See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/276181146

Risk and reliability assessment in chemical process industries using Bayesian

methods

Article in Reviews in Chemical Engineering · October 2014

DOI: 10.1515/revce-2013-0043

CITATIONS READS

13 3,525

3 authors, including:

Prashant Srivastava Shishir Sinha

Department of Science & Technology, Ministry of Science and Technology Govt. of… Indian Institute of Technology Roorkee
10 PUBLICATIONS 73 CITATIONS 74 PUBLICATIONS 1,297 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Prashant Srivastava on 24 April 2017.

The user has requested enhancement of the downloaded file.

 Rev Chem Eng 2014; 30(5): 479–499

Arnab Roy, Prashant Srivastava and Shishir Sinha*

Risk and reliability assessment in chemical

process industries using Bayesian methods
Abstract: Chemical process industries (CPI) are usually literature review of the application of Bayesian methods
home to a large number of complex systems and compo- through Bayesian parameter estimation techniques and
nents required for various operations involving hazardous Bayesian updating procedures in process industries. Both
chemicals. The intense operating conditions and com- these techniques have been extensively used in various
plex interactions between the systems make the chemi- aspects of risk analysis, which are very pertinent in CPI.
cal plants vulnerable to accidents. Quotidian incidents The purpose of the study is to produce an effective refer-
and mishaps can lead to catastrophic incidents. Thus, ence guide for scholars interested in applying Bayesian
the area of risk assessment and reliability analysis in CPI techniques to risk and reliability assessment in CPI.
has been of much interest to the research community. The
complexity of processes in CPI demands a risk assessment Keywords: data analysis and integration; decision making;
tool that can adapt itself to the dynamic environment failure dependency; risk analysis; subjective knowledge.
and can efficiently model the functional and sequential
dependencies between the components and the effects DOI 10.1515/revce-2013-0043
of external factors, component degradation, and vari- Received December 11, 2013; accepted June 13, 2014; previously pub-
ation in operating conditions. The risk assessment tool lished online August 5, 2014
must have applicability during the operational lifetime
of the system to serve as a platform for decision making
and risk management. The unavailability of empirical Abbreviations
data for some variables is another pertinent issue in risk
analysis and reliability assessment in CPI. Analysts often AI artificial intelligence
have to work with subjective information such as expert ASP accident sequence precursor
opinion. Bayesian statistical methods based on the Bayes BN Bayesian network
theorem are considered by many to be an effective tool CCF common cause failure
to address the above-mentioned issues. These methods CPI chemical process industry
are based on the subjective interpretation of probability CPT conditional probability table
that helps to model the epistemic uncertainties and eas- CTBN continuous time Bayesian network
ily propagates them through complex system models. DAG directed acyclic graph
The methods provide a formal systematic way to incorpo- DBN dynamic Bayesian network
rate subjective information into calculations. The inher- DCS distributed control system
ent updating property accoutres these methods with the DDT diagnostic decision tree
ability to deal with real-time changes. The opponents, DIF diagnostic importance factor
however, point out that the methods produce high over- DOOBN dynamic object-oriented Bayesian network
confidence and randomness in computed answers. In the DRA dynamic risk assessment
last two decades, an increased interest can be seen in the DTBN discrete time Bayesian network
research community toward the use of Bayesian methods EB empirical Bayes
in risk assessment. This paper presents a comprehensive ESD emergency shutdown
EVT extreme value theorem
ET event tree
*Corresponding author: Shishir Sinha, Department of FB fully Bayes
Chemical Engineering, Indian Institute of Technology, Roorkee, FMEA failure mode effects analysis
Roorkee-247667, India, e-mail: [email protected]
FFT fast Fourier transform
Arnab Roy: Department of Chemical Engineering, Indian Institute
of Technology, Roorkee, Roorkee-247667, India
FT fault tree
Prashant Srivastava: Department of CoEDMM, Indian Institute FTA fault tree analysis
of Technology, Roorkee, Roorkee-247667, India HAZOP hazard and operability study

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
480 A. Roy et al.: Risk assessment in process industries

HazID hazard identification study qualitative and quantitative properties. As Kirchsteiger

HRA human risk assessment (1997) pointed out, risk analysis is concerned with the
HMM hidden Markov models reconstruction of reality to identify the causes of accidents
LNG liquid nitrogen gas either in a quantitative or in a qualitative way. On the
LOPA layer of protection analysis qualitative front, methods such as hazard identification
MCAA maximum credible accident analysis study (HazID), failure mode effects analysis (FMEA), and
MCMC Monte Carlo Markov chains What-if checklist are well established. A detailed study of
MHDS multilevel heterogeneous data sets these techniques applied in chemical industries can be
MLE maximum likelihood estimation found in the work of Khan and Abbasi (1998).
MPE most probable event The quantitative aspect of risk assessment deals with
MSM Markov switch models the calculation of occurrence probability of accident sce-
nPEB nonparametric empirical Bayes narios and the reliability analyses of concerned systems.
OOBN object-oriented Bayesian network Quantitative risk assessment (QRA), as it is known in
ORA operational risk assessment chemical industries, is more cost-effective than its quali-
PEB parametric empirical Bayes tative deterministic counterpart and provides a more
PRA probabilistic risk analysis adaptable tool for decision-making purposes (Kirchsteiger
QRA quantitative risk assessment 1997). In this context, fault tree (FT) analysis (FTA) method
RBD reliability block diagram is most widely used because of its user-friendly properties
and its ability to present an intuitive representation of
the complex system (Boudali and Dugan 2005). However,
it fails to analyze the functional and sequential depend-
1 Introduction encies between the components (Khakzad et al. 2013b).
Modern online FT tools are, however, able to induct real-
The rise in demand of various chemicals accompanied by time information into cut set calculations. Ferdous et al.
the development of technology has led to the construction (2007, 2009) presented computer-aided FT techniques that
of huge and complex chemical process plants in the last can produce reliable results through repeated analysis.
two decades. Process industries nowadays are a complex Consequently, a number of modeling techniques have
technical environment home to a large number of sophis- been developed for assessing the reliability of complex
ticated systems and components. They usually contain a systems. Dynamic FTs (DFTs) (Bucci et al. 2008, Chiac-
diverse inventory of hazardous chemicals and complex chio et al. 2011) and Markov chains are notable mentions.
systems required to perform process operations such as Swaminathan and Smidts (1999) proposed a technique
storage, separation, reaction, and compression. Operat- to model dynamic characteristics of process operations
ing conditions are intense and complex interactions exist using event sequence diagrams. Luo and Hu (2013) used
between the systems, subsystems, and the components, the event sequence diagram model to identify the criti-
which make the plants vulnerable to accidents (Khan cal events and the evolution of risk in a system. Markov
and Abbasi 1998). A brief survey of the history of chemi- chains are extensively used for dependability analysis
cal process industries (CPI) will reveal that subsystem or and as an efficient tool for modeling dynamic behavior.
single-component failure has often led to accidents of However, their solutions often become intractable and
catastrophic consequence and huge financial loss such thus erroneous. Khan and Abbasi (2001) proposed a risk
as the Bhopal tragedy (Eckerman 2005), the vapor cloud assessment procedure named operational risk assessment
explosion at Flixborugh (Crowl and Louvar 1990), the fer- (ORA) for chemical industries, combining both quantita-
tilizer plant explosion at Toulouse, France, and the BP tive and qualitative analyses, using different tools. The
Texas City oil refinery accident (Kalantarnia et al. 2010). same authors combine the maximum credible accident
These accidents highlighted the need for a robust and analysis (MCAA) tool (Khan and Abbasi 2002) for worst-
well-defined risk assessment methodology applicable to case scenario analysis and the PROFAT II tool coded in
process industries, which has been a much devoted area Visual C+ (Khan and Abbasi 2000) for FT quantification to
of research in recent years (Knegtering and Pasman 2013). form a safety assessment procedure called safety credible
The reliability analysis of systems and risk assess- accident probabilistic (SCAP) FTA (Khan et al. 2001, 2002).
ment provides a useful tool for decision-making and risk The complexity of processes in chemical industries
management studies. Over the years, a large variety of requires a risk assessment tool that can adapt itself to the
risk assessment tools have been developed having both dynamics of the environment and properly portray the

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 481

temporal aspects (Labeau et al. 2000). These include mod- can only be represented formally by the subjective defini-
eling of sequential failures, dependencies between com- tion of probability. Bayesian techniques prove themselves
ponents, degradations, effects of exogenous constraints, suitable for the task. A browse through literature concern-
and changes in operating conditions (Weber et al. 2012). ing risk assessment and reliability analysis will reveal
To account for these variations, which occur almost reg- the popularity of these methods among the research fra-
ularly during the lifetime of the system, the tool should ternity. Researchers have incorporated the methods to
be able to induct new information (Khakzad et al. 2013c). modify and eliminate the disadvantages posed by tradi-
This makes it applicable not only at the design stage but tional risk analysis tools. Putting aside the debates over
also during the operational lifetime of the system, which the application of Bayesian techniques in risk analysis,
in turn forms a platform for real-time decision making and a bibliographical review is presented of studies where
risk management (Pasman and Rogers 2013). researchers have used the methods to their advantage.
The lack of formal data in conducting risk analysis The rest of the paper is organized as follows. A short
in process industries is quite a hindrance and has proven recapitulation of the Bayes theorem and the statistical
to be the biggest handicap in conducting a proper safety methods associated with it are given in Section 2. Section
assessment. Safety analysts have to use a large amount 2.1 highlights the different advantages of these methods.
of subjective information in their studies. Thus, the pre- Section 3 briefly explains the Bayesian statistical methods
ferred tool must be able to use this subjective information chiefly concerned with risk analysis in process industries.
formally to yield objective results. Section 4 provides a bibliographical review of application
Bayesian statistical methods based on the Bayes of Bayesian statistical methods in various aspects of risk
theorem have provided reliability engineers and safety analysis in separate subsections. Finally, Section 5 pre-
analysts with a tool for addressing the above-mentioned sents the conclusions arrived from the conducted wide
requirements considerably. The theorem is presented and literature survey.
explained in Section 2. The application of the Bayesian sta-
tistical method in risk analysis and assessment has been
a matter of much dispute among safety analysts as well
as among statisticians. The methods provide a formal sys-
2 Bayesian statistical methods
tematic way to incorporate subjective information into cal-
Bayesian analysis is a statistical framework for reasoning
culations. The inherent updating property qualifies these
and decision making under uncertainty. The backbone of
methods to integrate real-time changes in the process into
this approach is the Bayes theorem (Lindley 1965), which
the calculations. The critics, however, point out that the
states that the conditional probability of event E1 given
methods produce high overconfidence and randomness
event E2 has occurred can be written as
in computed answers. Ferson (2005), in his detailed report
submitted to the French Ministry for geological and mining P( E 1 ) P( E 2 | E 1 )
P( E 1 | E 2 ) = (1)
research, provided a thorough insight to the different P( E 2 )
aspects of application of Bayesian statistical methods in risk
analysis. This document elucidates the implications of the where P(E1) is the prior probability, which expresses the
usage of Bayesian statistical methods in risk analysis study analyst’s knowledge about E1 prior to observation of event
while drawing comparison with the two-dimensional Monte E2; P(E2|E1) is the likelihood function, which is the condi-
Carlo method and is a must read for interested scholars. tional probability of occurrence of event E2 given event E1
Over the years, the uses of Bayesian statistics in risk is observed; and P(E2) is the probability of event E2. It is
analysis and reliability studies have been generally dedi- called the normalization factor and can be estimated by
cated to satisfy the following purposes: the total probability law.
– to present a platform for decision making, which in The prior probability P(E1) is updated to give the pos-
turn aids risk and safety management; terior probability P(E1|E2) with the help of the likelihood
– to estimate the distributions of reliability indices, function P(E2|E1). It is the likelihood function through
such as failure rate and mean time to failure; and which new data are incorporated into the calculations,
– to parameterize and incorporate the input informa- which update the prior information. The law is based on
tion available. the axioms of total probability theorem. As pointed out by
Ferson (2005), the Bayes theorem is controversy free and
As pointed out by Siu and Kelly (1998), epistemic uncer- is widely accepted. It is the application of this law in risk
tainties, which plague most of the risk assessment studies, analysis study that is often debated.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
482 A. Roy et al.: Risk assessment in process industries

