Safety Science 70 (2014) 161–171
Contents lists available at ScienceDirect
Safety Science
journal homepage: www.elsevier.com/locate/ssci
Review
Fall risk assessment of cantilever bridge projects using Bayesian network
Tung-Tsan Chen a,⇑, Sou-Sen Leu b
a
b
Department of Civil Engineering and Engineering Management, National Quemoy University, No. 1 University Rd., Kinmen County 892, Taiwan, ROC
Department of Construction Engineering, National Taiwan University of Science & Technology, No. 43, Sec. 4, Keelung Rd., Taipei 106, Taiwan, ROC
a r t i c l e
i n f o
Article history:
Received 4 September 2013
Received in revised form 24 March 2014
Accepted 17 May 2014
Keywords:
Bayesian network
Bridge construction
Fault Tree
Fall risks
a b s t r a c t
Fall or tumble is one of the most common accidents in bridge construction. Failing to implement safety
management and training effectively may result in serious occupational accidents. Current site safety
management relies mostly on checklist evaluation; however, its effectiveness is limited by the experience
and the abilities of the evaluators, which may not consistently achieve the goal of thorough assessment.
Recently, several systematic safety risk assessment approaches, such as Fault Tree Analysis (FTA) and
Failure Mode and Effect Criticality Analysis (FMECA), have been used to evaluate safety risks at bridge
projects. However, these traditional methods ineffectively address dependencies among safety factors
at various levels that fail to provide early warnings to prevent occupational accidents. In order to
overcome the limitations of the traditional approach, in this paper a fall risk assessment model for bridge
construction projects is developed by establishing a Bayesian network (BN) based on Fault Tree (FT)
transformation. The model was found to provide much better site safety management ability by enabling
better understanding of the probability of fall risks through the analysis of fall causes and their relationships in a BN. The system has been used to analyze and verify safety practices at five cantilever bridge
construction projects currently under construction in Taiwan. It was found that BN analysis is consistent
with the conventional safety performance assessment. In practice, based upon the BN analysis by
inputting prior information about the site safety management, the probabilities of fall risks and their sensitive factors can be effectively assessed. Proper preventive safety management strategies could then be
established to reduce the occurrences of fall accidents at the bridge construction projects.
Ó 2014 Elsevier Ltd. All rights reserved.
Contents
1.
2.
3.
4.
5.
6.
7.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Literature survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Statistics of occupational accidents in Taiwan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Methods and transformation process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.
Bayesian network (BN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.
Conversion from a FT to a BN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.
Computation of CPT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Assessing BN-based fall risk of cantilever bridge construction projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.
Building FT framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.
Construction of BN from FT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.
CPT calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.
Assessment of prior probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Model validation and sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.
Model validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.
Sensitivity analysis and discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusions and future developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
⇑ Corresponding author. Tel.: +886 82313352; fax: +886 82313354.
E-mail addresses: tungtsan@nqu.edu.tw (T.-T. Chen), leuss@mail.ntust.edu.tw (S.-S. Leu).
http://dx.doi.org/10.1016/j.ssci.2014.05.011
0925-7535/Ó 2014 Elsevier Ltd. All rights reserved.
162
162
163
164
164
165
165
166
166
167
167
167
168
168
169
170
171
162
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
fundamental frameworks and BN is then developed by FT transformation. Section 6 discusses and verifies a fall risk assessment
model by effectively transforming a FT into a BN framework, and
Section 7 concludes the paper.
1. Introduction
A complete transportation network in high-speed railroad or
highway systems must rely on the use of bridges. Unfortunately,
the construction of bridges is often accompanied by occupational
accidents, such as fall and object collapse. The construction companies are required to take every safety measure to prevent occupational accidents. The current method to implement safety
management is to inspect regularly with checklists on unsafe
equipments and worker behaviors. However, this current method
is conducted under passive supervision, which fails to provide
warning in advance about likely occupational accidents. Recently,
several systematic safety risk-assessment approaches, such as
FTA and FMECA, have been used to evaluate safety risks at bridge
projects. However, these traditional methods ineffectively address
dependencies among safety factors at various levels that fail to
provide early warnings to prevent occupational accidents. Because
of that, several new approaches have been developed to address
the relationship among a variety of safety variables in order to
devise a preventive model. Structural equation models (SEM) and
BNs are some classical examples of the approaches (Paul and
Maiti, 2007; Martin et al., 2008). Using BNs, the most important
causes of site accidents can be identified and, most importantly,
the relationships among these causes can also be determined,
which may allow early and preventive safety measures to be
implemented.
In general, there are three approaches to construct a BN: (1)
learning from a large amount of training data; (2) basing analyses
upon the experiences of domain experts; and (3) hybrid method.
The second approach is usually used for practical BN construction
because the training data are often limited in engineering fields. In
addition, establishing a mutual relationship among nodes in the
network by directly incorporating the views of experts is generally
difficult and tedious. It could be more effective to build the BN by
FT transformation (Boudali and Dugan, 2005; Franke et al., 2009;
Marsh and Bearfield, 2007; Qian et al., 2005; Xiao et al., 2008).
