Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: Chemical process industries (CPI) are inherently hazardous complex systems where large inventory of
Received 3 May 2019 extremely flammable and explosive chemicals are processed and stored in a highly congested process
Received in revised form area. A reliable safety analysis method plays a significant role to measure risks and to develop preventive
4 October 2019
strategies in process industries. This paper proposed a novel Fuzzy Bayesian Network for dynamic safety
Accepted 5 October 2019
Available online 7 October 2019
analysis of process systems by incorporating Bayesian network (BN) with Fuzzy Best Worst Method
(Fuzzy-BWM). In the proposed approach a comprehensive and in-depth analysis of human and orga-
Handlig editor: Prof. Jiri Jaromir Klemes nizational factors (HOFs) involving in the accident scenario occurrence was also provided by integrating
Human Factor Analysis and Classification System (HFACS) and System Hazard Identification, Prediction
Keywords: and Prevention (SHIPP) methodology into the model. An ethylene storage tank was selected to verify the
Bayesian network applicability of the proposed approach and its application potential. The study also explained a com-
Fuzzy best worst method parison between the results of the proposed Fuzzy-BWM approach with the conventional BN approach
Process industries and a quantitative risk assessment (QRA) conventional technique such as bow-tie (BT). The findings
Safety analysis
revealed the capability of the proposed Fuzzy-BWM approach to provide high reliable results and to
detect risks that using the BT and BN approaches were not identified.
© 2019 Elsevier Ltd. All rights reserved.
1. Introduction methods are two important aspects of risk assessment (Han and
Weng, 2011), widely applied in chemical industries. Despite their
Chemical process industries (CPI) are much prone to cata- popularity as more adaptable and cost-effective tools for decision-
strophic accidents because containing large quantities of extremely making purposes than the qualitative methods (Roy et al., 2014),
combustible and explosive chemicals. The intense operating con- quantitative risk assessment (QRA) conventional methods (e.g. FTA,
ditions, the complex interactions among system components (Roy FMEA, …) have been criticized in presenting the accurate results
et al., 2015), and issues related to human and organizational fail- due to disadvantages such as being static (unable to update the
ures make CPI susceptible to process deviations and failures, which probability of events), using generic failure data (unable to handle
if not properly managed, may escalate small mishaps into cata- data uncertainty), and incapability to consider complex de-
strophic events (Al-Shanini et al., 2014). pendencies between system components (Zarei et al., 2017a).
The reliable risk analysis methodology plays a fundamental role In order to overcome the above mentioned limitations, during
to develop prevention plans and to control the failures during the last years many attempts have been undertaken to develop
process operations in the CPI. The qualitative and quantitative dynamic risk assessment (DRA) approaches (Villa et al., 2016). The
Bayesian networks (BNs) are one of the well accepted methods for
DRA that their superiority over QRA conventional methods has
been well proven in many of the previous studies (Zarei et al.,
* Corresponding author.
E-mail addresses: Jahangiri_m@sums.ac.ir, ja_me59@yahoo.com (M. Jahangiri).
2017a). Recently, in a comparative study on FTA and BN in
https://doi.org/10.1016/j.jclepro.2019.118761
0959-6526/© 2019 Elsevier Ltd. All rights reserved.
2 A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761
process safety analysis, Khakzad et al. (2011) explained various researchers have focused on developing models that is called Fuzzy
modeling features of BN helping to incorporate multi-state vari- Bayesian Networks (FBNs). For instance, Hou and Zhao (2016)
ables, dependent failures, functional uncertainty, and expert developed a new approach by combining a Fuzzy Analytic
opinion which are frequently encountered in safety analysis, but Network Process (FANP) decision-making method and BN to assess
cannot be considered by FTA. Li et al. (2016) developed a risk-based fire and explosion risks of oil wharf handling. Kabir et al. (2016)
accident model established based on a BN to analysis leakage fail- proposed a safety assessment model to evaluate oil and gas pipe-
ure of submarine oil and gas pipelines. Their investigation findings line failure by incorporating fuzzy logic into BN. Shan et al. (2017)
revealed that BN can provide a more case-specific and realistic presented a practical risk assessment method based on a FBN
analysis consequence compared to bow-tie (BT) method, since it approach incorporating BT model, fuzzy logic method and BN to
could consider the common cause failures and conditional de- quantitatively analysis the probability of the failure and corre-
pendency in the accident evolution process. Shan et al. (2017) have sponding consequences related to a natural gas pipeline leakage.
reported that while reasoning approaches in the QRA conventional Yazdi and Kabir (2018) combined fuzzy set theory, expert knowl-
methods such as BT, FTA, and Event Tree Analysis (ETA) are strictly edge, and evidence theory with BN to develop a QRA methodology
one way, BN also can provide a inversely inference or a two-way through FTA for risk assessment of process systems. Zhou et al.
analysis, not only to find the results from the causes, but also to (2018) integrated BN, fuzzy logic theory and Cognitive Reliability
find causes from the results. & Error Analysis Method (CREAM) to develop a quantitative Human
Recent investigations, however, revealed a limited capability of Reliability Analysis (HRA) model for the shipping industry. Zarei
BNs to handle all dimensions of uncertainty. According to Oztekin et al. (2019) developed a FBN framework combining BT model,
and Luxhøj (2008), while there are three dimensions of uncer- expert elicitation and fuzzy theory with BN for safety analysis of a
tainty in real-world applications including vagueness, ambiguity, city gate natural gas station.
