Accepted Manuscript

Title: A Fuzzy Bayesian Network approach for Risk Analysis

in Process Industries

Authors: Mohammad Yazdi, Sohag Kabir

PII: S0957-5820(17)30258-6
DOI: http://dx.doi.org/10.1016/j.psep.2017.08.015
Reference: PSEP 1146

To appear in: Process Safety and Environment Protection

Received date: 21-5-2017

Revised date: 21-7-2017
Accepted date: 13-8-2017

Please cite this article as: Yazdi, Mohammad, Kabir, Sohag, A Fuzzy Bayesian Network
approach for Risk Analysis in Process Industries.Process Safety and Environment
Protection http://dx.doi.org/10.1016/j.psep.2017.08.015

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
A Fuzzy Bayesian Network approach for Risk Analysis in Process Industries

Mohammad Yazdia, Sohag Kabirb,*

a
Department of Industrial Engineering, Eastern Mediterranean University, Mersin 10, Turkey
b
School of Engineering and Computer Science, University of Hull, Hull, HU6 7RX, UK

Corresponding author. Email: [email protected] (Mohammad Yazdi), [email protected]

(Sohag Kabir)

Highlights

 A comprehensive framework is proposed for risk assessment in process industries.

 A combination of expert knowledge and improved fuzzy AHP is used to address uncertainty in failure data.
 Bayesian Network is utilized for probabilistic reasoning about system reliability.
 Fault tree and Bayesian network are employed for the risk analysis of ethylene transportation line.
 Critical basic events are identified considering dependency among events.

Abstract

Fault tree analysis is a widely used method of risk assessment in process industries. However, the classical fault tree
approach has its own limitations such as the inability to deal with uncertain failure data and to consider statistical
dependence among the failure events. In this paper, we propose a comprehensive framework for the risk assessment
in process industries under the conditions of uncertainty and statistical dependency of events. The proposed
approach makes the use of expert knowledge and fuzzy set theory for handling the uncertainty in the failure data and
employs the Bayesian Network modeling for capturing dependency among the events and for a robust probabilistic
reasoning in the conditions of uncertainty. The effectiveness of the approach was demonstrated by performing risk
assessment in an ethylene transportation line unit in an ethylene oxide (EO) production plant.

Key words

Hazard Analysis, Fault tree analysis, Bayesian Networks, Fuzzy Set Theory, Process Industry

1. Introduction

In recent years, complex chemical plants have been rapidly developed to meet the increasing demand of process
industries. As these plants usually process hazardous materials, their failure can cause serious harm both to people
and the environment. For this reason, it is necessary to identify potential risks posed by such systems and then take
measures to minimize the likelihood of these risks. To deal with the large number of accidents, incidents, near
misses, and mishaps in process industries, different risk assessment approaches have been developed and widely
used to perform a hazard analysis, thus enabling the prevention of an inadvertent incident and planning of mitigative
actions (Khakzad et al., 2013a, 2011; Yan et al., 2016).

The risk assessment techniques deployed for identifying and controlling the risks in hazardous industries might
appear satisfactory to their users; however, in a real life scenario, there have been many cases of disastrous accidents
due to the failure of such preventive measure, such as deepwater horizon explosion and oil spill in 2010, Fukushima
disaster in 2011, storage tank explosion at loading port in Tianjin in 2015, the explosion of a natural gas (NG)
factory in Belgium in 2004, and conflagration caused by NG leakage in Paraguay in 2004, etc. (Escande et al., 2016;
Han and Weng, 2011; Khakzad, 2015; Taveau, 2010). The approaches used for hazard analysis in process industries
include, but are not limited to, Hazard and Operability Analysis (HAZOP) (Dunjó et al., 2010), Fault Tree Analysis
(FTA) (Vesely et al., 2002), and Bow-tie diagrams (De Dianous and Fiévez, 2006). FTA is the most popular among
all available techniques and it has been extensively used for risk analysis in several industries (Kabir, 2017; Khan et
al., 2008; Wang et al., 2002).

A fault tree (FT) can be analyzed both qualitatively and quantitatively. A qualitative analysis minimizes a fault tree
to a set of minimal cut sets (MCSs), which are the smallest combinations of events that are necessary and sufficient
to cause the top event (TE), i.e., a hazardous event. A quantitative analysis mathematically calculates the occurrence
probability of the top event and other relevant numerical indexes, given the failure rate/probability of an individual
element of a system. For this reason, the applicability of FTA for quantitative analysis largely depends on the
availability of failure data. However, for most of the large and complex systems, it is often very difficult to obtain
precise failure data due to the lack of knowledge, scarcity of statistical data, ambiguous component behavior, and
the operating environment of the system (Huang et al., 2004; Kabir et al., 2016; Omidvari et al., 2014). In such
cases, FTA cannot be used for risk assessment due to the lack of failure data. In addition to that, during the
quantitative analysis, different events (e.g. intermediate events representing MCSs) in the fault tree are generally
considered to be statistically independent. However, in real life systems, the events are not always statistically
independent. For example, two MCSs may become statistically dependent because of sharing of a common basic
event. For this reason, the statistical independence assumption about the events could lead to an inappropriate
estimation of system dependability.

In order to deal with the ambiguities and shortages of data in conventional FTA, extensive studies have been
performed by employing fuzzy set theory in different areas. In recent years, Celik et al. (2010) and Lavasani et al.
(2015a, 2015b, 2011) have used the triangular and trapezoidal fuzzy numbers to compute the failure probability (FP)
of the top event (TE) with respect to an expert judgment in different chemical industries. Fuzzy FTA was applied for
the reliability analysis of fire and explosion of a crude oil tank by Wang et al.(2013). Furthermore, Yazdi et al.
(2017) utilized the similarity aggregation method (SAM) in terms of fuzzy set theory to compute the failure
probability of a granule storage tank. Ferdous et al. ( 2009) used a computer aided fuzzy fault tree analysis for the
same purpose. However, their model cannot cope with the dependency and redundant events in a realistic system. A
potential remedy to this problem is to translate a static fuzzy fault tree to a different model, which is capable of
capturing dependency among events and can also model scenarios with redundant events.

Bayesian network (BN) is a well-known graphical inference method which expresses the causal relationships
between the causes and final outcomes in a system (Rausand, 2011). In addition, BN has also been widely used in
various engineering applications such as risk and reliability assessment (Abimbola et al., 2015; Bouejla et al., 2014;
Hänninen et al., 2014; Huang et al., 2006; Khakzad et al., 2013a, Yuan et al., 2015), improving the safety
performance of a system ( Herrero et al., 2013; John et al., 2016; Trucco et al., 2008), updating failure probability
(Leu and Chang, 2013; Musharraf et al., 2016; Wu et al., 2015), and mapping static or dynamic FTs into
corresponding BNs (Barua et al., 2016; Hänninen et al., 2014; Kabir et al., 2014; Khakzad et al., 2013a, 2013b;
Marvin et al., 2017; Pereira et al., 2016; Yeo et al., 2016; Zarei et al., 2017). In a BN model, both forward and
backward analysis could be performed. A forward analysis is performed to estimate the probability of unknown
variables by following the arcs of the network. On the other hand, the backward analysis is performed by following
the network arcs in opposite direction to update the probability of known variables based on some evidence.

The objective of this study was to find a new approach for performing risk analysis in a more consistent way under
the condition of uncertainty. In the proposed approach, a fault tree is used for qualitative analysis to identify the root
causes of the hazardous event and the fuzzy set theory, along with expert judgment, is used to obtain the unknown
failure data of basic events of the FT. The probability of the occurrence of the hazardous events and other related
reliability indexes are calculated by translating the fault tree into a Bayesian network. The proposed methodology
was applied for risk analysis of ethylene transportation line. The results obtained by the proposed approach are also
compared with those obtained by the classical fuzzy fault tree approach. This approach offers an improved
quantitative analysis of complex systems by eliminating the assumption of statistical independence among the
events. Moreover, unlike the traditional approaches, the proposed one can be used for diagnostic analysis of
systems, which is particularly important to determine maintenance strategies.