The rest of this paper is organized as follows. Section 2 reviews
the state of the art of safety risk assessment and BNs. Section 3
introduces basic statistics about occupational accidents at bridge
construction projects in Taiwan. Section 4 describes the basic
concepts of FT and BN. Section 5 discusses the BN development
process proposed in this study. Mainly, FT provides the
2. Literature survey
Common methods used to assess risks include in-depth interviews, Delphi, Factors Analysis, FTA, FMECA, etc. Moreover, quantitative risk-analysis methods, such as statistical inference,
reliability analysis, decision trees, and simulation, have also been
used for safety risk assessment in construction projects (Hartford
and Baecher, 2004; Ebeling, 1997; Rao, 1992; O’connor, 2002;
Kales, 2006). Nevertheless, limited data can generally be collected
during the life cycle of the construction projects, which constrains
the use of these quantitative methods in safety assessment because
of data availability. Furthermore, the methods, such as FTA and
FMECA, ineffectively address dependencies among safety factors
at various levels that fail to provide an early warning in preventing
occupational accidents. Thus in recent years, BN has become a
popular tool used for the risk assessment on uncertain causal
relationships among multi-dimensional factors. Bedford and
Gelder (2003) assessed safety of third parties during construction
in multiple places using BNs. Martin et al. (2008) used BN to analyze workplace accidents caused by falls from a height. In addition
to safety assessment, BNs have been also applied in several knowledge areas, such as medicine (Antal et al., 2007), ecology
(Adriaenssens et al., 2004), environmental assessment impact
(Baran and Jantunen, 2004; Marcot et al., 2001; Matias et al.,
2007), business risk (Marcot et al., 2001), and product life-cycle
analysis (Zhu and Deshmukh, 2003).
Generally, the knowledge of professionals is used in developing
BNs that illustrate problems with causal relationships between
nodes and their conditional probabilities. However, the direct construction of BN is more applicable to simple problems, even though
it is quite difficult to develop complicated BNs directly. At present,
some scholars have proposed several systematic approaches to BN
construction by FT transformation. The main techniques make use
of
and
logic transforms into BNs to perform
probabilistic analysis of event occurrences (Bobbio et al., 1999,
2001; Xiao et al., 2008; Boudali and Dugan, 2005; Marsh and
Bearfield, 2007; Qian et al., 2005; Franke et al., 2009). Some past
0.24
0.22
All industry
0.223
Manufacturing
0.21
0.2
Construction
0.188
Fatalities per 1000
0.18
0.172
0.175
0.161
0.16
0.14
0.129
0.131
0.128
0.12
0.123
0.109
0.1
0.08
0.06
0.077
0.069
0.065
0.063 0.067 0.059
0.05
0.044 0.045
0.04
0.041
0.02
2000
2001
2002
2003
0.038 0.033 0.035 0.034
0.03 0.038 0.035 0.034 0.036
2004
2005
2006
2007
2008
0.028
2009
0.036
0.035
2010
Year
Fig. 1. Fatalities per 1000 persons in construction industry and all industries (excluding deaths from occupational disease and traffic accidents), 2000–2010.
163
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
studies regarded both events and logic gates in FTs as nodes in the
BN. Nevertheless, these two have different definitions and purposes. Logic gates are used to describe the relationship between
events in a sequence. Therefore it is meaningless to convert them
into physical BN nodes. To overcome the above-mentioned conflict,
some transformation algorithms mapped events of a FT into nodes
of a BN while logical gates of the FT are embedded in conditional
probability tables (Khakzad et al., 2011). Moreover, it is generally
assumed that there are only two states inside events in the conventional FTA; i.e., either fault or normal. For the more complicated
risky situations, there is room for improvement in multi-state FT
(Graves et al., 2007). Therefore, this study first combines FT and
BN to develop a reasonable transformation process of a FT to a
BN. Afterward, a fall risk assessment model for bridge construction
projects is established based upon the transformation procedures
proposed in this study.
900
3. Statistics of occupational accidents in Taiwan
The rate of incidence of major occupational accidents in the
Taiwanese construction industry is rather high. According to the
Yearbook of Labor Statistics published by the Taiwanese Council
of Labor Affairs, the rate of fatalities per 1000 workers in the construction industry (excluding the deaths from occupational disease
and traffic accidents) was 0.109 in 2010, which is much higher
than that in other industries (Fig. 1). As shown in Fig. 2, of all occupational accidents in Taiwan, fall is the most common cause of
injury (48%) among all accident types over the past decade
(2000–2010). As shown in Fig. 3, the number of fall deaths
(2000–2010) in bridge construction projects was 15 deaths (up
to 22%). Although fall is the second accident cause in bridge construction projects, it generally results in deaths, compared with
820
(48%)
800
No of obs (percentage)
700
600
500
325
400
(19%)
300
144
(8%)
12
13
23
14
1
(1%)
(1%)
(1%)
(1%)
(0%)
ge
ts
1
(0%)
Pu
ea
A
br
as
ec
io
n,
Br
ts
A
re
O
bj
Fi
Pi
ka
en
w
id
cc
th
do
ll
Fa
O
Co
n
s
er
de
lli
ec
bj
lO
fu
m
ec
H
O
ar
bj
d
he
nc
ts
ng
n
lli
Fa
ts
Ex
ob
by
A
pl
je
os
ct
io
s
it
H
ts
e
en
ps
cc
Co
fic
ts
Tr
bj
O
af
ec
ec
El
id
lla
w
D
tri
ro
fic
Fa
at
ni
io
lin
n
ng
g
0
3
(0%)
e
27
(2%)
(1%)
s
(2%)
ok
(5%)
20
tre
34
Ch
(3%)
(3%)
91
82
(5%)
54
46
100
nc
200
Fig. 2. Occupational accidents of the construction industry in Taiwan (2000–2010).
60%
51%
50%
No of obs
40%
30%
22%
20%
10%
9%
10%
3%
3%
1%
1%
0%
0%
0%
0%
0%
0%
ha
cu
Co
nt
ac
tw
ith
be
d,
rm
t,
a
ts
Fig. 3. Occupational accidents of the bridge Construction in Taiwan (2000–2010).
