Professional Documents
Culture Documents
Article history: This paper presents a systemic decision support approach for safety risk analysis under uncertainty in tunnel construction.
Received 6 February 2014 Fuzzy Bayesian Networks (FBN) is used to investigate causal relationships between tunnel-induced damage and its in fluential
Received in revised form variables based upon the risk/hazard mechanism analysis. Aiming to overcome limitations on the current probability
16 May 2014 estimation, an expert confidence indicator is proposed to ensure the reliability of the surveyed data for fuzzy probability
Accepted 14 June 2014 Available assessment of basic risk factors. A detailed fuzzy-based inference procedure is developed, which has a capacity of
online 24 June 2014
implementing deductive reasoning, sensitivity analysis and abductive reasoning. The “3σ criterion” is adopted to calculate the
Keywords:
characteristic values of a triangular fuzzy number in the probability fuzzification process, and the α-weighted valuation
Construction safety
method is adopted for defuzzification. The construction safety analysis progress is extended to the entire life cycle of risk-
Fuzzy Bayesian networks (FBN)
Tunnel leakage prone events, including the pre-accident, during-construction continuous and post-accident control. A typical hazard
Risk analysis concerning the tunnel leakage in the construction of Wuhan Yangtze Metro Tunnel in China is presented as a case study, in
Case study order to verify the applicability of the proposed approach. The results demonstrate the feasibility of the proposed approach
and its application potential. A comparison of advantages and disadvantages between FBN and fuzzy fault tree analysis
(FFTA) as risk analysis tools is also conducted. The proposed approach can be used to provide guidelines for safety analysis
and management in construction projects, and thus increase the likelihood of a successful project in a complex environment.
1. Introduction already gridlocked city [5]. In China, the number of construction accidents
shows a rising trend in metro construction projects. On November 15, 2008,
Construction is one of the most dangerous industries in the world 21 people were killed as a result of a road cave-in above a metro tunnel under
[1]. The ramifications of construction accidents are growing with trends construction in Hangzhou [6]. Also, on December 25, 2012, eight people were
toward larger-scale and more complex construction projects [2], especially in killed and five others hurt in a fatal tunnel explosion in north China’s Shanxi
developing countries, like China. With the exploita-tion of urban underground province [7]. Tunnel construction entails to be a highly complicated project
space, underground construction has presented a powerful momentum for the with large potential risks, which can bring enormous dangers to public safety
development of a rapid economy worldwide in the past ten years. Due to [8]. Therefore, it is necessary to investigate the causal relationship and safety
various risk factors in complex project environments, safety violations occur risk mechanism of construction failures in tunnel construction by considering
frequently in metro construction. On January 12, 2007, the Pinheiros Station the accident scenario and real-time safety analysis, aiming to provide decision
on Metro Line Four at Sao Paulo’s Aquarium in Brazil collapsed, causing the support for assuring the safety of tunnel construction.
death of seven people [3]. On July 6, 2010, a tunnel collapse also took place
in Prague, Czech Republic, causing a 15-meter-wide sunken pit [4]. On
August 23, 2012, a metro line leak caused chaos in Warsaw, Poland. Water To avoid heavy casualties and property losses caused by safety violations,
flooded into the tunnel at the planned Powisle station, causing considerable innumerable studies have introduced risk-based analysis into safety
transportation problems in the management practice. Risk analysis can be divided into qualitative and
quantitative risk analysis [9]. The former includes fault tree analysis (FTA),
comprehensive fuzzy evaluation method (CFEM), check list, and others;
while the latter includes job risk analysis method (LEC), influence diagrams,
n Corresponding author.
Neural Networks (NN), support vector machines, decision trees, and others.
E-mail addresses: limao_zhang@hotmail.com (L. Zhang),
wxg0220@126.com (X. Wu), mirek@umd.edu (M.J. Skibniewski), The above risk-based analysis methods make a significant contribution to
jb_zhong@126.com (J. Zhong), bdgluy@nus.edu.sg (Y. Lu). safety risk
http://dx.doi.org/10.1016/j.ress.2014.06.006
0951-8320/& 2014 Elsevier Ltd. All rights reserved.