and randomness, BNs only address the uncertainty in the form of However, what is important is the fact that most of the
randomness. Furthermore, conventional BN analysis usually re- mentioned previous studies have obtained the probability of the
quires too much precise information and such information is event failure using fuzzy linguistic terms and have calculated the
available as crisp probabilities (e.g., prior probability and condi- fuzzy weights without considering pair-wise comparisons of the
tional probability of node variables) that are very difficult and often events probability that may lead to false or unrealistic fuzzy
challenging to obtain (Kabir et al., 2016). weights. Furthermore, the above mentioned previous models have
Fuzzy Set theory (FST) has been proposed as an appropriate applied conventional methods such as FTA and BT for accident
alternative to introduce the ambiguity (Oztekin and Luxhøj, 2008) scenario modeling. These conventional methods have been mainly
and vagueness (Ren et al., 2009) dimensions of uncertainty to the designed to identify process and technical failures and are not able
classical framework of BNs. Hence, in recent years, some of the alone to well consider role of the other important accident
A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761 3
contributory factors such as human and organizational factors acts (3) unsafe supervision, and (4) organizational influences
(HOFs) which play a signification contribution in accident occur- (Wiegmann and Shappell, 2003). Each level in the HFACS frame-
rence. It seems that integrating other methods, designed specif- work has a set of sub-levels and each sub-level is consisted of the
ically to analyze HOFs, into the mentioned models can provide a classification codes (nano-codes).
clearer and more comprehensive model of the accident occurrence. In this study, the developed HFACS version by Aas (2008) for oil
Considering the important role of HOFs in accident occurrence, and gas (O&G) industry, O&G HFACS, was used to identify and to
70%e90% of accidents in the chemical and process industries (Zarei categorize human and organizational failures leading to the acci-
et al., 2017b), developing safety analysis models providing an dent scenario occurrence. The mentioned O&G HFACS framework is
comprehensive analysis of HOFs leading to the occurrence of acci- a merged version of the developed HFACS from the US Naval Safety
dents is an essential priority to prevent catastrophic accidents in Center and the US DoD which is consisted of four main levels with a
the CPI. Thus, developing new FBN models characterized by a number of the specific sub-levels and nano-codes, considered
powerful capability in reducing uncertainties and a detailed specifically for oil and gas industry. The O&G HFACS has been
analytical property in analyzing human and organizational failures successfully validated and applied in analysis of the Norwegian
is a crucial task in order to present more objective results and to offshore accidents (Aas, 2008).
design corresponding effective preventive safety strategies more
accurately. 2.2. Bayesian networks (BNs)
In this context, the present study aimed to propose a novel FBN
model taking advantage of Human Factor Analysis and Classifica- BNs belong to a family of probabilistic graphical models, called
tion System (HFACS) and System Hazard Identification, Prediction directed acyclic graphs (DAGs) (Nielsen and Jensen, 2009). The
and Prevention (SHIPP) methodology to overcome the limitations network is composed of nodes, directed arcs, and condition prob-
of the previous models on analyzing HOFs in both of the occurrence abilistic tables (CPTs). Nodes representing the random variables,
of accidents and their consequences. Furthermore, in the proposed directed arcs describing conditional dependencies among the cor-
model a new fuzzy Multi-Criteria Decision-Making (MCDM) responding random variables, and CPTs determining the quantita-
method called Fuzzy Best-Worst Method (Fuzzy-BWM) is used to tive dependency strength from one random variable to others. The
remedy uncertainties raised from application of crisp probabilities nodes are including parent and child nodes. Parent nodes acting as
in BN. Fuzzy-BWM (Guo and Zhao, 2017) is an extended method independent variables which have a prior probability distribution,
from a new MCDM approach called Best Worst Method (BWM) and child nodes acting as independent variables that each of them
proposed by Rezaei (2015). It is a comparison-based approach and have a conditional probability distribution in condition of the
has been proven to provide more reliable and consistent results as parent nodes.
compared to other MCDM methods (Rezaei, 2015) applied in the In BNs, the conditional dependencies between variables are
previous models. determined by joint probability distribution (JPD). The relation-
Literature review shows that there is no a study on application ships represented by DAG allow JPD to be specified by the CPTs for
of BNs combined with Fuzzy-BWM in safety analysis of process each node (Zhang et al., 2014). Considering “pa (Ai)" as parent
systems. Considering the mentioned unique properties of Fuzzy- nodes of “Ai” in DAG, the CPTs of “Ai” is denoted by P(Ai|pa(Ai)).
BWM in providing higher reliable results and the robust potential Thus, the JPD a set of variables U ¼ {A1, A2,A3,…, An} can be written
of HFACS and SHIPP in analyzing the HOFs, it is expected that the as Eq. (1). (Nielsen and Jensen, 2009):
proposed model in this research can be an appropriate base to
Yn
overcome the difficulties associated with the lack of an in-depth PðUÞ ¼ PðAijpaðAiÞÞ (1)
i¼1
and comprehensive analysis of HOFs and the further reduction of
a high level of vagueness and uncertainties found in the previous BN is a flexible tool capable to make a two-way inference
models. including both predictive and diagnostic reasoning. The predictive
The rest of this article is arranged as follows. Section 2 dem- reasoning is performed by calculating the prior probability of Ai as
onstrates the used materials and methods in the study. Section 3 Eq. (2):
provides a description of the methodology application. A case X
study is presented on safety analysis of an ethylene storage tank to P ðAiÞ ¼ PðUÞ (2)
explain the feasibility of the proposed approach and its application U\Ai
potential. Through the case study a comparison between the results
Diagnostic reasoning is performed by updating the prior prob-
of the proposed FBN approach with the conventional BN approach
abilities of variables. Prior probabilities of variables are updated
and a QRA conventional technique such as BT has been also pro-
using Bayes theorem when new evidences or information of vari-
vided. In section 4, results and discussion are presented. Section 5 is
ables E are given (Zarei et al., 2017a). Accordingly, the posterior
dedicated to the conclusion.