Fi
re
br
as
io
fu
n
ls
ub
st
an
ce
Ex
pl
o
O
si
bj
on
ec
ts
ru
pt
ur
e
fly
in
g
er
th
O
O
bj
ec
led
ol
sr
St
ab
Ca
u
af
gh
t,
it
i
cc
id
en
ts
ti o
n
tr o
ec
el
fic
A
cu
ng
ro
w
ni
H
it
D
ol
ng
,r
Fa
lli
Tr
O
bj
ec
ts
co
lla
ps
e
lin
g
0%
164
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
Bayesian
Network
Fault Tree
Primary
Events
Mapping
Intermediate
Events
Mapping
T
FTA of Falling Risk of Bridge
Construction Projects
(Cantilever construction method)
Root Nodes
Intermediate
Nodes
Graphical
Mapping
Top Event
Numerical
Mapping
Event
Occurrence
Probability
Boolean Gates
Mapping
Mapping
Mapping
Leaf Node
Prior
Probability of
Root Nodes
Conditional
Probability
Tables
Fig. 4. Transformation flow chart from FT to BN.
object collapse (the most common cause in bridge construction).
Thus, preventing fall accidents from occurring becomes the most
crucial plan of action in the safety management program for
bridges construction.
To effectively plan and promote accident-mitigation strategies
and allocate resources to safety and health management, it is necessary to conduct a detailed analysis of the relationship between
fall accidents and the factors leading to such occurrences, such as
personnel, equipment, processes, and management. This paper
combines FTA and BN to establish a fall risk assessment model
for bridge construction projects in order to identify the risk potential of and other significant causes to fall accidents. With a better
understanding of fall risks and their significant causes, more effective accident-preventive measures can be taken to prevent the
occurrence of falling at bridge construction sites.
4. Methods and transformation process
The Fault Tree is transformed into a BN using the conditional
probability tables (CPTs) of the logic gates; in addition, expert
G1
Falling casued by
Pier Column
Construction
G2
Falling casued by
Section block
construction
G3
Falling casued by
Uncoordinated
behavior
Fig. 5. FT of falling accidents at cantilever bridge construction projects.
opinions are used to add supplemental links between nodes in
the BN. The simplified procedure of mapping FTs into BNs is
depicted in Fig. 4 (Khakzad et al., 2011). In the following paragraphs, the basics of BN and the transformation processes are
explained in detail.
4.1. Bayesian network (BN)
Combined with probability theory and graph theory, BNs
consist of nodes, joints among nodes, and CPTs. A BN has a high
efficiency and accuracy in uncertain inference, especially for complicated systems with highly-correlated elements (Qian et al.,
2005; Doguc and Ramirez-Marquez, 2009; Bobbio et al., 1999;
Xiao et al., 2008). In recent years, BNs have been widely used in
areas with high degrees of uncertainty or mutual influences
(Qian et al., 2005; Doguc and Ramirez-Marquez, 2009; Bobbio
et al., 1999; Xiao et al., 2008).
As mentioned above, making analysis with the experience of
domain experts is commonly used for practical BN construction
due to the constraint of data availability. Nevertheless, it is generally complicated to establish mutual relationships among nodes in
the network solely based on the knowledge of engineers and
experts. Therefore, several transformation processes from FTs to
BNs have been proposed (Bobbio et al., 1999, 2001; Xiao et al.,
2008; Boudali and Dugan, 2005; Marsh and Bearfield, 2007; Qian
et al., 2005; Franke et al., 2009). In some past transformation, the
conversion of logic gates in a FT into a BN is one-to-one; i.e., logic
gates in a FT are converted into corresponding physical nodes in a
Table 1
Experts’ profiles.
No.
Department
Title
Numbers
Seniority (yrs)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Total
Dept. of safety, health and environmental engineering
Industrial safety & QC dept.
Industrial safety & QC dept.
H/S office
Construction office
Construction office
Construction project site office
Construction project site office
Construction project site office
Construction project site office
Construction planning
H/S section
Labor safety & QC office
Labor safety office
Environmental safety office
Lecturer
Manager
Vice manager
Specialist
Executive V.P.
V.P.
Director
Vice director
Senior Engineers
Engineer
Junior manager
Section manager
Supervisor
Supervisor
Senior manager
2
2
2
2
1
1
5
5
10
5
2
1
2
1
1
42
24, 30
25, 28
15, 17
15, 19
26
32
21, 24, 25, 27, 33
13, 15, 16, 16, 18
15–30
10–20
16, 22
16
17, 21
23
16
165
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
BN. However, there are differences in the significance of an event
node in a BN and that in a logic gate in a FT. An event node is used
to represent a variable in the problem domain, while a logic gate is
used to describe the logical relationship between nodes. In the
transformation from a FT to a BN, the event nodes and the logic
gates should be treated separately. To overcome the above-mentioned conflict, some new-developed transformation algorithms
mapped events of a FT into nodes of a BN while logical gates of
the FT are embedded in conditional probability tables (Khakzad
et al., 2011).
intermediate events, and the top event of FT are represented as
root nodes, intermediate nodes, and the leaf node in the corresponding BN, respectively. The nodes of a BN are connected in
the same way as corresponding components in FT. In the numerical
mapping, the occurrence probabilities of the primary events are
assigned to the corresponding root nodes as prior probabilities.
For each intermediate node and leaf node, a conditional probability
table (CPT) is developed. The CPTs are developed according to the
type of gate (Bobbio et al., 2001). Fig. 4 summarizes the simplified
procedure of mapping FTs into BNs (Khakzad et al., 2011).
4.2. Conversion from a FT to a BN
4.3. Computation of CPT
The construction of the FT follows a top/down design. It’s first to
identify a particular undesired event as a top event, and then
proceed from the event to its causes until the primary events are
reached. The relationships between events and causes are commonly defined and represented by means of
or
logic
gates (Bobbio et al., 1999, 2001; Xiao et al., 2008; Boudali and
Dugan, 2005; Marsh and Bearfield, 2007; Qian et al., 2005;
Franke et al., 2009).