This paper is organized as follows. In Section 2, a new approach based on the fuzzy theory and Bayesian modeling
is introduced that computes the failure probability of TE. A numerical example is presented in Section 3 to
exemplify the feasibility and effectiveness of the proposed model. Finally, the conclusion and future remarks are
described in Section 4.

2. Material and Methodology

FT is a well-known method which is widely used in failure analysis (Yuan et al., 2015). It is a deductive
methodology that represents the interrelationship between basic events (BEs) and the root cause in a complex
system. These interrelationships are typically obtainable as logical AND/OR gates (Bobbio et al., 2001; Rausand,
2011). In order to find the probability of TE in a specific case, the failure rates of BEs are taken from standard
reliability data sources such as OREDA (2002). However, as mentioned earlier, there some limitations, which might
occur during the analysis, such as the lack of data for BEs or analyzing the dynamic system, which conventional
FTA is not capable of. This section, to deal with these limitations, provides a new approach. A graphical overview
of the approach is shown in Fig. 1. Hazard analysis, FT construction and data collection, Bayesian modeling, and the
calculation are the four key stages in the proposed method, which are presented in details as follows.

2.1. Hazard analysis

In the recent decade, several methods have been developed for hazard analysis, e.g., HAZOP and FMEA (Khakzad
et al., 2011). In this regard, all possibilities of failures need to be considered. Therefore, understanding the process
thoroughly is the vital step. After completing the collection of information process, the identification of all hazards
and menaces for any hazardous events that may cause damage to the equipment or harm to people and/or
environment needs to be taken into consideration (Rausand and Hoyland, 2004). Among all the mentioned analysis
techniques based on brainstorming methods, the critical and hazardous system and sub-system can be figured out by
employing a group of experts for further studies, such as causal and frequency analysis. In this study, the output of
HAZOP is selected as a worst-case event.

2.2. Fault tree construction and data collection

The construction of an FT always starts with a specified TE placed at the top of the tree and the rest of the tree is
constructed in the downward direction. The TE usually indicates an accident that can cause asset loss or safety
hazards (Lewis, 1996). In order to complete a tree, the BEs that are denoted as the lowest level (leaves) of the tree
should be known beforehand. In an FT, BEs are usually considered statistically independent and could be in any of
the two binary states (failed and non-failed).
As mentioned earlier, the exponentially distributed failure rate of known BEs can be obtained from OREDA (2002)
and subsequently is transferred to failure probability by Equation (1).

𝑃(𝑡) = 1 − 𝑒 −𝜆𝑡 (1)

where P denotes the probability of failure, λ represents failure rate (the number of failures per year), and t is the time
inspection interval.

On the other hand, the three methods-extrapolation, statistical, and expert judgment-can be employed to estimate the
probability of BEs with unknown or limited failure data (Preyssl, 1995). The extrapolation technique is based on an
estimation approach, which is applied on a standard reliability data source, while the statistical technique implies the
examination of data in a direct way to compute the probability of an event. Besides, the expert judgment method can
be engaged to estimate the probabilities with respect to experts’ opinions.

In this study, the expert judgment method as a scientific consensus approach is employed to compute the probability
of unknown BEs. Therefore, an integration of fuzzy set theory and subjective opinions to overcome any possible
ambiguity can help the assessors (Yazdi et al., 2017a).

The following section introduces a way showing how experts can be employed and weighted to obtain the failure
data.

2.3. Using expert judgment (Fuzzy AHP)

Fuzzy set theory has been used in several studies to aggregate the expert opinions to overcome the possible
ambiguities in failure data. The main aspect of the fuzzy set theory is how to quantify the qualitative index. The
subjectivity should be avoided in order to get more credible and reliable results. A simple averaging method,
regarding some criteria including personal experience, job tenure, and education level, has mostly applied in the
literature for expert weighting. This approach does not provide sufficiently high objectivity (Lavasani et al., 2015a,
2015b; Miri Lavasani et al., 2011; Ramzali et al., 2015; Yan et al., 2016; Yazdi et al., 2017b). Therefore, an
extension of fuzzy analytical hierarchy process (FAHP) can overcome subjectivity issues.

Conventional AHP is a well-known method, which is commonly used in multi-criteria decision-making problems. In
this case, as conventional AHP method cannot deal with subjective knowledge, FAHP has been developed to solve
the AHP problems (Gul and Guneri, 2016). In other words, the main purpose of AHP is collecting expert opinions,
though conventional AHP cannot reflect the human thinking. Several FAHP techniques have been proposed in the
last few decades. Of them, the two most important were introduced by Buckley (1985) and Chang (1996), which
used trapezoidal and triangular fuzzy membership function for a pairwise comparison scale, respectively.

In this study, an extension of Buckley's method was used for weighting the experts, due to the limitations in other
techniques, such as all fuzzy comparison matrices cannot be completely used. Besides, in Buckley's method,
illogical zero weight may also be obtained for the selection criteria (Chan and Wang, 2013).

The weight of each expert can be computed in a more reliable way on the basis of their knowledge and experience.
Therefore, the computed weights are vital in order to represent the relative superiority of the employed experts. The
next section introduces an approach based on the fuzzy theory to transform the linguistic possibilities of expert
opinions into a fuzzy probability to aggregate their opinions into a crisp probability value (Altunkaynak et al., 2005;
Duan et al., 2016; Mohsendokht, 2017; Shi et al., 2014).

2.4. Aggregation procedure

The aggregation procedure of expert judgment in the fuzzy logic system is divided into three stages as below.
Stage 1: Obtaining linguistic terms of unknown BEs based on expert judgment
Stage 2: Converting linguistic terms into corresponding fuzzy numbers
Stage 3: Converting fuzzy numbers into fuzzy possibility scores (FPS)

Further details of the mentioned stages are explained below.

2.4.1. Obtaining linguistic terms of unknown BEs based on expert judgment (stage 1)

The purpose of stage 1 was determining the failure likelihood of BEs considering the linguistic terms expressed by
the experts. Several experts were consulted for this purpose and were furnished with a questionnaire by email to
provide their judgment about the failure possibility of basic events. The linguistic terms, representing the probability
of BEs, were scaled at seven levels: very high (VH), high (H), fairly high (FH), medium (M), fairly low (FL), low
(L), and very low (VL). These levels were based on Saaty's approach (Saaty and Ozdemir, 2003), who discussed that
the proper number for expert judgment at a specified time is between five and nine or in the other words the
common capacity of human judgment is seven plus/minus two chunks.

2.4.2. Converting linguistic terms into corresponding fuzzy numbers (stage 2)

There are many applications of fuzzy set theory to deal with uncertainties and inaccuracy in expert judgments in
linguistic terms such as triangular, trapezoidal, intuitionistic, and Gaussian fuzzy membership function (Atanassov,
2012; Purba et al., 2014). The guarantee of the best membership function is based on realistic circumstances
(Markowski and Mannan, 2008). In earlier studies, triangular and trapezoidal fuzzy numbers have been found to be
effective for risk assessment purpose (Ferdous et al., 2013; Lavasani et al., 2015a, 2015b; Mardani et al., 2015;
Ramzali et al., 2015; Yazdi, 2017). Therefore, both triangular and trapezoidal fuzzy numbers were utilized to map
linguistic opinions to fuzzy membership function. The reason of using these two types of fuzzy numbers is that
under some weak assumptions, the defined membership functions can directly meet the appropriate optimization
criteria (Pedrycz, 1994).

In this way, a corresponding fuzzy number for each linguistic term could be obtained. Further, it is necessary to
aggregate subjective opinions of experts regarding identified BEs into a single opinion. Various techniques are
available in the literature to aggregate experts' opinions such as linear opinion pool, max-min Delphi, sum-product,
max-product, and similarity aggregation (Aqlan and Mustafa Ali, 2014; Ishikawa et al., 1993; Ross, 2009).
However, Liu et al.(2014) discussed that there is no way to show which method has superiority to the other one.

The sum-production algorithm, which is used in this study for the aggregation process could be presented as follows.