30 L. Zhang et al. / Reliability Engineering and System Safety 131 (2014) 29–39
analysis and management in complex engineering projects [10,11], however, consists of a directed acyclic graph (DAG) and an associated joint probability
they are confined to static control management [12]. Khakzad [10] indicated distribution (JPD). A BN model with n nodes can be represented as BoG, Θ4,
FTA unsuitable for complex problems with its limitation in explicitly where G stands for a DAG with n nodes, and Θ stands for the JPD of the BN
representing dependencies of events, updat-ing probabilities, and coping with model. The nodes {X1,…, Xn} in the graph are labeled by related random
uncertainties. Yang et al. [13] regarded LEC unsuitable in complex dynamic variables. The directed edges between nodes represent association
environments, resulting from the insufficiency in timely diagnosing and relationships among variables. DAG contains conditional independence
dealing with various problems. When associated parameters, such as assumptions, and the relations represented by DAG allow JPD to be specified
geological, design and construction parameters are changed, the locally by the conditional probability table (CPT) for each node. Assuming
aforementioned methods cannot accurately illustrate the updated features of π(Xi) is the parent nodes of Xi in DAG, the CPT of Xi is denoted by P(Xi|
dynamic environ-ments as the construction progress evolves. Nor can π(Xi)). The JPD of P(X1,…, Xn) can be written as Eq. (1).
professional supports or suggestions be provided in real time as the
parameters are updated. Recently, Bayesian networks (BN) has been proposed
to model the complexity in man-machine systems [14]. BN can describe
dependencies between variables both qualitatively and quantitatively, and is
2.2. Fuzzy set theory and Fuzzy Bayesian Networks
suitable for knowledge representation and reasoning [15]. Also, BN is
powerful in dealing with uncertainty information, and can be used for
Fuzzy set theory (FST) is first introduced by Zadeh [24] with an effort to
reliability and failure analysis in complex environments [16,17].
deal with uncertainty due to the imprecision and vagueness. FST provides a
basis to generate powerful problem-solving techniques with wide
applicability, especially in the field of decision making [25]. A fuzzy set P is
In conventional BN analysis, the occurrence probability of root nodes is
usually indicated by a tilde "⁓”, where X is characterized by a membership
always regarded as a crisp value [18]. However, in the construction
engineering fields, it is difficult or nearly impossible to obtain exact values of function Fp(x) with an interval [0,1]. The function Fp(x) represents the
probability due to a lack of sufficient data [19]. Thus, a group decision- member- ship value of x in P . In general, FST uses triangular, trapezoidal or
making technique is generally employed to assess the occurrence probability Gaussian fuzzy numbers to convert the uncertain numbers into fuzzy numbers
of root nodes. Hanss [20] indicated that the fuzzy set theory (FST) provided a [26]. Without loss of generality, triangular fuzzy numbers are often utilized to
successful tool to solve engineering problems under uncertainty. The provide more precise descriptions and to obtain more accurate results [27].
uncertainty can be taken into account in terms of intervals or fuzzy numbers Thus, in this paper, triangular fuzzy numbers are used for representing
[21]. Currently, FST and BN have both emerged as powerful and effective probabilities of root nodes in a BN model. A fuzzy number P ¼ (a, m, b) is
tools for knowledge reasoning in uncertainty environments [22]. Thus, it is called a triangular fuzzy number if its membership function is given by Eq.
certainly quite appropriate to investigate the amalgamation of FST and BN, (2). Herein, a, m, and b represent the lower least likely value, the most likely
which may well prove to provide an indispensable means of incorporating value, and the upper least likely value, respectively. The constants [a, b] give
uncertain factors/elements in a probabilistic risk analysis model domain [23]. the lower and upper bounds of the available area, reflecting the fuzziness of
This paper therefore investigates the possibility of merging BN and FST, the actual data. Assuming two triangular two triangular namely Ã1 =(a1,m1, b1)
which is Fuzzy Bayesian Networks (FBN), to provide an alternative means to and Ã2 =(a2,m2, b2) the operation between A1 and A2 can be defined by Eq (3)
facilitate the construction failure analysis in tunnel construction. Currently, a including the addition, subtraction, multiplication, and division (see the work
universally accepted standard regarding the safety risk analysis procedure has of Mentes and Helvacioglu [19]).
not been reached in tunnel construction. A systemic decision approach based
upon FBN is developed with step-by-step procedures in detail, aiming to
provide guidelines for safety manage-ment in tunnel construction throughout
the entire life cycle, with the pre-accident, during-construction continuous and
post-accident con-trol included. A typical hazard concerning the tunnel
leakage in the construction of Wuhan Yangtze Metro Tunnel in China is
presented as a case study. Results demonstrate the feasibility of the proposed
fuzzy decision approach and its application potential.
the fuzzy marginalization rule, and fuzzy Bayesian rule can be calculated by and uses them in subsequent computation [31]. Owing to the high potential
Eqs. (4) and (5), respectively. Herein, T stands for the leaf root, while Xi risks for the tunnel construction in complex environ-ments, the fuzzy
stands for the root nodes. Combing with Eq. (3), the FBN-based inference probability analysis should meet the highly required precision for safety
techniques can then be fulfilled. assurance in tunnel construction practice.