The transformation from FT into BN includes graphical and
numerical tasks. In the graphical mapping, primary events,
The CPT structure becomes complicated when a node in the BN
has several parent nodes, or when each parent node and child node
has several states. For example, consider a child node with three
parent nodes and the number of states is three. Then the total
number of CPT values will reach up to 34 (81). In addition, since
the CPT values are generally defined by experts based on their
experiences, the elicited probability values could be inconsistent,
especially for the complicated CPT condition. The software, AgenaRisk, is used to eliminate the above-mentioned difficulties in this
study (Agena, 2008). Through parameters defined in the software
T
G1
G3
G2
F2
F1
E2
E1
E3
F3
E4
E5
D3
D2
C2
C1
D1
C5
B13 B12
B5
D4
D5
D6
D7
B6
B6
B6
A4 A1 A2 A1
A1
A1
B10
B10
C4
C3
F4
B11
B17
B4
B11 B15 B6
B5
B12 A4 A3 A4 A1 A2 A1 A4
B3
B13 B12 B14 B7 B16 A1
A2 A4
B1
A3 A4
A3 A3 A4
B2
A1 A2 A3 A1 A2 A2
A4
B13 A1 A3 A3 A4
A4
A1 A4 A1 A4 A1 A4 A1 A4
C6
B2
B3
C7
C8
C9
C10
B2
B3
B9
B8
A1 A4 A1 A3 A2 A1 A4 A1 A3 A1 A1 A4 A1 A3 A1 A1 A4 A1 A3 A1 A1 A4 A1 A3 A2
A1
B9
B2
B3
B8
B2
B3
B8
B2
B3
B8
Fig. 6. Overall FT of falling accidents of cantilever bridge construction projects.
166
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
and the weights among nodes defined by experts, the probability
values in CPT can be calculated rather effectively.
In theory, the calculation of CPT in AgenaRisk follows rank
nodes (Fenton et al., 2007). Ranked nodes represent discrete variables, the states of which are expressed on an ordinal scale that
can be mapped onto a bounded numerical scale that is continuous
and monotonically ordered. The two basic components of rank
nodes are the weight values of parent nodes to children nodes
and weighted functions. The probability of a child node is defined
as a weighted rank function of the parent node values. Weighted
rank node functions are generally assumed to be the doubly truncated normal distribution, and its central tendency is controlled by
a function parameter (i.e., mean average, minimum, and maximum). The definition of the function parameters in AgenaRisk is
significant to the defining of CPT. In the selection of the function
parameter, minimum is selected if the corresponding logic gate
in the FT is
, whereas maximum is selected if the logic gate
is
.
After the function parameters in AgenaRisk are defined based
upon the logic gates in the FT, the weights are determined and
inputted based the opinion poll of experts on the contribution of
parent nodes to children nodes. The weight score ranges from 1
to 5; 1 meaning the least impact of one parent node on the child
node, 5 meaning the most impact. Once the above-mentioned data
are input into AgenaRisk, all the CPTs in a BN can be calculated
rather efficiently and effectively. In addition, the subsequent probabilities of the top event and of all intermediate nodes in a BN can
be inferred through AgenaRisk.
5. Assessing BN-based fall risk of cantilever bridge construction
projects
Based upon the proposed BN construction process, we assess
the BN-based fall risk of the cantilever bridge construction
projects. In order to obtain sound knowledge input, 42 specialists,
whose background information is listed in Table 1, were interviewed for the construction of the model. Furthermore, the validity
of the model was verified against five cantilever bridge construction projects. Finally, the causes affecting fall risks were assessed
by means of sensitivity analyses. The detail of the model development is explained below.
5.1. Building FT framework
Due to its frequent occurrence and severe impact in bridge construction projects, fall is selected as the top event of the FT in this
research. Based on the domino theory in safety management
(Heinrich et al., 1980), the causes of fall accidents at work can be
classified into: 1. accident locations and situation (such as falling
caused by pier column construction, falling caused by section block
construction, and falling caused by uncoordinated behavior) and
their detail; 2. indirect causes and their detail (such as unsafe
behaviors, unsafe equipment and unsafe environments); and 3.
primary causes (including improper safety plans, inappropriate
environment maintenance, insufficient safety training, and poor
safety management). The relations between these causes and the
top event are connected by logical gates as shown in Fig. 5.
The circumstances surrounding work tasks that could play a
role in triggering a fall accident can be analyzed by means of expert
interview and literature review. For example, at the pier column
construction, there are two main tasks in which a fall could
happen: (1) improper lifting operations; and (2) incorrect scaffolding setting. If necessary, these tasks can be broken down into more
detailed subtasks. Furthermore, the indirect causes that may
trigger a fall accident at the pier column construction can also be
analyzed step by step.
Based on the domino theory and safety management concepts
(Jitwasinkul and Hadikusumo, 2011; Lingard and Rowlinson,
2005), four primary causes which results in the occurrence of
occupational accidents are: insufficient safety training, poor site
environment management, poor safety and health management,
and improper health and safety plan. With the aid of occupational
accident records, safety theories, and expert interview, the interaction of these basic reasons and their indirect causes were identified
to form the overall FT. The completed FT of falling at bridge construction projects are shown in Fig. 6. As well as their codes and
definition are summarized in Table 2.
Table 2
Codes and definition of overall FT of falling accidents.
