Optimization
When the established evacuation model has been validated in terms of pedestrian
densities located at key evacuation nodes, it is believed that it is rational and reliable
to apply the established model to simulate the pedestrian evacuation process in the real
world. Aiming to improve the efficiency of the current route planning strategy, a novel
route planning strategy that is specified in Scenario II is developed. Thus, the route
planning strategy specified in Scenario I is replaced by the route planning strategy
specified in Scenario II. All passengers would choose an evacuation route that has
the shortest evacuation length, the least evacuation time, and the lowest pedestrian
density during the evacuation [see Eq. (9)]. The main purpose is to compare the
strategies specified in Scenarios I and II and identify which strategy is better during
pedestrian evacuation.
6 Route Planning Optimization 191
Fig. 11 Distribution of pedestrian densities in Scenario II (unit: p/m2 ). Note Exit F is moved to the
center to simplify the layout of the hall of the station
Figure 11 illustrates the distribution of pedestrian densities in Scenario II in a very

visual way, where the degree of the pedestrian density is represented by different
colors. The distribution of pedestrian densities at different locations of the metro
station can be monitored over time. To assess the performance of evacuation effi-
ciency in different route planning strategies, the proposed four KPIs, namely AveD,
AveL, AveT, and AveC, are applied. The results of the four KPIs are analyzed in detail
as follows.
6.1 Average Pedestrian Density
The pedestrian density is a very important KPI for the performance assessment of
the evacuation efficiency since the pedestrians cannot move forward if the pedes-
trian density is too high. More time would be wasted, and the “faster-is-slower”
phenomenon is likely to occur. Following the simulation of the pedestrian evacua-
tion process within six minutes in both Scenarios I and II, pedestrian densities at seven
key evacuation nodes are recorded every 0.01 min. Equation (11) is used to obtain
the average pedestrian density at each key evacuation node. Figure 12 illustrates the
statistic of pedestrian densities at those nodes in Scenarios I and II.
Taking Fig. 12a as an example, the average pedestrian density at the key evacuation
node of A is calculated to be 2.66 p/m2 in Scenario I and 1.94 p/m2 in Scenario II,
respectively. Generally, the pedestrian density in Scenario II is lower than that in
Scenario I at almost all the key evacuation nodes, except for the nodes of C and G.
At the nodes of C and G, the pedestrian density in Scenario II is slightly higher than
that in Scenario I. In terms of deviations displayed on those curves in Fig. 12, the
central tendency of pedestrian densities in Scenario II is more apparent than that in
140 200 110

A_I 190 B_I C_I
130
A_II 180 B_II 100 C_II
120 A_I _Ave=2.66 170 B_I _Ave=2.24 C_I _Ave=1.66
110 A_II_Ave=1.94 160 B_II_Ave=1.88 90 C_II_Ave=1.70
150
Frequency
Frequency
100 80
Frequency
140
90 130
70
120
80
110
60
70 100
90 50
60
80
50 70 40
40 60
50 30
30 40
20
20 30
20
10 10
10
0 0 0
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
Pedestrian density Pedestrian density Pedestrian density

(a) (b) (c)
200 160 120
190 D_I E_I F_I
150
180 D_II E_II 110 F_II
140
170 D_I _Ave=2.50 E_I _Ave=2.11 100
F_I _Ave=1.92
160 D_II_Ave=2.12 130 E_II_Ave=1.94 F_II_Ave=1.68
150 120 90
Frequency
Frequency
Frequency
140 110
130 80
100
120 70
110 90
100 80 60
90 70
80 50
60
70 40
60 50
50 40 30
40 30
30 20
20
20 10
10 10
0 0 0
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8
Pedestrian density Pedestrian density Pedestrian density

(d) (e) (f)
90
G_I
80
G_II
G_I _Ave=1.68
70
G_II_Ave=1.71
Frequency
60
50
40
30
20
10
0
0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8
Pedestrian density
(g)
Fig. 12 Statistic results of pedestrian densities in Scenarios I and II at seven key evacuation nodes
(unit: p/m2 ): a Turnstile A; b Exit B; c Exit C; d Exit D; e Turnstile E; f Exit F; g Exit G
Scenario I, indicating that a more balanced fashion regarding the crowded condition
is achieved in Scenario II, with that in Scenario I compared.
Given the above seven critical evacuation nodes taken into account, Eq. (10) is used
to obtain the average pedestrian density as a whole, and the results are shown in Table
3. Results indicate that the value of the average pedestrian density is reduced from
2.11 p/m2 in Scenario I to 1.83 p/m2 in Scenario II. In other words, the performance
Table 3 Comparisons of the KPIs during the pedestrian evacuation in Scenarios I and II
KPIs Scenario I Scenario II Improved Efficiency (%)
Average pedestrian density (p/m2 ) 2.11 1.83 0.28 13.27
Average evacuation length (m) 56.47 55.58 0.89 1.58
Average evacuation time (min) 1.55 1.26 0.29 18.71
Average evacuation capacity (p/min) 92 122 30 32.61
of the crowded condition during the evacuation is increased by 13.27% (i.e., (2.11–
1.83)/2.11) using the improved route planning strategy specified in Scenario II.
6.2 Average Evacuation Length
The evacuation length measures the distance that a pedestrian walks during the
evacuation. Along with the pedestrian evacuation simulation within six minutes as
mentioned above, the number of pedestrians evacuating by the rth evacuation route
(r = 1, 2, …, 12) is recorded in both Scenarios I and II. Equation (12) is used to
obtain the average evacuation length in Scenarios I and II, respectively. Figure 13
illustrates the distribution results of the evacuation length for all pedestrians during
the evacuation in Scenarios I and II.
It is clear to see that in Scenario I, a total of 552 pedestrians excavate from the
metro station within six minutes, and the length of all the evacuation routes in total
reaches up to 31,172.5 m. As a result, the average evacuation length in Scenario I
is 31,172.5/552 = 56.47 m. In Scenario II, a total of 728 pedestrians excavate from
the metro station within six minutes, and the length of all the evacuation routes in
total reaches up to 40,460.5 m. As a result, the average evacuation length in Scenario
II is 40,460.5/728 = 55.58 m. That is to say, compared with Scenario I, a distance
70
Scenario I
65 Scenario II
60
55
50
Evacuation length (m)
45
40
35 Scenario I
30 Evacuated pedestrians=552
Total evacuation length=31172.5 m
25
Average evacuation length=56.47 m
20 Scenario II
15 Evacuated pedestrians=728
Total evacuation length=40460.5 m
10
Average evacuation length=55.58 m
5
552 728
0
0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750
Pedestrians during the evacuation
Fig. 13 Distribution results of evacuation length for all pedestrians during the evacuation in
Scenarios I and II. Note The records of evacuation length of each pedestrian are sorted from smallest
to largest
of 56.47–55.58 = 0.89 m can be saved in terms of the average evacuation length

during the evacuation by using the improved route planning strategy as specified in
Scenario II. As shown in Table 3, from the perspective of the average evacuation
length, the efficiency of the route planning strategy in Scenario II is increased by
1.58%, compared to that in Scenario I.
6.3 Average Evacuation Time
The evacuation time measures the time that a pedestrian spends during the evacuation
from the start node to the end node. Along with the pedestrian evacuation simulation
within six minutes as mentioned above, the time that a pedestrian evacuates on
average by the rth evacuation route (r = 1, 2, …, 12) is recorded in both Scenarios
I and II. Figure 14 illustrates the statistic of average evacuation time among 12
evacuation routes during the pedestrian evacuation in both scenarios.
Results indicate that the values of the average evacuation time among the above
12 routes in Scenario II do not vary greatly and achieve a balance to some extent,
while the average evacuation time varies greatly in Scenario I. Given the above 12
routes taken into account, Eq. (13) is used to obtain the average evacuation time as a
whole. Results, as shown in Table 3, indicate that the value of the average evacuation
time is reduced from 1.55 min in Scenario I to 1.26 min in Scenario II. In other words,
0.29 min that is about 18 s can be saved for every pedestrian during the evacuation
using the improved route planning strategy specified in Scenario II.
2.6
Scenario I
2.4 Scenario II
2.2
Average evacuation duration (min)
2.0
1.8
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
1 2 3 4 5 6 7 8 9 10 11 12
Route paths
Fig. 14 Statistic results of average evacuation time among 12 evacuation routes during the
pedestrian evacuation in Scenarios I and II
6.4 Average Evacuation Capacity
The evacuation capacity measures the number of pedestrians who can be evacu-
ated within a certain period. During the pedestrian evacuation simulation within six
minutes, the number of pedestrians that evacuate by the rth evacuation route (r = 1,
2, …, 12) is recorded in both Scenarios I and II. Equation (14) is used to obtain the
average evacuation capacity in both scenarios I and II, respectively. Table 3 illustrates
the statistic of average evacuation capacity in both scenarios. Results indicate that
the value of the average evacuation capacity is increased from 92 p/min in Scenario I
to 122 p/min in Scenario II. That is to say, the performance of the evacuation capacity
during the evacuation is increased by 32.61% (i.e., (122–92)/92) using the improved
route planning strategy specified in Scenario II.
As mentioned above, pedestrians appear continuously at a flow rate of 30 p/min
at the Escalator 1, 25 p/min at Escalators 2 and 3, and 50 p/min at Escalator 4.
Generally, the input flow rate is 130 p/min in total. It is clear to see that 38 (i.e.,
130–92) and 8 (i.e., 130–122) pedestrians would be stranded at the station hall every
minute in Scenarios I and II, respectively. In general, the route planning strategy
as specified in Scenario II provides a better strategy for pedestrian evacuation than
that in Scenario I. In other words, the evacuation performance can be significantly
improved, when the route planning strategy specified in Scenario II is implemented
during the pedestrian evacuation.
7 Merits of Simulation and Optimization
In this chapter, a systematic simulation-based decision approach for improved evac-

uation route planning is developed with detailed step-by-step procedures, which is
consisting of evacuation network construction, influential factors identification, route
planning strategy design, and evacuation efficiency assessment. The capabilities of
the developed approach are discussed as follows.
(1) Simulating and validating pedestrian evacuation dynamics in a continuous
space. Simulating the pedestrian evacuation process is considered a challenging
task, particularly for the heavily over-crowded metro stations since the pedes-
trians’ motives and actions are dynamic, uncertain, and independent. In the past
years, several models, such as the cellular automaton model and the discrete
event simulation model, have been put forward to simulate the pedestrian
behavioral dynamics and uncertainties during evacuation. They have, however,
encountered a lot of constraints in their applications, including discrete state
space and the ignorance of the pedestrians’ interactions. This chapter employs
the social force model to describe the inner drive of pedestrians to escape
and excavate under stressful situations, which enables the simulation of the
pedestrian excavation process in a continuous space. Furthermore, the estab-
lished evacuation model is validated by comparing the simulation outputs with
observations in terms of pedestrian densities. Results indicate that the fitted

regression line between simulation outputs with observations has a gradient
of 0.98 approaching 1, indicating that the central tendency of the model simu-
lation outputs is not significantly different from that of the observations. In
this way, the developed approach is able to model the pedestrian evacuation
process in the real world.
(2) Optimizing the evacuation strategy. To find out a more reasonable and effi-
cient route for evacuation is the objective of the route planning during the
pedestrian evacuation. Most methods that have been developed in previous
literature attempt to model the evacuation process according to the principles
of least evacuation time and/or the shortest distance, leading to a tradeoff of the
“faster-is-slower” effect [7]. In this chapter, three factors, including the length
of evacuation route (L), time of evacuation process (T ), and density of pedes-
trian flow (D), have been considered and identified as significant factors that
influence the efficiency of the evacuation process, especially during rush hours
in a metro station. Two scenarios, Scenario I (in which L and T are considered)
and Scenario II (in which L, T, and D are all considered), are developed to
conduct what-if scenario analysis, in order to find out which route planning
strategy is better. Results, as shown in Table 3, indicate that the route planning
strategy as specified in Scenario II provides a better strategy for pedestrian
evacuation than that in Scenario I.
(3) Assessing evacuation performance. The developed approach can be used to
assess the performance of the evacuation efficiency on a quantitative basis.
Many route planning strategies have been proposed, and it is necessary to
determine which route planning strategy is more reasonable and efficient. Four
KPIs, namely average pedestrian density, average evacuation length, average
evacuation time, and average evacuation capacity, have been proposed in this
chapter to evaluate the evacuation performance within different route planning
strategies. In the case study part, the four KPIs are measured and compared
among different route planning strategies. Results (as shown in Table 3) indicate
the performance of the evacuation efficiency can be largely improved when
the improved route planning strategy specified in Scenario II is implemented
during the pedestrian evacuation. Given numerous route planning strategies as
alternatives, the developed approach can be used to test the performance of
each alternative strategy and identify the best one.
The findings in this chapter can give valuable theoretical and practical insights into
the field of pedestrian evacuation, especially in heavily over-crowded situations. A
deep understanding of route planning strategies during the pedestrian evacuation can
be provided to the general architectural community, when they are designing evacua-
tion systems without suitable formulas for engineering calculation. The improvement
of safety and economic objectives can then be achieved in the design or re-design
of metro evacuation systems. Besides, a supervisory control system with the simu-
lation model and the built-in prediction and optimization functions incorporated can
7 Merits of Simulation and Optimization 197
be developed to monitor and supervise pedestrian behaviors and evacuation perfor-

mance. In addition, the light emitting diodes (LEDs) act as the supervisory control
system, which can be installed at the key evacuation nodes to guide the selection of
evacuation routes for pedestrians in real-time.
8 Summaries
To improve the effectiveness of route planning strategies during the pedestrian evac-
uation in rush hours, a systematic simulation-based decision approach is developed
to investigate how to simulate, assess, and improve the performance efficiency of
the route planning strategy during the evacuation. Scenario I is used to simulate the
pedestrian evacuation process in the real world. Scenario II is used to improve the
performance efficiency of the pedestrian evacuation process. Four KPIs are proposed
to evaluate the evacuation performance within different route planning strategies.
One metro station located at the Wuhan metro system in China is used as a case
to demonstrate the applicability and effectiveness of the proposed route planning
strategy for improving the evacuation performance efficiency. Results indicate that
the developed approach is capable of (i) simulating and validating pedestrian evac-
uation dynamics given the complex interaction between pedestrians and the traffic
is taken into account, (ii) optimizing the route choice planning strategy during the
evacuation, and (iii) assessing the evacuation performance. All the three factors,
including the length of evacuation route (L), time of evacuation process (T ), and
density of pedestrian flow (D), are regarded as significant factors that affect the
efficiency of the evacuation process. Furthermore, valuable theoretical and prac-
tical insights can be provided into a deep understanding of route planning strategies
during the pedestrian evacuation, enabling to achieve the improvement of safety and
economic objectives in the design or re-design of the metro evacuation systems.
The major novelty of the proposed ABS approach in this chapter is that the uncer-
tainties and dynamics underlying the pedestrian behaviors during the evacuation
process are modeled, when the complex interactions between pedestrians and the
traffic can be taken into account. A better route planning strategy has been devel-
oped, which has been demonstrated to be effective in optimizing the pedestrian
evacuation performance. Some useful KPIs are proposed to compare and benchmark
the evacuation efficiency among different route planning strategies.
References
1. Alizadeh R (2011) A dynamic cellular automaton model for evacuation process with obstacles.
Saf Sci 49:315–323
2. Chen Y-Z, Shen S-F, Chen T, Yang R (2014) Path Optimization Study for Vehicles Evacuation
based on Dijkstra Algorithm. Proc Eng 71:159–165
3. Chung TH, Mohamed Y, AbouRizk S (2006) Bayesian updating application into simulation in
the North Edmonton Sanitary Trunk tunnel project. J Constr Eng Manag 132:882–894
4. Fang Z, Li Q, Li Q, Han LD, Wang D (2011) A proposed pedestrian waiting-time model for
improving space–time use efficiency in stadium evacuation scenarios. Build Environ 46:1774–
1784
5. Guo R-Y, Huang H-J, Wong S (2012) Route choice in pedestrian evacuation under conditions
of good and zero visibility: Experimental and simulation results. Transp Res Part B Methodol
46:669–686
6. Helbing D, Farkas I, Vicsek T (2000) Simulating dynamical features of escape panic. Nature
407:487–490
7. Helbing D, Farkas IJ, Molnar P, Vicsek T (2002) Simulation of pedestrian crowds in normal
and evacuation situations. Pedestrian Evacuation Dyn 21:21–58
8. Helbing D, Molnar P (1995) Social force model for pedestrian dynamics. Phys Rev E 51:4282–
4286
9. Jahangirian M, Eldabi T, Naseer A, Stergioulas LK, Young T (2010) Simulation in manufac-
turing and business: a review. Eur J Oper Res 203:1–13
10. Johansson A, Helbing D, Shukla PK (2007) Specification of the social force pedestrian model
by evolutionary adjustment to video tracking data. Adv Complex Syst 10:271–288
11. Kiesling E, Günther M, Stummer C, Wakolbinger LM (2012) Agent-based simulation of
innovation diffusion: a review. CEJOR 20:183–230
12. Lei W, Li A, Gao R, Hao X, Deng B (2012) Simulation of pedestrian crowds’ evacuation in a
huge transit terminal subway station. Phys A 391:5355–5365
13. Leng N, Corman F (2020) The role of information availability to passengers in public transport
disruptions: an agent-based simulation approach. Transp Res Part A Policy Pract 133:214–236
14. Lohmer J, Bugert N, Lasch R (2020) Analysis of resilience strategies and ripple effect in
blockchain-coordinated supply chains: an agent-based simulation study. Int J Prod Econ
228:107882
15. Marzouk M, Al Daour I (2018) Planning labor evacuation for construction sites using BIM and
agent-based simulation. Saf Sci 109:174–185
16. Musolino G, Vitetta A (2014) Calibration and validation of a dynamic assignment model in
emergency conditions from real-world experimentation. Proc Soc Behav Sci 111:498–507
17. Na HS, Banerjee A (2019) Agent-based discrete-event simulation model for no-notice natural
disaster evacuation planning. Comput Ind Eng 129:44–55
18. Railsback SF, Lytinen SL, Jackson SK (2006) Agent-based simulation platforms: review and
development recommendations. SIMULATION 82:609–623
19. Rozo KR, Arellana J, Santander-Mercado A, Jubiz-Diaz M (2019) Modelling building emer-
gency evacuation plans considering the dynamic behaviour of pedestrians using agent-based
simulation. Saf Sci 113:276–284
20. Skiena S (1990) Dijkstra’s algorithm. Implementing discrete mathematics: combinatorics and
graph theory with mathematica. Addison-Wesley, Reading, MA, pp 225–227
21. Wan J, Sui J, Yu H (2014) Research on evacuation in the subway station in China based on the
combined social force model. Phys A 394:33–46
22. Xiong C, Zhu Z, Chen X, Zhang L (2018) Optimal travel information provision strategies: an
agent-based approach under uncertainty. Transportmetrica B Transp Dyn 6:129–150
23. Xu M, Liu Y, Huang Q, Zhang Y, Luan G (2007) An improved Dijkstra’s shortest path algorithm
for sparse network. Appl Math Comput 185:247–254
24. Xu S, Duh H-L (2010) A simulation of bonding effects and their impacts on pedestrian
dynamics. IEEE Trans Intell Transp Syst 11:153–161
25. Xu Z (2015) Uncertain multi-attribute decision making: methods and applications. Springer
26. Yuen J, Lee E, Lam W (2014) An intelligence-based route choice model for pedestrian flow in
a transportation station. Appl Soft Comput 24:31–39
27. Zhang L, Liu M, Wu X, AbouRizk SM (2016) Simulation-based route planning for pedestrian
evacuation in metro stations: a case study. Autom Constr 71:430–442
References 199
28. Zhang L, Wu X, Chen Q, Skibniewski MJ, Hsu S-C (2014) Towards a safety management
approach for adjacent buildings in tunneling environments: case study in China. Build Environ
75:222–235
29. Zhong M, Shi C, Tu X, Fu T, He L (2008) Study of the human evacuation simulation of metro
fire safety analysis in China. J Loss Prev Process Ind 21:287–298
Expert Systems
1 Introduction
Tunnel construction entails a highly complicated project with large potential risks,
which can bring enormous dangers to public safety. Numerous accidents have led
to growing public concerns about prior risk identification and assessment in rela-
tion to tunnel construction safety. Risk identification plays an important role in the
safety assurance process, aiming to reveal the potential safety risk and determine
risk factors’ contribution to the occurrence of an accident. Critical potential risks
and risk factors can be discovered to assist project engineers in determining critical
safety checking points for the improved safety management [31].
Generally, the frequently used methods for risk identification can be divided into
qualitative and quantitative risk analysis tools, such as Safety Check List (SCL),
Delphi, Fault Tree Analysis (FTA), Event Tree Analysis (ETA), Hazard and Oper-
ability study (HAZOP), and others. Šejnoha et al. [25] conducted a quantitative risk
assessment of the tunneling excavation process in a Czech Republic tunnel project
using FTA. Hong et al. [14] employed ETA to provide a quantitative risk evaluation
in a tunnel project passed under the Han River in South Korea. Using these risk
analysis tools, domain experts are commonly invited to identify the potential safety
risks according to their expertise accumulated from years of experience. Generally,
the values of risk factors are captured by manually reading engineering drawing
documents.
There are three kinds of deficiencies existing in the current safety risk identifica-
tion process: (1) The domain experts’ expertise is excessively relied on due to a lack of
autonomous inference capacity in the current approach. Domain experts are generally
considered as a scarce resource that is unable to provide universal consultation and
real-time guidance due to time constraints. (2) The process of the engineering param-
eter extraction involves many time-consuming and error-prone activities, especially
in reading engineering drawing documents, and therefore, the accuracy and effec-
tiveness of the risk identification results can be significantly affected to some extent.
202 Expert Systems
(3) Empirical knowledge belongs to tacit knowledge, which is not organized in an