In context to the failure probability estimation, the Decision making: Decision making is integral to Bayes-
theorem can be written in the following form: ian methods. This is because Bayesian analysis follows
the axioms of decision theory by expressing uncertainties
f ( x | Data ) ∝ f ( x ) g ( Data | x ) (2)
in terms of probability. Many of the methods proposed by
different researches as expressed in the following bib-
where f(x) is the probability distribution function express-
liographical review can be implemented during the plant
ing the uncertainty in failure probability of the system,
operation timeline, thus helping in risk and safety man-
g(Data|x) is the likelihood function, and f(x|Data) is the
agement. Although the explicit application of Bayesian
improved probability distribution called posterior distri-
methods in decision making in the case of risk analysis
bution. The likelihood function is defined as the distribu-
is sparse in literature, the application of the methods in
tion of identically and independently distributed (i.i.d.)
the same has been much appreciated in various studies as
data collected from industries conditional upon failure
will be seen later in the literature review.
probability x. The normalizing factor, f(Data), which
Handling of uncertainty: Bayesian statisticians argue
should appear in the denominator of Equation (2) in the
that these methods are the most suitable for incorporat-
case of an equality sign, is the integral of the product of
ing uncertainty in calculations. In the case of risk analysis
the prior and the likelihood with respect to Data. The
and reliability studies in process industries, uncertainty
value of this factor is such that the posterior probability
undoubtedly presents a large drawback to a sound assess-
integrates to unity.
ment. Bayesian statistics are thus very much well suited
At the heart of the Bayesian approach lays the subjective
for the purpose. Meel and Seider (2006) mentioned that
definition of probability. This definition is opposed to the
Bayesian methods provide a smooth pathway for the prop-
classic frequentists’ definition of probability. Debates and
agation of uncertainty through complex system models.
comparisons between these two opposed definitions have
Flexibility: The Bayes theorem is based on the subjec-
been going on for a long time in the research world (Jaynes
tive definition of probability. This permits the incorpora-
and Bretthorst 2003). As the paper is more concerned with
tion of data from various sources rather than only from
the application of Bayesian statistics in risk analysis in
empirical sources. Plant-specific data are quite rare in
process industries and not mathematical debate, we will
process industries barring a few sources such as the Guide-
not encroach upon the subject further. Interested readers
lines for Process Equipment Reliability Data published
can go through a wide array of literature (Berger 1985, Vose
by the Center for Chemical Process Safety (CCPS) of the
2000) on the subject to satiate their curiosity.
American Institute of Chemical Engineers (AIChE) (1989),
To put in simple terms, proponents of the subjective
the Utility Requirements Document prepared by the Elec-
definition believe that probability can be applied to any
tric Power Research Institute (EPRI) (1995), the Standard
event that can be termed as uncertain. It measures the
500 published by the Institute of Electrical and Electron-
degree of belief of the analyst regarding the occurrence
ics Engineers (IEEE) (1984), and the Offshore Reliability
of the event. On the contrary, the opposing school of
Data (1997). Analysts have to rely on other sources of infor-
thought, the frequentists, maintain that probability can
mation such as expert judgment. Bayesian methods can
only be applied to situations that have equally likely pos-
not only incorporate such subjective information but also,
sible outcomes.
as researchers have shown, can work with different data
sampling distributions. Brief descriptions of these works
are mentioned in the subsequent sections. These methods
2.1 Advantages of Bayesian statistical also have the flexibility of data mining where the analyst
methods can study the generic data sets and can generally improve
their hypothesis from the observations.
The proponents of Bayesian statistical methods highlight Mainly due to these advantages, researchers and
various advantages to establish its superiority over tra- safety analysts have repeatedly used different Bayes-
ditional frequentist methods. Berry (1996) compared the ian statistical methods for risk assessment studies. The
two methods extensively and explicitly favors the former applications are in various fields of risk analysis. The next
for its various advantages. Ferson (2005) also delineated section briefly describes the statistical methods employed
the different advantages of Bayesian methods. The funda- by the researchers. It should be kept in mind that the
mental ones among others posed by Bayesian statistical section only provides an overview and is insufficient for
methods in risk analysis and reliability study in process proper understanding of the subject. Relevant literatures
industries are briefly listed below. are mentioned to aid those who are interested.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 483

3 Overview of Bayesian statistical are frequent (Bier and Yi 1995b). Thus, reliability engineers

methods in risk analysis have to rely on data from various levels of the systems and
also from different sources. The ability to infuse data of
different kinds, both statistical and nonstatistical, led to
Although the application of Bayesian methods in reliabil-
the use of Bayesian parameter estimation techniques in
ity and risk analysis are diverse, safety analysts chiefly
safety analysis of systems.
employ the following methodologies in the different
According to Siu and Kelly (1998), the Bayesian
aspects of safety analysis studies (Røed and Aven 2009).
parameter estimation method follows the following steps:
The reliability of a system is usually expressed as a
– identification of the parameter to be estimated,
function of reliability indices, which are various para-
– determination of the prior distribution,
meters such as failure rate, failure probability, and time to
– determination of an appropriate likelihood function
failures. The objective of Bayesian analysis is the determi-
based on available information, and
nation of these parameters with the help of the available
– derivation of the posterior distribution using the
data. Intuitively, this approach is named Bayesian param-
Bayes theorem.
eter estimation. A thorough study of this methodology and
its implications in risk analysis applications can be found
The prior distribution is determined using two alternative
in the work of Siu and Kelly (1998).
methods – the fully Bayes (FB) and the empirical Bayes
The second approach involves the future estimation of
(EB) estimation procedures. Interested readers can find
the defined parameters. Bayesian methods are employed
detailed explanations of both these methods in the work
for the prediction of some performance measures of these
of Robert (2001).
items based on available information. It mainly involves
the determination of the posterior density function and
is commonly referred as Bayesian updating method. This
3.1.1 FB approach
approach is more suitable for decision-making purposes
(Aven 2003).
The ability of Bayesian statistical methods to incorporate
These above-mentioned methods are closely related
different types of data encouraged researchers to employ
in view that they can be performed in a sequential order.
these techniques for reliability analysis of complex
Meel and Seider (2006) combined the methods to form
systems. The term FB can be interpreted in different ways,
the dynamic failure assessment methodology with the
but the most accepted definition is the Bayesian analysis
use of accident sequence precursor (ASP) data. Khakzad
in which the prior is specified. This permits the prior to be
et al. (2013b) modified dynamic risk assessment (DRA)
less influenced by the scarcely available and ambiguous
using bowties to depict failure scenarios and updating
data. To address the problems of data unavailability, vari-
the parameters with Bayesian networks (BNs). As will be
ability, and uncertainty, researchers were prompted to use
shown in the bibliographical review, researchers have
FB methods through a hierarchical model for the estima-
used these methodologies in varied facets of risk analysis.
tion of reliability parameters.
The application of FB methods hampered initially
due to the absence of computational tools. The advent of
3.1 Bayesian parameter estimation Monte Carlo Markov chain (MCMC) computational method
in the 1990s solved the computational problem of Bayes-
Quantitative methods for risk analysis have been preva- ian methods (Casella 1992, Chib and Greenberg 1995). The
lent in the nuclear industry from the early 1980s (Cooke freely available software package named Bayesian infer-
and Goossens 1990, Bier and Yi 1995a), where it was ence using Gibbs sampling (BUGS) contributed to the
known as probabilistic risk analysis (PRA). Early works rising popularity of Bayesian methods in different fields.
on the quantification of risks in chemical industries are More recent advances in this software, the WinBUGS
by Papazoglou et al. (1992) and Kafka (1996). Papazoglou (Lunn et al. 2000) and the open source OpenBUGS, have
et al. (1992) presented a systematic methodology for con- expanded the usage of the MCMC computational method
ducting a QRA study in CPI. (Kelly and Smith 2009).
The PRA and QRA methods are handicapped with the A supplementary work to Siu and Kelly (1998) and
unavailability of system failure data, as failure of whole Kelly and Smith (2009) gives a detailed study on the appli-
systems are quite rare events in both these industries. At cation of MCMC sampling-based WinBUGS software in
the same time, subsystem and component-level failures Bayesian methods employed to various aspects of risk and