(updated) probabilities can be calculated as Eq. (3):
makers. Then the weights of criteria are determined by deter- 3. Application of the methodology
mining preference of the best criterion over other criteria and
preference of all criteria on worst criterion. This section describes an application of the proposed method-
Fuzzy-BWM is a extended model of BWM, proposed by Guo and ology for safety analysis of a liquid ethylene storage tank located in
Zhao (2017). In the Fuzzy-BWM, the reference comparisons in BWM a petrochemical plant in southern areas of Iran.
are expressed in Triangular Fuzzy Numbers (TFNs). The nonlinearly Ethylene, C2H4 is an extremely hazardous material that forms
constrained optimization problem is then built to calculate the the major production of petrochemical industries. Ethylene is easily
optimal fuzzy weights of criteria and alternatives. The fuzzy ignited and can undergo a violent chemical reaction at elevated
ranking scores of alternatives can be derived from the fuzzy temperatures. Depending on its physical state, gas or liquid,
weights of alternatives with respect to different criteria multiplied ethylene is generally stored in specific tanks under high or atmo-
by fuzzy weights of the corresponding criteria, and then the crisp spheric pressure. Because of its unique physical features such as a
ranking score of alternatives can be calculated by employing the high vapor pressure, a low flash point and a low explosive limit,
Graded Mean Integration Representation (GMIR) method for even a little release of ethylene from a storage tank can form an
optimal alternative selection (Guo and Zhao, 2017). extremely flammable and explosive vapor cloud.
et al. (2017a), the likelihood and severity of each the identified factor engineering, and process and instrumentation engineers in
scenario was then specified and a qualitative risk analysis was order to identify the potential casual factors and consequences of
conducted by means of the MIL-STD-882E to identify the worst case the accident scenario occurrence.
accident scenario. Based on the risk analysis findings, ethylene As mentioned in the methodology section of this article, HFACS
release due to the leakage from the tank body was selected as the was used to make a comprehensive analysis of HOFs involving in
highest risk scenario (Table 1). the accident scenario occurrence. To this end, we used two re-
sources to identify the main HOFs that may lead to failure of the
4.2. Accident scenario modeling (BT model) desired storage tank. At the first, the reports of near misses and
accidents over five last years were reviewed and the main HOFs
A BT model (Fig. 3) was developed by several meetings to contributed to the tank failure were extracted. Then a compre-
interview a group of the experienced specialists including inspec- hensive interview was conducted to chairmen, managers, site-men,
tion and maintenance engineers, academicians active in human board-men, and operators who working on repair, maintenance,
6 A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761
Table 1
Results of the hazards analysis using HAZOP and MILeSTDe882E techniques.
and management of processes associated with the desired storage required action in a timely manner by operators. In this case, as is
tank. They were asked to describe the main human errors and shown in Fig. 3.h, failure of warning systems (X55 in Fig. 3.h) was
organizational failures leading to tank failure which they had categorized as a nano-code at level of ‘precondition for unsafe act’
experienced over their carrier. Finally, based on the interview and sub-level of ‘environmental factors’, and delayed an necessary
findings and the extracted results from the near misses/accident action (X46 in Fig. 3.h) by operators was categorized as nano-codes
reports, the identified HOFs were categorized into each of the O&G at level of ‘unsafe acts’ and sub-level of ‘error’, respectively. Fig. 3.h
HFACS levels, their corresponding sub-levels and classification depicts the FT model developed from O&G HFACS framework. The
codes (nano-codes) by two academicians who were active in hu- symbols and descriptions related to each of the basic events in the
man factor engineering and familiar to the O&G HFACS approach. developed FT model (the O&G HFACS’s nano-codes), X42-X89, are
In the next step, the O&G HFACS was mapped into the FT part of also presented in the Table 2.
the BT model. In the mapping procedure, the first level of O&G The ET part of the BT model was established based on the
HFACS, unsafe act level, was considered as top event of the FT sequential failure of safety barriers and the potential consequences
model and its two sub-levels including “errors” and “violation” resulting from the occurrence of accident scenario. In order to
were considered as intermediate event. Each other level of O&G integrate HOFs into the ET part of the BT model, safety barriers of
HFACS and its sub-levels were considered as intermediate event the ET model were developed based on the suggested conceptual
and at the lowest level, the O&G HFACS’s nano-codes, were accident model in the SHIPP methodology (Rathnayaka et al., 2011).
considered as basic event of the FT model. For example a human The SHIPP methodology is a new process accident model that uses a
and organizational failure that may cause release of ethylene from sequential modeling approach by applying safety barriers to
the tank by overfilling is failure of warning systems such as failure describe the accident process. The SHIPP’s framework is consisted
of a high level alarm (HLA) in combination with fail to take the of safety barriers, in which human factors and management and
A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761 7
organizational factors serve as a special barrier which affects the that because of the necessity of finding and diagnosing alarm sig-
whole process of accident evolution (Rathnayaka et al., 2011). nals, as well as judging and verifying the leakage occurrence, role of
Fig. 4 shows the developed safety barrier model for the desired the human, management and organization factors is more promi-
ethylene storage tank. nent when early detection and alarm barrier (ED&AB) fails.