explicit format and has low efficiency in knowledge transformation. This paper aims
to present an innovative approach that merges BIM and expert systems to address
these deficiencies.
BIM is a digital representation of the building process which can facilitate the
exchange and interoperability of project information and promote a collaborative
practice for the Architectural, Engineering, Construction, and Facilities Management
(AEC/FM) industry. Compared to the conventional two-dimensional (2D) draw-
ings, BIM provides a more realistic and enriched model beneficial to all phases
of the building life cycle. BIM offers a geometrically accurate three-dimensional
(3D) representation and enables to affiliate attributes and data to the components
and objects inside a model. The growing implementation of BIM provides potential
advantages for safety management in construction design and planning by facilitating
the integration of construction safety practices. BIM is one of the fastest-growing
tools that has gained acceptance in the AEC/FM industry worldwide, offering new
solutions to improve inefficiencies of the current paper-based processes [16].
BIM allows various types of analysis during early design phases to engage clients
and to support facility management in a project life cycle. Hadikusumo and Rowl-
inson [12] illustrated that BIM-based design review tools could be utilized to add
safety hazards to 3D geometry objects for hazard identification. Hardin [13] noted
that the implementation of BIM in construction practice promised to improve the
communication and collaboration between participants through higher interoper-
ability of data. Pan and Zhang [23] presents a closed-loop digital twin framework
that integrates BIM, IoT, and data mining to better understand, predict, and optimize
the physical construction operations. Opportunities are identified to promote safety
with BIM by involving partners such as designers, contractors, safety specialists, and
others.
Safety risk identification plays a significant role in construction safety manage-
ment, which is a very challenging task given the complexity and size of the building
systems designed. The construction safety risks are related to many influential
factors in tunnel construction, such as structures, construction techniques, geolog-
ical and hydro-geological conditions. In general, how to calculate and identify the
potential safety risk in construction projects remains a challenging problem, since
various factors are involved, especially in tunnel construction [32]. Expert system
(ES) provides a powerful tool for knowledge integration and autonomous infer-
ence mechanism [19]. ESs are designed to solve complex problems by knowledge-
driven reasoning, which can be used to significantly reduce dependence on domain
experts. With strong inference capacities, ESs have been widely applied in knowl-
edge management, automatic fault diagnosis, intelligent traffic management, and
other fields in the past two decades. Merging BIM and ESs provides an alternative
solution to realize the automated safety risk identification and management in tunnel
construction.
In this chapter, a BIM-based Risk Identification Expert System (B-RIES) for
tunnel construction is developed. It consists of three built-in modules: including
BIM extraction, knowledge base management, and risk identification modules. An
1 Introduction 203
integrated knowledge base is established to systematize the fragmented explicit and

tacit knowledge. B-RIES is used to systematize the fragmented knowledge and facil-
itate knowledge sharing and communication among dispersed clients and domain
experts. Finally, a realistic tunnel case in China is used as a case study for demonstra-
tion purposes. The results testify the applicability and effectiveness of the developed
B-RIES.
2 Expert Systems
An expert system (ES) is a computer system designed to solve complex problems by

emulating the decision-making process of human experts [27]. It was firstly inspired
by the experiment done by Newell et al. [21]. They analyzed the thinking activities
when people solve problems, and they found three basic procedures were followed in
the activities, including (a) developing a rough plan, (b) trying to solve the problem
according to the plan, and (c) constantly analyzing the problem-solving methods and
accordingly revising the plan. Early developed systems include DENDRAL system
[18] and MACSYMA system [24], which were established by Stanford University
and MIT, respectively. These early ESs already could interpret spectral data by them-
selves, i.e., to use the knowledge of the problem itself to solve the problem. Not only
this, these systems inspired researchers to start to use codes to represent the knowl-
edge and develop the knowledge-based systems to solve complex problems. The
successful development and applications of these systems made ESs accepted in
various fields and believed as a promising way to solve complex problems.
The ES normally consists of five parts, including the knowledge base, the
processing memory, the reasoning machine, the system interpreter, and the inter-
action interface. The knowledge base refers to the expertise stored, which can absorb
new knowledge. The knowledge is expressed in a way that the computer can under-
stand and apply. The memory part is used to store data for solving problems. The
reasoning machine is responsible for matching the data in the memory with the stored
knowledge in the systems. The interpreter of the system is to interpret the results of
the outputs from the system, such as to explain the reasoning results and evaluate the
correctness of them. The interaction interface is used for the operations by users.
The ES is usually sorted into four classes according to its working principles,
including the rule-based, the frame-based, the fuzzy logic-based, and the neural
network-based expert systems.
(1) Rule-based ES. The rule-based ES is one of the earliest ESs developed by
researchers. It usually consists of three parts, i.e., the database, control rule,
and the production rule. Because of this simple structure, the rule-based ES
presents many advantages, such as that it is very natural for a human to use.
(2) Frame-based ES. The frame-based ES was developed by Minsky and Winston
[20]. A frame is a developed data structure containing the concept related infor-
mation. This frame contains all information needed for solving problems, such
204 Expert Systems
as the concept name, the possible values of each attribute, and the procedural
information of the target problems. Thus, the frame-based ES owns the ability
to store and process descriptive and behavioral information. Due to this advan-
tage, the frame-based ES can deal with more complex problems, compared to
the rule-based system.
(3) Fuzzy logic-based ES. Integrating the fuzzy theory into systems, the fuzzy
logic-based ES is established. The fuzzy logic is the basis of the reasoning,
and it is used in the system to produce a rough fuzzy judgment for solving
problems. Because of the integration of the fuzzy logic, the fuzzy logic-based
ES has high robustness, and it can perform tentative and heuristic reasoning.
(4) Neural network-based ES. The neural network-based ES is based on the
rapid development of machine learning algorithms, and this feature makes
it fundamentally different from other ESs. With the integration of machine
learning algorithms, the neural-network-based ES can automatically acquire
the needed information, instead of manual processing. Along with the applica-
tion of machine learning in ESs, the reasoning performance of ESs is greatly
promoted.
The rapid development and wide applications of ESs provide effective tools to
solve complex problems, and accordingly, the demand and applications with ESs are
increasing in various domains, such as finance, education, public health, military,
transportation, and manufacturing industries. For instance, ESs are used in organiza-
tions for solving problems, such as financial analysis, consultant, strategic planning.
Bohlouli et al. [4] developed an ES for data information management in the orga-
nization and it was found that the developed ES could effectively solve complex
cases, where the management of them normally needs more time for human experts
to solve. Goksu [10] introduced ESs into the teaching and learning systems to create
a dynamic environment and help facilitate the learning experience. Abu-Naser and
Aldahdooh [1] integrated ESs into a diagnosis platform, which could identify disease
symptoms and provide an immediate, accurate response for the patients.
ESs have also been widely used in the construction industry. Chou [5] proposed a
linear model-based ES to estimate the cost of transportation projects. The proposed
system could effectively provide detailed information for the cost changing track for
the transportation projects. Similarly, Chou and Tseng [6] developed an ES for data
analysis and cost prediction in transportation projects, which could help process the
management information and produce preliminary budgets for the agencies. To deal
with the problem of construction disputes, Elziny et al. [8] presented an ES that can
evaluate the overall dispute settlement procedure for the construction projects, and
four cases studies were performed to test the effectiveness of the proposed system,
and it was found that the developed ES can efficiently facilitate the decision makings
in the dispute resolution process. Latif et al. [17] developed a knowledge-based ES
for the material cost control of building construction projects. The proposed system
could identify the impacts, the causes, and construct the relationships between the
influential factors and aims. A case study was conducted to present the advantage
of the proposed system for the improvement of the project monitoring and process
2 Expert Systems 205
control. Koulinas et al. [15] developed a simulation-based ES for the analysis of

schedule delays in construction projects. Yu and Chern [30] proposed an ES that
can provide a reasonable estimation on the tunnel deformation and present expert
advice on the decision making system and excavation procedures adopted during the
construction process. Grossauer et al. [11] developed an ES to realize the automatic
displacement monitoring for tunnel projects. Through case studies, the proposed
system presented promising results for effectively estimating delay risk in project
schedules and predicting the possibilities of in-time project completion. The advan-
tage of effectively solving complex problems has made the ES being widely applied
in the construction fields.
Risk management is also one of the most critical problems in construction engi-
neering and management, particularly for tunneling projects. Many influential factors
exist, making the underground conditions unpredictable and uncertain, which could
cause safety issues and make an adverse impact on tunneling projects [22]. However,
the adoption of the ES for safety and risk management in tunnel projects has not been
widely explored as expected. This may be due to the complex nature of underground
construction, where a high uncertainty exists during the tunneling process [29]. Very
limited information is available, and it is difficult to identify all the relevant factors
and to determine the causation between influential variables and the safety risk issues
for the tunneling projects. A hybrid ES, capable of collecting relevant information
and taking into account the experience and judgment of the experts for the safety
risk issues, can be expected to greatly promote the risk management in the tunnelling
fields.
3 Relationships Between BIM and Construction Safety
3.1 BIM for Safety
Safety is a major issue across global construction industries [28] noted that construc-
tion safety should be a prime consideration in the conceptual and preliminary design
phases, since the ability to influence construction site safety is progressively lost as
the project moves into the construction phase. Behm [3] investigated the relationship
between construction fatalities and design for construction safety. A total of 224
fatality investigation reports were examined, where the results suggested that 42%
of the fatalities reviewed were linked to design issues. Gambatese et al. [9] noted
that design for safety is one significant element in minimizing construction project
risk and enhancing worker safety through multi-level risk assessment and hazard
prevention mechanisms. Recent research indicates that designers need to indicate
knowledge and understanding of how risks and hazards affect health and safety can
arise in construction and how they can be avoided or reduced through design.
In the actual construction practices, contracts and regulatory requirements from
the Occupational Safety and Health Administration (OSHA) place the burden for
206 Expert Systems
worker safety solely on the constructor [3]. This approach is still widespread across
many countries, especially in developing countries. Zhou et al. [33] claimed that
designers began to appreciate each other’s concerns by working with their colleagues
who were responsible for the construction of the project in design-build firms.
Good ideas would be shared and applied in subsequent projects. Nevertheless, many
designers, who were not part of design-build firms, noted that they lack the skills and
training to address construction safety. The best construction practice highlights the
importance of multi-party collaboration for improved design for safety. However,
there is a lack of responsive tools and resources to assist designers in addressing
construction safety risks.
Managing construction safety is challenged due to the lack of efficient tools to
facilitate multi-party collaboration in construction management practices. To guar-
antee construction safety, a range of new tools have been developed. BIM is one of
the fastest-growing tools that has gained wide acceptance in the AEC/FM industry,
offering great solutions to address the inefficiencies of the current paper-based
processes. BIM is widely applied as a central technology for site hazard preven-
tion and safe project delivery. Hardin [13] noted that the implementation of BIM
in construction practice promised to improve the communication and collaboration
between participants through higher interoperability of data. Opportunities are iden-
tified to promote safety with the application of BIM by involving multi-partners. BIM
can be used as a support tool for constructability review, site management, safety
planning, and the identification and resolution of time–space conflicts.
3.2 BIM Data Reuse
Considered as the core of the construction safety management, safety risk iden-
tification is quite difficult given the complexity and size of the building systems
designed, such as the tunnel construction. Tunnel construction is characterized as
a highly complicated project with high potential risks, which integrates multiple
design disciplines, sub-projects, and interfaces. Information technology has proven
to be crucial to the success of a project by effectively controlling the safety risks.
Information flow from design to construction is critical and allows for design-build
and other integrated project delivery methods to be favored. BIM is regarded as
the information carrier of engineering characteristics, structure design, and structure
regulations. The information concerning structural component attributes, constraint
relations, and the interaction can be retrieved from a BIM model.
Most safety risks in tunnel construction are related to the engineering informa-
tion, and thus, the domain experts and engineers can identify risks and risk factors
through reading the conventional 2D engineering drawings. However, BIM itself
cannot identify the potential safety risks due to the lack of inference capacity, but
can provide the related information needed in the risk identification process.
The Industry Foundation Classes (IFC) standard has been developed as an open
standard for common data structures on information capturing and exchange. The IFC
3 Relationships Between BIM and Construction Safety 207
IfcGridPlacement
ObjectPlacement 1
ObjectPlacement
IfcLocalPlacement
IfcProduct
IfcPhysicalRepresentation
Representation
1
Represenataion IfcAbstractRepresentation
IfcMatrialRepresentation
Fig. 1 Relationships between the IfcProduct class and its related geometric properties
standard is supporting building elements, material, properties, geometry, and place-

ments, and addressing the broad scope of building design, engineering, construction,
and operation. It provides an extensive set of generic building object types, such as
beam, column, wall, and slab, with the associated attributes and properties. In addi-
tion, BIM offers numerous shape definition methods and means to depict relations
between objects. The intended role of IFC is to depict all information associated with
a building through construction and to support exchanges of this range of information.
Integration of the open IFC standard into BIM (i.e., so-called Open BIM) seeks
solutions to improve the productivity and efficiency of the building process by
enabling interoperability between AEC/FM and BIM applications. From the perspec-
tive of IFC-based BIM, all the building components are derived by IfcProduct enti-
ties. The IfcProduct entity refers to an abstract base class, consisting of several
hundred entities organized into an object-based inheritance hierarchy. The relation-
ships between this class and its related geometric properties are defined in Fig. 1. Two
main properties, ObjectPlacement and Representation, are included in one IfcProduct
entity. The ObjectPlacement property is used to determine its position information
either in the world coordinate system or the relative coordinate system. The repre-
sentation property aims to define the geometrical model of the building components,
including the physical objects (e.g., walls, beams, panels, and columns), abstract
objects (e.g., relationships, types, and groups), and material objects (e.g., concrete,
and steel). Those kinds of information are stored in open formats, enabling to make
the information accessible and readable in an easy manner.
3.3 Construction Safety Risks in Tunnels
Tunnel construction risk results from various kinds of uncertainty, which can lead to
economic loss, project delay, human injury, and damage. The key to the risk identi-
fication process is to identify risk events and their mutual relationships. Safety risks
are related to many influential factors in tunnel projects, such as structural forms,
208 Expert Systems
construction techniques, geological and hydro-geological conditions, and circum-