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
484 A. Roy et al.: Risk assessment in process industries

reliability assessment. Atwood and Kelly (2009) demon- inhabited (Pasman et al. 2009). Industrial areas and
strated the software’s capability to model both time- and industrial complexes are constructed, which are home to
demand-based common cause failure (CCF) rates. a large number of varied industries. Major accidents in the
past have shown the risks posed by the process industries
on the concerned population. The growth in industrializa-
3.1.2 EB tion and population over the years has thus increased the
risks posed by the probable accidents in these industries
EB is an alternative method to the traditionally used FB (Khan and Abbasi 1998).
hierarchical model for generating informative prior dis- Risk analysis is a tool for representing risks posed by
tributions (Berger 1985). The method is generally treated the activities of the process plants (Pasman et al. 2009).
as being an incomplete Bayesian approach because it According to Khan and Abbasi (1998), risk assessment
uses classic techniques for comparing the prior distribu- methodology in CPI is mainly concerned with
tions with the available data. Prior parameters are gen- – construction of accident forecasting methods (Zheng
erated from analyzing the data using methods such as and Liu 2009) and
maximum likelihood estimation (MLE), moment match- – analysis of consequences of likely accidents.
ing, and maximum entropy methods (Siu and Kelly 1998).
EB estimators can be classified into parametric EB (PEB) The potential hazard scenarios are identified using
and nonparametric EB (nPEB) procedures. The former methods such as hazard and operability study (HAZOP),
requires specifying a functional form for the prior distri- FMEA, and What-if analysis, and QRA methods are
bution. Sarhan (2003) gave a detailed commentary on the applied for the quantification of the risks involved due
application of nPEB in estimating failure rates and prior to the hazard scenarios. Quantification of risks gener-
parameter estimation. ally implies the construction of the likelihood of rare
The development of MCMC sampling methods accel- accident scenarios. This obviously requires the inclu-
erated the use of FB methods in different fields of reliabil- sion of engineering expert judgment into the analysis.
ity analysis in process plants. In comparison, the use of EB Standard statistical methods thus prove themselves
methods has been limited. The works mentioned in this inefficient for this purpose (Aven and Kvaløy 2002).
paper necessarily do not apply the methods in reliability Bayesian techniques are suitable for this purpose, as
analysis of systems in process plants, but the problems they employ subjective probability (Siu and Kelly 1998,
they address are similar to those faced by reliability and Aven and Kvaløy 2002).
risk analysts in CPI. Rather than focusing on the estimation of statistical
A specific branch of safety research community parameters such as failure rate and probabilities, Apeland
believes that the EB succeeds FB methods in reliabil- et al. (2002) proposed a fully subjective approach based on
ity assessment, as it uses empirical priors rather than Bayesian inference to predict the occurrence probability
subjective priors used in the latter (Quigley et al. 2011). of events and the magnitude of consequences. Aven and
Initially, EB methods were applied because of their Kvaløy (2002) provided a conceptual description of imple-
simplicity in application and estimation. Vaurio (1987) menting Bayesian concepts in risk analysis and illustrate
used the approach called parametric robust EB (PREB) the above-mentioned predictive approach.
using data from various plants to construct a plant-spe- Bayesian updating such as the proposed predictive
cific informative prior. The one-stage method is illus- method involves the process of updating the initial prob-
trated using a γ prior Poisson likelihood distribution ability estimates by incorporating operational data in the
and moment matching methods. Lahiri and Park (1992) Bayes theorem. Bayesian updating improves the preexist-
estimated the mean residual life of an operating compo- ing knowledge about the system and gives an actual oper-
nent by both PEB and nPEB procedures using a Dirichlet ational scenario of the system (Targoutzidis 2010). Lindley
process prior. (1965), on the contrary, criticized Bayesian updating for
yielding overestimated results in the case of complex
systems. However, the effectiveness of the method is
3.2 Risk analysis using Bayesian updating widely accepted, and major implementations can be ref-
and predicting approach erenced. The predictive and updating techniques provide
a sound basis for decision making under uncertainty.
Logistical and economic reasons often prompt the devel- Bayesian updating has been mainly performed with the
opment of process industries in locations that are densely help of ASP data and BNs.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 485

3.2.1 ASP data that the inclusion of ASPs is a must in databases to gener-
ate major event frequency estimates without large error.
Catastrophic events are usually rare events in process
industries defined by “low probability, high consequence
occurrence.” Abnormal events, on the contrary, occur fre- 3.2.2 BNs
quently and are indicators of system malfunctions. ASPs
are defined as such low consequence events that indi- BN or Bayesian belief net is an artificial intelligence (AI)
cate an increased likelihood of such an accident. “Near- tool for representing uncertainty. A BN can be defined
miss” situations and “incidents” are generally defined as a directed acyclic graph (DAG). The nodes represent
as accident precursors. These events are usually plant random variables, both continuous and discrete, and are
specific in nature, and their usage in the Bayesian updat- connected by directed arcs that express the dependen-
ing approach reduces the uncertainty bands around the cies and the causal relationships between the variables
parameters-expectation values. (Jensen 2001). As an illustration, a simple BN structure of
Phimister et al. (2003) gave a detailed and compre- the variables X = (X1, X2, X3) is given in Figure 1.
hensive study on near-miss situations in context to risk The node having an outgoing arc to another specific
analysis in chemical industries. The near-miss region was node is called the root or “parent” of that node. This second
identified as the bottom part of the safety pyramid-con- node in turn named as the “descendent” or the “child” of
taining events with no alarming consequences but having the parent node. Nodes having no incoming arcs, that is,
the potential to escalate to an accident. Oktem et al. without any parents, are “root nodes” and those without
(2010) broadened the definition by introducing events children are “leaf nodes.” In the figure, X3 is the leaf node
such as unobserved problems and unsafe conditions as and X1 and X2 are its parents. The conditional independ-
near-miss situations. Applications of near-miss manage- ence assumption underlining the qualitative part can be
ment in various industries such as CPI and in financial explained by the graph theoretic concept of “d-separation”
sectors have been illustrated by Kleindorfer et al. (2012). (Pearl 1988). The assertion states that, given the probabili-
Saleh et al. (2013) presented a thorough review of accident ties of all the “parents” of a variable X, its probability is
precursors, near misses, and warning signals in context to independent of other “child nodes” in the structure. Thus,
safety analysis. the “root nodes” have marginal probability distributions
The analysis of ASP data started in the 1980s in the associated with them and the other nodes have conditional
nuclear industry to estimate the frequency of rare events probability tables (CPT) associated with them.
(Cooke and Goossens 1990, Bier 1993). Mosleh et al. Taking the d-separation and the conditional inde-
(1987), Cooke and Goossens (1990), and Bier (1993) pre- pendence assumptions in consideration, the joint prob-
sented various statistical methods to estimate rare event ability distribution of all the variables can be expressed as
frequency using ASP data. It was argued that, in tra- n
ditional PRA methodologies, as the dependencies are P [ X 1 , X 2 , …, Xn ] = ∏ i=1 ≡ P [ Xi | Parent ( Xi )] (3)
explicitly modeled by the analyst, the results are in some
way biased. As ASP data reflect the information about the This factorization of the joint probability distribu-
system during the operating period of the plant, it implic- tion is the main factor for making BNs suitable for mod-
itly contains the interdependencies between the system eling complex systems (Weber et al. 2012). As stated by
and the components involved in the precursor event. Khakzad et al. (2011), due to the concept of d-separation,
The use of ASP data for risk analysis in CPI initially
suffered mainly due to the lack of databases. However,
major accidents in Flixborough, Seveso initiated the col-
lection of different databases both in the United States X2
[Nuclear Regulatory Commission (NRC); http://www.nrc.
gov/] and in Europe [Major Accident Reporting System
(eMARS)].
The shortcomings of conventional risk estimators in X3

estimating major accident frequency are demonstrated

X1
by Kirchsteiger (1997). The author performed a sensitiv-
ity analysis of a database reporting only major accidents
using Bayesian calculations and ASP data. It is concluded Figure 1 A simple BN.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
486 A. Roy et al.: Risk assessment in process industries

BNs represent a local model in place of global model rep- – Data integration into hierarchical tools,
resentation of Markov chains. – Data analysis and assessment of reliability indices,
BNs are principally used in risk and reliability anal- – Decision-making studies,
ysis as a tool to provide inference by updating the prior – Modeling of failure dependency, and
probabilities with new information applying the under- – DRA methodology.
lined Bayes theorem paradigm (Khakzad et al. 2011). Diag-
nostic inferences that can be performed by BNs include
(Bobbio et al. 2001) 4.1 Data integration into hierarchical tools
– posterior marginal probability distribution given each
component, Hierarchical tools such as FTs and reliability block dia-
– posterior joint probability distribution given a set of grams (RBDs) are very much prevalent in risk analysis in
components, and process industries. They can depict the process systems
– posterior joint probability distribution given some intuitively and can portray the causal relationships
evidence nodes. leading to an undesired event (Ferdous et al. 2007).
The use of Bayesian statistical methods in risk and
The advantages of BN modeling include the incorpora- reliability assessment started with the integration of data
tion of data from various sources (Langseth and Portinale into different levels of hierarchical tolls such as FTs and
2007), performing both prediction and diagnosis, ability RBDs. The first major contributions are from Mastran
to model interactions of human, technical, and opera- (1976), Mastran and Singpurwalla (1976), and Martz and
tional factors (Weber et al. 2012), and providing an effi- Wailer (1990). Mastran and Singpurwalla (1976) used
cient tool for decision making (Pasman and Rogers 2012). RBDs and an approximate Bayesian technique to estimate
A host of BN software tools are available online for ana- the reliability of a coherent system using both component-
lysts, such as HUGIN (http://www.hugin.com), BayesiaLab and system-level pass/fail test data. Martz and Wailer
(http://www.bayesia.com/en/products/bayesialab.php), (1990) and Martz and Almond (1997) proposed a bottom-
and AgenaRisk (http://www.AgenaRisk.com) (Agena Ltd. up approach to determine the posterior distribution of
2007). Murphy (2004) gave a detailed list of such tools. reliability of a system by combing available data with
Langseth and Portinale (2007) provided a detailed failure rate distributions of the subsystems in an RBD. The
study on the application of BNs in reliability analysis. A moments of the prior distributions of the components are
comprehensive bibliographical review of the use of BNs combined with the binomial data and propagated upward
in the areas of maintainability, risk analysis, and depend- to calculate the system-level reliability.
ability is given by Weber et al. (2012). Khakzad et al. (2011) The unavailability of computational tools initially
presented a comparison study between FTs and BNs in the led to the consideration of unwanted approximations,
field of safety analysis and accident modeling. The BNs which led to biased results. Due to the computational
prove to be a superior technique due to its inherent prob- constraints, a variation of hierarchical FB technique, the
ability updating and uncertainty reducing capabilities. two-stage Bayes method, was used for assessing reliabil-
The next section presents the bibliographical review. ity data. The two-stage method proposed by Kaplan (1985)
Although most of the works have not been performed in was implemented in the German ZEDB (Bunea et al. 2005)
CPI, their inherent characteristics make them relevant to project for collecting reliability data from nuclear facili-
risk assessment studies in process industries. The authors’ ties. The Swedish T-Book (T-Book 1987) project employed
purpose is to produce a substantial reference guide for the method proposed by Pörn (1990) on the same frame-
reliability and risk assessment in CPI. work. Hofer (1999) undermined the two-stage method for
having an incorrect chance mechanism in the likelihood
model and proposed an alternative approach. However,
the proposed procedure has also come under criticism
4 Bibliographical review (Hamada et al. 2004, Vaurio 2005).
In the area of data variability, the first major study
The review presents a wide array of works of application came from Johnson et al. (2003). The authors presented a
of the above-mentioned methods in varied fields of risk reliability model based on an FB treatment of a hierarchi-
analysis. As reviewed by the authors, Bayesian methods cal model, which is able to constantly adopt independent
have been chiefly applied in following aspects of risk data available at multiple levels of an RBD. The proposed
analysis: methodology is carried forward by Hamada et al. (2004)