Step (1) Risk/Hazard identification: Carry out the preliminary risk (1) In traditional expert investigation, all the collected survey data are
mechanism analysis for the construction failure, reveal potential obtained without any kind of data reliability evaluation. Actually, most
risks/risk factors and their causal relationships, and then identify interviewed individuals may have different confidence levels toward their
root nodes (RNs), intermediate nodes (INs) and their leaf nodes own subjective judgments due to the difference in education background,
(LNs). working experience and risk attitudes. Thus, a certain deviation exists in
Step (2) BN model construction: Identify potential failure- the data reliability among different interviewed individuals. For instance,
consequence scenarios of the target risk, develop failure logic by the reliability of the investigation data obtained from a project engineer
means of conditional probability distribution, and then build up a with 30 years of working experience is comparatively higher than that
topology of the failure network connecting all specific nodes. from an engineer with 5 years of working experience. It is therefore
Step (3) Fuzzy probability assessment: Carry out the expert necessary to carry out the reliability evaluation for the surveyed data,
investigation for probability estimation of root nodes, gather the with both the expert ability and subjectivity taken into account.
data, transform the linguistic and fuzzy expressions into fuzzy
numbers, and then calculate the fuzzy probability of root
nodes based upon the fuzzification process. (2) In regard to the division of probability intervals, five linguistic terms
Step (4) Fuzzy decision analysis: Take advantage of the [very low (VL), low (L), medium (M), high (H), and very high (VH)] are
advanced reasoning features of FBN, and carry out risk analysis, commonly used to assess the probability of occurrence. The current
including deductive reasoning, sensitive analysis and abductive probability span is excessively large within a single interval. For instance,
reasoning. Meanwhile, convert the calculating fuzzy probability the scope of “H” ranges from 21.5% to 67% (see Abdelgawad & Fayek,
into a crisp value based upon the defuzzification process, and [33]). Such rough intervals division cannot meet the required precision
then rank the results. for failure probability assess-ment of root nodes in tunnel engineering
practice.
Step (5) Decision making: Find out critical risks and risk factors
according to risk analysis results, propose corresponding con-trol
measures for risk response, and make up decisions in real time. 3.2. Expert confidence evaluation
Meanwhile, the implementation effect of the risk mon-itoring and
reviewing can provide feedback and suggestions for optimizations An expert confidence indicator is proposed to reveal the reliability of the
in previous steps. data acquired from interviews. On one hand, the expert judgment ability
should be first taken into account. In construction practice, it is generally
Steps (1), (2), and (5) rely on the elicitation of knowledge from experts considered that the judgment ability of individuals tends to become
increasingly sophisticated and stable with the accrual of educational
using a standard technique (e.g., Chapman [30]). The following sections
background and work-ing experience. The judgment ability level, denoted by
highlight steps 3 and 4, respectively.
ζ, is then increased accordingly. The judgment ability is divided into four
levels, represented by “I–IV” as seen in Table 1. The level “I” with a score ζ
3. Fuzzy probability assessment ¼ 1:00 stands for the highest reliability for the expert judgment ability. On
the other hand, the expert confidence indicator involves a kind of subjective
Fuzzification plays a crucial role in fuzzy decision analysis, which measurements. It is
attempts to define the basic data into a fuzzy probability set
Risk/Hazard identification BN model construction Fuzzy probability assessment Fuzzy decision analysis Decision making
Risk mechanism analysis Potential failure scenarios Expert investigation Forward reasoning Critical risks/risk factors
Reveal causal relationships Develop a failure logic Data gathering Sensitivity analysis Safety control measures
Build up a network
Identify basic nodes Probability fuzzification Backward analysis Monitoring and reviewing
topology
p i
> >
;
>
> 18
3 0.15 0.20 0.25 < ð þ
>
> ¼ þ
>
>
>
ð
5 0.35 0.40 0.45 1 Þ
>
>j k
6 0.45 0.50 0.55 >
7 0.55 0.625 0.70 :
8 0.70 0.775 0.85 3.5. Fuzzification
9 0.85 0.925 1.00
9
3.3. Probability interval division m ¼ EðPÞ ¼ ∑ ðci PiÞ ð9Þ
i¼1
9 ð10Þ
It is known that the size of the probability intervals can indicate the σ ¼ DðPÞ ¼ ∑ ½ðci EðPÞÞ 2 pi &
reliability of the estimation. Short intervals indicate that the statistic is pffiffiffiffiffiffiffiffiffiffi i ¼ 1
precisely known, while wide intervals indicate uncer-tainty. To reach the
highly required precision for the occurrence probability of the top event in the
tunnel construction practice, it is better to conduct a group of short intervals, a ¼ m 3σ; b ¼ m þ3σ ð11Þ
rather than that of wide intervals. However, according to Dawes' experiment, where ci refers to the mean of the ith probability interval as shown in Table 2;
the use of 5–9 point scales is likely to produce slightly higher mean scores a, m, and b refer to characteristic values of a triangular fuzzy number.