The model was set up according to Rathnayaka et al. (2011) by Therefore, an independent barrier called human and organizational
locating six main safety barriers in sequential order together with (H&O) responses was defined and located between the ED&AB and
two additional safety barriers including human factor barrier and automatic emergency shutdown (ESDa) barrier, as shown in Fig. 3.a.
management and organization barrier located common to all of the As can be seen from Fig. 3.a, the failure or function of the safety
main safety barriers to prevent escalation of the leakage event to barriers can lead to different consequences from a simple near miss
catastrophic accident. (C1) to a catastrophic event (C21) which involves undesirable
In this paper, according to Li et al. (2016) and based on judgment outcomes such as catastrophic pool fire, vapor cloud explosion
of safety engineers and human factors specialists, it was concluded (VCE), sever causalities, and catastrophic property damage.
8 A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761
Fig. 3. (continued).
A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761 9
Fig. 3. (continued).
10 A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761
Fig. 3. (continued).
4.3. BN/FBN modeling work, an example on the calculation of fuzzy prior probabilities of
parent nodes of a child node, “external corrosion” for instance (see
4.3.1. Mapping BT into BN framework Fig. 5) is explained in the following section. A step by step
The proposed algorithm by Khakzad et al. (2013) was used to description of the Fuzzy-BWM details for estimating the prior
map BT into the BN framework. The mapping procedure was probability of root events/nodes can be also found in our recent
including directly transform the basic events, intermediate events, study (Rostamabadi et al., 2019).
top event, safety barriers, and consequences of the BT into the
corresponding nodes in BN including root nodes, intermediate Step 1: Determine a set of decision criteria (C1, C2.....Cn).
nodes, pivot node, safety nodes and consequence nodes, respec-
tively. Software GeNIe 2.0 was applied to simulate and run the Considering external corrosion (IE.5) as a child node (see Fig. 5),
desired BN. Fig. 5 depict the established BN framework. In Fig. 5, four its parent nodes including anti-corrosion protective coating
symbols X indicate the root nodes (green circles); symbols IE failure (C1), erosion (C2), poor detection of corrosion (C3), and
indicate the intermediate nodes (purple circles); the pivot node is cathodic protection failure (C4) were determined as the decision
marked by red circle; and the blue and yellow circles marked the criteria (factors leading to external corrosion of the ethylene stor-
safety nodes and consequence nodes, respectively. age tank).
4.3.2. BN/FBN probability estimating Step 2: Determine the best (CB) and the worst (CW) criteria from
In this study, we used both of the crisp probabilities and fuzzy n alternatives by an expert or an expert team.
probabilities to determine prior probabilities of the root nodes and
safety nodes. Crisp probabilities extracted from literature and fail- To do this, a group consisted of five experts with different pro-
ure rate databases such as OREDA (OREDA, 2002) were used in both fessional position, service time, and education level were asked to
of the BT and BN probability estimating while the fuzzy probabili- from the defined criteria in the step 1, determine the most probable
ties obtained from the expert opinions based on Fuzzy-BWM were factor and the least probable factor leading to external corrosion of
used in the FBN probability estimating. The obtained crisp proba- the ethylene storage tank.
bilities of root events/nodes and safety nodes using BT and BN
approaches are presented in Tables 3 and 4, respectively. Step 3: Execute the fuzzy reference comparisons regarding the
In order to describe how Fuzzy-BWM was used in the current best criterion.
A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761 11
Table 2
The symbols and descriptions of human and organizational failures.
Symbol Description
X42 Omitting checklist items when inspect tank and its auxiliary facilities
X43 Ignoring warning alarms related to leakage sensing devices, overfilling/overpressure prevention instrumentation, and etc
X44 Not follow procedures when operate with tank
X45 Misreading gauges and dials of related to the tank pressure or ethylene level
X46 Delaying an necessary action such as long delay in the tank maintenance and inspection schedule
X47 Wrong response to a multitude of alarms based on an incomplete understanding of the actual problem
X48 Failure to set control switch
X49 Poor decision during operation
X50 Inadvertent operation of operator
X51 Poor teamwork
X52 Poor physical fitness of operator to work conditions
X53 Inadequate rest due to working in unscheduled hours or continuous night work
X54 Inadequate procedural/technical knowledge of system to operate with the tank and to quickly identify the specific causes leading to a leakage
X55 Failure of warning/control systems such as leakage detectors, overfilling/overpressure prevention instrumentation, etc.
X56 Unreliable measurement of instrumentation
X57 Workplace incompatible with human, for example visibility restrictions, presence of gases or lack of oxygen
X58 Personal equipment interference
X59 Inadequate instrument such as lack of manual actuators, not existing tank gauging redundancy, inadequate sensory feedback systems, and etc.
X60 Noise inference
X61 Uncomfortable temperature extremes
X62 Mental psychological fatigue of operator
X63 Operator awareness failure
X64 Disqualified medical condition of operator, e.g., drug effects, hangover, fatigue, etc.