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 487

for application in FTs. The nonoverlapping basic-level sources and proposed remedies in context to the system
and higher-level data in FTs are simultaneously combined model presented by Anderson-Cook et al. (2008). The
by using the minimum cut sets and the total probability aspect of best allocation of resources, the discrepancies
theorem. A generic algorithm is presented, which high- between component- and system-based approach, and
lights the optimum levels in an FT for use of additional the identification of possible weaknesses and omissions
resources to reduce uncertainty in top event probability. in the system model are discussed. Wilson et al. (2006)
In both these works, the data considered are binomial also echoed the significance and the possibilities of incor-
data. This method is generalized by Graves et al. (2008), porating MHDS.
who replace the β prior distribution by the Dirichlet prior Guo and Wilson (2010) generalized the reliability
distribution and the corresponding likelihood function assessment approaches modeled by Wilson et al. (2006)
from binomial distribution to multinomial distribution and Hamada et al. (2008). The FT of a complex system is
function. considered and data of different types are introduced at
The nonoverlapping constraint of Hamada et al. both basic and nonbasic event levels. Logit function is used
(2004) is overcome by Graves et al. (2008), which extends and the degradation model is based on the work of Wilson
the methodology for incorporating “overlapping” data, et al. (2006). Bayesian melding approach is adopted to
that is, simultaneous data in multiple levels of an FT. Each combine prior information given at different FT levels.
subsystem is represented by an FT involving lower-level Recently, Peng et al. (2013) presented a robust frame-
data. The FT for the whole system is generated by the con- work for reliability analysis using MHDS. The various
junction of all lower-level FTs associated with simultane- assumptions in system structures and data types in the
ously available lower-level data. Data integration is done above-mentioned works are overcome by the authors.
using the Bayesian approach and the posterior component Reliability indices such as time to failure are formulated
reliabilities are sampled using variable-at-a-time Metropo- for decision-making purposes, thus generalizing the
lis Hastings algorithm. The algorithm, which is an MCMC applicability of the method.
method, is used to draw out random samples from a prob- Wilson et al. (2011) utilized the Bayesian data integra-
ability distribution. The algorithm is very useful when tion methodologies developed over the years to perform a
direct sampling becomes difficult. Complimentary work case study of reliability analysis of a complex system. The
on overlapping data involving multiple demands on the case study illustrates the flexibility of Bayesian methods
system is presented by Jackson and Mosleh (2012), who for the integration of MHDS for reliability and risk analy-
used Bayesian inference methods in hierarchical logic sis studies. The authors point out that the ability of these
model such as FTs and RBDs to assess component reliabil- methods to quantify uncertainty at different levels aids
ity parameters with continuous variables. The likelihood the decision-making process. A summary of main contri-
function for the Bayesian inference is generated from the butions in this field is given in Table 1.
overlapping data. The authors illustrate the usefulness of
the additional information obtained from using overlap-
ping data. 4.2 Data analysis and assessment
Anderson-Cook et al. (2008) integrated binomial of reliability indices
pass/fail and lifetime data to assess component-level reli-
ability parameters using a series system model. System In risk assessment and reliability analysis of complex
reliability is assessed using the component-level reliabil- systems in process industries, the main stumbling block
ity estimations done through an FB approach. Hamada faced by analysts was the variability of available data.
et al. (2008) presented a methodology using Bayesian The development of Bayesian computational tools made
approach to induct pass/fail, lifetime, and degradation possible the application fully hierarchical Bayesian
data into system reliability analysis. Reliability assess- tactics in analyzing various data types. Sarhan (1999)
ment using lifetime data sets have also been investigated gave a Bayes parameter estimation of a general hazard
by Reese et al. (2011). model using Type II censored data. Simulations done
Anderson-Cook (2009) further stressed on the impor- using Monte Carlo methods reveal that the errors due to
tance of incorporating multilevel heterogeneous data Bayesian estimation are comparatively less than the mean
sets (MHDS) into reliability analysis of complex systems squared errors associated with traditional regression esti-
and also other scientific problems, a synthesis process mators such as least squares method. In Sarhan (2001),
he termed as “meta-analysis.” The author discussed the a Bayesian approach has been implemented to estimate
problems of collecting and inducting data from different the reliability indices of components in a multicomponent

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
488 A. Roy et al.: Risk assessment in process industries

Table 1 Summary of the main contributions in data integration in hierarchical tools employing different Bayesian statistical methods.

Area addressed Method employed Main contribution Objective Application demonstrated in

Data integration FB approach Johnson et al. (2003) Multilevel data in RBD Anti-aircraft missile system
into hierarchical Hamada et al. (2004) Multilevel data in a FT Nuclear industry
tools Graves et al. (2008) Nonoverlapping data in RBD Nuclear industry
Jackson and Mosleh (2012) Nonoverlapping data in FT Sensors
Anderson-Cook et al. (2008) Binomial pass/fail data in series system Military systems
model
Anderson-Cook (2009) MHDS Hypothetical example
Wilson et al. (2011) Reliability assessment using MHDS in Nuclear weapon design
RBD
Peng et al. (2013) Reliability indices determination from Satellite system
MHDS

system when masked system observations of life data are Acknowledging the advantage of EB methods in relia-
available. The estimations are closed-form solutions and bility analysis, Quigley et al. (2011) proposed an innovative
simulations are generated for different levels of masking. methodology to increase inference accuracy by homoge-
Regarding population variability analysis, Droguett et al. nizing the data pool to incorporate expert judgment in EB
(2004) provided models based on the FB approach, which methods. Homogenization measures are introduced in the
can use both observed data and expert judgment. They form of scaling factors. The epistemic uncertainty in these
use a nonconjugate lognormal-Poisson-lognormal likeli- factors is studied and formalized using moment matching
hood model that considers a continuous parameter space methods and analyzed using simulations.
made possible by the MCMC method. A novel hybrid technique combining both EB and FB
EB has found its way in analyzing censored data methods for reliability assessment at design stage is pre-
due to its ability to provide inference under sparse data. sented by Quigley and Walls (2011). The epistemic uncer-
Li et al. (2007) calculated EB estimators of reliability for tainties in the new design is estimated by analyzing data
analyzing Type II censored data. Burr XII distribution from earlier designs. nPEB is used to develop the distribu-
model is used to formulate the likelihood function from tion of time to failures, and an FB approach is taken to
the data. The distribution model resembles the lognormal model prior distributions of failure rates, thus taking full
distribution model and is suitable for various shapes of advantage of the sparse data.
hazard. Low-frequency failures can be effectively modeled Guikema (2007) provided a descriptive study on
with the distribution, as it has an algebraic tail instead of using different methodologies for formulating informa-
an exponential tail. The parameter estimation of shape tive priors from available data sets. Both FB and EB pro-
factor of the distribution and formulation of reliability cedures are employed to assess the reliability of a system
indices is done under an asymmetric loss function LINEX. using binomial data.
This follows from their earlier study of calculating the Hierarchical Bayesian technique is integrated with
approximate confidence limits of reliability indices of cold ASP-based approach by Yang et al. (2013) to provide a novel
standby systems (Shi et al. 2005). The Burr XII distribu- framework for frequency estimations of rare events. Hier-
tions have been used earlier used by Soliman (2002) to archical Bayesian technique was employed to update the
determine reliability estimators in generalized life model. hyper-priors of the failure probability distributions of safety
Bayes estimates under both symmetric asymmetric loss systems. The precursor-based estimator proposed by Bier
functions are obtained using Monte Carlo simulations. and Yi (1995a) is used to find the occurrence rate of oil spill
Most recently, Peng and Yan (2014) used the same Type accidents in offshore plants. A sensitivity analysis showed
II censored data set to obtain Bayesian reliability estimators hierarchical Bayes to be more sensitive to data changes
for a new extended Weilbull prior lifetime distributions. than the conventional Bayesian updating method.
This versatile distribution is widely used in the analysis of Returning to the associated two-stage method, Bunea
life data, as they can produce closed-form solutions when et al. (2005) presented a novel mathematical model based
used with standardized conjugate priors. The results of on two-stage Bayes methodology to analyze the reliabil-
MCMC simulations reveal the hazard rate function to be ity of a specific plant assimilating data from other similar
increasing and having an upside-down bathtub shape. plants. Bunea et al. (2005) used a γ prior with Poisson