jacent environments. According to Ding et al. [7], safety risks in tunnel construc-
tion can be divided into three categories: technical, geological, and environmental
risks. The information related to construction technology, geological conditions, and
surrounding environment should be extracted from BIM. Herein, IFC plays a bridging
role between the BIM data and tunnel safety risks. Table 1 presents the relationships
between the extracted information and the related risks.
(1) Technical risks. Technical risks are the risks caused by the complexity of a
project and the limitations of the construction methods used. For instance, the
risk of water gushing often occurs in tunnels with the shield method adopted,
while the risk of tunnel collapse occurs frequently when using the cut and cover
method.
(2) Geological risks. Geological risks are the risks caused by harmful geological
conditions, such as caves, water capsules, and harmful gases. For example,
underground caves can be encountered in tunnel excavation work. The presence
of harmful gases, such as methane, can result in casualties in the working site.
(3) Environmental risks. Environmental risks are the risks caused by construction
activities that lead to detrimental environmental impacts. For example, tunnel
construction can cause tilting, cracking, or even the collapse of surrounding
structures/buildings, and then, the leakage or fracture of underground pipelines
is very likely to occur.
4 Knowledge Base Development for Construction Safety
Historical resources provide massive prior knowledge for safety identification in

tunnel construction. Beckman [2] indicated that knowledge, experience, and expertise
should be formalized, distributed, shared, and applied before they can be transformed
into valuable organizational assets. In this chapter, an integrated knowledge base,
consisting of fact base, rule base, and case base, is established to systematize the
fragmented knowledge in tunnel construction through knowledge structuralization,
expression, and explanation. Figure 2 illustrates a conceptual framework for the
development of the integrated knowledge base.
4.1 Knowledge Resources
The prior expert expertise and historical cases can provide a wealth of knowledge
resources for safety risk analysis and identification. With the development of tunnel
construction practice worldwide, large amounts of scattered knowledge are accumu-
lated, including explicit and tacit knowledge. Sherehiy and Karwowski [26] claimed
that both explicit and tacit knowledge was equally important in knowledge base
4 Knowledge Base Development for Construction Safety 209
Table 1 Relationships between the extracted information and the related risks
Risk category BIM elements Engineering Risk factors Safety risks
parameters
Technical risks Station contour Outline length Large foundation Instability of steel
Standard width pit width strut
Number of tip
wells
Bored pile Pile diameter Poor effect of Seepage from side
Distance waterproofing wall of foundation
between piles Short seepage pit
paths
Jet grouting Pile diameter
pile Distance
between piles
Bored pile Pile length Bored pile not Flowing sand from
State of embedded into the bottom of
embedded into rocks foundation pit
rocks Embedded length Instability of
ratio < 0.7 support system
Embedded Soft ground not
length ratio reinforced
Supporting Steel strut stories Over-limited Instability of steel
system slenderness ratio strut
Structure Construction Special restriction Construction
design method of construction method risk
Structure safety method
grade
Foundation pit Reinforcing Soft ground not Flowing sand/soil
bottom stratum scope reinforcement from bottom of
Reinforcing foundation pit
depth
Geological risks Foundation pit Soil layer Soft ground of Sand/soil gush or
bottom stratum foundation pit upheaval from the
Artesian water Artesian water bottom bottom of
level Harmful geological foundation pit
conditions within Seepage of harmful
Harmful Silt stratum the scope of gas
geological Harmful gas excavation
conditions
Cave
Environmental Building Structural type Close to foundation Building tilt or
risks Foundation type pit cracking
Shallow foundation
Building height embedment
Foundation Bad structure
depth health
(continued)
210 Expert Systems
Table 1 (continued)
Risk category BIM elements Engineering Risk factors Safety risks
parameters
Distance to the
foundation pit
Underground Pipe type Close to foundation Underground pipe
pipe Pipe material pit cracking and
Pressure-bearing seepage
Distance to the pipes
foundation pit Rigid joint
Pipe direction
Pipe burial depth
Pipe connection
style
Standards Experience Fact Breakdown

Fact Base
Explicit
Mannuals Knowledge Cases
Rule Expression
Reports Tacit Simulations

Rule Base
Knowledge
Case Explanation
Case Base
Knowledge Resources Risk Mechanism Knowledge Base
Fig. 2 Conceptual framework in knowledge base development
construction, where a balance of both was key to the development of an effective

knowledge-based system.
During the knowledge base construction process, knowledge acquisition aims to
obtain relevant knowledge from the accumulated knowledge resources, which plays a
very important role. This involves developing new knowledge contents and updating
old contents through socialization, externalization, internalization, and combination.
Risk mechanism analysis makes efforts to reveal potential risks, risk factors, and
their causal relationships. The description of the two types of knowledge resources
is presented below.
1. Explicit knowledge. Explicit knowledge can be easily codified and transmitted
between individuals in documented and organized forms, such as standard spec-
ifications, technical manuals, and research reports. The initial risks and their
factors can be achieved by knowledge structuralization, expression, and expla-
nation. Taking the hazard of “water gushing at foundation pit” as an example,
according to the standards, the water gushing hazard refers to an accident that
occurs after the loose soil granules that are saturated by the underground water
Safety risk categories and events Safety risk factors
Construction complexities Safety-based knowledge Flowing sand of foundation pit (f1) Retaining pile embedded
into rocks (e1)
Large deformation of steel strut (f2)
Water head difference of
Complexity in Large deformation of steel strut (f3) foundation pit (e2)
Technical Risks
construction techniques
Explicit Knowledge Instability of support system (f4) Geological conditions of
foundation pit (e3)
Complexity in Instability of foundation pit (f5)
Geological Risks Reinforcement state of
geological conditions foundation pit (e4)
Unpredicted underground caves (f6)
Tacit Knowledge Embedded length ratio of
Complexity in Unpredicted harmful gases (f7)
Environmental Risks bored pile (e5)
surrounding environments
Large ground settlement (f8) Bed thickness of
foundation pit (e6)
Building tilt or cracking (f9)
Underground pipe cracking (f10)
Fig. 3 Results of risk mechanism analysis in tunnel construction
and when hydrodynamic pressure is equal to or greater than the submerged unit
weight of soil, given the existence of water head difference. Consequently, the
terms or keywords concerning the specific risk “water gushing at foundation
pit” can be preliminarily achieved.
2. Tacit knowledge. Tacit knowledge is generally developed by individuals
through years of working experience. Dialogues and communications among
individuals are basic means of knowledge sharing. Therefore, questionnaires,
depth interviews, and group decision-making methods can be used to reveal and
transfer the tacit knowledge in a reusable manner. Also, numerous researchers
built simulation models for the safety analysis, serving as a valid reference for
the relation discovery within various influential factors.
With the support of safety-based knowledge, including explicit and tacit knowl-
edge resources, the projects in tunnel construction are analyzed through the risk
mechanism process. Figure 3 lists the results of risk mechanism analysis in tunnel
construction. There are mainly 10 kinds of safety risks in total, among which
the hazard occurring in “Flowing sand of foundation pit (f 1 )” is considered as a
commonly occurring risk event. A total of six risk factors are identified to contribute
to the hazard occurrence. A case study concerning this safety risk is presented later
in this chapter.
4.2 Knowledge Representation
Regarded as a core component of an ES, the knowledge base contains a summary of

expert expertise and guidelines, which can be used to implement the smart control by
providing practical supports to the identified problems. In this research, an integrated
knowledge base is established, which is consisting of the fact base, rule base, and
case base.
(1) Fact base
Fact base uses a fact table to store the values of each risk factor, including
the fact code, attribute, and credibility factor (CF(e)). Different values of a
212 Expert Systems
risk related parameter lead to the different credibility degree of the factor
being considered as the evidence in accidence occurrence. CF(e) takes a value
of zero when the evidence is unreliable, while CF(e) takes a value of one
when the evidence is completely reliable. The more difficult to become the
evidence, the smaller its CF(e) is. In accordance with the safety-related facts
in tunnel construction, the facts are divided into two types, namely certain
and uncertain facts. Accordingly, the method to determine the corresponding
CF(e) of different types would vary as follows.
• Certain fact. A certain factor refers to the risk factor which has discrete
values. For a risk factor with a discrete value, its fact CF(e) is a discrete
value. Taking the state of “Retaining pile not embedded into rocks (e1 )” as
an example, CF(e1 ) is represented by “1” in the “Not embedded into rocks”
state, otherwise as “0”, as seen in Eq. (1).

1 N ot mbedded into r ocks
C F(e1 ) (1)
0 Embedded into r ocks
• Uncertain fact. An uncertain factor refers to the risk factor which has an
interval in measurement due to randomness, fuzziness, and greyness. For
a risk factor with a continuous value within an interval, its fact CF(e) can
be represented by a sectional function. The sectional threshold value can
be obtained from theoretical calculations or empirical formulas. Taking the
state of “Water head difference between the inside and outside of a foun-
dation pit (e2 )” as an example, the water head difference x is a continuous
value, which has an approximately linear relationship with the evidence
credibility of e2 . Therefore, the sectional function expression for CF(e2 ) is
given by Eq. (2).
⎧
⎨ 0.2 x < 2m
C F(e2 ) = 0.8 x−2 + 0.2 2m ≤ x < 4m (2)
⎩ x−2
0.8 x ≥ 4m
During the safety risk identification in tunnel construction, both certain and uncer-
tain facts exist in the establishment of the fact base. Taking the risk “Flowing sand
of foundation pit (f 1 )” as an example, the safety-related fact base is constructed
according to explicit and tacit knowledge, as seen in Table 2.
(2) Rule base
In tunnel construction projects, safety risk identification knowledge involves
various kinds of theories and practical knowledge, such as construction
methods, engineering geology, structural design, accident analysis, and reli-
ability theory. Most of the knowledge is empirical, displaying patterns of
diversity, hierarchy, inheritance, and uncertainty. Existing knowledge repre-
sentation methods consist of predicate logic, semantic network, production
Table 2 Safety-related fact base for the risk “Flowing sand of foundation pit (f 1 )”
Risk factors Fact code Certain? Fact attributes CF(ei )
Retaining pile FID.025 Yes Embedded 0
embedded into rocks Not embedded 1.0
(e1 )
Water head difference FID.104 No <2 m 0.2
of foundation pit (e2 ) 2–4 m 0.8(x − 2)/(4 − 2) + 0.2
≥4m 0.8
Geological conditions FID.100 Yes Medium-coarse sand 0.6
of foundation pit (e3 ) Fine sand 0.7
Silt sand 0.8
Powder sand 0.5
Others 0.2
Reinforcement state of FID.101 Yes Reinforcement 0
foundation pit (e4 ) Non-reinforcement 1.0
Embedded length ratio FID.026 No <0.4 0.6
of the bored pile (e5 ) 0.4–0.7 0.6(0.7 − x)/(0.7 − 0.4)
+ 0.2
≥0.7 0.4
Bed thickness of FID.102 No <400 mm 0.8
foundation pit (e6 ) ≥400 mm 0.8e−(x − 400)/400
rule, and frame method. Owing to that the production rule “IF (premise)
THEN (conclusion)” provides a powerful tool for knowledge representation
and reasoning under uncertainty, the rule can be employed to describe the
empirical knowledge for safety risk identification.
However, due to the lack of classification and relevance, this type of rule is
likely to cause combination explosion and low reasoning efficiency. Therefore,
an extended rule representation as seen in Eq. (3) is adopted, with the hierarchy
and uncertainty of rules being fully considered. The causal relationship between
the premise (e) and the conclusion (h) is represented by Eq. (4).
I f e T hen h (C F(h, e), λ) (3)
C F(h) = C F(h, e) × C F(e) ∀C F(e) ≥ λ (4)
where λ refers to the rule threshold. The rule can be activated if and only if
the rule evidence CF(e) ≥ λ. Generally, λ is defined within a range [0.5, 1.0],
depending on the importance of the project.
The rule credibility CF(h, e) with a range [0, 1] is related to the credibility
degree of the rule. CF(h, e) can be calculated by Eq. (5).
214 Expert Systems
Case code Premise Conclusion

Technical solutions
Construction methods Construction Project Responses &
Processing procedure
Environmental features (C1) Solutions (C4)
Implementation result
Geological features
Time & location Casualties
Triggering event Accident Economic losses

Accident (C2) Cases in tunnel
Accident symptoms Consequence (C5) Responsibility
construction
Event description Public safety
Technical problem Application range

Problems & Reasons Case Application
Environmental problem Application frequency
(C3) (C6)
Managerial problem Case evaluation
Fig. 4 Simplified structure for the case representation in tunnel construction

P(h|e )−P(h)
P(h|e ) ≥ P(h)
C F(h, e) = 1−P(h)
P(h|e )−P(h) (5)
P(h)
P(h|e ) < P(h)
where P(h) refers to the prior probability of the conclusion h, and P(h, e) refers
to the conditional probability of the conclusion h when the premise e is true.
(3) Case base
A large number of accident records are accumulated, which provide prior
knowledge for risk identification in tunnel construction. By matching the char-
acteristics of a specific project with cases in the case base, the very similar
accident event can be detected. Compared to rules, we do not need to construct
explicit rules as seen in Eq. (2) for case representation. Meanwhile, the case
base is an open system, which is easy to maintain. There is no need to perform
dependency and consistency checks when adding new cases into the case base.
To facilitate the efficiency during the case matching process, cases are required
to be structured in accordance with the characteristic of domain cases. There
are numerous methods for the case representation, including feature vectors,
object, frame, and category representation methods. Accident cases in tunnel
construction generally are characterized by multi-level and multi-attribute
patterns. Therefore, a combination of frame and category representation
methods can satisfy the demand, which is adopted to represent the acci-
dent cases in tunnel construction. A complete case record consists of several
attributes, including Construction Project Data (C 1 ), Accident Data (C 2 ), etc.
Each attribute can then be further refined. A simplified structure for a tunnel
case is presented in Fig. 4.
4.3 Knowledge Database Structure
Initially, either rules or cases are stored in text forms, which are likely to result in
inconvenience in retrieval and traversal operations for knowledge-based reasoning.
Due to the simple structures, high independence, and strong description ability, the
relational database can be widely employed to store various kinds of knowledge.
Rule
Rule premise Rule code A10 Rule conclusion
Rule premise desc A100
R_Premise code A10 R_Conclusion code A10
Worksite type A10
R_Premise name A100 R_conclusion desc A100
Construction method A10
Sign A10 CF conclusion A100
Related Rule type A10 Reasoning
Value A100 Memo A100
Rule group number A10
Fact R_Premise CF N4.2 Rule CF N4.2
Reference Rule threshold N3.1 R_Conclusion code <pi>
Fact code A10 R_Premise code <pi>
Conclusion code A10
Fact attribute A100
Y/N A100 Rule code <pi>
Explained
Fact CF N4.2
Fact SIM N5.2
Case
Fact code <pi> Case conclusion
Reference Case premise Case code A10
Case premise desc A100 C_Conclusion code A10
C_Premise code A10
Worksite type A10 C_conclusion desc A100
Construction data A100
Construction method A10 SIM conclusion A100
Accident data A100 Related Case type A10 Reasoning Risk name A200
Problem reasons A100
Case group number A10 Consequence A200
Case SIM N5.2
Case SIM N5.2 Response solution N5.2
C_Premise code <pi> Case threshold N3.1 Case application A200
Conclusion code A10
C_Conclusion code <pi>
Case code <pi>
Fig. 5 Database structure in the knowledge base
Therefore, to improve data matching and reasoning efficiency, rules and cases in
tunnel construction can be stored in a relational database, which can be retrieved
using a structured query language (SQL). The database structure of the knowledge
base is shown in Fig. 5.
Both rule and case premises are stored in the fact table, aiming to ensure the
consistency of terms in the knowledge base. The rule premise table describes the
various states of risk factors. The rule table describes a basic structure of a production
rule, representing the relation between the case premise and its conclusion. The case
premise table describes the various attributes of case records. The case table describes
the structure of a case, representing the relation between case characteristics and their
consequences. The case consequence table describes the consequence of the accident
case and provides explanation information for the risk identification results and the
corresponding control measures.
5 BIM-Based Risk Identification System (B-RIES)
5.1 System Architecture
To facilitate knowledge sharing and communication among dispersed clients and

experts during the safety risk identification, the merging of BIM and ESs provides a
solution to address the aforementioned deficiencies. Thus, a BIM-based Risk Identifi-
cation Expert System (B-RIES) is developed to systematize the fragmented explicit
and tacit knowledge in tunnel construction. A simplified logical framework of B-
RIES is presented in Fig. 6. B-RIES is a distributed multi-layer system consisting of
three main layers, namely the data layer, standard and protocol layer, and application
layer. The description of each layer is presented below.
216 Expert Systems
Application layer
BIM Extraction Knowledge Base Management Risk Identification

Subsystem (BES) Subsystem (KBMS) Subsystem (RIS)
Project Creation Fact Base Management Parameter Input

BIM Element
Recognition Rule Base Management
IFC Browser Case Base Management Hybrid Reasoning

Documentation
Parameter Extraction Management Interpretation Report
Standard protocol layer
Role-based Regulatory Security

IFC Standard
Access Protocols Protocols Protocols
Data layer
MS SQL
JAVA Tomcat
Server
Fact Base Rule Base Case Base Project Base
Fig. 6 The logical framework of B-RIES
(1) Data layer. B-RIES is developed in the JAVA programming language and
running in the Tomcat container. Mainstream data warehouses consisting of
the Microsoft SQL Server and Oracle database system are compatible with
the system. In addition, four built-in databases, namely Fact Base, Rule Base,
Case Base, and Project Base, are used to provide data support for the intelligent
inference for the safety risk identification process.
(2) Standard and protocol layer. This layer provides a unified platform solution
for data communication, processing, and the interface between the system and
external environments using four main types of protocols. To be specific, the
IFC standard is employed to extract the information of model components
from the input BIM of a specific project. The role-based access protocols aim
to set up authorities among dispersed clients, experts, and maintainers. The
regulatory and security protocols are used to ensure the security of the network
environment, protecting the system against external attacks.
(3) Application layer. This layer provides all functions and services for the system
application, mainly consisting of three subsystems, including BIM Extraction
Subsystem (BES), Knowledge Base Management Subsystem (KBMS), and
Risk Identification Subsystem (RIS). Specifically, BES aims to extract the
engineering parameters of BIM models, KBMS aims to acquire, represent,
5 BIM-Based Risk Identification System (B-RIES) 217
store, and update the risk-based knowledge in tunneling fields, and RIS aims
to calculate the risk occurrence probability and propose control measures in
real-time. Each subsystem is further composed of several functional modules.
B-RIES can be used to automatically determine construction safety risks and the
corresponding control measures. B-RIES is a Browser/Server (B/S) system. Figure 7
illustrates the overall workflow for the safety risk identification process in B-RIES.
In general, B-RIES will go through three steps during the safety risk identification
process, as elaborated below.
BIM
Convert
IFC XML
Geometry Structure information