relative to the highest possible attainable score, and a larger number is usually
impractical [34]. Therefore, in this research, the occurrence probability of root 4. FBN-based safety risk analysis
nodes is divided into 9 intervals, represented by “1–9”. As seen in Table 2, the
kth interval is defined by [ak, ak þ 1] together with a mean ck (1rkr9). During Each construction failure shows different status and character-istics along
the expert investigation, the purpose of questionnaires is to collect the with the construction progress, including before, during and after a specific
information of two types, namely the occurrence probability interval k and accident. Accordingly, safety control of risk-prone events can be divided into
subjectivity reliability level ψ. For the “Probability interval k”, interviewed three stages in the overall work process: namely pre-accident, during-
individuals are required to fill in a number ranging from 1 to 9 (as seen in construction, and post-accident control. Taking advantage of the powerful
reasoning features within FBN models, the deductive reasoning, sensitivity
Table 2). For the “Subjectivity reliability level ψ”, they are required to fill in a
analysis and background reasoning techniques can be used for safety analysis
number ranging from 1.0 to 0.6. If they fail to evaluate the data reliability
within a reliable degree of more than 0.6 by themselves, that field can be left and management in the above three stages, respectively. In this way, the real-
blank. time and effective support can
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L. Zhang et al. / Reliability Engineering and System Safety 131 (2014) 29–39 33
be provided for decision makers in the entire management process of failures factor when an accident or failure (T) occurs. Posterior probability
in tunnel construction. distribution can provide reliable references for fault diagnosis. Posterior
probability distribution of risk factor Xi, represented by P(Xi ¼xi|T¼t), can
4.1. Deductive reasoning be calculated by Eq. (16). Xi is more likely to become the direct cause of an
accident or failure (T) when P(Xi ¼xi|T¼t) is close to 1.
Deductive reasoning aims to predict the probability distribu-tion of the
risk event (or leaf node) T under the combination of risk factors (or root
nodes) X1, X2… Xn. The state of each risk factor is treated as evidence input Þ PðT ¼ 1jXi ¼ xiÞ&∅PðT ¼ 1Þ; i ¼ 1;
into a FBN model. Probability distribu-tion of T, represented by P(T¼1), is 2; :::; n ð16Þ
calculated by Eq. (12). P(T¼1) can be served as an indicator to evaluate the
potential safety level of the risk event (T), helping decision makers take 4.4. Defuzzification
proper preven-tions in advance.
In many applications of fuzzy set theory in the field of decision making,
2
n
decision makers are faced with a problem of selecting one from a collection
PðT ¼ 1Þ ¼ ∑½PðT ¼ 1jX1 ¼ x1; X2 ¼ x2; :::; Xn ¼ xnÞ of possible solutions. The selection process of the best solution may require
1
ranking or ordering fuzzy numbers. In the above fuzzy-based decision
PðX1 ¼ x1; X2 ¼ x2; :::; Xn ¼ xnÞ& ð12Þ analysis, the calculated results for the root or leaf node remain to be a fuzzy
where n stands for the number of the root nodes. Each root node X i has two triangular number, denoted by Pj ¼(aj, mj, bj) (j¼1,2,…,K). For the purpose
different states, denoted by “Yes/No” respectively. Thus, n root nodes of risk ranking in the Bayesian inference, it is therefore necessary to
n
contribute to 2 combinations. P(T¼1|X1 ¼x1, X2 ¼x2,…, Xn ¼xn) transform fuzzy values into a crisp value at the defuzzification stage.
represents the CPT of T, while P(X 1 ¼x1,X2 ¼x2,…,Xn ¼xn) represents the Specifically, the objective of defuzzification is to determine an exact value as
joint probability distribution of X. the representative of the fuzzy number. Currently, several defuzzification
methods have been developed, such as the center of gravity (COG) [36],
4.2. Sensitivity analysis mean of maxima (MOM) [37], center of maxima (COM) [38], and the height
method (HM) [39]. According to Detyniecki and Yager [40], some
Sensitivity analysis plays an important role in probabilistic risk information was lost during the transforming process in the above
assessment, aiming to illustrate the performance of each risk factor's defuzzification methods. Detyniecki and Yager [40] then proposed an α-
contribution to the occurrence of a construction failure. In this research, three weighted valuation method, and results indicated the proposed method was
key performance indicators (KPIs), REV, RRV, and AVG, are proposed to efficient in reducing the information loss. Therefore, the α-weighted valuation
measure the contribution of each risk factor Xi toward the risk event T. Key method is adopted for defuzzification, and presented in a case study later in
risk factors can then be identified as to help engineers determine the main this research.
checkpoints at the construction stage.