Code
Risk factors
Code
Risk factors
T
G1
G2
G3
F1
F2
F3
F4
E1
E2
E3
E4
E5
D1
D2
D3
D4
D5
D6
D7
C1
C2
C3
C4
C5
Falling Risk of bridge construction Projects
Falling caused by pier column construction
Falling caused by section block construction
Uncoordinated body movement
Improper lifting operations
Incorrectly scaffolding setting
Imbalance standing
Dangerous behavior
Operation error of steel wire rope
Collision of hanging material
Insufficient materials of level department
The construction of junction point is not solid
Sheets and the ground are not lock tight
Operator error of crane
Dangerous position
Stumble
Improper force
False step
Dangerous position
Walking on dangerous path
No intermediate supports
No use of safety belt
Wrong order
Operator’s error
No use of hauling rope
C6
C7
C8
C9
C10
B1
B2
B3
B4
B5
B6
B8
B9
B10
B11
B12
B13
B14
B15
B16
B17
A1
A2
A3
A4
Imbalance of both sides of work vehicles
Unstable gravity of working vehicle
Incorrect way of promoting business
Improper molding or operating
Improper form of removal operations
Poor control process
Incorrect operation process
Dangerous procedure or method
Placing unsafely
Unaware of surroundings
Improper behaviors or posture
Operator error
Without safety measure
Job-site environment is disorder
Unorganized material
No use of safety device
Unsafe equipment
No use of personal protective equipment
Waste not cleaned up
Fail to implement self-management
Lack of environmental labeling
Poor H/S training
Poor environmental maintenance
Poor H/S planning
Poor H/S management
167
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
5.2. Construction of BN from FT
According to the transformation process shown in Section 4, the
FT diagram was transformed to a BN. The overlapping nodes were
combined into one and meaningful supplementary arcs among the
BN nodes were added based upon the experts’ experiences. The
complete BN framework is shown in Fig. 7. The follow-up analysis
and the in-depth investigation were then explored.
5.3. CPT calculation
AgenaRisk was used to calculate CPT based on the constructed
BN framework. Either Maximum or minimum values were selected
for the function parameters based upon the definition of FT logic
gates. Furthermore, questionnaires were designed to collect the
information about the relative weights of parent nodes on their
child nodes. The questionnaire example is depicted in Table 3.
For example, there are three influencing factors (G1, G2, and G3)
of falling risk at the cantilever bridge projects (T). Based upon their
impacts to the falling risk, the weight scores in column 3 are
defined and filled (5 means the greatest influence, and 0 is minimal
influence). A total of 42 experts were invited to assess 97 questions
based upon their own practical experiences, and their answers
were statistically analyzed. Using the above-mentioned input data,
the CPTs for all the arcs in the BN were successfully calculated.
Lastly, the subsequent probabilities of the BN nodes (including
both top nodes and the intermediate nodes) are generated once
the probabilities of root causes (i.e. prior probabilities of BN) are
assessed and inputted into the complete BN.
5.4. Assessment of prior probabilities
As stated above, four significant primary causes were defined in
the model and they are insufficient safety training, poor site
T
G1
G3
G2
F2
F1
C6
E3
E1
F4
F3
C8
E2
C7
E5
D2
C9
D1
D6
D3
C1
E4
C5
D4
C10
C2
D5
D7
C4
C3
B13
B2
B12
B6
B16
B8
B9
B5
B14
B11
B3
B15
B4
B7
B10
B1
B17
A3
A4
A1
A2
Fig. 7. BN of falling accidents at cantilever bridge construction projects.
168
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
environment management, poor safety and health management,
and an improper health and safety plan. In order to assess the prior
probabilities of these causes, a safety performance evaluation table
was established in this study (sample as shown in Table 4). Significant safety performance checklist items of each primary cause
were summarized and listed in the table. The more items marked
based upon the site investigation, the higher probability of a poor
performance on the root cause evaluation is. By inputting prior
probabilities into BN, fall risks at bridge construction projects
and their significant causes can be identified through this model.
6. Model validation and sensitivity analysis
The results of BN inferences were validated against the actual
safety inspection records of five cantilever bridge construction projects currently under construction in Taiwan. Sensitivity analysis
was further performed to identify key causes of fall accidents at
the bridge construction projects. The results from the sensitivity
analysis may be used as an important reference in future diagnostic analysis and prevention programs for unsafely behaviors and
unsafe environment.
6.1. Model validation
From the results of the actual safety inspection records of five
cantilever bridge construction projects and the probabilities of
the top node in the BN, the proposed BN model was validated.
Basic information about the five bridge construction projects is
shown in Table 5. According five cantilever bridge construction
project monthly site safety inspection records, and the summary
of their actual safety inspection records assessment score are
depicted in column 4 of Table 6. In this study, the safety performance status of each project was evaluated by means of the safety
performance evaluation table. The prior probabilities of the four
root causes were then subjectively assessed and then input to
AgenaRisk for inferring the posterior probabilities of nodes in BN.
Table 6 compares the analytical results of the BN model to the
actual safety inspection records. It can be observed that the ranks
of probabilities from the BN model are highly consistent with the
ranks of safety performances obtained from the real site assessment records. However, in Project 1 the probability of fall risk from
the BN is 38.942 (%), and in Project 4 that of fall risk from the BN is
39.549 (%), the two risk percentages are very close, but the safety
performances obtained from the actual records shows a gap of
8.88 points.
Monthly safety inspection records of five cantilever bridge construction projects were looked into in detail. The top ranked project is indeed excellent in site safety management. Its record of
poor safety performances, such as safety penalties and near miss
incidents, is fewer in entries than those of the other four projects.
By contrast, the bottom-ranked project performed poorly on safety
management and the BN-based model indicated its fall risk was up
to 78.45%. Through the actual appraisal and verification of these
five bridge construction projects, the BN-based fall risk assessment
model has shown its accuracy and applicability, and potential as a
Table 3
Questionnaire example of relative weights of BN arcs.
AgenaRisk
id
T
Effect
FTA of Falling Risk of bridge construction
Projects
G1
Falling caused by pier column construction
G2
Falling caused by section block construction
G3
Falling caused by Uncoordinated behavior
F1
Improper lifting operations
F2
Incorrectly scaffolding setting
F3
Imbalance standing
F4
Dangerous behavior
E1
Operation error of steel wire rope
E2
Collision of hanging material
B14
Without personal protective equipment
B15
B16
B17
Waste not cleaned up
Fail to implement self-management
Lack of environmental labeling
Levels of influence
Low
Medium
High
0
2
4
1
3
Cause
AgenaRisk
code
Q&A
no.