𝑍𝑖 = ∑𝑛𝑗=1 𝑤𝑗 . 𝑓𝑖𝑗 , 𝑖 = 1,2, … , 𝑚 𝑗 = 1,2, … , 𝑛 (2)

where 𝑍𝑖 denotes the aggregated fuzzy number for BEi, 𝑤𝑗 represents the weight of experts j, and 𝑓𝑖𝑗 is corresponding
fuzzy number of BEi given by expert j. n and m are the number of experts and BEs respectively. The 𝛼-cut is a
commonly used method to operate the fuzzy membership function (Lowen, 1996; Mesiar, 2007; Ross, 2009; Zhang
et al., 2017).

2.4.3. Converting fuzzy numbers into fuzzy possibility scores (FPS) (stage 3)

FPS denotes a crisp value that is based on experts’ opinions aggregated for a possible event. In order to defuzzify a
quantifiable outcome in the fuzzy set theory, several common techniques are available including the center of area
(CoA), the center of the largest area, max-min, bisector, weighted average, mean max, and the center of sum
(Akkurt et al., 2004; Ross, 2009). In this study, the max-min aggregation methods, proposed by Chen and Hwang
(1992), was used for defuzzification process (Shi et al., 2014).The maximum and minimum fuzzy sets are expressed
as follows.

𝑥, (0 ≤ 𝑥 ≤ 1)
𝑓𝑚𝑎𝑥 (𝑥) = { (3)
0, (otherwise)

1 − 𝑥, (0 ≤ 𝑥 ≤ 1)
𝑓𝑚𝑖𝑛 (𝑥) = { (4)
0, (otherwise)

Subsequently, the right and left score of fuzzy set (Z) can be computed as follows, respectively.

𝑠𝑢𝑝 (1 − 𝑑)
𝐹𝑃𝑆𝑅𝑖𝑔ℎ𝑡 (𝑍) = 𝑥[𝑓𝑧 (𝑥) ⋀ 𝑓𝑚𝑎𝑥 (𝑥)] = ⁄[1 + (𝑑 − 𝑐)] (5)

𝑠𝑢𝑝 (1 − 𝑎)
𝐹𝑃𝑆𝐿𝑒𝑓𝑡 (𝑍) = 𝑥[𝑓𝑧 (𝑥) ⋀ 𝑓𝑚𝑖𝑛 (𝑥)] = ⁄[1 + (𝑏 − 𝑎)] (6)

In addition, the relationship between left and right side of fuzzy set (Z) is illustrated in Fig. 3.

Therefore, the fuzzy possibility scores of fuzzy number 𝑍𝑖 can be obtained by the following equation.

𝐹𝑃𝑆(𝑍𝑖 ) = [𝐹𝑃𝑆𝑅𝑖𝑔ℎ𝑡 (𝑍𝑖 ) + 1 − 𝐹𝑃𝑆𝐿𝑒𝑓𝑡 (𝑍𝑖 )]⁄2 (7)

Thus the aggregated FPS with respect to experts’ weight is computed for each BE. In the next section, the failure
probability calculation process and mapping of FT into BN for executing further actions are explained.

2.5. Calculation and Bayesian Modeling

2.5.1. The calculation of failure probability and TE probability

FPS are converted to failure probability by using following equation proposed by Onisawa (1990).

𝑘
Failure probability = { 1⁄10 𝐹𝑃𝑆 ≠ 0 (8)
0 𝐹𝑃𝑆 = 0

𝑘 = 2.301 × [(1 − 𝐹𝑃𝑆)/𝐹𝑃𝑆]1⁄3 (9)

The process of transferring linguistic terms into failure probabilities considering 𝛼-cut methods is completed at this
point. Therefore, by following the above processes failure probability of all BEs can be computed. Moreover, using
Boolean algebra and the failure probability of the BEs, the failure probability of TE can be computed (Banerjee,
2003). This process is known as the analytical method for quantifying FTs.

2.5.2. Bayesian Modeling

2.5.2.1. Bayesian Network

Similar to FT, BN is a probabilistic graphical technique, which is widely used for constructing system reliability
models based on uncertain knowledge (Khakzad et al., 2013, 2011; Zarei et al., 2017). BN results in a directed
acyclic graph that includes the set of nodes denoting the variables, which are connected by directed arcs. The arcs
represent the probabilistic causal relationship between the variables and Conditional Probability Tables (CPTs) are
set to the nodes in order to represent the conditional dependencies.
According to conditional dependency of variables and chain rules, BN denoted the joint probability distribution of
set of variables as follows (Jensen and Nielsen, 2007):

𝑃(𝑈) = ∏𝑛−1
𝑖=1 𝑃(𝑋𝑖 ∣ 𝑋𝑖+1 , … 𝑋𝑛 ) (10)

Where 𝑈 = {𝑋1 , 𝑋2 , … , 𝑋𝑛 } and 𝑋𝑖+1 is the parent of 𝑋𝑖 . Accordingly, the probability of 𝑋𝑖 can be computed as:

𝑃(𝑋𝑖 ) = ∑𝑈∖𝑋𝑖 𝑃(𝑈) (11)

The risk and reliability analysis is a well-known application of BN to update the prior failure probabilities of
specified events, given the new information in the form of evidence. This information is usually based on expert
knowledge or becomes available in the lifecycle of processes, such as accidents, incidents, near misses, mishaps, etc.
Therefore, BN allows reaching a better analysis in a dependable situation of a complex system including common
cause failure (CCF) and diagnostic reasoning (Bobbio et al., 2001).

Based on Bayes theorem, BN can be used to update the prior probability of an event (E):

𝑃(𝑈∩𝐸) 𝑃(𝑈∩𝐸)
𝑃( 𝑈 ∣ 𝐸 ) = =∑ (12)
𝑃(𝐸) 𝑈 𝑃(𝑈∩𝐸)

In the case of FT, it can be shown as follows:

𝑃(𝐵𝐸𝑘 ∩𝐸𝑖𝑛 ) 𝑃( 𝐸𝑖𝑛 ∣∣𝐵𝐸𝑘 )∙𝑃(𝐵𝐸𝑘 )

𝑗,𝑘 ∣
𝑃 ( 𝐵𝐸𝑘 ∣∣ 𝐸𝑖𝑛𝑗,𝑘 ) =
𝑗,𝑘
= (13)
𝑃(𝐸𝑖𝑛 ) 𝑃(𝐸𝑖𝑛 )
𝑗,𝑘 𝑗,𝑘

where
 𝐸𝑖𝑛𝑗,𝑘 represents the jth state of “given new information” denoting the effect of new events on 𝐵𝐸𝑘 . Besides,
𝑃 ( 𝐵𝐸𝑘 ∣∣ 𝐸𝑖𝑛𝑗,𝑘 ) is the posterior failure probability of 𝐵𝐸𝑘 given 𝐸𝑖𝑛𝑗,𝑘 ;
 𝑃(𝐵𝐸𝑘 ) is the prior failure probability of 𝐵𝐸𝑘 provided by expert judgment or standard reliability sources;
 𝑃(𝐸𝑖𝑛𝑗,𝑘 ) is the probability of state j of the k “given new information” which is estimated from the BN
 𝑃 ( 𝐸𝑖𝑛𝑗,𝑘 ∣∣ 𝐵𝐸𝑘 ) is denoted as degree of belief in the failure of 𝐸𝑖𝑛𝑗,𝑘 given the failure of 𝐵𝐸𝑘 . The value
of 𝑃 ( 𝐸𝑖𝑛𝑗,𝑘 ∣∣ 𝐵𝐸𝑘 ) can be found out by asking question from experts.