X65 Task/informational overload
X66 Fail to replace worn-out auxiliary equipment in a timely manner
X67 Poor personnel management
X68 Inadequate operation management
X69 Fail to find hidden dangerous timely
X70 Fail to report unsafe work practices
X71 Limited recent experience
X72 Fail to provide correct data
X73 Fail to identify risky practices
X74 Inadequate supervisory response to critical situation
X75 Inadequate provided guidance/policy
X76 Authorized unqualified person
X77 Authorized unnecessary hazards
X78 Directing individual to violate regulations
X79 Provided substandard equipment
X80 Fail to provide required training
X81 Lack of equipment/facility resources
X82 Poor informational resources
X83 Inadequate upgraded organizational policies
X84 Poor organizational safety culture
X85 Unclear organizational structure
X86 Inadequate risk management
X87 Poor working permit system
X88 Inadequate maintenance and inspection program
X89 Poor management of change
In this step, the experts were asked to determine their prefer- The fuzzy reference comparisons for the least probable factor
ence of the most probable factor (best criterion) over the other (worst criterion) were also executed, and the linguistic evaluations
factors (criteria) according to linguistic terms presented in Table 5. of one of the experts for the fuzzy preference of the other factors
Table 6 shows the linguistic terms of one of the experts for fuzzy (criteria) over the worst criterion are shown in Table 7.
preferences of the most probable factor (best criterion) over other Then, the fuzzy Others-to-Worst vector was obtained according
factors (criteria). to Table 5 and Eq. (5) as follows:
Then, the fuzzy Best-to-Others vector was obtained according to
Table 5 and Eq. (4) as follows:
For getting the optimal fuzzy weights of all the criteria, the
Step 4. Execute the fuzzy reference comparisons for the worst following nonlinearly constrained optimization problem was built
criterion. according to Eq. (6):
A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761 13
Table 3 Table 6
Crisp probabilities of root events/nodes using BT and BN approaches. Linguistic terms for fuzzy preferences of the best criterion over the other criteria.
criteria (parent nodes of the external corrosion node) including respectively. In this study, these obtained crisp weights were also
anti-corrosion protective coating failure (C1), erosion (C2), poor called as fuzzy possibilities (FPs). Similar calculations were per-
detection of corrosion (C3), and cathodic protection failure (C4) formed to obtain the FPs values of the above mentioned criteria
were calculated as w ~ 1¼ (0.466085, 0.510555, 0.524799); w ~2 ¼ from other experts. Finally, FPs for each criterion was calculated as
(0.095418, 0.170185, 0.170185); w ~ 3 ¼ (0.127639, 0.127639, the following equation:
0.228456); and w ~ 4 ¼ (0.085092, 0.170185, 0.32799), respectively.
The obtained optimal fuzzy weights were then transformed to EPs ¼ GMIRExpert1 WSExpert1 þ GMIRExpert2Þ þ
crisp weights by the Graded Mean Integration Representation
(GMIR), Rða~j Þ, as Eq. (7): þ GMIRExpert n WSExpert n
(10)
li þ 4mi þ ui
~j ¼
R a (7) Considering Eq. (10), the FPs values of criteria C1, C2, C3, and C4
6
were calculated as 0.39687, 0.21697, 0.17003, and 0.20614,
Finally, the crisp weights obtained for C1, C2, C3, and C4 were respectively. At last, these FPs values for each criterion were con-
calculated as 0.505517, 0.157724, 0.144441, and 0.182303, verted to fuzzy prior probability (FPr) by using the following
respectively. equations proposed by Onisawa (1988):
The above crisp weights were then multiplied by a weighting
8
factor representing the relative quality of the response of the ex- >
< 1 ; FPss0
perts. The weighting factor was calculated based on previous FPr ¼ 10k (11)
studies (Renjith et al., 2010; Zarei et al., 2019) by estimating weight >
:
0; FPs ¼ 0
scores each expert divided by weight scores all experts, as
following equations:
1
ð1 FPsÞ 3
k ¼ 2:301 (12)
Weight score of expert i FPs
Weight factor of expert i ¼ Pn
i¼1 Weight score of experts
Where K is a constant value.
(8)
Based on the Eq. (11) and Eq. (12), FPr for criteria C1, C2, C3, and
C4 were calculated as 2.26E-03, 3.00E-04, 1.30E-04, and 2.50E-04,
Weight score of Expert i ¼ Score of PP þ Score of E þ Score of respectively. It is noteworthy that criteria C1, C2, C3, and C4 are
EL þ Score of A (9) according to symbols of X10, X9, X11, and X8 (see Fig. 5), respec-
tively. The obtained FPs, K, and FPr values for other root nodes and
Where PP is professional position of expert i, E is experience time of safety nodes are presented in Table 10 and Table 11, respectively.
expert i, EL is education level of expert i, and A is age of expert of
expert i.
Score of each expert was calculated according to Table 8, as 4.3.3. BN/FBN inference
proposed by Renjith et al. (2010). Bayesian inference was conducted using predictive and diag-
Table 9 reveals corresponding weighting factors for the selected nostic reasoning. The results of predictive reasoning of the accident
experts. scenario occurrence and its main contributing factors using BT, BN,
Considering the weighting factor of the Expert No.4 (0.203; see and FBN approaches are presented in Table 12 (2, 3, and 4th col-
Table 9), the final crisp weights of criteria including C1, C2, C3, and umns) and Fig. 6.
C4 were calculated as 0.102817, 0.032079, 0.029378 and 0.37079, A comparison on the obtained probability values using the BT,
Table 8
Experts’ scores (adopted from Renjith et al. (2010)).
Table 9
Experts’ weighting score.