Falling caused by pier column construction
G1
Q1
Falling caused by section block construction
Falling caused by uncoordinated behavior
Improper lifting operations
Incorrectly scaffolding setting
Imbalance of both sides of work vehicles
Unstable gravity of working vehicle
Incorrect way of promoting business
Improper molding or operating
Improper form of removal operations
Imbalance standing
Dangerous behavior
Operation error of steel wire rope
Collision of hanging material
Insufficient materials of level department
The construction of junction point is not
solid
Sheets and the ground are not lock tight
Dangerous position
Stumble
Improper force
False step
Dangerous position
Walking on dangerous path
Without intermediate supports
Without safety belt
Operator error of crane
Without hauling rope
H/S planning
H/S management
H/S management
H/S management
Environmental maintenance
H/S planning
G2
G3
F1
F2
C6
C7
C8
C9
C10
F3
F4
E1
E2
E3
E4
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9
Q10
Q11
Q12
Q13
Q14
Q15
Q16
E5
D2
D3
D4
D5
D6
D7
C1
C2
D1
C5
A3
A4
A4
A4
A2
A3
Q17
Q18
Q19
Q20
Q21
Q22
Q23
Q24
Q25
Q26
Q27
Q92
Q93
Q94
Q95
Q96
Q97
5
169
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
Table 4
Checklist of construction safety performances.
Project:
Item
Date of inspection:
Level Inspection items
Health and Safety A1
Training
Environmental
Maintenance
A2
Done
(yes/
no)
1. Holding general H/S training workshop
2. Daily education training before workers
enters into job sites
3. Holding H/S training for special operation
workers
4. Popularize workers training and keep the
training record contact
5. Workers understand and are familiar with H/
S regulation and practice
6. Workers are fully aware of the consequences
of breaking H/S rules
7. Workers are able to comply with H/S Codes of
Conduct
8. Workers are able to perform their works
based on standard operating procedures
1. Materials are stacked organized
2. Job sites are clean and have no water pool
3. Workers are familiar with operating
environment
4. Completed construction moving path
5. Clear indication on job sites
Item
Level Inspection items
Health and Safety A3
Planning
Health and Safety A4
Management
5. Facilities meet H/S requirement
6. Personal protective equipments meet H/S
standard
7. Risk assessments conducted before high risk
operation
8. Materials are in place, not causing problems
while construction
9. Exact planning of construction moving path
10. Completed emergency response and medical
care plan
1. H/S organization develops in accordance with
H/S rules
2. Site access control
3. Auto check mechanism
4. Regular workplace inspection
5. Improvement and tracking data
6. H/S management records
7. Workers use helmets and protective
equipment
8. Construction scaffolds are set in right place
9. A-type ladder meets the standard
10. Protective measures are taken in open part
space on the job site such as safety net
11. Proper approach is taken to prevent objects
from falling
12. The pitch and strength of support frame
meets the construction code
13. Construction machinery has been inspected
and meets the requirement
14. Wires on wet ground are elevated
15. Installation of leakage circuit breakers
16. Reward and punishment system are
developed
6. Good lighting and construction moving path
7. Height over 1.5 m and with hoist device
8. Clean up waste on time
9. Completed safety equipments
10. Functional fire-fighting facilities
Health and Safety A3
Planning
Done
(yes/
no)
1. Clear H/S objective and feasible policies
2. Sufficient and reasonable H/S budget
3. H/S plans and SOP are developed completely
4. Materials and construction methods are in
compliance with the regulation
Table 5
Basic information of five cantilever bridge construction projects.
Project no.
Scope of works
Size
1
2
3
4
5
5743 m
5970 m
4545 m
6062 m
5677 m
Bridge
Bridge
Bridge
Bridge
Bridge
(STA.45K + 000–50K + 743)
(STA.45K + 000–50K + 970)
(STA.50K + 970–53K + 315)
(STA.52K + 943–59K + 005)
(STA.53K + 315–58K + 992)
useful tool for the fall-risk assessment of cantilever bridge construction projects.
6.2. Sensitivity analysis and discussions
To further examine the key causes of fall risk in the cantilever
bridge construction projects, sensitivity analysis was performed
5305 m/Embankment
4839 m/Embankment
2722 m/Embankment
4975 m/Embankment
4585 m/Embankment
438 m
1131 m
1823 m
1087 m
1092 m
Period (days)
Cost (NT$)
915
973
1031
1038
1046
2,887,256,002
2,640,812,916
3,352,337,437
2,292,635,110
2,055,633,135
and the key causes are summarized as follows. The most significant
direct cause of falling is section blocks construction resulting in the
fall and the lack of the safety facilities. Indirect cause include
dangerous procedures or methods, improper behavior or posture,
job-site environment disorder, unsafe equipment, poor selfmanagement, no intermediate support, improper molding or
operating. Finally, the root factors are safety and health planning
Table 6
Comparison between BN and real site assessment.
Project no.
Fall risk (%) from BN
Risk rank By BN
Real site assessment (score)
Safety performance rank
1
2
3
4
5
38.942
60.851
78.454
39.549
19.362
4
2
1
3
5
85.24
72.35
65.28
76.36
87.67
2
4
5
3
1
170
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
70%
62%
No of obs
60%
50%
40%
30%
22%
20%
7%
6%
10%
3%
0%
0%
Pay no improper use
attention to of personal
the stand at protective
the
equipment
environment
Not use the Do not use Improper use
Not
insurance
personal of equipment prohibited
policy or
protective
entering the
ignore the equipment
lifting and
warning
operating
range
Fig. 8. Unsafe behaviors of occupational accidents of bridge construction projects.