~
REV(Risk Expansion Value) is used to assess the performance of risk Assuming a triangular fuzzy number P ¼(aj, mj, bj) (see Fig. 2), the
REV ~
expansion for each risk factor Xi, represented by I (Xi). The membership function of P, represented by F P(x), is given by previous Eq. (2).
REV
higher the value of I (Xi), the more responsibility Xi has in risk expansion Using the α-weighted valuation method, a generalized formulation of the
REV
of T. I (Xi) can be calculated by Eq. (13). RRV (Risk transformed exact value is derived by Eq. (17). Herein, Fα¼{x|F(x)Zα} is the
Reduction Value) is used to assess the performance of risk reduc-tion for each α-level set of F(x), Average(Fα) is an average of the α-level set, and f(α) is
RRV RRV
risk factor Xi, represented by I (Xi). The higher the value of I (Xi), the the α-weighted valuation function. Generally, Average(F α) can be calculated
RRV by Eq. (18).
more responsibility Xi has in risk reduction of T. I (Xi) can be calculated ð
by Eq. (14). AVG (Average Sensitivity ValðFÞ ¼ R 1f α dα ð17Þ
1
0 Average FαÞ f ðαÞdα
Measure) is used to measure the average sensitivity of risk factor X i,
AVG AVG
represented by I (Xi). The higher the value of I (Xi), the
SM R0 ð Þ
more responsibility Xi has in risk sensitivity of T. I (Xi) can be calculated
by Eq. (15). Average F uα þvα 18
REV ð αÞ¼ 2 ð Þ
I ðXiÞ ¼ ½MaxfPðT ¼ 1jXi ¼ xiÞg⊖PðT ¼ 1Þ&∅PðT ¼ 1Þ; i ¼ 1; 2; :::; n
where uα represents the lower bound of the α-level set, and can be calculated
ð13Þ by Eq. (19); vα represents the upper bound of the α-level set, and can be
RRV
I ðXiÞ ¼ ½PðT ¼ 1Þ MinfPðT ¼ 1jXi ¼ xiÞg&∅PðT ¼ 1Þ; i ¼ 1; 2; :::; n calculated by Eq. (20), respectively. Usually, we set f (α)¼1, and thus the
transformed exact value can be defined by Eq. (21).
ð14Þ
AV G 1 REV RRV uα ¼ ðmj ajÞ αþaj ð19Þ
I ðXiÞ ¼ 2 ½I ðXiÞ I ðXiÞ&; i ¼ 1; 2; :::; n ð15Þ
Among the above three KPIs, one can be used in conjunction with the vα ¼ bj ðbj mjÞ α ð20Þ
other two, and be verified by their results as well. Normally, the identified 1 1
½ð Þ þ
factors at the top of the ranking list of each indicator are considered as the key ValðFÞ ¼
2
R
aj α
1
0 dα
m
0 j aj þbj
risk factors for safety control and management in real projects.
1 j j R j jÞ j α&d
j þ ðmajj b2m bjα
m a m b¼ þ þ ð21Þ
Compared to the traditional risk-based methods, such as FTA, CFEM and 5. Case study
NN, the feature of abductive reasoning technique is unique and matchless in
FBN inference [10]. Abductive reasoning aims to obtain the posterior Wuhan Yangtze Metro Tunnel (WYMT), known as “the first metro tunnel
probability distribution of each risk across the Yangtze River in China”, is an important
34 L. Zhang et al. / Reliability Engineering and System Safety 131 (2014) 29–39
route connecting two large cities, comprising the metropolitan area of Wuhan, structure. A Bayesian network model is established and offered as a decision
namely Wuchang and Hankou. It is a double-spool tunnel with a total length support tool for safety assurance, providing real-time and effective support for
of almost 6200 m. The tunnel construc-tion started in April, 2008, and the decision makers in the overall drilling process.