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 489

likelihood and compared the results with those from distribution with upper and lower bounds to express epis-
Kaplan (1985) who used a lognormal prior distribution. temic uncertainties. The learning parameters introduced
They concluded that the lognormal model is more suit- determine the ability of the model to induct knowledge
able for reliability assessment using data analysis, as it from available information. The epistemic uncertainty
allows truncation of domain of integrals to concentrate modeling of both the α factors and the failure rates of
the results of likelihood around the maximum. individual components provides a complete sensitivity
In the modeling of CCFs, EB finds its application in analysis of CCF rates.
the work of Vaurio (2002). The proposed methodology The application of BNs to address epistemic uncer-
takes into account the epistemic and assessment uncer- tainty in data in risk and reliability assessment is pro-
tainties in quantification of CCF rates using plant-specific vided by Simon et al. (2008). A tool to estimate reliability
and multiple plant failure data. The rate distributions is presented, which takes into account random and epis-
and their moments are obtained in terms of impact vector temic uncertainties. Evidential networks are formulated
weights using probability generating functions. Effec- from BNs using Dempster-Shafer theory to construct basic
tive plant-specific data inputs are expressed as weights belief assignments corresponding to the incoherency
and are generated using moment matching methods. The between data. The incompleteness of the data is addressed
method is applied by Vaurio (2005) for quantification of by evidence theory taking advantage of the modeling
FTs by estimating the CCF probabilities of the basic events. power of BNs. Besides interval-based probabilities, the
Vaurio and Jänkälä (2006) pointed out that numeri- tool is capable in incorporating fuzzy valued probabilities
cal simulations using MCMC of FB methods, especially the (Simon and Weber 2009a) and multistate systems (Simon
two-stage method, can become intractable. The proposed and Weber 2009b). Table 2 catalogues some of the perti-
PREB model by Vaurio (1987) is revised and the improved nent works in this field.
technique is used to estimate failure rate and probabilities
using γ and β priors, respectively. Combined models are
able to accommodate both time- and event-based failure 4.3 Decision-making studies
data. Effective data inputs accounting for uncertainties
in data assessment are generated using the CCF model of In the field of human risk assessment (HRA) and ORA, BNs
Vaurio (2002). have been extensively used. Kim and Seong (2006) built up
Jones et al. (2010) used BNs for conducting a delay a quantitative method using BNs to observe human inter-
time analysis study of a factory producing carbon black. actions and IC control systems in nuclear power plants.
The failure rate is estimated using BNs. Several scenarios Mathematical models suggested by the same authors (Kim
have been created to study the propagation of evidence and Seong 2004) are used in the procedure. The method
and identify the most problematic area. is suggested as an alternative to contemporary HRA tech-
Failure data collected from industries are often niques such as ATHENA and CREAM. Cai et al. (2013)
laden with uncertainty due to various reasons (Martz described a dynamic BN (DBN) structure to model the
and Hamada 2003). A host of Bayesian techniques have different human factors responsible for offshore blowout
been proposed to assess failure probability under uncer- events. Here, pseudo-FTs are introduced, which removes
tainty in data observed (Siu and Apostolakis 1986, Martz the binary restrictions of FTs by using discrete probabili-
and Picard 1995). Among early works, Atwood and Gen- ties in place of Boolean logic gates. An algorithm based
tillon (1996) used Jeffery’s rule for prior construction for on noisy or formalism is presented to map the pseudo-FTs
modeling uncertainty in CCF models. Tan and Xi (2003) into BNs. Martins and Maturana (2013) used BNs to study
described three Bayesian frameworks to update the failure the human factors involved in oil tanker operations.
probabilities under uncertainty in specific conditions. Recently, efforts have been made by Pasman and
Martz and Hamada (2003) presented an FB treatment of Rogers (2012) to implement BNs in the field of deci-
uncertainty in data collected in assessing the event occur- sion making such as land-use planning and emergency
rence rate using a Poisson model. The model is capable response planning. The authors used BNs to perform risk
of accommodating nonconjugate prior distributions and analysis study on spillage scenario of hydrogen distribu-
MCMC is employed for posterior distribution sampling. tion and transportation system. The annual expected
In Troffaes et al. (2013), a robust Bayesian model is event frequency, the fatality distribution, and the mon-
presented for the treatment of data uncertainty in mod- etary damage is calculated. The data generated are used
eling CCFs using the α factor modeling approach. The α to measure the societal risk expressed as F-N curve. The
factors are represented by an imprecise Dirichlet prior Bayesian net is extended to form an influence diagram

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 490 A. Roy et al.: Risk assessment in process industries

by the addition of decision and cost nodes. According to

Hypothetical electrical network data

Pasman and Rogers (2013), the authors conducted a layer
of protection analysis (LOPA) study by mapping bowties
Application demonstrated in

into hybrid BNs.

The maintainability studies involving degradation

Hypothetical example
Hypothetical system modeling and repair time estimations have benefited from
Nuclear industry

Nuclear industry

Aircraft design
the emergence of the computational tools. A suitable model
Hypothetical

for assessing the monitoring intervals is presented by

Gelman et al. (1995). The probabilistic modeling approach
is based on stochastic filtering that explicitly takes into
account the uncertainties in degradation process of the
Reliability analysis of specific plant using data from similar plants

component and the condition monitoring measurement

data. Pulkkinen and Simola (2000) proposed Bayesian
aging models to compute the stochastic failure rate and
the failure on demand probability. MCMC method using
Failure probability estimation under data uncertainty

Reliability analysis of system with data uncertainty

Gibbs sampler algorithm is used to compute the posterior

Expert data pool homogenization for EB methods
Table 2 Summary of the main contributions in data analysis and assessment of reliability indices using different Bayesian methods.

distributions. The sensitivity of the models to observed

data is increased by ignoring monotonous ageing trend.
Data uncertainty in failure rate estimation

Reliability estimation of complex systems

Determination of plant specific CCF rates

Aging indicators based on the posterior failure probability

and failure rates are also presented. In the field of damage
Estimation of rare event frequency
FT quantification under CCF data

accumulation modeling, Jiang et al. (2013) presented a

Bayesian framework to develop a nonlinear probability
hazard model of a particular failure mode for assessing
Data uncertainty in CCF

system reliability in condition-based maintenance.

Hu et al. (2010) expanded the role of BNs in risk
analysis by presenting a systematic method for safety
Objective

prewarning in a complicated industrial scenario. The

systematically defined method uses BN inference to form
the basis of safety analysis and prognosis steps. The BN
is generated using HAZOP of the system and the degrada-
Martz and Hamada (2003)

Quigley and Walls (2011)

tion models are derived from Markov analysis. It is also

Troffaes et al. (2013)

Quigley et al. (2011)

worth mentioning here that the BN is updated using real-

Simon et al. (2008)

Bunea et al. (2005)

Main contribution

Yang et al. (2013)

Tan and Xi (2003)

time data from a conditioning monitoring system. Staying

Vaurio (2002)
Vaurio (2005)

in the field of fault diagnosis, Duan and Zhou (2012) for-

mulated a technique that uses FTs for modeling and BNs
for inference purposes. The diagnostic importance factor
(DIF) is generated using BNs. A diagnostic decision algo-
rithm generated the diagnostic decision tree (DDT), which
Combining EB and Bayes

integrated with ASP data

helps in proper identification of failing components. Real-

Hierarchical Bayes is
Method employed

time data from sensors are used to update the probabili-

Two-stage Bayes

ties in the BNs.

Although the trend started recently, researchers are
methods

adapting Bayesian tools in various aspects of decision-

making studies. Table 3 lists some of them.
BN

EB

EB
FB

Reliability assessment

Rare events frequency

4.4 Modeling of failure dependency

Data uncertainty

homogenization
Area addressed

at design stage

estimation

The earliest sourced literature concerning the use of BNs

Data pool

in reliability analysis is of Almond (1992). He defines a

CCFs

GRAPHICAL-BELIEF tool to model complex systems and

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 491

estimate reliability indices. The contemporary works of

Hydrogen transportation and storage stations

Hydrogen transportation and storage stations

Arroyo et al. (1998) and Torres-Toledano and Sucar (1998)
address the same objectives. The SERENE project of Bouis-

Transportation in oil and gas industry

Seepage data from subway tunnels

sou et al. (1999) tries to induct expert analysis into mod-
eling the dependability in complex systems with BNs.
Offshore oil and gas industry
Application demonstrated in

Bobbio et al. (2001) pointed out that standard FTs

are incapable of modeling the functional and sequen-
Nuclear power plants

tial dependencies between the components. The authors

answer this problem by presenting a holistic technique for
Aircraft engine

NRC database
mapping FTs into corresponding BNs. BNs are shown to be
just as effective as FTs in calculating the predictive prob-
Industry

abilities. The additional diagnostic power of BNs is dem-

onstrated by estimating the most probable event (MPE)
given any evidence.
Human factors influencing offshore blowout events
Quantitative study of human factors and IC system

System reliability in condition-based maintenance

Using BNs in LOPA accident modeling technique

Weber et al. (2004) formalized the prevalent 2TBN

Using BNs in land-use planning and emergency

method (Boyen and Koller 1998) for the construction of

Table 3 Summary of the main contributions indifferent aspects of HRA and maintainability studies using different Bayesian methods.

Integrated method for safety prewarning

DBNs to model system dependability under the influence

of exogenous constraints, that is, external factors that
Fault diagnosis based on FT and BN

Monitoring maintenance intervals

cause degradation and system failures. The authors dem-

onstrate equivalence between BNs and Markov chains by
HRA in oil tanker operations

structuring the 2TBN on Markov switch models (MSM) and

Monitoring aging trend

hidden Markov models (HMM).

response planning

Arroyo et al. (1998) extended standard BNs to model

the temporal dimension of a system in reliability analysis.
ORA in CPI

The probabilistic framework known as temporal node BN

Objective

failures

(TNBN) is able to represent the dynamics of the opera-

tional conditions of system under uncertainty. Following
the event/interval-based approach (Arroyo et al. 1998),
Martins and Maturana (2013)

Pulkkinen and Simola (2000)

Boudali and Dugan (2005) presented the discrete time BN

Pasman and Rogers (2012)

Pasman and Rogers (2013)

(DTBN) formalism for the reliability assessment of dynamic

Duan and Zhou (2012)
Kim and Seong (2006)

Gelman et al. (1995)

systems. A hierarchical and modular method is followed

Main contribution

Jiang et al. (2013)

Meel et al. (2007)

for mapping DFTs into DBNs. The procedure employs a

Cai et al. (2013)

Hu et al. (2010)

static time discretization technique. DTBNs are generalized

to form continuous time BNs (CTBN) in the work of Boudali
and Dugan (2006) to accommodate continuous variables
by increasing the number of time intervals (discretization)
to its limit. Continuous variables are expressed as pdfs,
methodology using ASP data

and the joint probability distributions are expressed as

joint density functions. In the absence of CPTs, functional
Dynamic risk analysis
Method employed

and sequential dependencies are expressed by using unit

step function and the impulse function. Simultaneously,
Bearfield and Marsh (2005) proposed a novel approach to
model event trees (ETs) into BNs.
BNs

BNs
BNs

BNs
BNs

BNs

BNs

The DTBN technique has, however, come under fire

FB
FB
FB

due to its tendency to generate huge multidimensional

Maintainability studies

CPTs when applied to complex systems (Montani et al.