2.5.2.2. Mapping procedure

The mapping of FT into BN is based on graphical and numerical functions (Bobbio et al., 2001). In graphical
mapping, top event, intermediate events, and basic events of FT are denoted as leaf nodes, intermediate nodes, and
root nodes in an equivalent BN, respectively. In an FT, the top event and the intermediate events are always
represented as logic gates. For this reason, to be able to distinguish the nodes representing the basic events from the
nodes representing the logic gates in a BN, we used single circles to represent the basic event nodes and double
circles to represent the top and intermediate event nodes. In numerical mapping, the failure probabilities of BEs are
given to the related nodes as prior failure probabilities. CPTs can be generated for all intermediate nodes based on
the logic gates they represent ( see details in Bobbio et al. (2001)). Fig. 4 shows the simplified procedure of mapping
of FTs into BNs presented by Khakzad et al. (2011).
2.5.3. Importance measures and Critical nodes

In FTA, there are many importance measures available to measure the relative importance of each BE or MCS in
terms of their impact on the occurrence of the top event (Modarres, 2006; Rausand, 2014). In parallel way,
according to Bayesian theorem, the importance measures can be represented in the new approach and computed by
their logical relationships in BN (Van Der Borst and Schoonakker, 2001).

Birnbaum Measure (BM): BM of an event is obtained by taking the difference between the probability of the top
event by setting the event probability to 1 and 0, respectively.

In BN model, BM can be obtained as:

𝐵𝑀
𝐼𝐵𝐸𝑖
= 𝑃( 𝑇𝐸 ∣ 𝐵𝐸𝑖 = 1 ) − 𝑃( 𝑇𝐸 ∣ 𝐵𝐸𝑖 = 0 ) (14)

Where 𝑃( 𝑇𝐸 ∣ 𝐵𝐸𝑖 = 1 ) and 𝑃( 𝑇𝐸 ∣ 𝐵𝐸𝑖 = 0 ) denoted the conditional probability of the TE for occurrence and
nonoccurrence of 𝐵𝐸𝑖 , respectively.

Risk Reduction worth (RRW): RRW shows the effect of BE on TE with respect to non-occurrence of 𝐵𝐸𝑖 . RRW
represents the decrease of TE probability when a given BE is assured not to occur. It can be calculated from the BN
as:

𝑅𝑅𝑊
𝐼𝐵𝐸𝑖
= 𝑃(𝑇𝐸) − 𝑃( 𝑇𝐸 ∣ 𝐵𝐸𝑖 = 0 ) (15)

The criticality of the nodes can be obtained by computing above mentioned importance measures. The rank of
critical nodes is obtained as follows.

𝐼
𝑅 ∗ = ∑𝑛 𝑚 × 100 (16)
𝑖=1 𝐼𝑚

where, 𝐼𝑚 is the mth importance measures and n is the number of BE or MCS. 𝑅∗ denotes the normalized weight of
importance measures.

The total weighing of importance measures can be computed as follows:

∗ ∗ ∗
𝑅𝑇𝑜𝑡𝑎𝑙 = 𝑅𝐵𝑀 +𝑅𝑅𝑅𝑊 (17)

Finally, the BEs is ranked in ascending order from most critical to the least critical BEs.

3. Application of the proposed methodology

The proposed methodology was applied to an ethylene oxide (EO) plant for an ethylene transportation line unit. A
risk assessment study was conducted by Khan et al. (2002) for an EO production plant. In the study, most hazardous
units, such as the EO storage unit, reaction unit, ethylene EO distillation column, ethylene transportation line, and
ethylene reboiler, were recognized. It was suggested that further details of risk assessment are required for the
mentioned units. In addition, ethylene transportation line unit was identified as the third major hazard in the present
units. Moreover, the most probable accident scenario that can be expected in this case is the ethylene release due to a
leak or rupture, causing a vaporized cloud to be formed and reaching to an ignition source resulting in fire and
explosion. To find an optimal maintenance strategy, another study was performed on the above case study in (Khan
and Haddara, 2004) based on risk-based maintenance (RBM) method (Khan and Haddara, 2003). RBM
methodology combines several pre-existing methods and tools such as MCAS (Khan, 2001), MAXCRED (Khan and
Abbasi, 1999a), PROFAT (Khan and Abbasi, 1999b) and PROFAT II (Khan and Abbasi, 2000) to efficiently reach
a maintenance decision.

3.1. Overview of the Process

EO is produced by the oxidation of ethylene and pure oxygen. The completed process flow diagram (PFD) of EO
plant is illustrated in Fig.5. Ethylene and oxygen are reacted at 10-30 bar and 150-260ᵒ C in a fix bed catalysis
reactor and has been transported from the storage tanks to isolated vicinity to the reaction unit through pipeline (see
Khan et al. (2002) for more details).

3.2. Probabilistic Risk Assessment

3.2.1. FT Development

The TE of the FT was selected as an ignition of vapor cloud which may lead to a fireball. The developed FT is
illustrated in Fig. 6. The identified 25 basic events, which contribute directly and/or indirectly to the specified TE,
are shown in Table 1. As mentioned earlier, the BEs with known failure rates are separated from the ambiguous
ones. The failure rates of some BEs are determined by employing OREDA (2002) and the failure rates of the
ambiguous ones are estimated based on the expert judgment. As shown in Table 1, failure rate data for 9 BEs were
obtained from the reliability database, whereas the failure rates could not be identified for 16 BEs from the available
data.

3.2.2. Obtaining Failure Rates of BEs

As mentioned earlier, to perform a proper risk analysis, it is necessary to access the historical failure rate data to
allocate them to BEs. Thus, the probability of hazards with identified failure rate can be computed. The failure
probability of the BEs with known failure rates were computed on per year basis (8760 h) and are presented in Table
2.

The process of obtaining the failure probability of the basic events with unknown data is described in the following
subsections.

3.2.3. Rating stage

As mentioned above, to compute the FP of ambiguous BEs, an expert judgment is employed. The expert knowledge
is biased by individual visions and purposes (Ford and Sterman, 1998); thus, it is very difficult to obtain an impartial
expert opinion. The main point here is the selection of both heterogeneous specialists (e.g., workers and experts) and
homogenous specialists (in this case it includes only experts).

It has been suggested that the impact of individual experience is smaller in a homogenous group compared to a
heterogeneous group as a result of differences in the experience. Therefore, the main advantage of a heterogeneous
group compared to the homogenous group is that all possible opinions of heterogeneous specialists can be
considered. Thus, in this study, a heterogeneous group of experts including four specialists with different
backgrounds was employed to compute the FP of the 16 ambiguous BEs.

Expert No. 1: An experienced safety auditor and risk assessor working as consultant for complex chemical plant.
Expert No. 2: An experienced technician working in different kinds of process industry.
Expert No. 3: A senior chemical process designer from process engineering department with master certificate.
Expert No. 4: An experienced safety officer working in a complex process plant with safety engineering certificate.
Fuzzy AHP method was used to compute each expert’s capability and assigning the respective weights. The system
of expert information is illustrated in Fig. 7, and the expert profile and weights are shown in Table 3.

Let's take, as an example, that BE23 valve fails to open. With respect to the definition of fuzzy numbers presented in
Fig. 2, the qualitative terms, given by four experts, fall into “VL”, “FL”, “FL”, and “FH” categories. The integrated
fuzzy number is attained as follows:

𝑓(𝑥) = 𝑤𝐸1 × 𝑓𝑉𝐿 (𝑥) + (𝑤𝐸2 + 𝑤𝐸3 ) × 𝑓𝐹𝐿 (𝑥) + 𝑤𝐸4 × 𝑓𝐹𝐻 (𝑥)

The membership function of aggregated fuzzy number can be obtained as:

𝑥 + 0.128
, 0.188 < 𝑥 ≤ 0.288
0.16
1, 0.288 < 𝑥 ≤ 0.363
𝑓(𝑥) =
−𝑥 + 0.463
, 0.363 < 𝑥 ≤ 0.463
0.1
{ 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

The right and left utility scores of fuzzy number f(x) are computed using equations (5)-(7) and provided as follows:

𝐹𝑃𝑆𝐿𝑒𝑓𝑡 (𝑥) = 0.738

𝐹𝑃𝑆𝑅𝑖𝑔ℎ𝑡 (𝑥) = 0.488

𝐹𝑃𝑆(𝑥𝑖 ) = 0.375

Using equations (8) and (9), the FP of BE23 is computed to be 0.001874.