hydrogeological section of WYMT is shown in Fig. 3. A slurry shield
machine with Φ of 6.52 m in cutter diameter is utilized to push the tunnel
from Jiyu Bridge Station to Jianghan Road Station. Fig. 4 illustrates the 5.1. Step 1: Risk/hazard identification in tunnel leakage
working site of the construction of WYMT. In construction practice, the
tunnel waterproof works are generally regarded as a highly complicated Tunnel leakage is considered a big potential threat to the quality of tunnel
project associated with large potential risks. Statistics showed that the construction, as well as its subsequent opera-tional safety. For one thing, the
phenomenon of water leakage existed in more than half of the existing tunnels underground enclosure structure can be softened due to the leakage, leading to
[41]. A case concerning the risk analysis of the tunnel leakage in the the structure deformation and damage. Also, the leakage can cause subsidence
construction of WYMT is presented in this research. A specific monitoring and even cracks in the foundation, resulting in excessive uneven settlement.
section is chosen to present the detailed fuzzy-based risk analysis, where the On the other hand, the seepage water can accelerate erosion rates of built-in
tunnel exactly passes under the foundation of Jiangbei Levee with a height of equipment, such as steel rails, commu-mication and lighting facilities. Thus,
25.3 m from the foundation base to the tunnel roof. Jiangbei Levee is the service life of power lines and equipment can be reduced at the
regarded as a significant life line for flood control in Wuhan, concerning the operational stage, associated with the increased operation and maintenance
safety of nearly 10 million lives. Thus, no fault is permitted while the tunnel costs.
is crossing under the existing
Due to complexities in the underground construction techniques and the
surrounding environments, the occurrence of the tunnel leakage is a
complicated process where numerous risk factors interact together and
contribute to the construction failure (T). In order to secure the waterproof
works in tunnel construction, it is therefore necessary to investigate risk
Fp(x)
mechanism of the tunnel leakage first. Based upon engineering practices and
1.0 theoretical analysis, four types of influential variables involved in the tunnel
leakage are revealed as follows: (1) Poor Geological Conditions (B 1): The
tunnel excavation works can cause a change to the seepage field of the
α underground water, leading to a disturbed zone around the tunnel. The
underground water then drains toward the tunnel excavation. The tunnel
leakage is very likely to occur in poor geological conditions; (2)
Unreasonable Design Parameters (B2): For a long time in tunnel construction
0 practice, designers concentrate almost all their energies on the structure
a uα m vα b x safety, regardless of the
~
Fig. 2. Membership function of a triangular fuzzy number P .
Fig. 4. Working site of the construction of WYMT: (a) slurry shield machine; (b) tunnel segment; and (c) tunnel structure.
L. Zhang et al. / Reliability Engineering and System Safety 131 (2014) 29–39 35
importance of the tunnel waterproofing and drainage system design [42]; (3) 5.3. Step 3: Probability assessment
Poor Construction Quality (B3): The construction quality plays a crucial role
in the waterproof works, and the poor construc-tion quality is generally During the expert investigation, questionnaires are adminis-tered to a
considered as the direct cause of the tunnel leakage. For instance, the tunnel group of experts. A total of 100 questionnaires are distributed, and 85 are
leakage frequently occurs where the tunnel segments are damaged or badly returned (the return rate is about 85%), of which 8 are invalid and 77 are
valid. Among 77 valid question-naires, the reviewed experts consist of 52
installed; (4) Improper Management (B4): Effective management can be
domain experts with at least of five years of working experience, and of 22
beneficial to the risk prevention, and vice versa. The leadership, human
research workers in this field. According to their experience and knowl-edge,
resources and management capacity are called “soft” factors which are the experts individually evaluate occurrence probability intervals of root
positively related to project successes in total quality management practices nodes together with their subjective reliability levels. The expert confidence
[43]. For a specific project, the above four influential variables can be further
indicator θ among individuals is then calculated by Eq. (6). In addition, the
refined.
reliable confidence is distributed among 9 intervals using Eq. (7). For
instance, during the occurrence probability assessment for the root node X 8
¼“damage in tunnel segments”, one expert with a judgment ability ζ ¼0.8
considers the probability interval was within 3th interval, associated with a
subjectivity reliability ψ ¼0.7. That is to say, the most likely occurrence
5.2. Step 2: Establishment of a BN model
probability for x1 is between 15% and 25% (see Table 2) with a completely
Between 2006 and 2013, researchers at Huazhong University of Science reliable confidence
and Technology have developed safety control systems for metro construction
and operation tasks for Shenyang, Zhengzhou, Shenzhen and Wuhan metro
systems. The researchers also have developed early warning web-based θ¼0.56 using Eq. (5). The residual reliable confidence
systems for safety control of each project. A large number of safety related
knowledge resources have been accumulated during the progress on these
projects [44,45]. On a basis of the previous risk mechanism analysis in tunnel Table 3
Descriptions of nodes in TLBN.