Table 10 well treated in the form of common failure causes using BN, a
FPs, K, and FPr values for corresponding root node/nano-code. unique capability that BT does not have (Li et al., 2016; Zarei et al.,
Root Events Fuzzy probabilities Root Events Fuzzy probabilities 2017b). On the other hand, the observed significant differences
FPs K FPr FPs K FPr
between the results of the FBN with BT and BN is due to the fact that
the BN (and the BT) approach can only deal with crisp sets (Ren
X1 0.301 3.044 9.00E-04 X46 0.207 3.599 2.51E-04
et al., 2009), whereas using the FBN approach the domain experts
X2 0.169 3.907 1.30E-04 X47 0.168 3.919 1.20E-04
X3 0.284 3.131 7.40E-04 X48 0.183 3.781 1.65E-04 can define ranges rather than crisp values under “vague” conditions
X4 0.141 4.198 6.33E-05 X49 0.202 3.634 2.32E-04 (Zhang et al., 2016). Therefore, while BT and BN approaches suffers
X5 0.093 4.902 1.25E-04 X50 0.101 4.768 2.70E-05 from uncertainty and insufficient data on the failure data, FBN
X6 0.351 2.821 1.50E-03 X51 0.329 2.916 1.21E-03
approach makes it possible for decision makers to have more
X7 0.638 1.903 1.24E-02 X52 0.170 3.902 1.25E-04
X8 0.206 3.606 2.50E-04 X53 0.308 3.012 9.42E-04 detailed and more accurate data on the failure probability.
X9 0.216 3.529 3.00E-04 X54 0.182 3.795 1.60E-04 In this study, the probability of tank failure using FBN (8.24E-02)
X10 0.396 2.645 2.26E-03 X55 0.225 3.470 3.40E-04 was more than triple, when compared with the BN (2.43E-02) (see
X11 0.170 3.903 1.30E-04 X56 0.199 3.655 2.20E-04 Table 12). These results are almost similar to findings of a recent
X12 0.307 3.017 9.55E-04 X57 0.101 4.755 1.75E-05
research conducted by Zarei et al. (2019) which found a fourfold
X13 0.257 3.272 5.34E-04 X58 0.100 4.769 1.70E-05
X14 0.173 3.870 1.35E-04 X59 0.181 3.801 1.60 E04 increase in the calculated probability values using FBN than BN on
X15 0.321 2.950 1.12E-03 X60 0.080 5.171 6.74E-06 safety analysis of a Iranian natural gas station.
X16 0.215 3.536 2.91E-04 X61 0.099 4.797 1.59E-05 According to the Table 12, the predictive reasoning results reveal
X17 0.207 3.594 2.54E-04 X62 0.238 3.387 4.10E-04
that near misses (C1) were the most likely consequences resulting
X18 0.543 2.170 6.76E-03 X63 0.303 3.033 9.26E-04
X19 0.446 2.473 3.36E-03 X64 0.109 4.629 2.34E-05
from the scenario occurrence using BT, BN, and FBN approaches.
X20 0.221 3.497 3.18E-04 X65 0.308 3.011 9.74E-04 However, comparing the obtained probability values (see 2, 3, and
X21 0.268 3.211 6.14E-04 X66 0.211 3.569 2.69E-04 4th columns), it can be concluded that FBN represents a higher
X22 0.148 4.117 7.63E-05 X67 0.202 3.632 2.38E-04 probability value for near misses (C1) and a lower probability value
X23 0.155 4.041 9.38E-05 X68 0.212 3.560 2.75E-04
for all other consequences (C12eC14, and C17) than BT and BN
X24 0.157 4.027 9.38E-05 X69 0.277 3.164 6.84E-04
X25 0.176 3.848 1.41E-04 X70 0.086 2.348 4.49E-03 approaches. This may be due to a high level of the implemented
X26 0.087 5.033 9.26E-06 X71 o.454 2.446 3.57E-03 safety preventive measures in place that by means of the experts’
X27 0.057 5.857 1.39E-06 X72 0.535 2.193 6.41E-03 opinion well reflected in FBN approach, whereas the probability of
X28 0.093 4.905 1.24E-05 X73 0.474 2.381 4.15E-03
consequences in the BT and BN are based on generic databases such
X29 0.113 4.560 2.75E-05 X74 0.260 3.258 5.51E-04
X30 0.352 2.817 1.52E-03 X75 0.255 3.289 5.14E-04
as OREDA which are constant over time and by means of them it is
X31 0.268 3.216 6.07E-04 X76 0.208 3.593 2.55E-04 not possible to consider the recent improvements in the reliability
X32 0.269 3.208 6.19E-04 X77 0.534 2.196 6.35E-03 of the system components and the real conditions of the system
X33 0.098 4.809 1.55E-05 X78 0.247 3.334 4.63E-04 safety (Zarei et al., 2019).
X34 0.124 4.406 3.92E-05 X79 0.184 3.772 1.69E-04
Results of the diagnostic reasoning revealed that in both the FBN
X35 0.095 2.115 7.67E-03 X80 0.353 2.813 1.53E-03
X36 0.097 2.098 7.97E-03 X81 0.137 4.243 5.71E-05 and the BN approaches, HOFs were the most contribution factor to
X37 0.101 2.071 8.47E-03 X82 0.313 2.986 1.03E-03 the ethylene storage tank failure while different results was found
X38 0.208 1.560 2.75E-02 X83 0.294 3.077 8.37E-04 regarding mechanical failures and process failures (Table 12). In
X39 0.105 2.039 9.12E-03 X84 0.427 2.537 2.90E-03 other word, using BN approach, process failures were identified as
X40 0.257 1.424 3.76E-02 X85 0.267 3.217 6.06E-04
X41 0.548 2.156 6.70E-03 X86 0.337 2.879 1.32E-03
the second highest contributing factor whereas by means of FBN
X42 0.319 2.961 1.09E-03 X87 0.164 3.951 1.12E-04 approach, mechanical failures were the second highest contrib-
X43 0.162 3.975 1.05E-04 X88 0.247 3.334 4.63E-04 uting factor, a result that has been also found in the Zarei et al.