35%
When correlations the result of sensitivity analysis to that of
statistical surveys in indirect factors. Table 7 shows the indirect
factors, dangerous procedures or methods, improper behavior or
posture, poor self-management, and improper molding or operating four factors are highly correlation. Job-site environment is
disorder and unsafe equipment two factors are moderate correlation. No intermediate support is low correlation. The root factors,
poor H/S training and poor H/S planning two factors are highly
correlation. Demonstrate both result of sensitivity analysis and
statistical surveys have correlation.
In summary, this study identifies sensitive causes of fall
accidents by means of sensitivity analyses. Based upon the
above-mentioned analysis, project managers could propose preventive safety measures in advance to reduce falling occurrences.
Moreover, the fall risk assessment and sensitivity analysis allow
us to allocate resources in advance to places of critical safety concerns so that the risk of falling could be significantly reduced.
30%
30%
7. Conclusions and future developments
No of obs
25%
23%
23%
20%
15%
10%
10%
12%
5%
1%
0%
Unsafe
Mismanagement Use dangerous
The employer Unsafe operating
does not labor to environment equipment and and instructions methods or
materials
procedures
use personal
protective
equipment
Others
Fig. 9. Unsafe conditions of occupational accidents of bridge construction projects.
and training which play an important role in reducing fall risk in
bridge construction projects.
On the other hand, according to the occupational hazards data
from the Institute of Occupational Safety and Health in Taiwan
and accident logs from project site, the following results are summarized and analyzed (MOL, 2013). In bridge construction, the
main causes of falling are: limited construction pedal board, safety
net unable to be erected, and the lack of ideal fixed-points of lifelines and safety belts. The most sensitive indirect cause of falling is
unsafe behavior and unsafe condition. That could easily lead to
construction accidents. As illustrated in Fig. 8, the statistical survey
of occupational accidents of bridge construction indicates that no
use of the insurance policy and ignorance of the warnings are crucial factors, in which the occurrence percentage of occupational
accidents during the bridge construction process is the highest
(62%). In addition, based on the statistics of occupational accidents
at the bridge construction projects, approximately half of the occupational accidents during the bridge construction stage are caused
by mismanagement and instructions, dangerous methods or procedures, and insufficient use of personal protective equipments (as
shown in Fig. 9).
This paper illustrated an effective process to build a BN-based
fall risk assessment model for cantilever bridge construction projects. The assessment model started with the formation of a FT
based upon the problem domain. Then, a basic BN was obtained
by transforming the FT to a BN. Furthermore, experts’ inputs of
meaningful supplementary arcs among nodes were inserted into
the BN in order to derive a complete BN framework. Finally, logic
gates in the FT were converted into CPTs in the BN in a logical
transformation approach. A safety performance checklist was then
created to subjectively assess the prior probabilities of root safety
causes. The inference results of the BN were validated against five
cantilever bridge construction projects in Taiwan. Through the
analysis and comparison, it was found that the results of BN
analysis are consistent with safety records of five cantilever bridge
construction projects. This shows that the transformation process
from a multi-state FT to a BN can effectively assess a realistic
and accurate fall risk. Additional, according to the model assessment and the sensitivity analysis, site project managers can take
preventive safety measures and allocate resources in advance to
reduce the risks of falling safety.
While this study is exploratory in nature, further research needs
to continue in this area. A BN model can be constructed from raw
data. If complete and sound safety data are available, objective BN
framework and parameters can be explored and established.
Furthermore, the transformation mechanism from a FT to a BN
has been well examined in this study, and the use of a BN, nevertheless, relies on the input of experts’ experiences on arcs and CPTs
in the BN. Data provided by different experts will directly affect the
accuracy and the assessment quality of a BN. Thus, more attention
should be paid to expert experience elicitation in the future study.
Finally, there are other occupational accidents taking place at
bridge construction projects, such as object collapses, collisions,
and drowning. It may be applicable to extend the BN scope to cover
these accidents and make use of BN for the overall safety diagnosis
at bridge construction projects.
Table 7
Comparison of sensitive factors from BN and actual statistics.
Accident type
Indirect factors of sensitive analysis from BN
Actual statistics
Fall
(B3) Dangerous procedures or methods
(B6) Improper behavior or posture
(B10) Job-site environment is disorder
(B13) Unsafe equipment
(B16) Poor self-management
(C1) No intermediate support
(C9) Improper molding or operating
High
High
Middle
Middle
High
Low
High
Root factors of sensitive analysis from BN
Actual statistics
(A1) Poor H/S training
(A3) Poor H/S planning
High
High
T.-T. Chen, S.-S. Leu / Safety Science 70 (2014) 161–171
References
Adriaenssens, V., Goethals, P.L.M., Charles, J., Pauw, N.De, 2004. Application of
Bayesian belief networks for the prediction of macro-invertebrate taxa in rivers.
Ann. Limnol. Int. J. Limnol. 40 (3), 181–191. http://dx.doi.org/10.1051/limn/
2004016.
Agena, 2008. <http://www.agenarisk.com/products/free_download.shtml>.
Antal, P., Fannes, G., Timmerman, D., Moreau, Y., Moor, B.De, 2007. Bayesian
applications of belief networks and multilayer perceptions for ovarian tumour
classification with rejection. Artif. Intell. Med. 29 (1), 39–60. http://dx.doi.org/
10.1016/S0933-3657(03)00053-8.
Baran, E., Jantunen, T., 2004. Stakeholder consultation for Bayesian decision support
systems in environmental management. In: Proc. of the Regional Conference on
Ecological and Environmental Modeling (ECOMOD 2004), Penang, Malaysia, pp.
15–16.
Bedford, T., Gelder, P.V., 2003. Safety and reliability. In: Proc. of European Safety and
Reliability Conference 2003 (ESREL 2003), Maastricht, Netherlands, pp. 15–18.
Bobbio, A., Portinale, L., Minichino, M., Ciancamerla, E., 1999. Comparing fault trees
and Bayesian networks for dependability analysis. Lect. Notes Comput. Sci.,
Comput. Saf., Reliab. Security 1698, 310–322. http://dx.doi.org/10.1007/3-54048249-0_27.