The qualitative terms based on experts’ opinions and FP of unknown BEs are shown in Table 4.

3.2.4 Bayesian Modeling and Analysis

Once the data for all basic events are obtained, the fault tree (Fig. 6) is mapped to a Bayesian Network (Fig. 8). The
prior probability values of the root nodes of the BN are defined based on the values shown in Tables 2 and 4. The
conditional probability values of each intermediate node of the BN are populated based on the type of logic gate it
represents.

In this BN, G1 is the node corresponds to the TE of the FT. Now running a query on this node would return the
value of system unreliability. The value of system unreliability for the system obtained from the BN model is
3.623E−09. The system unreliability was also calculated using the analytical approach and the value obtained was
3.713E−09. This value is 2.48% higher than the value estimated by the BN based approach. This is due to the fact
that the analytical approach does not consider the statistical dependence among the events. However, it can be seen
from the BN model that some events are statistically dependent on each other. For example, the events represented
by nodes G4 and G5 are statistically dependent on each other as they share a common basic event BE1. For a similar
reason, nodes G9 and G15, and G17 and G18 are also statistically dependent. The effect of these dependences also
propagates through the network to the node representing the TE.

One important aspect of probabilistic risk assessment of a system is to determine the critical components based on
their contribution to the occurrence of the system failure. This information can help in improving the system
reliability by taking the necessary measures such as by putting more design efforts on the weakest part of the system
or by improving the system design by introducing fault tolerant strategies. The criticality of the basic events of the
FT in Fig. 6 is calculated following the procedure described in section 2.5.3 using both the analytical and the BN-
based approaches. The results of the evaluation are shown in Table 5. According to the results shown in Table 5,
BE6 (Ignition source present) contributes the most to the top event probability, hence, ranked as the most critical
component. The second most critical event is BE1 (Flammable gas detector fails). If we compare the ranking of the
events, we can see that both BN-based and the analytical approaches agree on the ranking of most of the
components. However, there are some disagreements between the two approaches regarding the ranking of events.
For example, the BN-based approach ranked BE14 as the least critical event and BE16 as the second least critical,
whereas the analytical approach ranked them in opposite order. We believe that these disagreements are due to the
statistical independence assumption of the events in the analytical approach.

We used the predictive reasoning on the BN model to obtain the system unreliability by following the directions of
the network arcs. For predictive analysis, we used the information (failure probability) about the causes (component
failure) to obtain the new belief (unreliability value) about the effect (system failure). With the help of an evidence-
based analysis, we can also perform a diagnostic analysis for the BN, i.e., reasoning from symptoms to causes. For
instance, when an analyst observes that the system has failed, depending on this observation, his/her belief regarding
the failure probability of the basic events can be updated. This means a posterior probability distribution for the
components can be obtained given the status of the system. The posterior probability values for the root nodes of the
BN (Fig. 8) are obtained by providing the evidence that the top event has occurred (i.e., observing the BN node G1
to be true). The point to be noted is that this reasoning is performed backwardly, i.e., opposite to the direction of the
arcs. Now, if we perform a predictive analysis based on this updated belief about the probability distribution of the
basic events, then the value of system unreliability will also be updated. For this case study, the updated value of the
system unreliability is 0.5407, previously it was 3.623E−09.

4. Conclusion

In this paper, we proposed a comprehensive framework to combine fuzzy set theory and expert knowledge with
FTA through Bayesian Network modeling to enable risk assessment of complex systems with ambiguous failure
data. The proposed method addresses three important issues in probabilistic risk assessments, namely, the challenges
of unavailability of failure data, the dependency of failure events, and the uncertainty. The use of expert elicitation
and fuzzy set theory allows handling of the issue of insufficient failure data and also explicitly highlights the areas
of uncertainty in the data. We used Bayesian Network for probabilistic reasoning to obtain system reliability related
indexes and also to capture the dependencies among the events. The effectiveness of this approach was
demonstrated by applying it to the risk assessment of an ethylene transportation line unit in an ethylene oxide (EO)
plant and by comparing the results with the results obtained by the analytical approach. The proposed approach was
found to be more robust and the results obtained were more accurate, as this approach does not estimate the system
reliability under the unrealistic assumption of statistical independence of events.

In the present work, we consider the failure rate of basic events as exponentially distributed. In future, we have a
plan to extend this work by considering non-exponentially distributed failure data and the dynamic behavior of
systems.
References

Abimbola, M., Khan, F., Khakzad, N., Butt, S., 2015. Safety and risk analysis of managed pressure drilling
operation using Bayesian network. Saf. Sci. 76, 133–144. doi:10.1016/j.ssci.2015.01.010

Akkurt, S., Tayfur, G., Can, S., 2004. Fuzzy logic model for the prediction of cement compressive strength. Cem.
Concr. Res. 34, 1429–1433. doi:10.1016/j.cemconres.2004.01.020

Altunkaynak, A., Özger, M., Çakmakcı, M., 2005. Fuzzy logic modeling of the dissolved oxygen fluctuations in
Golden Horn. Ecol. Modell. 189, 436–446. doi:10.1016/j.ecolmodel.2005.03.007

Aqlan, F., Mustafa Ali, E., 2014. Integrating lean principles and fuzzy bow-tie analysis for risk assessment in
chemical industry. J. Loss Prev. Process Ind. 29, 39–48. doi:10.1016/j.jlp.2014.01.006

Atanassov, K.T., 2012. On the Concept of Intuitionistic Fuzzy Sets. pp. 1–16. doi:10.1007/978-3-642-29127-2_1

Banerjee, S., 2003. Industrial hazards and plant safety. Taylor & Francis.

Barua, S., Gao, X., Pasman, H., Mannan, M.S., 2016. Bayesian network based dynamic operational risk assessment.
J. Loss Prev. Process Ind. 41, 399–410. doi:10.1016/j.jlp.2015.11.024

Bobbio, A., Portinale, L., Minichino, M., Ciancamerla, E., 2001. Improving the analysis of dependable systems by
mapping Fault Trees into Bayesian Networks. Reliab. Eng. Syst. Saf. 71, 249–260. doi:10.1016/S0951-
8320(00)00077-6

Bouejla, A., Chaze, X., Guarnieri, F., Napoli, A., 2014. A Bayesian network to manage risks of maritime piracy
against offshore oil fields. Saf. Sci. 68, 222–230. doi:10.1016/j.ssci.2014.04.010

Buckley, J.J., 1985. Fuzzy hierarchical analysis. Fuzzy Sets Syst. 17, 233–247. doi:10.1016/0165-0114(85)90090-9

Celik, M., Lavasani, S.M., Wang, J., 2010. A risk-based modelling approach to enhance shipping accident
investigation. Saf. Sci. 48, 18–27. doi:10.1016/j.ssci.2009.04.007

Chan, H.K., Wang, X., 2013. Fuzzy Extent Analysis for Food Risk Assessment, in: Fuzzy Hierarchical Model for
Risk Assessment. Springer London, London, pp. 89–114. doi:10.1007/978-1-4471-5043-5_6

Chang, D.-Y., 1996. Applications of the extent analysis method on fuzzy AHP. Eur. J. Oper. Res. 95, 649–655.
doi:10.1016/0377-2217(95)00300-2

Chen, S.-J., Hwang, C.-L., 1992. Fuzzy Multiple Attribute Decision Making, Lecture Notes in Economics and
Mathematical Systems. Springer Berlin Heidelberg, Berlin, Heidelberg. doi:10.1007/978-3-642-46768-4

De Dianous, V., Fiévez, C., 2006. ARAMIS project: A more explicit demonstration of risk control through the use
of bow-tie diagrams and the evaluation of safety barrier performance. J. Hazard. Mater. 130, 220–233.
doi:10.1016/j.jhazmat.2005.07.010

Duan, Y., Zhao, J., Chen, J., Bai, G., 2016. A risk matrix analysis method based on potential risk influence : A case
study on cryogenic liquid hydrogen filling system. Process Saf. Environ. Prot. 102, 277–287.
doi:10.1016/j.psep.2016.03.022