Property Environment information
Relation Resource information Step 1

Export
BIM Inputs
Extracted
IFC Browser Information
Fact Matching
Fact Base
Credibility
Factor CF(e)
Reasoning
Strategies
CBR RBR
Add new cases Add new rules

Case matching
Case Base Rule Base
Rule matching Step 2
Optional cases Optional rules
Knowledge
Based
Similarity
(E, C) ≥ ψ ?
No No
CF(e) ≥ λ ? Reasoning
Expert
Yes Feedback Yes
Rule execution and
Target case
No conclusions
Consistent?
Yes
Output:
Calculated risk results
Control
Risk Category
Measures
Technical risks Reinforcement measures
Geological risks Management measures
Step 3
Environmental risks Prevention measures Results Output
Forewarning Extremely Very
Risk Ranking Dangerous Dangerous
Releasing
Dangerous Safe
Fig. 7 The overall workflow for the safety risk identification process in B-RIES
218 Expert Systems
5.2 Engineering Parameters Extraction
IFC is a schema developed to define an extensible set of consistent data representa-

tions of building information for exchange between AEC/FM software applications.
All IFC models provide a common general building spatial structure for building
elements, which can be used to organize all object information into a hierarchy of
Project-Site-Building-Storey-Space. BIM is considered as an information carrier of
engineering parameters which are related to construction safety risks, and the IFC
standard plays a bridge role between BIM models and construction safety risks.
When a BIM model of a specific project is input into B-RIES, the IFC standard
is employed to extract the information of model “components”, including geometry,
property, and relation parameters. Then, the parameter information is exported into
the Extensible Mark-up Language (XML) format which is compatible with B-RIES.
In the meantime, the recognition results of “components” saved in XML formats are
in correspondence with the fact base. When the results are matched with the fact base,
the “components” in XML files correspond with the column “fact name” in the risk
fact table. Consequently, the fact base is matched to calculate evidence credibility
CF(e) of the risk-related engineering parameters. In this way, B-RIES can extract
original and objective engineering parameters from BIM models automatically.
5.3 Knowledge-Based Reasoning
The knowledge reasoning mechanism endows B-RIES with artificial intelligence

(AI) capabilities. Generally, the reasoning in an ES consists of two approaches,
namely rule-based reasoning (RBR) and case-based reasoning (CBR). RBR is known
as a reasoning technique with powerful deductive inference capabilities, which can be
employed to deal with complicated problems, such as goal programming, scheduling,
and budgeting. However, an RBR system is required to traverse the entire rule base
during every reference process, leading to a long retrieval time and a low efficiency in
rule retrieval, especially when the rule base is very large. In contrast, a CBR system
attempts to seek an approximatively similar case by using an analogous reasoning
technique. The reasoning process in CBR is fast and efficient at the expense of in-
depth analysis. Also, a CBR system has self-learning and self-improvement ability
by adding new cases continually as the system grows. However, due to a lack of
deductive inference capacity, the CBR system has deficiencies in conformity with
strict logical inference, leading to poor interpretability of reasoning results.
In tunnel construction, complex interactions exist among various risk factors,
which produce a high demand for both rule-based and case-based knowledge during
risk identification. The integration of RBR and CBR techniques provide an optimum
solution to address the above dilemma. It is beneficial to improve the flexibility
and interpretation of system reasoning capabilities simultaneously. Thus, a hybrid
reasoning approach composed of four main procedures is proposed in B-RIES.
5 BIM-Based Risk Identification System (B-RIES) 219
(1) CBR: Initially, various historical safety-related cases in tunnel construction

projects are accumulated and stored in the case base. The objective of CBR is
to discover the most similar case as the target case from the case base. At first,
the risk related information regarding engineering parameters is obtained from
BIM and then entered into B-RIES as evidence input. Then, some alternative
cases are selected after being matched with the case base. Finally, the target
case is obtained if and only if the similarity is less than the given threshold ψ.
Assuming each case has n attributes to describe the information of engineering
parameters, the specific project is denoted by E = (CF(e1 ), CF(e2 ), …, CF(en )),
and one optional case is denoted by C = (CF(c1 ), CF(c2 ), …, CF(cn )). The
Euclidean distance measures the similarity to calculate the similarity between
E and C. As seen in Eq. (6), wi refers to the weight of the ith parameter based
on the expert estimation and construction practice. The threshold ψ is usually
given a value ranging from 0 to 0.5.

n

Similarit y(E, C) =
wi [C F(ei ) − C F(ci )]2 ≤ ψ, i = 1, 2, . . . , n
i=1
(6)
where CF(ei ) (or CF(ci )) refers to the credibility factor of the ith parameter
for the specific project (or the optional case) while being matched with the fact
base.
(2) RBR: RBR is incorporated into the knowledge reasoning mechanism, aiming
to find out valid rules which can be executed in the reasoning process. At first,
some alternative rules are selected from the rule base using rule matching.
Next, the valid rule is determined if and only if CF(e) is greater than the
given threshold λ. In general, the premise evidence e is a combination of risk
factors ei (i = 1,2, …, n), including disjunction, conjunction, and weighted
combinations. Accordingly, CF(e) is calculated by Eqs. (7)–(9), respectively.
Finally, the valid rule is executed to calculate the credibility degree of the
conclusion of the risk event using the former Eq. (4).
C F(e1 ∨ e2 ∨ . . . ∨ en ) = max(C F(e1 ), C F(e2 ), . . . , C F(en )) (7)
C F(e1 ∧ e2 ∧ . . . ∧ en ) = min(C F(e1 ), C F(e2 ), . . . , C F(en )) (8)

n
C F(e1 (w1 ) ∧ e2 (w2 ) . . . ∧ en (wn )) = wi × C F(ei ) (9)
i=1
where CF(e1 ∨ e2 ∨ …∨ en ) represents the disjunction of various pieces of

evidence, CF(e1 ∧ e2 ∧…∧ en ) represents the conjunction of various pieces
220 Expert Systems
of evidence, and CF(e1 (w1 ) ∧ e2 (w2 ) …∧ en (wn )) represents the weighted

conjunction of various pieces of evidence.
(3) Expert intervention and feedback: Given two approaches (i.e., CBR and
RBR) are included in the knowledge reasoning mechanism, the result calcu-
lated by one approach can be testified by the other. However, when the results
are not consistent between these two approaches, domain experts should be
involved to solve the problem. Besides that, existing rules stored in the rule
base might be modified according to the actual situation. Later on, the verified
project cases or records can also be added to the case base as new cases. With
a growth of the case base and rule base in B-RIES, the accuracy and reliability
of the system inference mechanism will be continuously improved with the
development of the system implementation.
(4) Reasoning strategies: As aforementioned, the computational efficiency of the
RBR approach can be reduced, especially when the size of the rule base is very
large. To keep a balance between computational efficiency and high accuracy,
the reasoning strategy can be adjusted, depending on the timeline of the system
implementation. Specifically, in the initial stage of the system implementation,
the size of rules (or cases) stored in the rule base (or case base) is relatively
small due to a lack of sufficient data. Also, the reliability of the initial rules
needs to be verified and improved by comparing them with real cases. Thus,
RBR and CBR can work in a parallel way in this situation, contributing to
the growth of the integrated knowledge base using expert intervention and
feedback. When the knowledge base grows larger, these two approaches can
work in sequence for the consideration of computational efficiency. Generally,
CBR can be activated first to save time and avoid the low efficiency in rule
retrieval. Only the target case is not reached in CBR, the RBR would then be
activated.
5.4 Risk Analysis and Control
Based on the calculated results through the knowledge reasoning mechanism, the
safety risk analysis and corresponding control measures can be worked out. To eval-
uate the degree of the identified safety risk, during the CBR approach, the degree
of the identified risk is equal to the risk level in the target case. During the RBR
approach, the conclusion credibility CF(h), which can be calculated by Eq. (2), is
used as an indicator for the risk evaluation.
With regard to CF(h) with a range of [0, 1], we divide its calculated value into
the following four levels, namely “I (Safe, [0, 0.25)), II (Dangerous, [0.25, 0.5)),
III (Very Dangerous, [0.5, 0.75)) and IV (Extremely Dangerous, [0.75, 1.0))”. These
four levels are defined by four different colors, namely, “Green, Yellow, Orange and
Red”, respectively. According to the risk identification results, some corresponding
safety control measures can be determined, and early-warning signals can be released
in a visualization scenery.
6 System Application 221
6 System Application
To relieve the pressure of urban traffic jams across the Yangtze River, the construc-
tion of Wuhan Metro Line Two (WMLT) in the city of Wuhan, China, started on
August 28, 2006. The 27.7 km route, with 21 stations and a total investment of nearly
US $3.2 billion, passed underground on a northwest-southeast alignment between
the Hankou and Wuchang districts. Due to complexities in construction environ-
ments, high potential safety risks existed during the excavation progress of WMLT.
In this section, B-RIES is implemented to identify safety risks at the pre-construction
stage of tunnel construction, aiming to provide guidelines for safety assurance at the
construction stage. A realistic case in the Mingdu station located at WMLT in China
is used as an example for demonstrating the system application.
6.1 Project Profile
The Mingdu station is an underground 2-story station. This station was started on
November 26, 2008, with an outlined length of 241.3 m, an average width of 18.88 m,
and a total floor area of 11,932.1 m2 . A foundation pit with a depth of around 15 m was
constructed using the cut and cover excavation method. The aerial view of the BIM
model of Mingdu station is shown in Fig. 8. The retaining structure was composed
of bored piles and jet grouting piles. Two high-rise buildings were adjacent to the
deep foundation excavation. Specifically, the Baoli Huadu Building was located in
the north of the foundation pit, while the Tibetan Middle School was located in the
south. The excavation space was limited due to the narrow working site. The working
site of the foundation pit excavation of the Mingdu station is shown in Fig. 9.
Fig. 8 Aerial view of BIM for the Mingdu station in Wuhan, China
222 Expert Systems
(a) (b)
Fig. 9 Working site of the excavation of the Mingdu Station: a steel support; b foundation pit
6.2 Safety Risk Identification
Following the aforementioned safety risk identification procedures in B-RIES, the

results of safety risk identification at the Mingdu station were conducted, as shown
in Table 3. There were six safety risks identified. To be specific, one risk (i.e., C.101)
was rated in a level of III (Very Dangerous), since CF(h) of this risk fell into a range of
[0.50, 0.75]. Two risks (i.e., C.105 and C.109) were rated in a level of II (Dangerous).
Three risks (i.e., C.113, C.203, and C. 309) were rated in a level of I (Safe). In this
case, the top risk (i.e., “water gushing at foundation pit bottom” (C.101)) was taken
as an example to illustrate the detailed computational procedures.
Table 3 Safe risk identification results at the Mingdu Station

Risk type Risk code Rule code Risk description Conclusion credibility
CF(h)
Technical risk C.101 R1-26 Water gushing at 0.608
foundation pit
C.105 R1-13 Instability of retaining 0.466
structure
C.109 R1-25 Instability of steel 0.256
structure
C.113 R1-30 Instability of the 0.216
support system
Geological risk C.203 R2-16 Harmful geological 0.206
conditions within the
scope of the excavation
Environmental risk C.309 R3-02 Underground pipe 0.196
cracking and seepage
(1) Information extraction for engineering parameters

BIM (in the format of Revit) of the Mingdu station was first input into B-
RIES. The BIM model was translated into an IFC model, where all the defined
objects were composed of the relevant object type and the associated geom-
etry, relation, and properties. Some risk-related parameter information can be
extracted from the IFC files. Next, the captured engineering parameters were
stored in an XML format which can be read by the B-RIES system. Finally, the
captured engineering parameters information was matched with the fact base
to calculate fact credibility CF(e) of the risk-related parameters.
Concerning the risk of “water gushing at foundation pit bottom”, the fact
matching results for relevant risk factors are presented in Table 4. Some risk
factors were easy to be identified directly during the BIM extraction process,
such as the water head difference of the foundation pit (e2 ), the embedded
length ratio of the bored pile (e4 ), and the bed thickness of the foundation pit
bottom (e6 ). However, the information relating to the other risk factors can
be extracted through certain intermediaries. Taking the state of the retaining
pile embedded in rocks (e1 ) as an example, the strata at the bottom of the
retaining pile acted as an intermediary, and its properties were extracted. If the
bottom strata were identified to be the rock, it meant that the retaining pile was
embedded in rocks, and then the fact credibility of CF(e1 ) should be recognized
to be 1.0. Otherwise, the fact attribute should be termed “Not embedded”, and
then the fact credibility of CF(e1 ) should be recognized as 0.
(2) CBR calculation
According to the extracted risk-related parameter information (see Table 4),
the case base was matched until the target case was figured out during the case
retrieval. Equation (6) is used to compare the similarity between the specific
case and the potential target case. With regard to the risk of water gushing,
the target case was identified to be the Guicheng station, since the similarity
between the specific project and the target was calculated to be 0.192 (that is
less than the given threshold ψ = 0.3). The case matching result is presented
in Fig. 10. Actually, the Guicheng station was one station of the Guanzhou
metro system in Guangzhou, China. On July 31, 2008, great quantities of water
swarmed into the Guicheng station under construction, resulting in serious
Table 4 Fact matching results for the risk of water gushing

Fact code Fact description Fact attribute Fact credibility
FID.025 (e1 ) State of retaining pile embedded in rocks Not embedded 1.0
FID.036 (e2 ) Water head difference of foundation pit 3m 0.6
FID.100 (e3 ) Geological condition Silt 0.8
FID.101 (e4 ) Reinforcement state of the foundation pit Not reinforced 1.0
bottom
FID.126 (e5 ) Embedded length ratio of bored pile 0.8 0.4
FID.128 (e6 ) Bed thickness of foundation pit bottom 510 mm 0.6
224 Expert Systems
Similarity (E, C) = 0.192

1.0
0.8
0.6
0.4
0.2
0.0
FID.025 FID.036 FID.100 FID.101 FID.126 FID.128
Fact Code
Fig. 10 Case matching result for the risk of water gushing
ground subsidence and construction delay. The hazard level of this accident
was evaluated to be a level of III during the accident investigation. Based on
the causal relationship and control process of the accident in the target case,
the project engineers can have a deep understanding of the potential safety risk
at the Mingdu station. Also, the risk level is identified and some relevant safety
control measures can be proposed.
(3) RBR calculation
Based on the aforementioned procedures in RBR, the rule base was first
matched to select suitable rules. It is known that the worksite type was
“station"(represented by No. 10) while the construction method was “cut and
cover” (represented by No. 1110). Table 5 listed some alternative rules for the
risk of water gushing. Using Eqs. (7)–(9), CF (e) of R1-11 and R1-14 was
less than the threshold λ, and therefore, both were discarded. Next, R1-26 was
activated and CF(e) was calculated to be 0.76 as follows:
6 System Application
Table 5 Rule matching results for the risk of water gushing

Rule Rule premise description Worksite Construction Rule Threshold
code type method credibility λ
R1-11 ¬FID.025∨ ¬FID.036∨ ¬FID.101 10 1110 0.5 0.9
R1-14 FID.025 ∧ FID.036 ∧ FID.100 ∧ FID.101 ∧ FID.126 ∧ FID.128 10 1110 1 0.8
R1-26 FID.025(0.2) ∧ FID.036(0.2) ∧ FID.100(0.3) ∧ FID.101(0.1) ∧ FID.126(0.1)∧FID.128(0.1) 10 1110 0.8 0.5
225
226 Expert Systems
C F(e) = 1.0 × 0.2 + 0.6 × 0.2 + 0.8 × 0.25 + 01.0 × 0.3

+ 0.4 × 0.1 + 0.6 × 0.1 = 0.76
Thus, CF(e) > λ = 0.5, indicating that the rule R1-26 was valid. Accordingly,
Eq. (4) was used to calculate the conclusion credibility as follows:
C F(h) = C F(h, e) × C F(e) = 0.76 × 0.8 = 0.608
As a result, the risk of “water gushing at foundation pit bottom” was rated at
a risk level of III (Very dangerous).
(4) Risk identification and report
The results can be testified when RBR and CBR worked in a parallel way. In this
case, the results were consistent between these two approaches. The safety risk
analysis and control measures can then be determined, including the risk type,
description, location, possible consequences, risk level, and control measures.
The report of the risk of “water gushing at foundation pit bottom” at the Mingdu
station is shown in Fig. 11. It is clear to see that the possible consequences
of this risk event include the instability of the retaining structure and the steel
structure, as well as the casualties of workers. As a response, retaining piles
should be embedded into rocks, the water head difference should be reduced,
and the bottom of the foundation pit should be reinforced.
Fig. 11 Report on the water gushing risk at the Mingdu station