leakage, the potential risk/risk factors and their causal relationships are
revealed. In the meantime, the Delphi technique is used to reach consensus Nodes Descriptions
among different experts' opinions [46]. First, the risk engineer is interviewed
to establish the structure of the BN model for tunnel leakage, and then sent to T Tunnel leakage
B1 Poor Geological Conditions
the senior risk coordinator for feedback. Modifications as recom-mended by B
2 Unreasonable Design Parameters
the senior risk coordinator are then presented to the risk engineer for further B3 Poor Construction Quality
review and feedback. With prior expert knowledge taken into account, the B4 Improper Management
logic diagram is subsequently used to build up the failure-consequence X1 Shallow underground water
X
scenario from the top to bottom nodes using a directed acyclic graph (DAG). 2 High water content of soil
X3 High slurry pressure
In accordance with the aforementioned Steps 1 and 2 in the proposed fuzzy
X4 Large percolation coefficient
decision approach, a tunnel leakage Bayesian network (TLBN) is established X5 Low curvature of tunnel segments
(see Fig. 5); where 14 root nodes contribute to the failure of the leaf node X6 Low waterproof capability
(tunnel leakage). The descriptions of all nodes are illustrated in Table 3. X7 Improper grouting ration
Meanwhile, the joint probability distribution (JPD) represents relations among X8 Damage in tunnel segments
X9 Poor installation of tunnel segments
various nodes. For instance, the fuzzy JPD of T in TLBN is presented, as seen X
10 Poor grouting quality
in Table 4. X
11 Poor sealing quality
X
12 Incomplete surveying documents
X
13 Unqualified sub-contractor
X
14 Improper schedule scheme
X7 X8
X6 X9
X5 X10
X4 X11
X3 B2 B3 X12
X2 X13
B1 B4
X1 X14
T
Table 4 Table 5
Fuzzy JPD of the leaf node (T) in TLBN. Results of fuzzy probability assessment for root nodes in TLBN.
B
1 B2 B3 B4 P(T¼t|B1,B2,B3,B4), t ¼Yes/No Root nodes Fuzzy probability Root nodes Fuzzy probability
assessment assessment
T¼No T¼Yes
a m b a m b
No No No No 1 0
No No No Yes 0.2 0.8 X1 0.07 0.08 0.09 X8 0.12 0.14 0.16
X
No No Yes No 0.2 0.8 2 0.05 0.06 0.07 X9 0.07 0.08 0.09
X
No No Yes Yes 0.7 0.7 X3 0.04 0.06 0.08 10 0.08 0.09 0.10
X X
… … … … … … 4 0.10 0.12 0.14 11 0.11 0.12 0.13
X
Yes Yes No No 0.4 0.6 X5 0.09 0.11 0.13 12 0.08 0.10 0.12
X
Yes Yes No Yes 0.6 0.4 X6 0.17 0.19 0.21 13 0.10 0.12 0.14
Yes Yes Yes No 0.8 0.2 X
X7 0.07 0.08 0.09 14 0.07 0.08 0.09
Yes Yes Yes Yes 1 0
Table 6
Probability distribution of the occurrence of tunnel leakage (T).
are prepared in advance to cope with the construction failure which is very
likely to occur. With effective safety control measures adopted, the
underground construction of WYMT goes smoothly, and the slurry shield
machine successfully completes its journey across the Yangtze River on 22
September, 2011.
6. Discussion
(b) show that X8 ¼Yes (with a 32.84% chance) turns out to be the most (1) From a perspective of deductive reasoning, both FBN and FFTA can be
unfavorable factor, which should be the focus of practical diagnosis for next used to calculate the occurrence probability of an undesired event using
fault diagnosis until the accident is under control. In this way, the route for the deductive reasoning technique. Each FFTA can be mapped to its
fault diagnosis in tunnel leakage can be extracted. corresponding FBN when transforming AND-gates and OR-gates into
CPT, and thus produce similar results during predictive analysis as long
as basic events are independent of each other. However, a FBN does not
5.5. Step 5: Implementation effects necessarily have an equivalent FFTA due to its different causal
relationships (specified in CPT) rather than simple Boolean functions
In accordance with the FBN-based failure analysis results, several rational such as OR-gate and AND-gate. Moreover, FBN and FFTA are both able
suggestions measures can be proposed for the safety assurance in the to handle uncertainty using data obtained from expert knowledge, in
construction of WYMT. At the design and pre-construction stage, the owner order to over-come the limitation due to a lack of sufficient statistic data
pays much more attention on securing the quality of the tunnel waterproofing in the construction industry.