Dunjó, J., Fthenakis, V., Vílchez, J.A., Arnaldos, J., 2010. Hazard and operability (HAZOP) analysis. A literature
review. J. Hazard. Mater. 173, 19–32. doi:10.1016/j.jhazmat.2009.08.076

Escande, J., Proust, C., Le Coze, J.C., 2016. Limitations of current risk assessment methods to foresee emerging
risks: Towards a new methodology? J. Loss Prev. Process Ind. 43, 730–735. doi:10.1016/j.jlp.2016.06.008

Ferdous, R., Khan, F., Sadiq, R., Amyotte, P., Veitch, B., 2013. Analyzing system safety and risks under uncertainty
using a bow-tie diagram: An innovative approach. Process Saf. Environ. Prot. 91, 1–18.
 doi:10.1016/j.psep.2011.08.010

Ferdous, R., Khan, F., Veitch, B., Amyotte, P.R., 2009. Methodology for computer aided fuzzy fault tree analysis.
Process Saf. Environ. Prot. 87, 217–226. doi:10.1016/j.psep.2009.04.004

Ford, D.N., Sterman, J.D., 1998. Expert knowledge elicitation to improve formal and mental models. Syst. Dyn.
Rev. 14, 309–340. doi:10.1002/(SICI)1099-1727(199824)14:4<309::AID-SDR154>3.0.CO;2-5

García-Herrero, S., Mariscal, M.A., Gutiérrez, J.M., Toca-Otero, A., 2013. Bayesian network analysis of safety
culture and organizational culture in a nuclear power plant. Saf. Sci. 53, 82–95. doi:10.1016/j.ssci.2012.09.004

Gul, M., Guneri, A.F., 2016. A fuzzy multi criteria risk assessment based on decision matrix technique: A case study
for aluminum industry. J. Loss Prev. Process Ind. 40, 89–100. doi:10.1016/j.jlp.2015.11.023

Han, Z.Y., Weng, W.G., 2011. Comparison study on qualitative and quantitative risk assessment methods for urban
natural gas pipeline network. J. Hazard. Mater. 189, 509–518. doi:10.1016/j.jhazmat.2011.02.067

Hänninen, M., Valdez Banda, O.A., Kujala, P., 2014. Bayesian network model of maritime safety management.
Expert Syst. Appl. 41, 7837–7846. doi:10.1016/j.eswa.2014.06.029

Huang, H.-Z., Tong, X., Zuo, M.J., 2004. Posbist fault tree analysis of coherent systems. Reliab. Eng. Syst. Saf. 84,
141–148. doi:10.1016/j.ress.2003.11.002

Huang, H.Z., Zuo, M.J., Sun, Z.Q., 2006. Bayesian reliability analysis for fuzzy lifetime data. Fuzzy Sets Syst. 157,
1674–1686. doi:10.1016/j.polymdegradstab.2005.12.004

Ishikawa, A., Amagasa, M., Shiga, T., Tomizawa, G., Tatsuta, R., Mieno, H., 1993. The max-min Delphi method
and fuzzy Delphi method via fuzzy integration. Fuzzy Sets Syst. 55, 241–253. doi:10.1016/0165-
0114(93)90251-C

Jensen, F. V., Nielsen, T.D., 2007. Bayesian Networks and Decision Graphs, Journal of Physics A: Mathematical
and Theoretical. doi:10.1007/978-0-387-68282-2

John, A., Yang, Z., Riahi, R., Wang, J., 2016. A risk assessment approach to improve the resilience of a seaport
system using Bayesian networks. Ocean Eng. 111, 136–147. doi:10.1016/j.oceaneng.2015.10.048

Kabir, S., 2017. An overview of Fault Tree Analysis and its application in model based dependability analysis.
Expert Syst. Appl. 77, 114–135. doi:10.1016/j.eswa.2017.01.058

Kabir, S., Walker, M., Papadopoulos, Y., Rüde, E., Securius, P., 2016. Fuzzy temporal fault tree analysis of dynamic
systems. Int. J. Approx. Reason. 77, 20–37. doi:10.1016/j.ijar.2016.05.006

Kabir, S., Walker, M., Papadopoulos, Y., 2014. Reliability Analysis of Dynamic Systems by Translating Temporal
Fault Trees into Bayesian Networks, in: Ortmeier, F., Rauzy, A. (Eds.), Model-Based Safety and Assessment,
Lecture Notes in Computer Science. Springer International Publishing, Cham, pp. 96–109. doi:10.1007/978-3-
319-12214-4

Khakzad, N., 2015. Application of dynamic Bayesian network to risk analysis of domino effects in chemical
infrastructures. Reliab. Eng. Syst. Saf. 138, 263–272. doi:http://dx.doi.org/10.1016/j.ress.2015.02.007

Khakzad, N., Khan, F., Amyotte, P., 2013a. Quantitative risk analysis of offshore drilling operations: A Bayesian
approach. Saf. Sci. 57, 108–117. doi:10.1016/j.ssci.2013.01.022

Khakzad, N., Khan, F., Amyotte, P., 2013b. Dynamic safety analysis of process systems by mapping bow-tie into
Bayesian network. Process Saf. Environ. Prot. 91, 46–53. doi:10.1016/j.psep.2012.01.005

Khakzad, N., Khan, F., Amyotte, P., 2011. Safety analysis in process facilities: Comparison of fault tree and
Bayesian network approaches. Reliab. Eng. Syst. Saf. 96, 925–932. doi:10.1016/j.ress.2011.03.012
Khan, F., 2001. Maximum credible accident scenario for realistic and reliable risk assessment. Chem. Eng. Prog. 97,
55-67.

Khan, F., Haddara, M., Krishnasamy, L., 2008. A new methodology for risk-based availability analysis. IEEE Trans.
Reliab. 57, 103–112. doi:10.1109/TR.2007.911248

Khan, F., & Abbasi, S. A., 1999a. MAXCRED- a new software package for rapid risk assessment in chemical
process industries. Environ. Model. Software, 14, 11–25.

Khan, F., & Abbasi, S. A., 1999b. PROFAT: a user-friendly systemfor probabilistic fault tree analysis. Process Saf.
Prog., 18, 42–49.

Khan, F., Abbasi, S.A., 2000. Analytical simulation and PROFAT II: A new methodology and a computer
automated tool for fault tree analysis in chemical process industries. J. Hazard. Mater. 75, 1–27. doi:10.1016/S0304-
3894(00)00169-2

Khan,F., Haddara, M. R., 2003. Risk-based maintenance (RBM): a quantitative approach for maintenance/inspection
scheduling and planning. J. Loss Prev. Process Ind. 16, 561-573.

Khan, F., Haddara, M.R., 2004. Risk-based maintenance of ethylene oxide production facilities. J. Hazard. Mater.
108, 147–159. doi:10.1016/j.jhazmat.2004.01.011

Khan, F., Husain, T., Abbasi, S.A., 2002. Design and evaluation of safety measures using a newly proposed
methodology “SCAP.” J. Loss Prev. Process Ind. 15, 129–146. doi:10.1016/S0950-4230(01)00026-2

Lavasani, S.M., Ramzali, N., Sabzalipour, F., Akyuz, E., 2015a. Utilisation of Fuzzy Fault Tree Analysis (FFTA)
for quantified risk analysis of leakage in abandoned oil and natural-gas wells. Ocean Eng. 108, 729–737.
doi:10.1016/j.oceaneng.2015.09.008

Lavasani, S.M., Zendegani, A., Celik, M., 2015b. An extension to Fuzzy Fault Tree Analysis (FFTA) application in
petrochemical process industry. Process Saf. Environ. Prot. 93, 75–88. doi:10.1016/j.psep.2014.05.001

Leu, S.-S., Chang, C.-M., 2013. Bayesian-network-based safety risk assessment for steel construction projects.
Accid. Anal. Prev. 54, 122–33. doi:10.1016/j.aap.2013.02.019

Lewis, E.E., 1996. Introduction to reliability engineering, 2nd ed. J. Wiley, New York.