Fig. 12 Risk event of water gushing at the Mingdu Station: a location of water gushing, and b a
precipitation well
6.3 Implementation Effects
According to the risk identification results from B-RIES, the designers gained insights
into a better understanding of the safety risk in the construction of the Mingdu station.
The contractor was suggested to strictly implemented the work of dewatering on
the working site during the construction stage, controlled the water head difference
between the inside and outside of the foundation pit, and reinforced the foundation pit
bottom to reduce the risk limit. Also, emergency supplies, equipment, and personnel
were prepared in advance to cope with the risk events which were very likely to
occur.
On September 14, 2009, a risk event of water gushing was witnessed at the
construction site of the Mingdu station, as shown in Fig. 12a. The volume of the
gushing water was small, and its impacts on the safety of nearby retaining piles
and adjacent buildings were limited, since some prevention measures were adopted
ahead of time. Subsequently, a precipitation well was built up at a location of the
water gushing and continued to be effective until the main structural work of the
Mingdu Station was completed, as shown in Fig. 12b. As a matter of fact, with effec-
tive safety control measures implemented, the construction of the Mingdu station
went smoothly, and no accidents were reported during the excavation construction
process. Finally, the main underground structural work was completed without any
project delay on August 9, 2010.
7 Summaries
This chapter describes an innovative approach that integrates BIM and ESs to address
deficiencies in the traditional safety risk identification in tunnel construction. B-RIES
is consisting of three main built-in subsystems, namely BIM extraction, knowledge
base management, and risk identification subsystems, providing real-time support
228 Expert Systems
for decision making in the tunnel risk management. Firstly, engineering parameter
information related to risk factors should be acquired from BIM models of a specific
project, where the IFC standard acts as an intermediary between the BIM data and
risk-related information. Then, an integrated knowledge base, composed of fact, rule
and case bases, is built up to systematize the fragmented explicit and tacit knowledge
for knowledge sharing and communication among dispersed clients and domain
experts. A hybrid reasoning approach with CBR and RBR combined can be used to
improve the flexibility and compatibility of the system reasoning capacity. Finally, the
results of safety-related risks and their occurrence probability and control measures
can be obtained in real-time. B-RIES can be used by practitioners in the industry as
a decision support tool by providing guidelines on risk assessment and management
for multi-parties in tunnel construction, enabling to increase the likelihood of a
successful project in a complex environment.
Traditionally, only experienced engineers and experts can identify potential safety
risks and be aware of unexpected dangerous situations and accidents. This system
provides a solution to automate safety risk identification in tunnel construction. One
benefit of the developed system is that the dependence on domain experts can be
reduced, and the knowledge sharing and communication in construction management
among dispersed clients and domain experts can be facilitated. Furthermore, the low-
efficiency in traditional information extraction from 2D drawing documents can be
significantly improved. Due to the implementation of the developed B-RIES and the
efforts of all participants, no accidents are reported during the construction of the
WMLT.
A large number of rules serving safety risk identification were obtained from
domain experts for system development. Numerous domain experts participated in
the collection, editing, reorganizing work of the safety-related knowledge resources,
making an essential contribution to securing a qualified rule base for the development
of B-RIES. Automatic knowledge acquisition and decision rule generation regarding
different knowledge resources should be of significance in developing a real-time
expert system for improved decision making.
References
1. Abu-Naser SS, Aldahdooh R (2016) Lower back pain expert system diagnosis and treatment.
J Multidisc Eng Sci Stud 4:441–446
2. Beckman TJ (1999) The current state of knowledge management. CRC Press
3. Behm M (2005) Linking construction fatalities to the design for construction safety concept.
Saf Sci 43:589–611
4. Bohlouli M, Mittas N, Kakarontzas G, Theodosiou T, Angelis L, Fathi M (2017) Competence
assessment as an expert system for human resource management: a mathematical approach.
Expert Syst Appl 70:83–102
5. Chou J-S (2009) Generalized linear model-based expert system for estimating the cost of
transportation projects. Expert Syst Appl 36:4253–4267
6. Chou J-S, Tseng H-C (2011) Establishing expert system for prediction based on the project-
oriented data warehouse. Expert Syst Appl 38:640–651
References 229
7. Ding L, Yu H, Li H, Zhou C, Wu X, Yu M (2012) Safety risk identification system for metro

construction on the basis of construction drawings. Autom Constr 27:120–137
8. Elziny AA, Mohamadien MA, Ibrahim HM, Abdel Fattah MK (2016) An expert system to
manage dispute resolutions in construction projects in Egypt. Ain Shams Eng J 7:57–71
9. Gambatese JA, Behm M, Rajendran S (2008) Design’s role in construction accident causality
and prevention: Perspectives from an expert panel. Saf Sci 46:675–691
10. Goksu I (2016) The evaluation of the cognitive learning process of the renewed bloom taxonomy
using a web based expert system. Turkish Online J Educ Technol TOJET 15:135–151
11. Grossauer K, Schubert W, Lenz G (2008) Automatic displacement monitoring data interpre-
tation–the next step towards an expert system for tunneling. In: 42nd US Rock Mechanics
Symposium and 2nd US-Canada Rock Mechanics Symposium, San Francisco, CA, USA
12. Hadikusumo BHW, Rowlinson S (2002) Integration of virtually real construction model and
design-for-safety-process database, 5th edn. Elsevier, Amsterdam, pp 501–509
13. Hardin B (2011) BIM and construction management: proven tools, methods, and workflows.
Sybex
14. Hong E-S, Lee I-M, Shin H-S, Nam S-W, Kong J-S (2009) Quantitative risk evaluation based
on event tree analysis technique: application to the design of shield TBM. Tunn Undergr Space
Technol 24:269–277
15. Koulinas GK, Xanthopoulos AS, Tsilipiras TT, Koulouriotis DE (2020) Schedule delay risk
analysis in construction projects with a simulation-based expert system. Buildings 10:134
16. Ku K, Mills T (2010) Research needs for building information modeling for construction safety.
In: International Proceedings of Associated Schools of Construction 45nd Annual Conference,
Boston, MA
17. Latif Y, Abidin I, Trigunarsyah B (2008) Knowledge-based material cost control for building
construction project using expert system approach. In: Proceedings of the CIB International
Conference on Building Education and Research Building Resilience. School of the Built
Environment, University of Salford, United Kingdom, pp 125–133
18. Lederberg J (1965) DENDRAL-64-A system for computer construction, enumeration and
notation of organic molecules as tree structures and cyclic graphs. Part II-Topology of cyclic
graphs Interim report
19. Martín A, León C, Luque J, Monedero I (2012) A framework for development of integrated
intelligent knowledge for management of telecommunication networks. Expert Syst Appl
39:9264–9274
20. Minsky M, Winston PH (1975) The psychology of computer vision
21. Newell A, Shaw JC, Simon HA (1959) Report on a general problem solving program. IFIP
congress. Pittsburgh, PA, p 64
22. Nezarat H, Sereshki F, Ataei M (2015) Ranking of geological risks in mechanized tunneling by
using Fuzzy Analytical Hierarchy Process (FAHP). Tunn Undergr Space Technol 50:358–364
23. Pan Y, Zhang L (2021) A BIM-data mining integrated digital twin framework for advanced
project management. Autom Constr 124:103564
24. Petrick SR (1971) Proceedings of the second ACM symposium on Symbolic and algebraic
manipulation. ACM
25. Šejnoha J, Jarušková D, Špačková O, Novotná E (2009) Risk quantification for tunnel excava-
tion process. In: Proceedings of World Academy of Science, Engineering and Technology, pp
101–109
26. Sherehiy B, Karwowski W (2006) Knowledge management for occupational safety, health,
and ergonomics. Hum Factors Ergon Manuf 16:309–319
27. Suryono S, Surarso B, Saputra R, Sudalma S (2019) Real-time decision support system for
carbon monoxide threat warning using online expert system. J Appl Eng Sci 17:18–25
28. Szymberski R (1997) Construction project safety planning. Tappi J 80:69–69
29. Wang F, Li H, Dong C, Ding L (2019) Knowledge representation using non-parametric Bayesian
networks for tunneling risk analysis. Reliab Eng Syst Saf 191:106529
30. Yu C, Chern J (2007) Expert system for D&B tunnel construction. Underground Space 4th
Dimension of Metropolises, London, England
230 Expert Systems
31. Zavadskas EK, Turskis Z, Tamošaitiene J (2010) Risk assessment of construction projects. J
Civ Eng Manag 16:33–46
32. Zhang L, Wu X, Ding L, Skibniewski MJ, Lu Y (2016) Bim-Based Risk Identification System
in tunnel construction. J Civ Eng Manag 22:529–539
33. Zhou W, Whyte J, Sacks R (2012) Construction safety and digital design: a review. Elsevier,
Amsterdam, pp 102–111
Computer Vision
1 Introduction
Because of structure degradation, thermal movement, and other reasons, unwanted

cracks will unavoidably appear in various types of structures, such as slabs, beams,
columns, walls, etc. Unfortunately, harmful and corrosive chemicals, water, and salts
will penetrate concrete layers through these existing cracks to exert negative impacts
on structural integrity and durability [1]. Thus, it is believed that cracks will speed
up the degradation and aging of structure, bringing unreliability and failures in the
structural systems. For instance, there are more than 100,000 bridges in the United
State with the occurrence of early-age bridge-deck cracking [19]. In a survey about 98
bridges in New York state, 46% are structurally deficient before the collapse [8]. The
past 40 years have witnessed more than 510,000 new bridges that have been built in
China, and issues rooting in cracks have caused 15% of bridges to reach their service
lives prematurely [21]. Accordingly, cracks can act as a critical indicator to evaluate
structure damages, and thus crack detection will have high practical application
values in triggering early warnings for structure maintenance, safety assurance, and
loss reduction.
It is known that human visual inspection is a relatively simple and popular method
of crack detection largely based upon subjective judgment, specialist experience, and
empirical knowledge. However, this manual approach is labor-intensive and time-
consuming, which tends to suffer from low accuracy. Besides, inspectors are unable to
recognize all superficial defects. Such shortcomings are obvious, particularly in large-
scale projects. Another thing to note is that the requirement of onsite diagnosis is to
close urban transit systems, which will lead to inconvenience in public transportation.
Servicing as solutions to tackling these limitations of human-based methods, several
automated visual inspection approaches have been proposed to replace the subjective
and inefficient manual process [18]. Since computer vision techniques have rapidly
developed in the past few years, more and more focus has been paid on the image
processing methods to check specific structural deterioration, like crack, spalling,
232 Computer Vision
and others, aiming to boost efficiency and reliability in the domain of structural
health monitoring.
In general, classical approaches in computer vision can be considered as feature
engineering containing two major aspects. The first one is to extract features from
images, which can be represented by a compact feature vector. The other is feature
matching to establish feature correspondence between two images about the same
object. However, the quality of hand-crafted extracted features will greatly influ-
ence prediction performance, and thus such a method is still limited by subjectivity
and domain expertise. Also, it requires a great number of preprocessed images to
make the detection process unadaptable, tedious, and inefficient. Even worse, these
hand-crafted features have no adequate ability to differentiate cracks and complex
backgrounds in low-level image cues, which are less applicable in images with large
variations.
Up to date, deep learning techniques are driving advances in computer vision
for resolving weaknesses of classical crack detection models. It aims to realize an
automated process in capturing intricate structures of large-scale data by models
with multiple processing layers [22]. Remarkably, one of the most important models
for automated feature learning and supervised classification is called the convolu-
tional networks (CNNs/ConvNets), which belong to a kind of neural network to take
images as inputs. Its unique properties in the partial connections, sharing weights,
and pooling layers make CNNs to automatically capture the grid-like topology of
images under fewer computations, and then generate promising detection results [4].
Implementation of CNN-related methods has been proven effective in classifying
multiple damage types of structural deterioration [24]. Moreover, some more compli-
cated CNN-based architectures have been developed and applied in the task of crack
detection, including U-net, SegNet, CrackNet, and others. Among these advanced
models, fully convolutional networks (FCNs) [16] have attracted great attention,
which can be useful in transforming image pixels to pixel categories. The core idea
behind FCNs is to convert the last fully connected layer into CNNs to a convolution
layer for training an end-to-end pipeline. FCNs have shown their promising perfor-
mance of semantic segmentation in some works [12]. However, it is likely for FCNs
to ignore the relationships between pixels, and then to poorly influence the reliability
of scene segmentation.
To address the above-mentioned issue from FCNs, the attention mechanism that
has brought great breakthroughs from natural language processing (NLP) can be
deployed in the computer vision task. More specifically, the role of the attention
mechanism is to assign different weights to features according to their importance,
allowing for making the network quickly focus on the more relevant parts of the
input. Therefore, the emerging self-attention is good at catching long-range depen-
dencies with no increase of parameter number, contributing to boosting the perfor-
mance of image classification, object detection, semantic segmentation, and others
[3]. Inspired by the successful application of FCN and attention mechanism, this
chapter attempts to design a hybrid network structure for a more credible process of
1 Introduction 233
feature extraction and pixel labeling. It is expected to semantically represent relation-

ships among features and aggregate multi-scale context features, which eventually
helps in detecting cracks in an automated and accurate manner.
In this chapter, the spatial-channel hierarchical network (SCHNet), a novel hybrid
deep learning approach, is developed for cracks image segmentation under various
scenarios. Its main goal is to extract meaningful information from images through
modeling rich global contextual dependencies among features, which are then used
to build a robust model for significantly improving the segmentation performance. As
expected, visual feature dependencies in both the spatial and channel dimensions will
be obtained and fused to improve the quality of feature representation, contributing
to a more reliable crack detection process to make a distinction between crack and
non-crack images at the pixel level.
2 Deep Learning and Computer Vision
Inspired by the structure and function of biological neural networks in human brains,
deep learning encompassing multiple processing layers enables the computer to
perform a process of learning and problem solving. The promise of deep learning has
been highlighted in capturing intricate structures of large-scale data and addressing
complicated problems with no human intervention in an end-to-end learning frame-
work. Currently, there is an increasing interest in applying deep learning to drive a
great revolution in traditional computer vision. A variety of deep learning algorithms
have developed to support the machine to process and interpret the unstructured visual
data from digital images or videos automatically, and thus, data representations with
different levels of abstraction can be extracted and learned effectively resulting in
improved accuracy and robustness [13]. In general, the deep learning-based computer
vision task (i.e., classification, detection, and segmentation) follows four key steps,
including feature extraction, forward propagation, loss calculation, and backpropa-
gation. Some recent deep learning-based studies with great performance have been
listed in Table 1, which are divided into three categories, namely the ConvNet-based
model, Region-based model, and FCN-based models.
(1) ConvNet-based models
The fundamental and mainstream deep learning model named ConvNet has
three kinds of neural layers (convolutional layers, pooling layers, and fully
connected layers). which has reached remarkable achievements in computer
vision. For one thing, ConvNet shows a great capability of automatically
extracting meaningful features from a great number of raw images. For another,
since ConvNet needs less image pre-processing, it can be easily trained and it
owns fewer parameters and connections compared to the standard feedforward
neural networks. Generally speaking, the ConvNet-based models regard crack
detection as a binary classification problem, where each pixel of the image is
labeled by one class. However, an issue to note is that the process of training the
234 Computer Vision
Table 1 Example of computer-vision based studies for automatic crack detection

Method Limitation Example
Authors Remarks
ConvNet-based models Slow training process Zhang et al. [25] ConvNet is built to
learn discriminative
features directly from
raw image patches,
and then to achieve
automatic detection of
pavement cracks. It is
trained in 500 images
of size 3264 × 2448
Cha et al. [5] The classical ConvNet
architecture is
established to detect
concrete cracks with a
sliding window
technique to cut
images into small
patches, which is
(277 images with
4928 × 3264 pixel
resolutions and 55
with 5888 × 3584
pixel
resolutions)
Li and Zhao [15] Crack detection is
based on the CNN
architecture of
binary-class outputs
by modifying the
AlexNet, which can
be applied by a
smartphone
application. The
method is trained in
205 images with 3120
× 4160 pixel
resolutions
Region-based models Low accuracy Cha et al. [5] A faster
R-CNN-based
structural visual
inspection approach is
performed to detect
five types of surface
damages, which is
trained in a dataset of
2366 images with 500
× 375 pixels
(continued)
2 Deep Learning and Computer Vision 235
Table 1 (continued)
Authors Remarks
Cheng and Wang [7] Faster RNN is
designed to detect
sewer pipe defects
from closed-circuit
television (CCTV)
images, which is
with the resolution
between 1440 × 720
and 320 × 256 pixels
Deng et al. [10] Concrete cracks from
real-world images
with complex
backgrounds are
identified
automatically based
on Faster R-CNN. 160
images in 3264 ×
2448 pixels are used
for training and testing
FCN-based models Lack of consideration Yang et al. [23] The FCN model with
of relationships VGG 19 is
implemented to
identify and measure
diverse cracks
concurrently at the
pixel level, whose
performance is
examined in 800
images with the
number of pixels
varying from 1 to 100
Dung and Anh [12] An FCN-based crack
detection method
under VGG16 encoder
is developed for
semantic
segmentation on 500
concrete crack images
in 227 × 227 pixels
(continued)
236 Computer Vision
Table 1 (continued)
Authors Remarks
Bang et al. [2] The encoder-decoder
network based on
FCN is built as a
pixel-level detection
method to identify
road cracks from 527
images with a
resolution of 2560 ×
1440, which are taken
from black-box
cameras
ConvNet-based model could be relatively slow due to the overlapping patches.