2008, Marquez et al. 2010) and problems in adopting
Human reliability

Decision making
Area addressed

Fault diagnosis

nonexponential distributions. To overcome this, Khakzad

Data analysis
assessment

et al. (2013a) proposed a new general formalism to model

the failure distribution of input variables yielding closed-
form analytical solutions for majority of pdf forms. An

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
492 A. Roy et al.: Risk assessment in process industries

innovative algorithm is proposed to reduce the dimension model the dependency between system components in
of CPT by an order of one magnitude and is applied to which bowties are unable to do so. The FT part of a bowtie
model probabilistic dynamic gates such as PAND gate and is mapped using the Bobbio et al. (2001) technique and
also conventional static AND/OR gates. It must be noted the ET part is mapped by the Bearfield and Marsh (2005)
here that the proposed algorithm can only be applied for method. In the study apart from using probability updat-
predictive purposes and not for diagnostic analysis. ing to find the MPE states, the authors implement the
Weber and Jouffe (2006) described an object-oriented probability adapting approach. Khakzad et al. (2013c)
approach for modeling system reliability using DBNs. In the implemented this technique in conducting a QRA study
proposed dynamic object-oriented BN (DOOBN) method, of offshore drilling operations. OOBNs (Kjaerulff and
the structural specification of the DBN is modeled using Madsen 2008) are constructed comprising individual net-
the functional analysis and the malfunctional analysis of works, which help to construct a large BN as a hierarchical
the system. The equivalence with Markov chains shown in structure of many subnetworks at different levels.
the work of Weber et al. (2004) is maintained here. Oliver and Yang (1990) described a Bayesian analysis
A descriptive investigation of the different inference of precursor data using ETs and influence diagrams. On
algorithms developed for hybrid BNs (containing both this platform, Bier and Yi (1995a,b) proposed an improved
discrete and continuous variables) in context to reliabil- method that can incorporate intersystem dependencies
ity analysis is given by Langseth et al. (2009). Compari- overlooked by Oliver and Yang (1990). Bier accomplished
sons are drawn between the discretization technique, the this by the use of natural conjugate distributions to
variational approximation method, the mixture of trun- analyze partial correlation probabilities analytically.
cated exponentials (MTE), and the MCMC algorithms. The BNs have thus been used extensively in risk analysis
authors concluded that MTE is best suited for approximat- and assessment to model the failure dependency. As can
ing the tails of the distributions accurately. be seen in Table 4 the works have developed over time
The popular discretization method was rejected by with each influencing the other.
Langseth et al. (2009) because it remained static through-
out the stages of Bayesian inference despite the presence
of new evidence. The trade-off between long computa- 4.5 DRA methodology
tional times and accuracy is a burden to any reliability
engineer. Neil et al. (2008) stated that hybrid BNs are Meel and Seider (2006) provided the first major contribu-
essential for efficient modeling of complex systems. As tion on integrating ASP data for conducting risk assess-
an alternative to the static discretization followed by ment in CPI. The presented technique, termed as DRA,
Boudali and Dugan (2005, 2006), dynamic discretization involves the accident modeling of the consequences using
algorithm is presented by Neil et al. (2008). The iterative an ET and updating the occurrence and failure probabili-
scheme is combined with standard inference algorithms ties through Bayesian updating approach. A Bayesian
to produce a definitive reliability tool. The advancement predictive method is employed to determine the future
of inference algorithms make modeling of non-Gaussian number of events in the nest time interval. As ASP data
distributions possible. Applications of the algorithm are generated quite often and are plant specific in nature,
are presented in the work of Marquez et al. (2010). Neil the methodology is dynamic in nature and at the same
et al. (2008) further extended the technique to model the time giving plant-specific results. In one of the proposed
availability of repairable systems in the work of Neil and Bayesian models, the dependency between the safety
Marquez (2012) using hybrid BNs. The same hierarchical systems is expressed as the joint posterior distribution
model and the propagation algorithms are used to esti- of the failure probabilities using copula theory. These
mate the repair times, logistic delay times, and scheduled joint distributions are then updated using available ASP
delay time distributions. data. This effectively portrays the correlations between
Montani et al. (2008) provided a software tool the failure probabilities at different branch points of the
RANDYBAN that analyzes a DFT and automatically converts ET and also those between the safety systems. Kalantar-
it to its corresponding DBN. The basic algorithm follows the nia et al. (2009) used the proposed DRA methodology
2TBN structure. To reduce the size of the CPTs divorcing is in safety analysis of a storage tank containing hazard-
employed by dividing the parent set of anode into subsets. ous chemicals. A comparison study with standard PRA
Bowties have long been recognized as an effective techniques demonstrates the efficiency of the method to
tool to model accident scenarios. Khakzad et al. (2013b) model real-time failure frequency during the lifetime of
developed a procedure for mapping bowties into BNs to the process unit. The efficiency of the DRA tool is further

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 493

Application illustrated in
witnessed in the work of Kalantarnia et al. (2010) where it

Chemical Safety Board

is used to assess the data from the BP Texas City incident.

Case study from U.S.

Hypothetical system
Hypothetical system
Hypothetical system
Hypothetical system
Hypothetical system
Hypothetical system
Meel et al. (2007) investigated and analyzed the NRC
database using the above-mentioned methodology. The
NRC database contains reports about oil, chemical, and
biological environmental discharges. The predictive fre-

CPI
quency of incidents in different chemical and petrochemi-
cal plants is calculated using the predictive Bayesian
analysis. In appropriate cases, the negative binomial prior
is shown to give better predictions than the γ prior. Equip-
ment failure and operator error probabilities are calculated
Analysis of precursor data using ETs and influence diagrams

Reliability analysis of systems with continuous variables

and examined by the Bayesian updating method. The loss

severity distribution of the incidents generated using the
extreme value theorem (EVT) helps to model rare and
Mapping DFTs into BNs with continuous variable
Mapping DFTs into BNs with discrete variable

extreme events. Fast Fourier transform (FFT) is conducted

to model the loss distribution. Thus, applying Bayesian
QRA study in offshore drilling operation

updating and predictive approach, a thorough operational

risk analysis is conducted in chemical industries.
The Bayesian model with copula proposed by Meel
and Seider (2006) is used to perform a real-time risk analy-
Mapping bowties into BN

sis for a continuous ethyl benzene process. ASP data from

Mapping ETs into BNs
Mapping FTs into BNs
Table 4 Summary of the main contributions in modeling of failure dependency using different Bayesian methods.

warning systems such as alarms are used to update the

failure probabilities of the safety systems. Transfer func-
tions are used to model the propagations of the abnor-
Objective

mal events through the safety systems to the end states.

Real-time simulations of the propagations are done using
SIMULINK in MATLAB. This provides a dynamic modeling
Bearfield and Marsh (2005)

of the safety systems. According to Pariyani et al. (2010),

Boudali and Dugan (2006)
Boudali and Dugan (2005)

dynamic data collected from distributed control systems

Khakzad et al. (2013b)

Khakzad et al. (2013c)

Bobbio et al. (2001)

(DCS) and emergency shutdown (ESD) systems are

Bier and Yi (1995b)
Main contributions

Neil et al. (2008)

used as ASP data for dynamic risk analysis of an indus-

trial fluid catalytic cracking unit (FCCU). Pariyani et al.
(2012a,b) used the dynamic failure analysis methodology
by using the data recorded in ESD and DCS to increase
process safety and product quality. The extensive research
carried over 6 months and published in two parts forms
Bayesian analysis of precursor data

a compact data set through a new set theoretic structure,

which in turn increases the efficiency of the probabilistic
calculations.
Rathnayaka et al. (2011a) focused on the accident
modeling part of the DRA technique and proposed the
Method employed

system hazard identification prediction and prevention

(SHIPP) methodology. The technique takes a preventive
OOBNs

approach and stresses on the implementation of safety

analysis at the design stage. The accident modeling pro-
BN

cedure is developed on the lines of the one proposed by

Modeling of failure

Kujath et al. (2010). The accident sequence is represented

Area addressed

with respect to safety barrier failure in an ET. The failure

dependency

of the safety barrier failure is analyzed using FTs. The inte-

gration of hazard and barrier analysis provides a robust
basis for decision making. ASP data are used to update the

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
494 A. Roy et al.: Risk assessment in process industries

Application illustrated in
failure frequencies, the abnormal event occurrence prob-

High-sulfur gas station

ability. In subsequent papers (Rathnayaka et al. 2011b,
2012), the authors applied the technique in a liquid nitro-
gen gas (LNG) facility. A dramatic variation of occurrence

LNG facility
probabilities and probabilities of degradation of system
barriers is observed with the incorporation of new data

FCCU
CPI
CPI
CPI
using Bayesian analysis.
An improved version of the SHIPP technique is
recently presented by Tan et al. (2013). The accident
modeling approach is extended to model the toxicity of
high sulfur gathering station. Event sequence diagrams

Dynamic accident modeling with ESD and failure probability updating

are used to model the consequences arising from failure
of safety barriers. The dependent barriers are modeled
through BNs to calculate consequence occurrence prob-

Accident modeling and updating the failure probabilities

ability. Bayesian updating is employed with real-time
plant data. Table 5 lists some of the works employing this
methodology.

DRA using data recorded in ESD and DCS

5 Conclusion and discussion

Plant-specific DRA using ASP data

Economic reasons prompt the construction of process

Real-time risk analysis

plants in locations from where raw materials, human

ORA using database

resource, and energy are easily accessible. These locations
for the obvious reasons are populated areas. The process Objective
industry has to work with large quantities of hazardous
materials, which can be toxic, inflammable, or explosive,
thus posing a risk to the surrounding population as well
as its employees. However, at the same time, the growth
of population has inadvertently led to the rise in demand,
which in turn has led to the growth of industrialization.
Process industries are an inevitable part of today’s civili-
Rathnayaka et al. (2011a,b)
Table 5 Summary of the main contributions using DRA methodology.

zation. From producing quotidian items such as cooking

Pariyani et al. (2012a,b)
Meel and Seider (2006)

Meel and Seider (2008)

oil and paints to providing fuel for rockets, process indus-

Main contributions

Meel et al. (2007)

tries are related to every human in one way or another.

Tan et al. (2013)

Thus, the construction and operation of a process plant

involves many decision-making aspects.
Risk assessment emerged as an effective decision-
making tool in the 1970s in the nuclear industry. System-
atic risk assessment is performed routinely for nuclear
power plants and the corresponding procedural steps have
reached a certain stage of maturity. The risk assessment
process is simple. It is an examination of the work and
Predictive accident modeling

workplace to identify what could cause harm to people

and assessment of the chance, high or low, that somebody
could be harmed by the hazards identified, together with
Area addressed

an indication of how serious the harm could be.