and drainage system design, as well as the accuracy and completeness of
surveying documents. At the construction stage, the contractor strictly
controls the selection of tunnel segment suppliers in the bidding process. (2) From a perspective of sensitivity analysis, FBN produces a more reliable
Clear provisions concerning the producing, sto-rage, transport, and unloading measure of performance of each risk factor’s contribution than FFTA by
of tunnel segments are specified in the supply contract. Also, the contractor providing the most probable configuration of primary events leading to an
performs the field experiments of the grouting and sealing quality ahead of accident. Particularly, among the proposed three KPIs in the FBN
time, in order to determine optimal parameters in the real grouting and sealing methodology, one can be used in conjunction with the other two, and be
process. In addition, the frequency of seepage monitoring is increased, verified by their results as well. The identified factors at the top of the
ensuring the feedback analysis of the field data in real time. Furthermore, ranking list of each indicator should be considered as the key risk factors
emergency supplies, equipment and personnel for safety control and management in real projects. In addition, compared
to FFTA, sensitivity analysis in the FBN inference does not need to
38
0.5 L. Zhang et al. / Reliability Engineering and System Safety 131 (2014) 29–39
0.45
Probability
0.4
0.25
0.35 0.5
Posterior
0.3 0.45
Probability
0.4
0.2 0.25
0.35
0.15
Posterior
0.3
0.1
0.05
0.2
0
0.15
0.1
0.05
0
X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X5 X7 X8 X9 X10 X11 X12 X13 X14 X14
Risk Factor Risk Factor
Fig. 8. Fault diagnosis in safety control of tunnel leakage: (a) P(Xi|T¼1); and (b) P(Xi|T¼1, X6 ¼Yes).
get minimal cut set, and the computational efficiency is therefore greatly support tool to provide guidelines for failure analysis in tunnel construction,
increased. and is worth popularizing in other similar projects.
(3) From a perspective of abductive reasoning, one of the unique The holistic FBN methodology proposed in this paper also has some
characteristics of FBN for safety risk analysis is its ability for obtaining limitations. During the risk/hazard identification and BN model construction
posterior probabilities in the light of actual observa-tions of root nodes. process, large amounts of scattered knowledge are accumulated from the
Unlike prior probabilities, the posterior probabilities have been updated tunnel construction practice, including explicit and tacit knowledge.
using the accident’s latest information, and thus can reflect the Numerous domain experts have participated in the collection, editing,
characteristics of the accident more specifically than prior probabilities. reorganizing work of the safety related knowledge resources, making an
FBN can repetitively substitute the posteriors for priors in the accident essential contribu-tion to securing a qualified conceptual causal frame for the
re-analysis when a new set of accident related information is observed. development of the cause-effect network model. However, this process is
This substitution not only continuously reduces the data uncertainty, but laborious and relies greatly on domain experts. In addition, we should
it also provides the accident scenario with real-time and up-to-date improve the conditions of validity and robustness of the probability
analysis. However, this abduc-tive reasoning capacity is not possessed in distribution (prior probabilities and conditional probability) of variables in the
FFTA-based risk analysis approach. proposed FBN model by data from a lot of event reports and experiments. Our
subsequent research goal will focus on automatic knowledge acquisition
regarding different knowledge resources, as well as adopting an Expect
7. Conclusions and future works System (ES) technique to develop a real-time expert system for knowledge
management.
Tunnel construction is typically a highly complicated project associated
with large potential risks. In recent years, safety risk analysis and
management of tunnel construction have attracted broad attention because of Acknowledgments
its close relation with public safety. Due to a lack of sufficient data, it is
difficult to have an exact estimation of the failure rate of the occurrence The National Science and Technology Support Plan (No. 51378235),
probability of undesired events. A decision approach based upon FBN for Wuhan City Construction Committee Support Project (Nos. 201334 and
failure analysis in tunnel construction is developed in this research. A typical 201207) and China Scholarship Council (CSC) are acknowledged for their
hazard concerning the tunnel leakage in the construction of WYRT in China financial support of this research.
is used to verify the applicability of the proposed approach. Results
demonstrate the feasibility of the proposed method, as well as its application
potential.
Also, there are some other projects encountering the similar situation,
where the statistical data is insufficient and high potential risks exist in
complex environments, such as coal mining, dam monitoring, nuclear power
plants and others. Specifically, during the fuzzy decision analysis, there
increases the need for precise failure probabilities for the purpose of failure
analysis in project management practice. To reach the highly required preci-
sion for the fuzzy decision analysis, the expert confidence indi-cator can be
first proposed to ensure the reliability of collected data during the fuzzy
probability assessment, with the expert judgment ability and subjectivity
being fully considered. With the capacity of deductive reasoning, sensitivity
analysis and abductive reasoning in Bayesian inference, the safety assurance
progress can be extended to the entire life cycle of risk-prone events,
including the pre-accident, during-construction and post-accident control,
accordingly. The proposed method can be used as a decision