Also, the trade-off between the number of patches and classification accuracy
will negatively affect model performance. More specifically, a larger size of
patches requires more max-pooling, leading to low classification accuracy. A
smaller size of patches makes the network only concentrate on local features.
(2) Region-based models
For higher accuracy of object detection and localization, the region convolu-
tional neural network (R-CNN) is proposed containing three different stages,
including CNNs, support vector machine (SVM), and bounding box regressor.
As a clearer explanation, features are extracted automatically by CNNs, and
then fed into SVM and regressor during the forward propagation. Neverthe-
less, its calculation process is slow, expensive, and difficult due to the unshared
parameters. To speed up R-CNN, the spatial pyramid pooling network (SPP-
net) is proposed relying on a regional proposal method that can generate a
region of interests (RoIs) through sharing computation, but it is unable to guar-
antee high accuracy. To fix the slow speed and accuracy, Fast R-CNN warps
RoIs into one single layer rather than generating a pyramid of layers, which can
train all parameters together. Additionally, a better model termed Faster-RNN
is developed under the combination of a region proposal network (RPN) and
Fast R-CNN. It no longer depends on selective search for determining region
proposals. However, Faster R-CNN still uses the ROIs pooling layer, which
is possible to cause the loss of translation invariance in the extracted features.
The multiple down-sampling of the anchor will also weaken performance in
detecting small targets.
(3) FCN-based models
FCN is an end-to-end and pixel-to-pixel convolutional network specialized in
semantically recognizing and segmenting pixels with different scales, which
can return segmentation results represented by a single-pixel width skeleton.
The following two parts are incorporated in FCN: one is the downsample
process with convolutional layers, pooling layers, and dropout layers, and the
2 Deep Learning and Computer Vision 237
other is the upsample process in deconvolutional layers. To be more specific,

multi-layer deep convolutional networks, such as VGG 16 and VGG 19, are
often taken as the backbone network (base net) in the downsampling. It is
expected to learn the spatial hierarchy of features from these pre-trained
networks. The upsampling allows for integrating the local and global informa-
tion to realize object detection and localization. Although the context fusion
displays a better performance in detecting different scale objects, FCN-based
models do not consider relationships among objects and stuff that could be
somewhat important for scene representation.
Apart from the above-mentioned limitations, another concern is that most of the
research works for exploring crack images in Table 1 are comparatively clean with
little noise. But it is inevitable for cameras to take many low-quality images with
disturbance from holes, shadows, rough surfaces, and others. Thus, a more robust
model is needed to distinguish cracks from various obstructions in a larger dataset.
Besides, we refer to a novel attention mechanism to selectively pay attention to the
critical part within a context while ignoring others, which can realize a more paral-
lelizable and efficient training process. Stemming from NLP, the attention mecha-
nism is known to be practical for text representation, natural language inference, and
sentence embedding. However, the extended application of attention to the computer
vision field has not been fully explored. This chapter attempts to incorporate the atten-
tion module into the deep learning frameworks in pursuit of higher defect detection
accuracy from images at a pixel level.
3 Computer Vision Framework
This chapter proposes a new structure termed SCHNet to distinguish crack and non-
crack images at the pixel level. The novelty of SCHNet lies in the combination of
the VGG base net and a self-attention mechanism, which helps in better capturing
long-range contextual information in both the spatial and channel dimensions. On
the one hand, VGG is a deep convolutional neural network model supporting up to
19 layers. Its performance in object recognition has been proven outstanding on the
ImageNet dataset. On the other hand, a self-attention mechanism displayed in Fig. 1
consists of three parallel modules, including a feature pyramid attention module, a
spatial attention module, and a channel attention module. That is to say, after the
process of inputting raw images into the feature pyramid attention module, mean-
ingful features will be extracted automatically and then fed into both the channel and
spatial attention module. Subsequently, the two output attention maps are integrated
using an element-wise summation operation, resulting in segmentation images in the
size of input images. In short, these three attention modules are the key parts of the
proposed SCHNet, contributing to the success of the high-level feature learning of
crack images. The detailed description of these modules along with the evaluation
metric and training process are provided below.
238 Computer Vision
Feature
Feature pyramid attention pyramid
module
Input
attention attention
matrix matrix
reshape reshape
Output
CONV
Spatial Channel
Attention Attention
Module Module
Fusion
Fig. 1 The framework of the proposed SCHNet approach for automated crack detection. Note
CONV represents the convolutional layer
3.1 Feature Pyramid Attention Module
It is known that many recent methods adopt skip connections or pyramidal shapes
to associate low-level feature maps for improving segmentation performance. But
these methods are prone to suffer from low efficiency and require high memory, since
independent predictions are made at all levels. To deal with the problem, a feature
pyramid attention module is deployed to quickly extract and combine features by
a top-down pathway and lateral connections, resulting in strong semantics at all
scales. As shown in Fig. 2, the output features from the base net VGG are input
into a convolutional layer in a kernel size of 1 × 1. Then, it will be multiplied
with the pyramid attention features in three different pyramid scales of 3 × 3, 5
× 5, 7 × 7, which are created by a U-shape structure. In other words, different
scale context features can be fused in the feature pyramid attention module, which
are then multiplied with original feature maps pixel-wisely. Eventually, the pixel-
level attention information for high-level feature maps can be built, contributing to
improving the feature extraction.
3 Computer Vision Framework 239
Basenet
1X1
CONV 7X7 7X7
CONV CONV
Global downsample
pooling upsample
1X1
X + 5X5 5X5
16X16 CONV CONV
CONV
downsample
8X8 +
upsample 4X4 3X3 3X3
upsample CONV CONV
+ output
Fig. 2 Feature pyramid attention module structure. Note All the convolution layers are followed
by batch normalization. 4 × 4, 8 × 8, where 16 × 16 means the resolution of the feature map. Red
and blue color represent the process of upsampling and downsampling, respectively
3.2 Spatial Attention Module
Notably, the context relationship in FCNs plays a key role in feature representation,
which helps to capture global spatial dependencies regardless of locations. Never-
theless, it is more likely for local features of traditional FCNs to misclassify objects
and stuff [9]. To improve the effectiveness in modeling rich contextual dependen-
cies over local feature representation, a spatial attention module in Fig. 3 is utilized,
which can encode a wider range of contextual information into local features. To

be more specific, after the feature map A ∈ R c ×h ×w is fed into three convolution
layers with the same kernel size of 1 × 1 and batch normalization, three new feature

maps B, C, and D ({B, C, D ∈ R c ×h ×w }) are generated. In particular, B and D
n×c
are reshaped to R , where n = h × w is the number of pixels in each channel.
Concurrently, the transposition of the new feature map labeled by C is multiplied by
B to get a self-attention map S ∈ R n×n . Then, the result of S × D goes through a

softmax layer to calculate the self-spatial attention maps E ∈ R c ×h ×w (Fu et al.). As
for the softmax function, it aims to map the non-normalized output to a normalized
probability distribution over predicted classes, which has been widely adopted in the
last layer of a neural network-based classifier.
Herein, an example is provided based on an image that is in a size of 64 × 64 × 3
pixels and contains three channels called RGB (R is red, G is green, and B is blue).
After the convolutional operations with a kernel size of 1 × 1, the output feature
map is reshaped to a two-dimensional matrix of 3 × 4096. Thus, the new shape
derived from the reshape operation is compatible with the original shape to reserve
all information.
240 Computer Vision
input feature maps

A
1X1 1X1 1X1

CONV CONV CONV
reshape reshape transpose reshape
B C D
+
X
output feature maps self-attention map
F
X
self-spatial
attention maps E
Fig. 3 Structure of the proposed spatial attention module. Note that the capital letter in bold italics
is the label of feature maps
3.3 Channel Attention Module
A channel attention module shown in Fig. 4 is designed to model interdependencies

between channels. Unlike the spatial attention module, the channel attention map
input feature maps

A
1X1 1X1 1X1

CONV CONV CONV
reshape reshape transpose reshape
B C D
X +
channel
attention map output feature maps
X F
Fig. 4 Structure of the proposed channel attention module. Note that the capital letter in bold italics
is the label of feature maps
3 Computer Vision Framework 241

X ∈ R c ×c can be directly calculated from the original feature maps A ∈ R c ×h ×w .

That is to say, A is reshaped to C ∈ R c ×n , and then it is multiplied with the trans-
position D. Finally, a softmax layer expressed in Eq. (1) is utilized to output the

channel attention map X ∈ R c ×c [14]. After the multiplication of X and B, the result
c ×h ×w
E is reshaped to R . As formulated Eq. (2), the final output F ∈ R c ×h ×w
is generated based on the element-wise sum operation of A and E scaled by the
parameter β. In other words, the final representation feature of each pixel can be
obtained via a weighted sum of the features of all pixels and original features, which
is helpful in modeling the rich contextual dependencies over image regions. Besides,
it can be employed to strengthen class-dependent feature maps and boost feature
discriminability.
exp(Am · An )
xnm = C (1)
m=1 exp(Am · An )

C

Fj = β x q p A p + Aq (2)
p=1
where x nm refers to the mth pixel’s impact on the nth pixel, and β gradually learns a
weight from 0.
3.4 Learning Process and Evaluation Metrics
Regarding the training process, features in the spatial and channel dimensions can
be captured from the two parallel attention modules. Then, these obtained features
are aggregated to model long-range dependencies, resulting in the improved feature
representation for pixel-level segmentation. More specifically, outputs of two atten-
tion modules are fused through an element-wise summation, which will then be put
into a convolutional layer in the kernel size of 3 × 3 to learn the final prediction
map. Cross-entropy loss (also known as log loss) based on the natural logarithm is
employed as the loss function to measure the difference in two probability distribu-
tions from the predictions and ground truth. Since there are only two classes in the
image (the crack and the background), we rely on the binary cross-entropy in Eq. (3)
to deal with the binary classification task in this chapter. The value of cross-entropy
decreases when the predicted class probability is much closer to the actual label. Our
goal is to minimize the cross-entropy during the model optimization process and
continuously update the model parameters through gradient descent.
1
N

J (θ ) = − yi log yi + (1 − yi )log(1 − yi ) (3)

N i=1
242 Computer Vision
IoU = 0.91 IoU = 0.82 IoU = 0.67
Ground True Prediction Ground True Prediction Ground True Prediction

(a) (b) (c)
Fig. 5 An example of computing IoU for various bounding boxes: a Excellent segmentation (with
an IoU of 0.91); b good segmentation (with an IoU of 0.82); and c poor segmentation (with an IoU
of 0.67)
where N is the total number of pixels, yi is the predicted value of the ith pixel, yi is
the corresponding actual value of the ith pixel.
For quantitatively assessing the segmentation performance of the developed
model, the metric called Intersection over Union (IoU) in the form of a Jaccard
index is adopted rather than the commonly used metrics like precision and recall in
the traditional machine learning-based classification tasks. That is because the (x,
y)-coordinates of the predicted bounding box cannot exactly match the ground-truth
bounding box, meaning that the simple consideration of object detection as a binary
problem is irrational. For this concern, IoU is defined in Eq. (4) as a ratio of the over-
lapping region in two boxes to their total combined regions, aiming to effectively
measure the similarity in the predicted bounding box and the ground truth bounding
box. A larger IoU can, therefore, represent the greater segmentation performance.
As an example, three pairs of ground truth and predictions under different values of
IoU are displayed in Fig. 5, which numerically indicate that how much the predicted
bounding box is close to the actual cracks.
Ar ea o f Overlap
I oU = (4)
Ar ea o f U nion
Besides, when cracks are detected in a large number of images, another metric
termed Mean IoU is formulated in Eq. (5) to evaluate the segmentation accuracy
among various images averagely.

N

Mean I oU = (1/N) n i_crack / n i_crack + n i_noncrack (5)
i=1
where N refers to the number of input images, ni_crack is the number of pixels that are
correctly segmented into cracks in each image, and ni_noncrack represents the number
of crack pixels that are segmented into the background.
4 Computer Vision for Automated Crack Detection 243
4 Computer Vision for Automated Crack Detection
4.1 Data Resource
For the validation of the proposed SCHNet, a series of comprehensive experiments are
conducted on an open concrete crack image dataset named SDNET2018 [17]. There
are over 56,000 cracked and non-cracked images about concrete bridge decks in this
adopted dataset. The collection of original full images is based on a 16 MP Nikon
camera at a working distance of 500 mm without zoom, where the International
Organization for Standardization (ISO) sensitivity is 125, the image resolution is
4068 × 3456 px, and the surface illumination is between 1500 and 3000 lx. Each
original full image actually presents a physical area of about 1000 mm × 850 mm,
which is then segmented into several 256 × 256 px images on behalf of a physical
area in approximately 60 mm × 60 mm.
Since the early-age cracking in bridge deck concrete is a severe problem causing
the reduction of bridge life, this case study mainly concentrates on images of bridge
decks for enhancing bridge structure monitoring and assessment. To this end, a total
of 11,000 cracked and non-cracked images about the bridge deck at 256 × 256 px are
selected from SDNET2018. The length of cracks in images is in the range from 0.06
to 25 mm. Notably, various obstructions, like shadows, surface roughness, scaling,
edges, holes, and background debris, are inevitable in images from the dataset to
raise difficulties of crack detection.
Another step in data preparation is manual annotation, Herein, we rely on an online
annotation tool LabelMe to annotate a new object by clicking along the boundary of
the desired object, aiming to make images own semantic information. Some examples
of images used in the concrete crack dataset are given in Fig. 6, where the first,
second, and third column provides original images, images with manual labels, and
the corresponding ground truth, respectively. The main goal of this chapter is to
segment cracks on images under the condition that all cracks on images are treated
as one class.
4.2 Model Training
For improving the robustness of the network, two key steps are conducted in the
model training process. One is to fine-tune all layers by backpropagation through
the whole network, and the other is data augmentation. For the first steps, the initial
learning rate is updated by a poly learning rate policy, which is multiplied by a factor
0.9
1 − total−iter
iter
after each iteration. This given factor expressed by a polynomial
decay function helps to drive the initial learning rate to the end learning rate within
the given decay steps. In this case, the initial value of three important parameters
named learning rate, momentum, and weight decay coefficients are set as 0.1, 0.9,
0.0001, respectively.
244 Computer Vision
Original Images Crack labels Ground Truth
(a)
(b)
(c)
(d)
Fig. 6 Examples of images used in the concrete crack dataset with a holes; b shadows; c rough
surfaces; d smooth surfaces
To perform further optimization in the developed model, a cross-entropy loss is

used to compare the model’s prediction with the true probability distribution. Clearly,
the cross-entropy decreases as the prediction accuracy boosts. Besides, two data
augmentation techniques including random cropping and random left–right flipping
are applied during the data training process in the ablation study for SDNET2018. At
the end of the contracting path, there are dropout layers to perform further implicit
data augmentation. Another preparation of model training is to randomly divide the
dataset into the training and testing categories by an 80–20% split. That is, 80%
of the SDNET2018 dataset serves for training the complex neural networks, while
20% is responsible for verifying the model performance. The parameters are zero-
initialized since the random initialization will bring no significant impact on the
performance and convergence of the model. Besides, the dropout is included in the
original classifier nets.
The developed model is implemented by Tensorflow and trained on three NVIDIA
GFORCE 1080TI GPUs with an 11 GB memory. In avoidance of the out-of-memory
problem, the batch size is set to be less than 16 during the training and tuning process.
In experiments, the resolution of images and ground truths is downsampled to be 256
× 256 px. The approximately estimated training time is about 1–2 days.
The proposed SCHNet model has demonstrated a fast convergence speed and great
stability. To choose the best initial learning rate for configuring the neural network,
we set the initial value as 10–2 , 10–3 , 10–4 , and 10–5 in sequence. Figure 7 displays the
loss of the developed approach under the four candidate learning rates. Observably,
the loss all converges after 1000 training epochs, indicating a fast training speed and
high effectiveness of SCHNet. In particular, the lowest loss can achieve based on the
learning rate of 1 × 10–2 , which is smaller than half of the learning rate 1 × 10–5 . It
reveals that the adoption of a too small learning rate is improper, since a relatively
2.2 2.2 learning rate 1e-3

learning rate 1e-2
2.0 2.0
1.8 1.8
1.6 1.6
1.4 1.4
Loss
1.2
Loss
1.2
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Epoch Epoch
(a) (b)
2.4
2.2 learning rate 1e-4 learnin rate 1e-5
2.0 2.2
1.8
2.0
Loss
1.6
Loss
1.8
1.4
1.2
1.6
1.0
1.4
0.8
0.6 1.2
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Epoch Epoch
(c) (d)
Fig. 7 Cross entropy loss under different initial learning rates: a 10–2 ; b 10–3 ; c 10–4 ; d 10–5
246 Computer Vision
long training time will be taken to reach the minimum of the loss function. Therefore,
we set the initial learning rate of 1 × 10–2 as the optimal value, enabling the model
to learn the problem stably and efficiently.
4.3 Analysis of Results
To validate the effectiveness of the proposed SCHNet approach, we implement it in

the given dataset for crack detection at the pixel level. Figures 8, 9 and 10 intuitively
display the segmentation results and the learned features. Table 2 compares the
detection speed of the proposed method with other popular models. The analysis of
SCHNet performance is summarized from the view of the attention modules, learning
speed, and learned features as follows.
(1) The attention modules help to boost the segmentation performance through
selective fusion over local features. Figure 8 visualizes the effects of the
channel and spatial attention modules, where the first and second rows display
the segmentation performance on simple-crack and complex-crack images,
respectively. The red boxes highlight the different areas of the segmentation
results. It can be observed that the fusion over local features from the channel
and spatial attention modules contributes to making details and the boundary
discrimination much clearer. In other words, rich contextual information can
be easily captured using the integration of the spatial and channel attention
maps.
(2) The running speed of SCHNet is competitive with state-of-the-art methods. To
highlight the trade-off between detection performance and speed, a comparison
50
100
0 50 100 0 50 100 0 50 100 0 50 100 0 50 100

0
50
100
0 50 100 0 50 100 0 50 100 0 50 100 0 50 100

(a) (b) (c) (d) (e)
Fig. 8 Visualization of the differences in segmentation results with feature pyramid attention
module (FPA), channel attention module (CAM), and spatial attention module (SPM): a orig-
inal image; b without any attention module; c with FPA and CAM; d with FPA, CAM, and SPM;
e Ground truth. Note X- and Y-axis represents the width and length of an image (unit: pixel),
respectively
0 0 0 0
First
Layer
100 100 100 100
0 0 0 0
Third
Layer
100 100 100 100
0 0 0 0
Fifth
Layer
100 100 100 100
0 0 0 0
Eighth
Layer
100 100 100 100
0 100 0 100 0 100 0 100