Disastrous accidents led to the introduction of risk
analysis tools in CPI. Over the years, a variety of tools
DRA

have been developed for conducting risk assessment in

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 495

chemical industries. Many tools prevalent in the nuclear Almond RG. An extended example for testing graphical belief.
industry have been effectively adopted into chemical Technical Report 6, Statistical Sciences, Inc., 1992.
Anderson-Cook C. Opportunities and issues in multiple data type
process plants. As Kirchsteiger (1997) pointed out, risk
meta-analyses. Qual Eng 2009; 21: 241–253.
analysis involves the reconstruction of reality to identify Anderson-Cook CM, Graves T, Hengartner N, Klamann R, Wiedlea
the causes of accidents either in a quantitative or in a ACK, Wilson AG, Anderson G, Lopez G. Reliability modelling
qualitative way. using both system test and quality assurance data. Military
The extensive literature review revealed that Bayesian Oper Res 2008; 13: 5–18.
Apeland S, Aven T, Nilsen T. Quantifying uncertainty under a
methods might be useful in accomplishing these tasks. Over
predictive, epistemic approach to risk analysis. Reliab Eng Syst
the years, researchers have developed procedures related Saf 2002; 75: 93–102.
to the Bayes theorem, which have addressed the different Arroyo G, Sucar L, Villavicencio A. Probabilistic temporal reasoning
problems related to risk assessment in process industries. In and its application to fossil power plant operation. Expert Syst
the Bayesian edifice, whenever new information arrives, it Appl 1998; 15: 317–324.
is added to the analysts’ knowledge. The posterior obtained Atwood CL, Gentillon CD. Bayesian treatment of uncertainty in
classifying data: two case studies. In: Proceedings of the
from previous calculations is treated as the prior.
ESREL ‘96/PSAM-III International Conference on Probability
As seen, the use of Bayesian methods in risk analysis Safety Assessment and Management, 1996.
started to overcome the problem of sparse availability of Atwood CL, Kelly DL. The binomial failure rate common-cause model
data. The advent of computer software opened new doors with WinBUGS. Reliab Eng Syst Saf 2009; 94: 990–999.
for the use of Bayesian methods in the field of safety assess- Aven T. Foundations of risk analysis – a knowledge and decision-
oriented perspective. Chichester, England: Wiley, 2003.
ment in process industries. Software such as WinBUGS not
Aven T, Kvaløy JT. Implementing the Bayesian paradigm in risk
only computes the refined posterior distributions but also analysis. Reliab Eng Syst Saf 2002; 78: 195–201.
has solved the problem of MHDS integration into risk and Bearfield G, Marsh W. Generalising event trees using Bayesian
reliability analysis. The AI tool BN can perform both predic- networks with a case study of train derailment. Lecture Notes
tive and diagnostic operations for risk assessment studies. Comput Sci 2005; 3688: 52–66.
Critics of the methods point out the overuse of subjec- Berger JO. Statistical decision theory and Bayesian analysis, 2nd
ed., New York: Springer-Verlag, 1985.
tive knowledge in Bayesian methods. This view is very jus-
Berry DA. Statistics: a Bayesian perspective. Belmont, CA: Duxbury
tified as pointed out by Ferson (2005). While adopting any Press, 1996.
Bayesian technique, researchers thus should not lose the Bier VM. Statistical methods for the use of accident precursor data
sight of objectivity. Objectivity is essential in any scientific in estimating the frequency of rare events. Reliab Eng Syst Saf
analysis because it is this that makes the whole analysis 1993; 41: 267–280.
Bier VM, Yi W. The performance of precursor-based estimators for
relevant. Echoing views of Ferson (2005), the authors rec-
rare event frequencies. Reliab Eng Syst Saf 1995a; 50: 241–251.
ommended the practice of sensitivity analysis to justify Bier VM, Yi W. A Bayesian method for analyzing dependencies in
the choice of priors in the calculations. precursor data. Int J Forecast 1995b; 11: 25–41.
The Bayesian paradigm provides a full and compre- Bobbio A, Portinale L, Minichino M, Ciancamerla E. Improving the
hensive framework for risk analysis of a complex system. analysis of dependable systems by mapping fault trees into
Initially, the research fraternity concentrated in the devel- Bayesian networks. Reliab Eng Syst Saf 2001; 71: 249–260.
Boudali H, Dugan JB. A discrete-time Bayesian network reliability
opment of proper methodologies and appropriate algo-
modeling and analysis framework. Reliab Eng Syst Saf 2005;
rithms (Kelly and Smith 2009) and applications in CPI 87: 337–349.
were less. However, the advent of user-friendly software Boudali H, Dugan JB. A continuous-time Bayesian network reliability
has initiated the use of these procedures for risk analysis modeling, and analysis framework. IEEE Trans Reliab 2006; 55:
of complex systems in process industries. Analysts have 86–97.
Bouissou M, Martin F, Ourghanlian A. Assessment of a safety
conducted case studies (Kalantarnia et al. 2010, Wilson
critical system including software: a Bayesian belief network
et al. 2011) successfully with these methods yielding for evidence sources. In: Reliability and Maintainability
satisfactory results. Symposium (RAMS’99), Washington, 1999.
Boyen X, Koller D. Tractable for complex stochastic processes.
In: Proceedings of the 14th Annual Conference on Uncertainty
in AI (UAI), Madison, Wisconsin, 1998: 33–42.
References Bucci P, Kirschenbaum J, Mangan LA, Aldemir T, Smith C, Wood T.
Construction of event-tree/fault-tree models from a Markov
Advanced Light Water Reactor Utility Requirements Document approach to dynamic system reliability. Reliab Eng Syst Saf
(Volume III)-Chapter 1, Appendix A, PRA Key Assumptions and 2008; 93: 1616–1627.
Ground Rules, Electric Power Research Institute, Inc., USA, 1995. Bunea C, Charitos T, Cooke RM, Becker G. Two-stage Bayesian
Agena Ltd. AgenaRisk software package, 2007. http://www. models – application to ZEDB project. Reliab Eng Syst Saf
AgenaRisk.com. 2005; 90: 123–130.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
496 A. Roy et al.: Risk assessment in process industries

Cai B, Liu Y, Zhang Y, Fan Q, Liu Z, Tian X. A dynamic Bayesian Hamada MS, Wilson AG, Reese C, Martz MF. Bayesian reliability.
networks modeling of human factors on offshore blowouts. New York: Springer, 2008.
J Loss Prev Process Ind 2013; 26: 639–649. Hofer E. On two-stage Bayesian modeling of initiating event
Casella G. Illustrating empirical Bayes methods. Chemometrics frequencies and failure rates. Reliab Eng Syst Saf 1999; 66:
Intell Lab Syst 1992; 16: 107–125. 97–99.
Chiacchio F, Compagno L, D’Urso D, Manno G, Trapani N. Dynamic Hu J, Zhang L, Ma L, Liang W. An integrated method for safety
fault trees resolution: a conscious trade-off between analytical pre-warning of complex system. Saf Sci 2010; 48: 580–597.
and simulative approaches. Reliab Eng Syst Saf 2011; 96: Jackson C, Mosleh A. Bayesian inference with overlapping data for
1515–1526. systems with continuous life metrics. Reliab Eng Syst Saf 2012;
Chib S, Greenberg E. Hierarchical analysis of SUR models with 106: 217–231.
extensions to correlated serial errors and time-varying Jaynes ET, Bretthorst GL. Probability theory: the logic of science.
parameter models. J Economet 1995; 68: 339–360. Cambridge University Press, 2003.
Cooke R, Goossens L. The accident sequence precursor Jensen FV. Bayesian networks and decision graphs. New York, NY:
methodology for the European Post-Seveso era. Reliab Eng Springer, 2001.
Syst Saf 1990; 27: 117–130. Jiang X, Yuan Y, Liu X. Bayesian inference method for stochastic
Crowl DA, Louvar JF. Chemical process safety: fundamentals with damage accumulation modeling. Reliab Eng Syst Saf 2013; 111:
application. Englewood Cliffs, New Jersey: Prentice Hall, Inc., 126–138.
1990. Johnson V, Graves T, Hamada M, Reese CS. In: Bernardo JM, Bayarri
Droguett EL, Groen F, Mosleh A. The combined use of data and MJ, Berger JO, Dawid AP, Heckerman D, Smith AFM, West M,
expert estimates in population variability analysis. Reliab Eng editors. A Hierarchical model for estimating the reliability
Syst Saf 2004; 83: 311–321. of complex systems, Bayesian Statistics 7. London: Oxford
Duan R, Zhou H. A new fault diagnosis method based on fault tree University Press, 2003: 199–213.
and Bayesian networks. Energy Procedia 2012; 17: 1376–1382. Jones B, Jenkinson I, Yang Z, Wang J. The use of Bayesian network
Eckerman I. The Bhopal saga: causes and consequences of the modelling for maintenance planning in a manufacturing
world’s largest industrial disaster. India: Universities Press industry. Reliab Eng Syst Saf 2010; 95: 267–277.
(India), 2005. Kafka P. Probabilistic safety assessment: quantitative process to
Ferdous R, Khan FI, Veitch B, Amyotte PR. Methodology for balance design, manufacturing and operation for safety of plant
computer-aided fault tree analysis. Process Saf Environ Protect structures and systems. Nucl Eng Des 1996; 165: 333–350.
2007; 85: 70–80. Kalantarnia M, Khan F, Hawboldt K. Dynamic risk assessment using
Ferdous R, Khan F, Veitch B, Amyotte PR. Methodology for computer failure assessment and Bayesian theory. J Loss Prev Process
aided fuzzy fault tree analysis. Process Saf Environ Protect Ind 2009; 22: 600–606.
2009; 87: 217–226. Kalantarnia M, Khan F, Hawboldt K. Modelling of BP Texas City
Ferson S. Bayesian methods in risk assessment. Unpublished refinery accident using dynamic risk assessment approach.
Report prepared for the Bureau de Recherches Geologiqueset Process Saf Environ Protect 2010; 88: 191–199.
Minieres (BRGM), New York, 2005. Kaplan S. The two-stage Poisson-type problem in probabilistic risk
Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. analysis. Risk Anal 1985; 5: 227–230.
London: Chapman & Hall, 1995. Kelly DL, Smith CL. Bayesian inference in probabilistic risk
Graves TL, Hamada MS, Klamann SR, Koehler AC, Martz HF. A fully assessment – the current state of the art. Reliab Eng Syst Saf
Bayesian approach for combining multi-level information in 2009; 94: 628–643.
multi-state fault tree quantification. Reliab Eng Syst Saf 2007; Khakzad N, Khan F, Amyotte P. Safety analysis in process facilities:
92: 1476–1483. comparison of fault tree and Bayesian network approaches.
Guidelines for Process Equipment Reliability Data with Data Tables. Reliab Eng Syst Saf 2011; 96: 925–932.
Center for Chemical Process Safety of the American Institute of Khakzad N, Khan F, Amyotte P. Risk-based design of process
Chemical Engineers, New York, NY, 1989. systems using discrete-time Bayesian networks. Reliab Eng
Guide to the Collection and Representation of Electronic, Syst Saf 2013a; 109: 5–17.
Sensing Component, and Mechanical Equipment Reliability Khakzad N, Khan F, Amyotte P. Dynamic safety analysis of process
Data for Nuclear Generating Stations. IEEE Standard 500-1984. systems by mapping bow-tie into Bayesian network. Process
Institute of Electrical and Electronics Engineers, New York, NY, Saf Environ Protect 2013b; 91: 46–53.
1984. Khakzad N, Khan F, Amyotte P. Quantitative risk analysis of offshore
Guikema SD. Formulating informative, data-based priors for failure drilling operations: a Bayesian approach. Saf Sci 2013c; 57:
probability estimation in reliability analysis. Reliab Eng Syst 108–117.
Saf 2007; 92: 490–502. Khan FI, Abbasi SA. Techniques and methodologies for risk analysis
Guo J, Wilson AG. Bayesian Methods for Estimating the Reliability of in chemical process industries. J Loss Prev Process Ind 1998;
Complex Systems Using Heterogeneous Multilevel Information. 11: 261–277.
Iowa State University Department of Statistics, Technical Khan FI, Abbasi SA. Analytical simulation and PROFAT II: a new
Report 2010-10, 2010. methodology and a computer automated tool for fault tree
Hamada M, Martz HF, Reese CS, Graves T, Johnson V, Wilson AG. analysis in chemical process industries. J Hazard Mater 2000;
A fully Bayesian approach for combining multilevel failure 75: 1–27.
information in fault tree quantification and optimal follow-on Khan FI, Abbasi SA. Risk analysis of a typical chemical industry
resource allocation. Reliab Eng Syst Saf 2004; 86: 297–305. using ORA procedure. J Loss Prev Process Ind 2001; 14: 43–59.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 497