Liu, Y., Fan, Z.P., Yuan, Y., Li, H., 2014. A FTA-based method for risk decision-making in emergency response.
Comput. Oper. Res. 42, 49–57. doi:10.1016/j.cor.2012.08.015

Lowen, R., 1996. Fuzzy Set Theory : Basic Concepts, Techniques and Bibliography. Springer Netherlands.

Mardani, A., Jusoh, A., Zavadskas, E.K., 2015. Fuzzy multiple criteria decision-making techniques and applications
- Two decades review from 1994 to 2014. Expert Syst. Appl. 42, 4126–4148. doi:10.1016/j.eswa.2015.01.003

Markowski, A.S., Mannan, M.S., 2008. Fuzzy risk matrix. J. Hazard. Mater. 159, 152–157.
doi:10.1016/j.jhazmat.2008.03.055

Marvin, H.J.P., Bouzembrak, Y., Janssen, E.M., van der Zande, M., Murphy, F., Sheehan, B., Mullins, M.,
Bouwmeester, H., 2017. Application of Bayesian Networks for hazard ranking of Nanomaterials to Support
Human Health Risk Assessment. Nanotoxicology 5390, 1–34. doi:10.1080/17435390.2016.1278481

Mesiar, R., 2007. Fuzzy set approach to the utility, preference relations, and aggregation operators. Eur. J. Oper.
Res. 176, 414–422. doi:10.1016/j.ejor.2005.07.016

Miri Lavasani, M.R., Wang, J., Yang, Z., Finlay, J., 2011. Application of fuzzy fault tree analysis on oil and gas
offshore pipelines. Int. J. Mar. Sci. Eng 1, 29–42.
Modarres, M., 2006. Risk analysis in engineering : techniques, tools, and trends. Taylor & Francis.

Mohsendokht, M., 2017. Risk assessment of uranium hexafluoride release from a uranium conversion facility by
using a fuzzy approach. J. Loss Prev. Process Ind. 45, 217–228. doi:10.1016/j.jlp.2017.01.004

Musharraf, M., Smith, J., Khan, F., Veitch, B., MacKinnon, S., 2016. Assessing offshore emergency evacuation
behavior in a virtual environment using a Bayesian Network approach. Reliab. Eng. Syst. Saf. 152, 28–37.
doi:10.1016/j.ress.2016.02.001

Omidvari, M., Lavasani, S.M.R., Mirza, S., 2014. Presenting of failure probability assessment pattern by FTA in
Fuzzy logic (case study: Distillation tower unit of oil refinery process). J. Chem. Heal. Saf. 21, 14–22.
doi:10.1016/j.jchas.2014.06.003

Onisawa, T., 1990. An application of fuzzy concepts to modelling of reliability analysis. Fuzzy Sets Syst. 37, 267–
286. doi:10.1016/0165-0114(90)90026-3

OREDA, 2002. Offshore Reliability Data Handbook, 4. ed. ed. Trondheim.

Pedrycz, W., 1994. Why triangular membership functions? Fuzzy Sets Syst. 64, 21–30. doi:10.1016/0165-
0114(94)90003-5

Pereira, J.C., Fragoso, M.D., Todorov, M.G., 2016. Risk Assessment using Bayesian Belief Networks and Analytic
Hierarchy Process applicable to Jet Engine High Pressure Turbine Assembly. IFAC-PapersOnLine 49, 133–
138. doi:10.1016/j.ifacol.2016.07.563

Preyssl, C., 1995. Safety risk assessment and management-the ESA approach. Reliab. Eng. Syst. Saf. 49, 303–309.
doi:10.1016/0951-8320(95)00047-6

Purba, J.H., Lu, J., Zhang, G., Pedrycz, W., 2014. A fuzzy reliability assessment of basic events of fault trees
through qualitative data processing. Fuzzy Sets Syst. 243, 50–69. doi:10.1016/j.fss.2013.06.009

Ramzali, N., Lavasani, M.R.M., Ghodousi, J., 2015. Safety barriers analysis of offshore drilling system by
employing Fuzzy event tree analysis. Saf. Sci. 78, 49–59. doi:10.1016/j.ssci.2015.04.004

Rausand, M., 2014. Reliability of safety-critical systems.

Rausand, M., 2011. Risk assessment : theory, methods, and applications. J. Wiley & Sons.

Rausand, M., Hoyland, A., 2004. System Reliability Theory: Models, Statistical Methods, and Applications.
doi:10.1109/WESCON.1996.554026

Ross, T.J., 2009. Fuzzy Logic with Engineering Applications. doi:10.1002/9781119994374

Saaty, T.L., Ozdemir, M.S., 2003. Why the magic number seven plus or minus two. Math. Comput. Model. 38, 233–
244. doi:10.1016/S0895-7177(03)90083-5

Shi, L., Shuai, J., Xu, K., 2014. Fuzzy fault tree assessment based on improved AHP for fire and explosion accidents
for steel oil storage tanks. J. Hazard. Mater. 278, 529–538. doi:10.1016/j.jhazmat.2014.06.034

Taveau, J., 2010. Risk assessment and land-use planning regulations in France following the AZF disaster. J. Loss
Prev. Process Ind. 23, 813–823. doi:10.1016/j.jlp.2010.04.003

Trucco, P., Cagno, E., Ruggeri, F., Grande, O., 2008. A Bayesian Belief Network modelling of organisational
factors in risk analysis: A case study in maritime transportation. Reliab. Eng. Syst. Saf. 93, 823–834.
doi:10.1016/j.ress.2007.03.035

Van Der Borst, M., Schoonakker, H., 2001. An overview of PSA importance measures. Reliab. Eng. Syst. Saf. 72,
241–245. doi:10.1016/S0951-8320(01)00007-2
Vesely, W.E., Stamatelatos, M., Dugan, J., Fragola, J., Minarick, J., Railsback, J., 2002. Fault tree handbook with
aerospace applications, NASA office of safety and mission assurance, Washington DC.

Wang, D., Zhang, P., Chen, L., 2013. Fuzzy fault tree analysis for fire and explosion of crude oil tanks. J. Loss Prev.
Process Ind. 26, 1390–1398. doi:10.1016/j.jlp.2013.08.022

Wang, Y., Teague, T., West, H., Mannan, S., 2002. A new algorithm for computer-aided fault tree synthesis. J. Loss
Prev. Process Ind. 15, 265–277. doi:10.1016/S0950-4230(02)00011-6

Wu, W.-S., Yang, C.-F., Chang, J.-C., Château, P.-A., Chang, Y.-C., 2015. Risk assessment by integrating
interpretive structural modeling and Bayesian network, case of offshore pipeline project. Reliab. Eng. Syst.
Saf. 142, 515–524. doi:10.1016/j.ress.2015.06.013

Yan, F., Xu, K., Yao, X., Li, Y., 2016. Fuzzy Bayesian Network-Bow-Tie Analysis of Gas Leakage during Biomass
Gasification. PLoS One 11, e0160045. doi:10.1371/journal.pone.0160045

Yazdi, M., 2017. The Application of Bow-Tie Method in Hydrogen Sulfide Risk Management Using Layer of
Protection Analysis (LOPA). J. Fail. Anal. Prev. 17, 291–303. doi:10.1007/s11668-017-0247-x

Yazdi, M., Daneshvar, S., Setareh, H., 2017a. An extension to Fuzzy Developed Failure Mode and Effects Analysis
( FDFMEA ) application for aircraft landing system. Saf. Sci. 98, 113–123. doi:10.1016/j.ssci.2017.06.009

Yazdi, M., Nikfar, F., Nasrabadi, M., 2017b. Failure probability analysis by employing fuzzy fault tree analysis. Int.
J. Syst. Assur. Eng. Manag. 1–17. doi:10.1007/s13198-017-0583-y

Yeo, C.T., Bhandari, J., Abbassi, R., Garaniya, V., Chai, S., Shomali, B., 2016. Dynamic risk analysis of offloading
process in floating liquefied natural gas (FLNG) platform using Bayesian Network. J. Loss Prev. Process Ind.
41, 259–269. doi:10.1016/j.jlp.2016.04.002