(a) (b) (c) (d)
Fig. 9 Learned features at different convolutional layers from four testing images: a Image 1;
b Image 2; c Image 3; d Image 4. Note X- and Y-axis represents the width and length of an image
(unit: pixel), respectively
is made among SCHNet, DeepLab-v3 [6], PSPNet [26], and FCN [16]. The
runtime speed is measured on a single GTX1080TI and summarized in Table
2. All the models are evaluated without any pre- or post-processing to ensure
fairness. As listed in Table 2, DeepLab-v3 shows the slowest computational
speed (8.5 frames per second (FPS) due to the ultra-deep VGG model as the
backbone network. On the contrary, FCN runs the fastest (24.1 FPS) because of
its small model size. But Table 5 finds out that the segmentation performance
of FCN is relatively poorer than SCHNet. Therefore, it can be concluded that
SCHNet can run at an acceptable speed (16.3 FPS) to reach the highest Mean
IoU (85.31%).
(3) SCHNet can learn features automatically and reliably through updating the
weights of receptive fields, which does not require additional feature extraction
techniques that are the necessity in other image-based studies for segmentation.
Figure 9 takes four crack images as examples to visualize the receptive field,
which is acknowledged as the learned features from networks. The number of
receptive fields of the designed architecture is 24 with 64 × 64 × 3 dimensions.
The visualized features indicate whether the network needs more training and
what kinds of features are recognized by the trained network. It can be seen
248 Computer Vision
Fig. 10 Some qualitative crack segmentation results by using SCHNet: a original images; b ground
truth; c segmentation
Table 2 Comparison of
Models FPS
running time with different
models DeepLab-v3 8.5
PSPNet 10.2
Dilated FCN 24.1
SCHNet 16.3
Note that FPS is the abbreviation of frames per second
from Fig. 9 that spots and lines become clearer as the convolutional layer
goes deeper. Herein, eight convolutional layers in the developed network are
suitable to train the network well in this experiment. Thus, features in the last
row of Fig. 9 can be rationally taken as crack features, while features in the
second and third rows are speculated as concrete surface cavities or aggregates
in the training data set. Besides, although receptive fields of a well-trained
network generally have smooth patterns, noisy features with various colors are
unavoidable due to the complex and arbitrary patterns in concrete surfaces. But
the segmentation results in the last column of Fig. 10 are very consistent with
the ground truth in the middle column, which implies that the proposed model
is able to well deal with noise to reduce the negative effects from noise.
4.4 Comparison with Existing Models
To demonstrate the outstanding performance of SCHNet, the proposed model is

compared to some state-of-the-art models based on the metric called Mean IoU.
Table 3 validates the effectiveness of attention modules in raising Mean IoU. Table 4
reveals that how the optimization techniques, including data augmentation and multi-
grid method, can influence the detection performance. Table 5 verifies the superiority
of SCHNet over other models. The results are analyzed in detail as follows.
(1) The three attention modules bring benefits in improving segmentation perfor-
mance. Table 3 lists models in different settings along with their resulted Mean
IoU values. It is clear that the baseline using FCN with the base net VGG16
leads to the Mean IoU of 70.03%. Then, a single feature pyramid attention
module is added to the baseline, yielding a result of 72.63%. Moreover, the
Table 3 Results of the ablation study on the concrete crack validation set
Models Base net FPA SAM PAM Mean IoU (%)
Dilated FCN VGG16 70.03
√
SCHNet VGG16 72.63
√ √
SCHNet VGG16 76.62
√ √
SCHNet VGG16 77.37
√ √ √
SCHNet VGG16 80.13
Dilated FCN VGG19 72.54
√
SCHNet VGG19 74.16
√ √
SCHNet VGG19 79.31
√ √
SCHNet VGG19 79.92
√ √ √
SCHNet VGG19 83.42
Note FPA represents feature pyramid module; SAM represents spatial attention module; PAM
represents channel attention module
250 Computer Vision
Table 4 Performance comparison between different strategies on the concrete crack validation set
Models Base net DA Multi-grid MS Mean IoU (%)
SCHNet VGG19 83.42
√
SCHNet VGG19 84.31
√ √
SCHNet VGG19 84.86
√ √ √
SCHNet VGG19 85.31
Note DA represents data augmentation with random scaling. Multi-grid represents employing a
multi-grid method, and MS represents multi-scale inputs during inference
Table 5 Segmentation
Models Basenet Mean IoU (%)
results using different models
on the concrete crack U-net [20] – 71.4
validation set DeepLab-v3 [6] – 75.7
PSPNet [26] VGG19 77.8
Ding [11] VGG19 71.60
Dilated FCN [16] VGG16 70.30
Dilated FCN [16] VGG19 72.54
SCHNet VGG16 80.13
SCHNet VGG19 85.31
Note “–” indicates the base net is not applied in the model
combination of the feature pyramid and spatial attention (or channel atten-
tion) modules can increase the value of Mean IoU to 76.62% (or 77.37%).
When the three attention modules work together, the Mean IoU can reach
up to the highest (80.13%). Moreover, when the base net VGG19 is used
to initialize the parameters, the network with three attention modules can
further enhance Mean IoU to peak at 83.42%. In contrast to the new base-
line model (FCN with VGG19), the feature pyramid attention, the spatial
attention, and the channel attention module under VGG 19 can individually
increase Mean IoU by 1.62% (74.16–72.54%), 5.15% (79.31–74.16%), and
5.76% (79.92–74.16%), respectively.
(2) Three new strategies, including data augmentation, multi-grid method, and
maps, help to increase the Mean IoU of SCHNet to 85.31%. From experimental
results in Table 4, it suggests that the developed approach experiences a 0.89%
increase by using the data augmentation with random scaling. In other words,
enriching the scale diversity of training data can act as an effective way to
improve network performance. Besides, an additional 0.55% improvement in
Mean IoU can be brought by the multi-grid strategy. The idea behind such a
method is to employ a hierarchy of grids with different sizes in the last VGG
net, and thus the greater feature representation can be gained from a pre-trained
network.
(3) The performance of SCHNet is superior to the state-of-the-art models on the

test set. In terms of the Mean IoU in Table 5, a descending sort can be obtained
as SCHNet with the base net VGG19 > SCHNet with the base net VGG16 >
PSPNet [6] > DeepLab-v3 [6] > Ding [11] > U-net [20] > Dilated FCN [16].
That is to say, SCHNet with the base net VGG19 performs best, whose Mean
IoU exceeds other models by at least 7.51%. It is believed that the improvement
mainly roots in the important roles of attention modules, which can capture
rich global dependencies explicitly for better feature representation.
5 Merits of the Proposed Computer Vision Approach
Apart from the accuracy of the proposed model in crack detection, it is necessary
to focus on two other crucial aspects including flexibility and robustness. To eval-
uate SCHNet more comprehensively, the concrete crack dataset is divided into four
scenarios, including Scenario I with holes on surfaces, Scenario II with shadows on
surfaces, Scenario III with rough surfaces, and Scenario IV with smooth surfaces.
Figure 11 presents some typical images in each scenario. Figure 12 provides the
segmentation results for images in the first column of Fig. 11. Discussions are
presented below.
(1) The developed SCHNet model is robust to different levels of noise, which
means that it is flexible and valid in different scenarios with holes, shadows,
and rough surfaces. Based on results from Table 6, the segmentation perfor-
mance in the four scenarios can be ranked as follows: Scenario IV > Scenario I
> Scenario III > Scenario II. More specifically, Scenario IV yields the highest
Mean IoU (88.67% with the base net VGG19 and 87.43% with the base net
VGG16, respectively). This duction is consistent with the reality, since images
in Scenario IV are more smooth with less noise. Inversely, the value of Mean
IoU is prone to be smaller in Scenario I, Scenario II, and Scenario III, which
is around 10% less than Scenario IV. It means that holes, shadows, and surface
roughness do exert a certain impact on crack segmentation. However, the Mean
IoU for the four scenarios all remains above 76%, meaning that influence from
noise is quite limited. Figure 12 gives examples about the segmentation perfor-
mance under four different scenarios in the test dataset, where the first, second,
and third column is the original image, segmentation results, and ground truth,
respectively. Even in Scenario I, II, and III under the relatively lower Mean
IoU, the developed model can reliably provide promising segmentation results
(see Fig. 12a–c). Therefore, the robustness of the model in crack detection is
validated.
(2) The developed SCHNet model with the base net VGG19 owns a greater gener-
alization ability in crack detection than the base net VGG16, as outlined in
Table 6. Taking Scenario IV as an example, its value of Mean IoU with the
base net VGG19 turns out to be 88.67%, while that with the base net VGG16
is 1.24% lower (87.43%). In a comparison of the segmentation performance
252 Computer Vision
Fig. 11 Typical images in the dataset under four different crack scenarios: a Scenario I (with
holes on surfaces); b Scenario II (with shadows on surfaces); c Scenario III (with rough surfaces);
d Scenario IV (with smooth surfaces)
between VGG16 and VGG19 under the same scenario, the value of Mean
IoU with the base net VGG19 is increased by 0.31% in Scenario I, 0.86% in
Scenario II, 0.74% in Scenario III, and 1.24% in Scenario IV, respectively. That
is due to the multi-nonlinear convolution layers in VGG19 that help to extract
more complex and expressive features. Therefore, SCHNet with the base net
VGG19 is capable of obtaining more representative features of images, which
eventually generate a more satisfying segmentation.
5 Merits of the Proposed Computer Vision Approach 253
Original image Segmentation result Ground truth

0
Width (pixel)
50
100
0 50 100 0 50 100 0 50 100

Length (pixel) Length (pixel) Length (pixel)
(a)
0
Width (pixel)
50
100
0 50 100 0 50 100 0 50 100

(b)
0
Width (pixel)
50
100
0 50 100 0 50 100 0 50 100

(c)
0
Width (pixel)
50
100
0 50 100 0 50 100 0 50 100

(d)
Fig. 12 Segmentation performance in the dataset under four different crack scenarios in the dataset:
a Scenario I (with holes on surfaces); b Scenario II (with shadows on surfaces); c Scenario III (with
rough surfaces); d Scenario IV (with smooth surfaces)
254 Computer Vision
Table 6 Segmentation
Scenario Basenet Mean IoU (%)
results of Mean IoU on the
testing set under four Scenario I VGG16 77.93
different crack scenarios Scenario I VGG19 78.24
Scenario II VGG16 76.31
Scenario II VGG19 77.17
Scenario III VGG16 77.12
Scenario III VGG19 77.86
Scenario IV VGG16 87.43
Scenario IV VGG19 88.67
6 Summaries
A novel computer vision-based hierarchical approach named SCHNet is developed

for concrete cracks image segmentation at the pixel level. Its unique feature comes
from a self-attention mechanism composed of three parallel modules, namely a
feature pyramid attention module, a spatial attention module, and a channel attention
module, to adaptively integrate local semantic features. To be more specific, the role
of the feature pyramid attention module is to decrease the amount of computation by
concentrating on regions and ignoring clusters in the input image. The spatial atten-
tion module and the channel attention module help to model global dependencies in
the spatial and channel dimensions, respectively. In the end, the output representa-
tions of the spatial and channel attention modules are fused using an element-wise
sum, aiming to capture the long-range dependencies for a better scene understanding.
The experimental study in a public dataset of concrete cracks (SDNET2018)
verifies the effectiveness and feasibility of our proposed model. Several conclusions
can be drawn up as: (1) The three attention modules can effectively capture long-range
contextual information in pursuit of better feature representation. The segmentation
performance of SCHNet is proven outstanding over other state-of-the-art models.
Evaluated by Mean IoU, SCHNet under VGG19 with the feature pyramid attention
module yields a result of 74.16%, which is 1.62% larger than the baseline FCN
(with VGG19). Similarly, the model with the spatial attention (or channel attention)
module individually exceeds the baseline by a 5.15% (or 5.76%) gain. (2) Three
new strategies, including data augmentation, multi-grid, and multi-scale strategies,
can work together to further improve Mean IoU. (3) SCHNet based on the base
net VGG19 can reach the highest Mean IoU of 85.31% in crack detection, which
is superior to other state-of-the-art models, including U-net, DeepLab-v2, PSPNet,
Ding, Dialted FCN. Moreover, SCHNet is very robust to noise in the form of shadows,
rough surfaces, and holes.
To sum up, the novelty of this deep learning-based computer vision approach for
semantic segmentation is the adoption of the self-attention mechanism that is made
up of three parallel attention modules, including a feature pyramid attention module, a
spatial attention module, and a channel attention module. Remarkably, the advantage
6 Summaries 255
of SCHNet lies in its ability to not only consider the semantic interdependencies in the
spatial and channel dimensions but also adaptively integrate local features into their
global dependencies. Except for the high value of Mean IoU, it should be noted that
SCHNet is robust to noises with a better generalization ability. Due to the promising
performance in crack detection, the application of SCHNet can be further extended
to the practical engineering project rather than the public dataset to make the model
more practical and effective. In addition, unmanned aerial vehicles (UAVs) can be
deployed to prepare the customized dataset. Plenty of photos or videos about the
cracks in various structural systems can be taken by UAVs, which will then be fed
into the developed deep learning approach for pixel-level automated crack detection.
References
1. Adhikari R, Moselhi O, Bagchi A (2014) Image-based retrieval of concrete crack properties

for bridge inspection. Autom Constr 39:180–194
2. Bang S, Park S, Kim H, Kim HJCAC, Engineering I (2019) Encoder–decoder network for
pixel-level road crack detection in black-box images. Comput Aided Civil Infrastruct Eng
34:713–727
3. Bello I, Zoph B, Vaswani A, Shlens J, Le QV (2019) Attention augmented convolutional
networks. In: Proceedings of the IEEE international conference on computer vision, pp 3286–
3295
4. Cha YJ, Choi W, Büyüköztürk O (2017) Deep learning-based crack damage detection using
convolutional neural networks. Comput Aided Civil Infrastruct Eng 32:361–378
5. Cha YJ, Choi W, Suh G, Mahmoudkhani S, Büyüköztürk O (2018) Autonomous structural
visual inspection using region-based deep learning for detecting multiple damage types.
Comput Aided Civil Infrastruct Eng 33:731–747
6. Chen L-C, Papandreou G, Kokkinos I, Murphy K, Yuille AL (2018) Deeplab: semantic image
segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs.
IEEE Trans Pattern Anal Mach Intell 40:834–848
7. Cheng JC, Wang M (2018) Automated detection of sewer pipe defects in closed-circuit
television images using deep learning techniques. Autom Constr 95:155–171
8. Cook W, Barr PJ (2017) Observations and trends among collapsed bridges in New York State.
J Perform Constr Facil 31:04017011
9. Dai J, Li Y, He K, Sun J (2016) R-FCN: object detection via region-based fully convolutional
networks. In: Advances in neural information processing systems, pp 379–387
10. Deng J, Lu Y, Lee VCS (2020) Concrete crack detection with handwriting script interferences
using faster region-based convolutional neural network. Comput Aided Civil Infrastruct Eng
35:373–388
11. Ding H, Jiang X, Shuai B, Qun Liu A, Wang G (2018) Context contrasted feature and gated
multi-scale aggregation for scene segmentation. In: Proceedings of the IEEE conference on
computer vision and pattern recognition, pp 2393–2402
12. Dung CV, Anh LD (2019) Autonomous concrete crack detection using deep fully convolutional
neural network. Autom Constr 99:52–58
13. Fang W, Love PE, Luo H, Ding L (2020) Computer vision for behaviour-based safety in
construction: a review and future directions. Adv Eng Inform 43:100980
14. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H., 2019. Dual attention network for
scene segmentation, Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 3146–3154.
256 Computer Vision
15. Li S, Zhao X (2019) Image-based concrete crack detection using convolutional neural network
and exhaustive search technique. Adv Civil Eng 2019:1–12
16. Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation.
In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 3431–
3440
17. Maguire M, Dorafshan S, Thomas RJ (2018) SDNET2018: a concrete crack image dataset for
machine learning applications
18. Pan Y, Zhang G, Zhang L (2020) A spatial-channel hierarchical deep learning network for
pixel-level automated crack detection. Autom Constr 119:103357
19. Prasanna P, Dana KJ, Gucunski N, Basily BB, La HM, Lim RS, Parvardeh H (2016) Automated
crack detection on concrete bridges. IEEE Trans Autom Sci Eng 13:591–599
20. Ronneberger O, Fischer P, Brox T (2015) U-net: convolutional networks for biomedical image
segmentation. In: International conference on medical image computing and computer-assisted
intervention. Springer, Berlin, pp 234–241
21. Shurun F (2018) Historic achievements of highway and bridge construction in China. http://
www.chinahighway.com/news/2018/1191607.php. Access Date: 19/11/2018
22. Voulodimos A, Doulamis N, Doulamis A, Protopapadakis E (2018) Deep learning for computer
vision: a brief review. Comput Intell Neurosci 2018:1–13
23. Yang X, Li H, Yu Y, Luo X, Huang T, Yang X (2018) Automatic pixel-level crack detection and
measurement using fully convolutional network. Comput Aided Civil Infrastruct Eng 33:1090–
1109
24. Zhang A, Wang KC, Li B, Yang E, Dai X, Peng Y, Fei Y, Liu Y, Li JQ, Chen C (2017) Automated
pixel-level pavement crack detection on 3D asphalt surfaces using a deep-learning network.
Comput Aided Civil Infrastruct Eng 32:805–819
25. Zhang L, Yang F, Zhang YD, Zhu YJ (2016) Road crack detection using deep convolutional
neural network. In: 2016 IEEE international conference on image processing (ICIP). IEEE, pp
3708–3712
26. Zhao H, Shi J, Qi X, Wang X, Jia J (2017) Pyramid scene parsing network. In: Proceedings of
the IEEE conference on computer vision and pattern recognition, pp 2881–2890
Conclusions and Future Directions
1 Main Conclusions
Recent decades have witnessed the rapid development of digital technology and
the growth of big data in the construction industry. In particular, AI implemen-
tation, which attempts to equip machines with human-like intelligent behavior and
reasoning, has gained a lot of attention. Currently, various AI techniques have created
tremendous value in revolutionizing the construction industry, leading to a more
reliable, automated, self-modifying, time-saving, and cost-effective process of CEM
[12]. In contrast to traditional computational methods and expert judgments, the
distinct advantage of AI lies in tackling complex and dynamic problems under great
uncertainty and intensive data, which is more likely to return accurate and convincing
results for improved decision-making. In general, AI-based CEM follows the key
steps including data acquisition and preprocessing, data mining based on appro-
priate models, and knowledge discovery and analysis. As a result, AI provides deeper
insights into big data, which commonly achieves value for itself from the following
three points.
(1) Modeling and pattern recognition
Modeling is a process of creating conceptual models in a standard, consistent, and
predictable manner, which can be known as a key prerequisite for further knowl-
edge interpretation and reasoning to resolve complex construction problems. It is
essential to ensure the high quality and reliability of the established model, since
the analytical results heavily rely on the model. There have been a wide variety of
knowledge representation methods based on rules, logic, and schemas that help to
construct models in an understandable form for computers. Also, rich information
of the real project, including declarative, procedural, structural, meta, and heuristic
knowledge, should be necessarily incorporated into the model to make it meaningful
and authentic. Apart from modeling, another way to extract information from data is
258 Conclusions and Future Directions
called pattern recognition. It mainly simplifies the complicated problem by decom-