X44 0.508 2.275 5.30E-03 X89 0.239 3.380 4.17E-04 (2019) study.
X45 0.127 4.374 4.22E-05 e e e
In this research, the obtained higher probability values for me-
chanical failures in FBN approch can be explained by the nature of
the desired industry where various extremely corrosive and reac-
Table 11
tive materials are present and most of the process equipment
FPs, K, and FPr values of safety barriers/nodes.
operate under rigorous operating conditions, including extremely
Safety barrier Fuzzy probabilities low temperature and high pressure conditions that may because of
FPs K FPr or accelerate various mechanical failure modes such as fracture,
ED&AB 0.064 5.60 2.49E-06
creep, corrosion, wear, fatigue and so on. On the other hand, pro-
H&O response barrier 0.234 3.415 3.84E-04 cess operations in petrochemical plants are performed under rigid
ESDa 0.142 4.185 5.62E-05 controlled conditions. Various high reliable control systems are
IIB 0.122 4.442 3.61E-05 used to make sure industrial processes are carried out within
ESDm 0.147 4.127 7.46E-05
specified limits and with as little deviation as possible. Therefore, in
DIB 0.128 4.351 4.44E-05
ERB 0.149 4.106 7.83E-05 such a situation, process operations are often carried out safely and
a lower rate of the process failures than mechanical failures is
expected.
BN and FBN reveals that while there is a little different in the
4.3.4. Criticality analysis
calculated prior probabilities values by BN than those by the BT, the
As a crucial step to provide an effective safety decision-making,
calculated prior probabilities values by FBN are significantly
it is necessary to determine the most critical basic (root) events
different from those calculated by the BN and by the BT. The reason
contributing to the occurrence of the top event (Kabir et al., 2016).
for the different probability values using BT and BN is related to the
To this end, the determination of the posterior probabilities of basic
conditional dependencies among variables (nodes). As can be seen
events given the scenario occurrence has been used in most of the
from Fig. 5, in the developed BN model, most of the root nodes have
previous studies. However, new investigations (Zarei et al., 2017b,
conditional dependencies to each other. Such dependencies can be
2019) have explored that focusing on the posterior probabilities for
16 A. Rostamabadi et al. / Journal of Cleaner Production 244 (2020) 118761
Table 12
The Results of the predictive and diagnostic reasoning.
Fig. 6. Predictive reasoning for the accident scenario and its main contributing factors
using BT, BN, and FBN.
p ðXi Þ q ðXi Þ
RoV ðXi Þ ¼ (13)
q ðXi Þ
5. Conclusion
Declaration of competing interest Congress of International Council of the Aeronautical Sciences, Anchorage,
Alaska.
Rathnayaka, S., Khan, F., Amyotte, P., 2011. SHIPP methodology: predictive accident
The authors declare that there is no conflict of interests modeling approach. Part I: methodology and model description. Process Saf.
regarding the publication of this paper. Environ. Prot. 89, 151e164. https://doi.org/10.1016/j.psep.2011.01.002.
Ren, J., Jenkinson, I., Wang, J., Xu, D., Yang, J., 2009. An offshore risk analysis method
using fuzzy Bayesian network. J. Offshore Mech. Arct. Eng. 131, 041101 https://
Acknowledgement doi.org/10.1115/1.3124123.
Renjith, V., Madhu, G., Nayagam, V.L.G., Bhasi, A., 2010. Two-dimensional fuzzy fault
This article was extracted from the thesis written by Akbar tree analysis for chlorine release from a chlor-alkali industry using expert
elicitation. J. Hazard Mater. 183, 103e110. https://doi.org/10.1016/
Rostamabadi, a PhD student of Occupational Health Engineering j.jhazmat.2010.06.116.
and was financially supported by the Shiraz University of Medical Rezaei, J., 2015. Best-worst multi-criteria decision-making method. Omega 53,
Sciences, Iran (Grant No.1396-01-04-15176). 49e57. https://doi.org/10.1016/j.omega.2014.11.009.
Rostamabadi, A., Jahangiri, M., Zarei, E., Kamalinia, M., Banaee, S., Samaei, M.R.,
2019. A novel fuzzy bayesian network-HFACS (FBN-HFACS) model for analyzing
References human and organizational factors (HOFs) in process accidents. Process Saf.
Environ. Prot. https://doi.org/10.1016/j.psep.2019.08.012.
Aas, A.L., 2008. The human factors assessment and classification system (HFACS) for Roy, A., Srivastava, P., Sinha, S., 2014. Risk and reliability assessment in chemical
the oil & gas industry. In: International Petroleum Technology Conference. In- process industries using Bayesian methods. Rev. Chem. Eng. 30, 479e499.
ternational Petroleum Technology Conference. https://doi.org/10.1515/revce-2013-0043.
Al-Shanini, A., Ahmad, A., Khan, F., 2014. Accident modelling and analysis in process Roy, A., Srivastava, P., Sinha, S., 2015. Dynamic failure assessment of an ammonia
industries. J. Loss Prev. Proc. 32, 319e334. https://doi.org/10.1016/ storage unit: a case study. Process Saf. Environ. Prot. 94, 385e401. https://
j.jlp.2014.09.016. doi.org/10.1016/j.psep.2014.09.004.