Bobbio, A., Prauzy, A., Minichino, M., 2001. Improving the analysis of dependable
systems by mapping fault tree into Bayesian network. Reliab. Eng. Syst. Saf. 71
(3), 249–260. http://dx.doi.org/10.1016/S0951-8320(00)00077-6.
Boudali, H., Dugan, J.B., 2005. A discrete-time Bayesian network reliability modeling
and analysis framework. Reliab. Eng. Syst. Saf. 87 (3), 337–349. http://
dx.doi.org/10.1016/j.ress.2004.06.004.
Doguc, O., Ramirez-Marquez, J.E., 2009. An efficient fault diagnosis method for
complex system reliability. In: Proc. of the 7th Annual conference on System
Engineering Research 2009 (CSER 2009), Session 12 – Systems Assessment &
Assurance.
Ebeling, C.E., 1997. An Introduction to Reliability and Maintainability Engineering,
first ed. McGraw-Hill Inc., New York, 486 p.
Fenton, N.E., Neil, M., et al., 2007. Using ranked nodes to model qualitative
judgments in Bayesian networks. IEEE Trans. Knowl. Data Eng. 19 (10), 1420–
1432.
Franke, U., Flores, W.R., Johnson, P., 2009. Enterprise architecture dependency
analysis using fault trees and Bayesian networks. In: Proc. 42nd Annual
Simulation Symposium (ANSS 2009), San Diego, CA, pp. 209–216. doi:http://
dx.doi.org/10.1145/1639809.1639866.
Graves, T.L., Hamada, M.S., Klamann, R., Koehler, A., Martz, H.F., 2007. A fully
Bayesian approach for combining multi-level information in multi-state fault
tree quantification. J. Reliab. Eng. Syst. Saf. 92 (10), 1476–1483. http://
dx.doi.org/10.1016/j.ress.2006.11.001.
Hartford, D.N.D., Baecher, G.B., 2004. Risk and Uncertainty in Dam Safety, first ed.
Thomas Telford Ltd., London, doi:http://dx.doi.org/10.1680/rauids.32705.
Heinrich, H.W., Petersen, D., Roos, N., 1980. Industrial Accident Prevention, fifth ed.
McGraw-Hill Inc., New York.
171
Jitwasinkul, B., Hadikusumo, B.H.W., 2011. Identification of Important
organizational factors influencing safety work behaviours in construction
projects. J. Civ. Eng. Manage. 17 (4), 520–528. http://dx.doi.org/10.3846/
13923730.2011.604538.
Kales, P., 2006. Reliability: For Technology, Engineering, and Management. Pearson
Education Taiwan, Prentice Hall Ltd., Taipei.
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 (8), 925–932.
Lingard, H., Rowlinson, S., 2005. Occupational Health and Safety in Construction
Project Management, first ed. Spon Press, London.
Marcot, B.G., Holthausen, R.S., Raphael, M.G., Rowland, M.M., Wisdom, M.J., 2001.
Using Bayesian belief networks to evaluate fish and wildlife population viability
under management alternatives from an environmental impact statement. For.
Ecol.
Manage.
153
(1–3),
29–42.
http://dx.doi.org/10.1016/S03781127(01)00452-2.
Marsh, W., Bearfield, G., 2007. Representing parameterised fault trees using
Bayesian networks. Comput. Saf., Reliab., Security, Lect. Notes Comput. Sci.
4680 (2007), 120–133. http://dx.doi.org/10.1007/978-3-540-75101-4_13.
Martin, J.E., Matias, J.M., Rivas, T., Taboada, J., 2008. A machine learning
methodology for the analysis of workplace accidents. Int. J. Comput. Math. 85
(3), 559–578. http://dx.doi.org/10.1080/00207160701297346.
Matias, J.M., Rivas, T., Ordonez, C., Taboada, J., 2007. Assessing the environmental
impact of slate quarrying using Bayesian networks and GIS. In: Proc. of the
International Conference on Computational Methods in Science and
Engineering 2007 (ICCMSE 2007), Corfu, Greece, pp. 1285–1288. doi:http://
dx.doi.org/10.1063/1.2835985.
MOL, 2013. 2013 Yearbook of Labor Statistics. Institute of Labor, Occupation Safety
and Health, Ministry of Labor, Taiwan. <http://www.iosh.gov.tw/IAKP/
ReportIndex.aspx?NID=575>.
O’connor, Patrick D.T., 2002. Practical Reliability Engineering, fourth ed. John Wiley
& Sons Inc., New York.
Paul, P.S., Maiti, J., 2007. The role of behaviour factors on safety management in
underground mines. Saf. Sci. 45 (4), 449–471. http://dx.doi.org/10.1016/
j.ssci.2006.07.006.
Qian, G., Zhong, S., Cao, L., 2005. Bayesian network based on a fault tree and its
application in diesel engine fault diagnosis. In: Proc. of the International Society
for Optical Engineering, Chongqing, China, pp. 60421P-1–60421P-6. doi:http://
dx.doi.org/10.1117/12.664626.
Rao, S.S., 1992. Reliability-Based Design, first ed. McGraw-Hill Inc., New York.
Xiao, L., Haijun, L., Lin, L., 2008. Building method of diagnostic model of Bayesian
networks based on fault tree. In: Proc. of the International Society for Optical
Engineering, Beijing, pp. 71272c-1–71272c-6. doi:http://dx.doi.org/10.1117/
12.806736.
Zhu, J.Y., Deshmukh, A., 2003. Application of Bayesian decision networks to life
cycle engineering in green design and manufacturing. Eng. Appl. Artif. Intell. 16
(2), 91–103. http://dx.doi.org/10.1016/S0952-1976(03)00057-5.