Yuan, Z., Khakzad, N., Khan, F., Amyotte, P., 2015. Risk analysis of dust explosion scenarios using bayesian
networks. Risk Anal. 35, 278–291. doi:10.1111/risa.12283

Zarei, E., Azadeh, A., Khakzad, N., Aliabadi, M.M., Mohammadfam, I., 2017. Dynamic safety assessment of
natural gas stations using Bayesian network. J. Hazard. Mater. 321, 830–840.
doi:10.1016/j.jhazmat.2016.09.074

Zhang, K., Duan, M., Luo, X., Hou, G., 2017. A fuzzy risk matrix method and its application to the installation
operation of subsea collet connector. J. Loss Prev. Process Ind. 45, 147–159. doi:10.1016/j.jlp.2016.11.014
 1. Collecting process
information

2. Selecting worst-case
Hazard Analysis
event (TE)

3. Preparing FT diagram

FT construction
and Data
Collection
4. Identify Basic Events (BEs)
of FT

5. Obtaining Failure Rates of 6. Rating state: highlighting BE

BEs with known failure rates with unknown failure rate
Using reliability data source

7. Using expert judgment

(Fuzzy AHP)

8. Aggregation procedure
(Max-min method)

9. Computing the failure

probability of each BE
Converting FTs to
Bayesian network
Bayesian
Modelling 10. Computing TE probability

11. Computing importance 13. Perform Diagnostic

measures Analysis
Calculation

14. Update Basic Event Failure

12. Confirming critical nodes
Probability

Fig. 1. The framework of the proposed method

 Very Low Low Fairy low Medium Fairly high High Very High
1

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 2. Fuzzy membership functions

b c
1

𝐹𝑃𝑆𝑅𝑖𝑔ℎ𝑡 (𝑍) 𝑦=𝑥

0.5
𝐹𝑃𝑆𝐿𝑒𝑓𝑡 (𝑍)
𝑦 = 1−𝑥

0
a d
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 3.The schematic computation of right and left FPS

Basic Intermediate Failure Probability

Fault Tree Top Event Logic Gates
events events of BEs

Mapping

Bayesian Intermediate prior probabilities

Root Nodes Leaf Node CPT
Network Nodes of Nodes

Fig. 4. Mapping of FT to BN
Fig. 5. Process flow diagram of the EO plant
Fig. 6. Fault tree for the ethylene transportation line (modified after Khan and Haddara(2004))

Expert capabilities

Job field Experience Education level

Expert 1 Expert 2 Expert 3 Expert 4

Fig. 7. Fuzzy AHP index system of respective expert capabilities

Fig. 8. Bayesian Network of the fault tree in Fig. 6
Table 1. Details of the BEs of FT of Fig. 6 (modified after Khan and Haddara (2004))

FT Tag BEs Description Failure Reference

BE1 Flammable gas detector fail Reliability data source
BE2 Gas out of run Expert judgment
BE3 Inert gas release mechanism failed Expert judgment
BE4 Flame arrestor A failed Expert judgment
BE5 Flame arrestor B failed Expert judgment
BE6 Ignition source present Expert judgment
BE7 Mechanical failure due to corrosion Expert judgment
BE8 Leak from valves (two valves) Reliability data source
BE9 Leak from bends (four bends) Reliability data source
BE10 Leak from joints (10 joints) Reliability data source
BE11 Flow sensor failed Reliability data source
BE12 Pressure sensor failed Reliability data source
BE13 Pipeline chocked Reliability data source
BE14 Valve chocked Reliability data source
BE15 High inlet flow Expert judgment
BE16 High inlet pressure Expert judgment
BE17 Pressure controller/trip failed Expert judgment
BE18 High inlet temperature Expert judgment
BE19 External heat source present Expert judgment
BE20 Side reaction Expert judgment
BE21 Temperature controller/trip failed Expert judgment
BE22 Phase change Expert judgment
BE23 Valves fails open (two valves) Expert judgment
BE24 Corrosion Reliability data source
BE25 Mechanical damage Expert judgment
Table 2. Failure probability of basic events with known failure rates

BEs Tag FP of BEs

BE1 0.005155
BE8 0.000613
BE9 0.018486
BE10 0.000438
BE11 0.028154
BE12 0.009937
BE13 0.003186
BE14 0.001313
BE24 0.005917

Table 3. Experts’ profile and related decision weights

No Job field Experience (years) Education level Weighing scores

Expert 1 Risk assessor 7 BSc 0.249
Expert 2 Process technician 5 BSc 0.126
Expert 3 Process designer 9 MSc 0.495
Expert 4 Safety officer 2 MSc 0.128

Table 4. Experts’ opinions, corresponding fuzzy number and corresponding failure probabilities of BEs

BEs Fuzzy corresponding Defuzzification of

(E1 E2 E3 E4) Corresponding FPs
reference number subjective BEs
BE2 (H L L FH) (0.300,0.400,0.413,0.513) 0.403 0.002391
BE3 (FH FL FL FH) (0.313,0.388,0.512,0.612) 0.357 0.001581
BE4 (M VL VL FH) (0.164,0.263,0.276,0.376) 0.403 0.002391
BE5 (M VL VL FH) (0.164,0.263,0.276,0.376) 0.403 0.002391
BE6 (H VL VL H) (0.264,0.364,0.364,0.464) 0.409 0.002509
BE7 (FL L L FL) (0.138,0.237,0.275,0.375) 0.392 0.002172
BE15 (M L L M) (0.213,0.313,0.313,0.413) 0.409 0.002509
BE16 (M L L M) (0.213,0.313,0.313,0.413) 0.409 0.002509
BE17 (FL FL FL L) (0.187,0.287,0.374,0.473) 0.370 0.001783
BE18 (FL VL VL FL) (0.075,0.175,0.323,0.386) 0.369 0.001764
BE19 (M L L H) (0.251,0.351,0.351,0.451) 0.409 0.002509
BE20 (FL FL FL L) (0.187,0.287,0.374,0.473) 0.370 0.001783
BE21 (FL L L L) (0.125,0.225,0.249,0.349) 0.398 0.002283
BE22 (VL L L L) (0.075,0.17,0.175,0.275) 0.409 0.002509
BE23 (VL FL FL FH) (0.188,0.288,0.363,0.463) 0.375 0.001874
BE25 (L M M L) (0.286,0.386,0.386,0.486) 0.409 0.002509
Table 5. Importance measures of the basic events based on the proposed approach and the analytical approach

Fuzzy Bayesian Approach Analytical Approach

Basic Events
Importance Measure Rank Importance Measure Rank
BE6 61.11595518 1 60.78163054 1
BE1 42.64527082 2 42.40267113 2
BE4 31.40830058 3 31.27437432 3
BE5 31.40830058 4 31.27437432 4
BE11 13.17872795 5 13.26074705 5
BE9 9.075048388 6 9.222933652 6
BE12 5.513108605 7 5.652468777 7
BE7 2.330717673 8 2.409440162 8
BE8 1.697741564 9 1.758326646 9
BE10 1.626812341 10 1.685240449 10
BE25 4.1572E-06 11 4.91542E-06 13
BE24 2.74737E-06 12 2.08435E-06 14
BE2 2.29668E-06 13 0.148964989 11
BE3 1.98417E-06 14 0.128815433 12
BE13 1.05226E-06 15 1.1711E-06 15
BE15 9.46526E-07 16 1.1711E-06 16
BE22 8.30669E-07 17 1.10664E-06 17
BE23 7.43748E-07 18 1.10664E-06 18
BE17 5.49543E-07 19 3.70671E-07 19
BE21 3.85666E-07 20 2.25626E-07 20
BE19 1.50746E-07 21 8.59526E-08 22
BE20 1.32713E-07 22 8.59526E-08 23
BE18 1.32241E-07 23 8.59526E-08 21
BE16 1.17332E-07 24 6.44645E-08 25
BE14 9.42206E-08 25 6.44645E-08 24