posing it into small pieces and returns outputs only based on characteristics of data
itself, aiming to detect and retrieve hidden patterns and regularities behind large data
with ease and automaticity. Up to now, pattern recognition has shown its applica-
tion potential in the field of process mining, computer vision, NLP, and others. For
instance, process discovery algorithms can automatically define and map the end-to-
end construction process into a clear process model, and then the extracted digital
representation can serve as a foundation for fact-based analysis to provide sugges-
tions for process improvement. Besides, pattern recognition is especially effective
in feature extraction from images or videos. It is expected to automatically identify
damage-like, crack-like unsafety condition-like patterns to realize more efficient and
reliable condition and performance assessment of facilities.
(2) Prediction
The AI-powered analytic based on machine learning is typically a prediction task,
which makes precise predictions for new observations by learning given sets of
historical data. For supervised learning in the task of classification or regression, it
can assign the class label to data or predict future data values and trends. For unsuper-
vised learning in the clustering work, its goal is to divide data points without labels
into several meaningful clusters, and thus data in the same group owns similar traits.
For time series analysis, it explores a sequence of data points over time to identify
trends, seasonality, cyclicity, and noise. In the construction industry, the prediction is
an important strategic task for project control instead of empirical methods. There-
fore, AI applications are supposed to classify, quantify, and forecast potential risks
related to project performance and their relevant impacts, which can ultimately realize
promising diagnosis and analysis ahead of time concerning broad aspects of the
project performance, like planning, constructability, safety, productivity, and others.
In particular, the predicted results are considered as the baseline to guide proactive
management and drive the project towards its goals effectively and reliability.
For instance, since undesired delays will inevitably cause poor efficiency, cost
overruns, and other negative effects, the meaning of predicting possible delays in the
construction process lies in uncovering key factors related to bottlenecks and pursuing
high-accuracy estimation in project duration. Also, at both the construction and O&M
phases with uncertainty and dynamics, AI-related methods can assist in reliably
perceiving the safety risk of project systems in advance by taking various factors
along with their interrelationships into consideration. Based on the risk prediction,
immediate actions can be taken to reduce the likelihood of project failures.
(3) Optimization
Optimization can be understood as a decision-making process for seeking and deliv-
ering practical sustainable solutions to the construction project. It attempts to drive
a process perfectly adhere to a set of criteria and constraints by maximizing the
expected effects. It is known that popular optimization algorithms, like genetic algo-
rithm (GA), particle swarm optimization (PSO), and others, are suitable in mini-
mizing time and cost and maximizing productivity and safety, which have been
1 Main Conclusions 259
applied across the entire lifecycle of the project. For example, when the project
objectives are optimized at the planning stage, proper plans concerning strategy,
operation, and schedule can be developed to act as an important premise for the
project’s success. As for the phase of executing construction tasks in a compli-
cated site, optimization demonstrates its great value in better allocating resources,
arranging staff, determining layouts of facilities, and making corresponding adjust-
ments in a reasonable and timely manner. Lastly, O&M optimization intends to both
run the day-to-day operational tasks responsibly and perform maintenance measures
suitably at the optimal cost, and thus the infrastructure is likely to be kept in a satis-
factory state. Also, it contributes to reducing waste and building energy consumption
to support sustainability.
2 Future Directions
In the future, it is believed that CEM will continue to undergo rapid digital transfor-
mation. Therefore, more and more advanced technologies inspired by AI will serve
as the technological innovator in the domain of CEM. They will be implemented
and spread to the full lifecycle of the project ranging from the phase of design to
construction to O&M, which can potentially pave a more affordable and effective way
to relieve the burden on manual labor and facilitate smart construction management.
For future studies, we hold the opinion that the following six important directions that
have already been or tend to be hotspots will become the research focus, contributing
to further addressing various issues existing in laborious, complex, or even dangerous
tasks within the CEM domain.
(1) Smart robotics
Smart robotics have been progressing rapidly to drive a wide range of semi- or fully-
autonomous construction applications. Notably, ground robots and aerial robots are
two main types of robotics [1]. For the ground robots in different functions, they
have been developed to automate some manual processes and take over repeatable
tasks, such as brick-laying, masonry, prefabrication, model creation, rebar tying,
demolition, and others. They can also protect workers from work-related injuries
and accidents. To sum up, these robots can provide opportunities to solve issues
of labor shortage, lower operation costs, and ensure overall quality, productivity,
and safety. For the aerial robots, the representative is the unmanned aerial vehicles
(UAV) equipped with image acquisition systems (i.e., camera, laser scanner, go-
pros). Demands of aerial robots are going up to make the procedure of land survey,
site monitoring, and structure health monitoring easier, safer, more efficient and
affordable. More specifically, UAVs fly over the construction site or even fly into
the building structure in place of manual inspecting, which assist in taking high-
resolution images, capture real-time videos, and conduct laser scanning remotely to
maintain the safety of employees and detect structural defects, like cracks, erosion,
blister, spall, and others. Besides, robots can be trained by various machine learning
algorithms, and thus they can be equipped with the talent to learn from data and
conditions.
(2) Cloud virtual and augmented reality (Cloud VR/AR)
The evolutionary path of VR/AR is towards the cloud. Due to the development of
the fifth-generation (5G) networks and edge cloud technologies, cloud VR/AR solu-
tions have been generated to accelerate VR/AR applications, leading to better users’
experience. On the one hand, VR/AR as the information visualization technology
can realize easier interactions between the physical and cyber worlds, where VR
simulates the whole situation and AR integrates the information about the real enti-
ties with computer-generated images. Particularly, a distinct advantage of VR/AR
is to offer an engaging and immersive environment. Therefore, VR/AR has been
tentatively adopted to simulate hazardous construction scenarios, aiming to help
managers quickly identify dangers and issues in the working environment and then
formulate reasonable plans and measures ahead of accidents in a visual and interac-
tive way [8]. In addition, construction engineering education and training is another
common application area of VR/AR [17], which well trains workers based on both
visualization and experience in real time instead of courses taught by professionals.
It is expected to strengthen workers’ cognitive learning and safety consciousness
and even raise the overall productivity. On the other hand, the 5G evolution is fast
enough to stream VR and AR data from the cloud, bringing about great benefits in
cloud computing and interactive quality networking. In other words, the 5G tech-
nology can effectively enhance the data processing capability from the local computer
to the cloud and then make real-time perception along with responsive interactive
feedback. Real-time information exchange will also promote collaboration among
different participants.
(3) Artificial Intelligence of Things (AIoT)
The new generation of IoT is called AIoT, incorporating AI techniques into IoT
infrastructure to facilitate more efficient IoT operation and data analysis. IoT can
be defined as a network of interconnected physical devices, like sensors, drones,
3D laser scanners, wearable and mobile devices, and radio frequency identification
devices (RFID). As reviewed, various IoT devices have attached to construction
resources to collect real-time data about the operational status of the project. Smart
IoT-based sensing systems have become a research focus in many existing studies to
feasibly track the progress, monitor the worksite, which eventually support contin-
uous project improvement and accident prevention [6]. In the meantime, a huge
amount of recorded data can be shared over a network. To deeply explore such
rich data sources, various AI methods can be implemented to offer actional insights
for better supervision and decision making. That is to say, AIoT solutions for the
construction industry largely rely on real-time data transformation and instantaneous
data analysis. Empowered by AI, AIoT is, therefore, superior over the traditional IoT
to effectively conduct analysis and control functions for intelligent decision making.
Through synthesizing and analyzing data collected via IoT infrastructure in unprece-
dented volumes and rates, AIoT can achieve a lot of functionality, such as automating
2 Future Directions 261
the real-time decision making at an operational level, controlling the construction

worksite remotely, optimizing the project performance, and predicting future condi-
tions for the maintenance planning [5, 9]. However, the practical use of AIoT is
still in the startup phase, since this new technology still has some wrinkles to work
out, like the edge computing issue, security issue, and others. Besides, the BIM-IoT
integration has been an important technology trend, which is increasingly benefi-
cial in several prevalent domains, like construction operation and monitoring, health
and safety management, construction logistics and management, facility manage-
ment [15]. Accordingly, it can be envisioned that the synergy between AIoT and
BIM under 5G wireless communication will constantly develop to be the hot spot
in future works, which can considerably improve efficiency in the process of data
collection, transmission, and processing towards smart home, smart city, and smart
construction industry [11].
(4) Digital twin
The digital twin is a realization of the cyber-physical system for visualization,
modeling, simulation, analyzing, predicting, and optimizing. There are three crit-
ical parts to create a practical loop in a digital twin, including the physical entity,
virtual entity, and connection of data [10]. Typically, a digital twin works in two ways
of dynamic mapping [14]. For one thing, inspection data is collected in the physical
world and then sent to the virtual world for further analysis. For the other, processes
in terms of simulation, prediction, and optimization are carried out in the virtual
model to offer immediate solutions that can guide the realistic process and make it
adapt to the changeable environment [4]. In the construction industry, more attention
has been paid to develop the digital twin towards smarter construction services under
the inclusion of BIM, IoT, and data mining techniques [2, 13]. To be more specific,
the role of BIM is a digital representation, which can be the start point of the digital
twin. The data collection relies on different IoT devices, and thus a large amount of
gathered data will then enrich the BIM model. The digital twin can gain easy access
into both the as-built and as-designed models, enabling the information to be contin-
uously exchanged and synchronized between two key parts. To make full use of data,
a wide range of data mining techniques are leveraged for different purposes, such
as the automated monitoring of site progress, early detection of potential problems,
optimization of construction logistics and scheduling, evaluation of structure health,
and others.
(5) 4D printing
4D printing is an emerging technology by adding the fourth dimension of time into 3D
printing, aiming to make 3D printed objects change their shape and behavior over time
in response to external stimuli, like heat, light, temperature, and others. Notably, 4D
printing inherits all merits from 3D printing (also known as additive manufacturing)
that can produce objects layer-by-layer based on a computer-controlled machine
guided by digital 3D models. It has been investigated that 3D printing can bring about
some promising advancements as follows [7]. Firstly, since the 3D printer does not
require too much human surveillance, it can work automatically and efficiently with
great potential in shortening the project duration from weeks or months to a matter of
hours or days. Secondly, the 3D printer is more flexible in creating curved walls, roofs,
and other unstandardized shapes, endowing the printer objects with the characteristic
of great geometric freedom. Thus, more sophisticated designs and structures can be
easily built and customized with no restriction of the usual fabrication. Thirdly, the
materials for 3D printing possess special features, such as great tensile strength,
corrosion resistance, high-temperature resistance. This kind of optimized material
with controlled mechanical behavior helps to raise the durability of buildings. Lastly,
most adopted materials are recycled, organic, and eco-friendly, contributing to great
sustainability. It is possible to estimate materials costs and quantities more accurately,
which is helpful to avoid unnecessary waste and decrease material costs. However,
some problems do exist to impede the promotion of 3D printing, such as how to
control structural safety, how to conduct architectural paradigm shift, and how to
develop a rational and digital design workflow, which need to be solved at first.
Currently, 3D printing is moving towards 4D printing under the consideration of time
dimension and intelligent behavior, aiming to make breakthroughs in transforming
configurations for self-assembly, multi-functionality, and self-repair. The research
status of 4D printing in construction is still at its experimental stage, and thus some
new challenges will arise, such as the great demand for digitally savvy engineers,
advanced computational analysis, and structure verification in new designs [3].
(6) Blockchain
A nascent technology called blockchain is a powerful shared global infrastructure,
whose original goal is to simplify and secure transactions among parties [16]. To
make blockchain easier to understand, its basic concept can be explained as a veri-
fied chain with blocks of information, where each block saves process-related data
in a trusted environment. All historical data along with modifications can be there-
fore stored across a network and protected by cryptographic technology. Due to the
distributed ledger built in the blockchain, the stored digital information can be acces-
sible to users of the network simultaneously. Once a block is entered and verified, any
modifications are forbidden in the information. When the application of blockchain
is extended to construction, it provides two value opportunities: one is to gather
and consolidate data in a shared dashboard, while the other is to convert the project
management systems into a more transparent and secure practice with the following
potential applications [18]. For instance, blockchain can work together with BIM to
support life-cycle project management. It allows for collecting large data from various
stages of the project and sharing data securely among stakeholders. Also, the BIM
model can possibly be updated in time when it receives the next block of information.
As expected, such integration of blockchain and BIM will drive the project delivery
to be more automated and streamlined along with improving productivity, trustwor-
thiness, and cost. Besides, another key application of blockchain lies in creating a
smart contract, aiming to enforce and monitor the expected behavior by itself. Only
when certain criteria are satisfied, the process will be executed. This contributes
to high accuracy, compliance, transparency, cost-effectiveness, and collaboration in
activities, like payment, contract administration, and others. It is also important to
2 Future Directions 263
notice that blockchain can provide open and trustworthy information for improving
the supply chain. Therefore, all participants are allowed to make audits on issues
from the downstream supply chain.
References
1. Ardiny H, Witwicki S, Mondada F (2015) Construction automation with autonomous mobile

robots: a review. In: 2015 3rd RSI international conference on robotics and mechatronics
(ICROM). IEEE, pp 418–424
2. Boje C, Guerriero A, Kubicki S, Rezgui Y (2020) Towards a semantic construction digital
twin: directions for future research. Autom Constr 114:103179
3. Buchanan C, Gardner L (2019) Metal 3D printing in construction: A review of methods,
research, applications, opportunities and challenges. Eng Struct 180:332–348
5. Cheng JC, Chen W, Chen K, Wang Q (2020) Data-driven predictive maintenance planning
framework for MEP components based on BIM and IoT using machine learning algorithms.
Autom Constr 112:103087
6. Kanan R, Elhassan O, Bensalem R (2018) An IoT-based autonomous system for workers’
safety in construction sites with real-time alarming, monitoring, and positioning strategies.
Autom Constr 88:73–86
7. Labonnote N, Rønnquist A, Manum B, Rüther P (2016) Additive construction: state-of-the-art,
challenges and opportunities. Autom Constr 72:347–366
8. Li X, Yi W, Chi H-L, Wang X, Chan AP (2018) A critical review of virtual and augmented
reality (VR/AR) applications in construction safety. Autom Constr 86:150–162
9. Louis J, Dunston PS (2018) Integrating IoT into operational workflows for real-time and
automated decision-making in repetitive construction operations. Autom Constr 94:317–327
10. Min Q, Lu Y, Liu Z, Su C, Wang B (2019) Machine learning based digital twin framework for
production optimization in petrochemical industry. Int J Inf Manage 49:502–519
11. Mo Y, Zhao D, Du J, Syal M, Aziz A, Li H (2020) Automated staff assignment for building
maintenance using natural language processing. Autom Constr 113:103150
12. Pan Y, Zhang L (2021) Roles of artificial intelligence in construction engineering and
management: a critical review and future trends. Autom Constr 122:103517
13. Pan Y, Zhang L (2021) A BIM-data mining integrated digital twin framework for advanced
project management. Autom Constr 124:103564
14. Qi Q, Tao F (2018) Digital twin and big data towards smart manufacturing and industry 4.0:
360 degree comparison. IEEE Access 6:3585–3593
15. Tang S, Shelden DR, Eastman CM, Pishdad-Bozorgi P, Gao X (2019) A review of building
information modeling (BIM) and the internet of things (IoT) devices integration: present status
and future trends. Autom Constr 101:127–139
16. Turk Ž, Klinc R (2017) Potentials of blockchain technology for construction management.
Procedia Eng 196:638–645
17. Wang P, Wu P, Wang J, Chi H-L, Wang X (2018) A critical review of the use of virtual reality
in construction engineering education and training. Int J Environ Res Public Health 15:1204
18. Wang J, Wu P, Wang X, Shou W (2017) The outlook of blockchain technology for construction
engineering management. Front Eng Manag 67–75

Artificial Intelligence in Construction Engineering and Management

Uploaded by

Copyright:

Available Formats

Artificial Intelligence in Construction Engineering and Management

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Artificial Intelligence in Construction Engineering and Management

Uploaded by

Copyright:

Available Formats

Lecture Notes in Civil Engineering

To submit a proposal or request further information, please contact the appropriate

More information about this series at http://www.springer.com/series/15087

Xianguo Wu Mirosław J. Skibniewski

ISSN 2366-2557 ISSN 2366-2565 (electronic)

highlight the benefits of AI techniques in providing intelligent solutions by learning

Singapore Limao Zhang

Introduction to Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Fuzzy Modeling and Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.3 Process Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

5.2 Engineering Parameters Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

The term Artificial Intelligence (AI) is a branch of computer science to make

then, researchers have emphasized adopting algorithms to solve mathematical prob-

Fig. 1 An overview of common AI techniques

Expert systems: An expert system is a straightforward and understandable method

Machine learning: Machine learning is a great step of AI to enable machines to

2.1 Significance of CEM

2.2 Activities in CEM

CEM is consisting of a series of physical activities for constructing infrastructures,

(3) Operation and Maintenance (O&M)

2.3 Characteristics of CEM

Since a construction project is unique, temporary, and progressive in nature for

3 How AI Benefits CEM

Although AI stands out as transformational technology to potentially bring about

advanced analytics help managers to better understand the construction project

4 Organization of the Book

method, Structural Equation Model (SEM), is chosen to discover casual relationships

Construction is one of the most dangerous industries worldwide, leading to a common

2 Safety Leadership in Construction

3 Knowledge Representation, Learning, and Discovery

In the construction industry, knowledge is represented in various forms and can be

SEM. Wu et al. [29] proposed a systematic SEM-based approach to the evaluation

4 SEM-Enabled Knowledge Discovery

4.1 Survey and Participants

4.2 Hypothesis on Causal Relationships

To clarify relationships among different stakeholders in construction plays a vital role

contractor according to a competitive bidding system. The DBB mode is extensively

Sub-contractors ... Suppliers

A useful hypothesis enables predictions and analysis by applying deductive

• Hypothesis H6b : The safety management performance of the supervisor positively

4.3 Knowledge Learning and Discovery

where Ai j refers to exogenous observable variables, T and Ai are endogenous latent

5 Knowledge Learning from Data and Validation

5.1 Data Collection

Table 2 Descriptions of 26 items affecting construction safety performance on sites

Table 3 Statistics of valid respondents within different stakeholders in the survey

The KMO measurement value of the collected data is calculated to be 0.922.

5.2 Common Method Bias

When a self-reported questionnaire is used to measure variables, there is a possibility

5.3 SEM Development

Table 5 Results of the

5.4 Measurement Model Evaluation

5.5 Structural Model Evaluation

(1) Hypothesis testing

Hypothesis testing is a primary task in the SEM-based applications. The results

Table 6 Results of regression weights in the established SEM

Fig. 3 Fitted SEM for modeling construction safety leadership

Table 7 Results of the goodness-of-fit verification in the established SEM

6 Knowledge Discovery in Safety Leadership