Guo, S., Zhao, H., 2017. Fuzzy best-worst multi-criteria decision-making method and Shan, X., Liu, K., Sun, P.L., 2017. Risk analysis on leakage failure of natural gas
its applications. Knowl. Based Syst. 121, 23e31. https://doi.org/10.1016/ pipelines by fuzzy Bayesian network with a bow-tie model. Sci. Program. 1e11
j.knosys.2017.01.010. https://doi.org/10.1155/2017/3639524.
Han, Z., Weng, W., 2011. Comparison study on qualitative and quantitative risk Villa, V., Paltrinieri, N., Khan, F., Cozzani, V., 2016. Towards dynamic risk analysis: a
assessment methods for urban natural gas pipeline network. J. Hazard Mater. review of the risk assessment approach and its limitations in the chemical
189, 509e518. https://doi.org/10.1016/j.jhazmat.2011.02.067. process industry. Saf. Sci. 89, 77e93. https://doi.org/10.1016/j.ssci.2016.06.002.
Hou, Z., Zhao, P., 2016. Based on fuzzy bayesian network of oil wharf handling risk Wang, Y.F., Roohi, S.F., Hu, X.M., Xie, M., 2011. Investigations of human and orga-
assessment. Math. Probl. Eng. 2016, 1e10. https://doi.org/10.1155/2016/ nizational factors in hazardous vapor accidents. J. Hazard Mater. 191 (1e3),
6532691. 69e82. https://doi.org/10.1016/j.jhazmat.2011.04.040.
Kabir, G., Sadiq, R., Tesfamariam, S., 2016. A fuzzy Bayesian belief network for safety Wiegmann, D.A., Shappell, S.A., 2003. A Human Error Approach to Aviation Accident
assessment of oil and gas pipelines. Struct. Infrastruct. Eng. 12, 874e889. Analysis: the Human Factors Analysis and Classification System. Ashgate,
https://doi.org/10.1080/15732479.2015.1053093. Aldershot, Great Britain. https://trove.nla.gov.au/version/28337285.
Khakzad, N., Khan, F., Amyotte, P., 2011. Safety analysis in process facilities: com- Yazdi, M., Kabir, S., 2018. Fuzzy evidence theory and Bayesian networks for process
parison of fault tree and Bayesian network approaches. Reliab. Eng. Syst. Saf. 96, systems risk analysis. Hum. Ecol. Risk. Assess. 24, 1e30. https://doi.org/10.1080/
925e932. https://doi.org/10.1016/j.ress.2011.03.012. 10807039.2018.1493679.
Khakzad, N., Khan, F., Amyotte, P., 2013. Dynamic safety analysis of process systems Zarei, E., Azadeh, A., Aliabadi, M.M., Mohammadfam, I., 2017a. Dynamic safety risk
by mapping bow-tie into Bayesian network. Process Saf. Environ. Prot. 91, modeling of process systems using bayesian network. Process Saf. Prog. 36,
46e53. https://doi.org/10.1016/j.psep.2012.01.005. 399e407. https://doi.org/10.1002/prs.11889.
Kidam, K., Hurme, M., 2013. Method for identifying contributors to chemical pro- Zarei, E., Azadeh, A., Khakzad, N., Aliabadi, M.M., Mohammadfam, I., 2017b. Dy-
cess accidents. Process Saf. Environ. Prot. 91, 367e377. https://doi.org/10.1016/ namic safety assessment of natural gas stations using Bayesian network.
j.psep.2012.08.002. J. Hazard Mater. 321, 830e840. https://doi.org/10.1016/j.jhazmat.2016.09.074.
Li, X., Chen, G., Zhu, H., 2016. Quantitative risk analysis on leakage failure of sub- Zarei, E., Khakzad, N., Cozzani, V., Reniers, G., 2019. Safety analysis of process sys-
marine oil and gas pipelines using Bayesian network. Process Saf. Environ. Prot. tems using Fuzzy Bayesian Network (FBN). J. Loss Prev. Proc. 57, 7e16. https://
103, 163e173. https://doi.org/10.1016/j.psep.2016.06.006. doi.org/10.1016/j.jlp.2018.10.011.
Nielsen, T.D., Jensen, F.V., 2009. Bayesian Networks and Decision Graphs, second ed. Zhang, L., Wu, X., Skibniewski, M.J., Zhong, J., Lu, Y., 2014. Bayesian-network-based
Springer Science & Business Media, New York. safety risk analysis in construction projects. Eng. Syst. Saf. 131, 29e39. https://
Onisawa, T., 1988. An approach to human reliability in man-machine systems using doi.org/10.1016/j.ress.2014.06.006.
error possibility. Fuzzy Sets Syst. 27, 87e103. https://doi.org/10.1016/0165-0114 Zhang, L., Wu, X., Qin, Y., Skibniewski, M.J., Liu, W., 2016. Towards a fuzzy bayesian
(88)90140-6. network based approach for safety risk analysis of tunnel-induced pipeline
OREDA, 2002. Offshore Reliability Data Handbook, fourth ed. DNV, Trondheim, damage. Risk Anal. 36, 278e301. https://doi.org/10.1111/risa.12448.
Norway. Zhou, Q., Wong, Y.D., Loh, H.S., Yuen, K.F., 2018. A fuzzy and Bayesian network
Oztekin, A., Luxhøj, J.T., 2008. Hazard, Safety Risk, and Uncertainty Modeling of the CREAM model for human reliability analysiseThe case of tanker shipping. Saf.
Integration of Unmanned Aircraft Systems into the National Airspace. 26th Sci. 105, 149e157. https://doi.org/10.1016/j.ssci.2018.02.011.