Khan FI, Abbasi SA. A criterion for developing credible accident Martz HF, Hamada MS. Uncertainty in counts and operating time in
scenarios for risk assessment. J Loss Prev Process Ind 2002; estimating Poisson occurrence rates. Reliab Eng Syst Saf 2003;
15: 467–475. 80: 75–79.
Khan FI, Iqbal A, Ramesh N, Abbasi SA. SCAP: a new methodology Martz HF, Picard RR. Uncertainty in Poisson event counts and exposure
for safety management based on feedback from credible time in rate estimation. Reliab Eng Syst Saf 1995; 48: 181–190.
accident-probabilistic fault tree analysis system. J Hazard Martz HF, Wailer RA. Bayesian reliability analysis of complex series/
Mater 2001; 87: 23–56. parallel systems of binomial subsystems and components.
Khan FI, Husain T, Abbasi SA. Design and evaluation of safety Technometrics 1990; 32: 407–416.
measures using a newly proposed methodology “SCAP”. J Loss Mastran DV. Incorporating component and system test data into
Prev Process Ind 2002; 15: 129–146. the same assessment: a Bayesian approach. Op Res 1976; 24:
Kim MC, Seong PH. A quantitative model of system-man interaction 491–499.
based on discrete function theory. J Korean Nucl Soc 2004; 36: Mastran DV, Singpurwalla ND. A method for reliability estimation of
430–449. logical structures. Eng Fracture Mech 1976; 8: 229–237.
Kim MC, Seong PH. A computational method for probabilistic safety Meel A, Seider WD. Plant-specific dynamic failure assessment using
assessment of I&C systems and human operators in nuclear Bayesian theory. Chem Eng Sci 2006; 61: 7036–7056.
power plants. Reliab Eng Syst Saf 2006; 91: 580–593. Meel A, Seider WD. Real-time risk analysis of safety systems.
Kirchsteiger C. Impact of accident precursors on risk estimates from Comput Chem Eng 2008; 32: 827–840.
accident databases. J Loss Prev Process Ind 1997; 10: 159–167. Meel A, O’Neill LM, Levin JH, Seider WD, Oktem U, Keren N.
Kjaerulff UB, Madsen AL. Bayesian networks and influence Operational risk assessment of chemical industries by
diagrams. New York: Springer, 2008. exploiting accident databases. J Loss Prev Process Ind 2007;
Kleindorfer P, Oktem UG, Pariyani A, Seider WD. Assessment of 20: 113–127.
catastrophe risk and potential losses in industry. Comput Montani S, Portinale L, Bobbio A, Codetta-Raiteri D. Radyban: a
Chem Eng 2012; 47: 85–96. tool for reliability analysis of dynamic fault trees through
Knegtering B, Pasman H. The safety barometer: how safe is my plant conversion into dynamic Bayesian networks. Reliab Eng Syst
today? Is instantaneously measuring safety level utopia or Saf 2008; 93: 922–932.
realizable? J Loss Prev Process Ind 2013; 26: 821–829. Mosleh A, Bier VM, Apostolakis G. Methods for the Elicitation
Kujath MF, Amyotte PR, Khan FI. A conceptual offshore oil and gas and Use of Expert Opinion in Risk Assessment: Phase 1–4
process accident model. J Loss Prev Process Ind 2010; 23: Critical Evaluation and Directions for Future Research. NUREG/
323–330. CR-4962, US Nuclear Regulatory Commission, 1987.
Labeau PE, Smidts C, Swaminathan S. Dynamic reliability: towards Murphy KP. Software packages for graphical models/Bayesian
an integrated platform for probabilistic risk assessment. Reliab networks, 2004. http://people.cs.ubc.ca/~murphyk/Software/
Eng Syst Saf 2000; 68: 219–254. bnsoft.html.
Lahiri P, Park DH. Nonparametric Bayes and empirical Bayes Neil M, Marquez D. Availability modelling of repairable systems
estimators of mean residual life at age t. J Stat Plan Infer 1992; using Bayesian networks. Eng Appl Artif Intell 2012; 25:
29: 125–136. 698–704.
Langseth H, Portinale L. Bayesian networks in reliability. Reliab Eng Neil M, Tailor M, Marquez D, Fenton N, Hearty P. Modelling
Syst Saf 2007; 92: 92–108. dependable systems using hybrid Bayesian networks. Reliab
Langseth H, Nielsen TD, Rumí R, Salmerón A. Inference in hybrid Eng Syst Saf 2008; 93: 933–939.
Bayesian networks. Reliab Eng Syst Saf 2009; 94: 1499–1509. Offshore Reliability Data Handbook, OREDA-97, SINTEF Industrial
Li X, Shi Y, Wei J, Chai J. Empirical Bayes estimators of reliability Management, Norway, 1997.
performances using LINEX loss under progressively type-II Oktem UG, Wong R, Oktem C. Risk and regulation [Special Issue on
censored samples. Math Comput Simul 2007; 73: 320–326. Close Calls, Near-Misses and Early Warnings]. LSE Publication,
Lindley DV. Introduction to probability and statistics from a 12-13, 2010.
Bayesian viewpoint. Part 2: inference. UK: Cambridge Oliver RM, Yang HJ. Bayesian updating of event tree parameters to
University Press, 1965. predict high risk incidents. In: Oliver RM, Smith JQ, editors.
Lunn DJ, Thomas A, Best N, Spiegelhalter D. WinBUGS – a Bayesian Influence Diagrams, Diagrams, Belief Nets and Decision
modelling framework: concepts, structure, and extensibility. Analysis. Chichester, UK: Wiley, 1990.
Stat Comput 2000: 325–337. Papazoglou IA, Nivolianitou Z, Aneziris O, Christou M. Probabilistic
Luo P, Hu Y. System risk evolution analysis and risk critical event safety analysis in chemical installations. J Loss Prev Process
identification based on event sequence diagram. Reliab Eng Ind 1992; 5: 181–191.
Syst Saf 2013; 114: 36–44. Pariyani A, Seider WD, Oktem UG, Soroush M. Improving process
Marquez D, Neil M, Fenton N. Improved reliability modeling using safety and product quality using large databases. In 20th
Bayesian networks and dynamic discretization. Reliab Eng Syst European Symposium on Computer Aided Process Engineering
Saf 2010; 95: 412–425. (ESCAPE), 2010.
Martins MR, Maturana MC. Application of Bayesian belief networks Pariyani A, Seider WD, Oktem UG, Soroush M. Dynamic risk
to the human reliability analysis of an oil tanker operation analysis using alarm databases to improve process safety and
focusing on collision accidents. Reliab Eng Syst Saf 2013; 110: product quality: part I – data compaction. AIChE J 2012a; 58:
89–109. 812–825.
Martz HF, Almond RG. Using higher-level failure data in fault tree Pariyani A, Seider WD, Oktem UG, Soroush M. Dynamic risk
quantification. Reliab Eng Syst Saf 1997; 56: 29–42. analysis using alarm databases to improve process safety and

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
498 A. Roy et al.: Risk assessment in process industries

product quality: part II – Bayesian analysis. AIChE J 2012b; 58: Sarhan A. Non-parametric empirical Bayes procedure. Reliab Eng
826–841. Syst Saf 2003; 80: 115–122.
Pasman HJ, Rogers WJ. Risk assessment by means of Bayesian Shi Y, Shi X, Xu Y. Approximate confidence limits of reliability
networks: a comparative study of compressed and liquefied performances for cold standby series system. J Appl Math
H2 transportation and tank station risks. Int J Hydrogen Energy Comput 2005; 19: 439–445.
2012; 37: 17415–17425. Simon C, Weber P. Imprecise reliability by evidential networks. Proc
Pasman H, Rogers W. Bayesian networks make LOPA more effective, Inst Mech Eng Part O J Risk Reliab 2009a; 223: 119–131.
QRA more transparent and flexible, and thus safety more Simon C, Weber P. Evidential networks for reliability analysis and
definable! J Loss Prev Process Ind 2013; 26: 434–442. performance evaluation of systems with imprecise knowledge.
Pasman HJ, Jung S, Prem K, Rogers WJ, Yang X. Is risk analysis a IEEE Trans Reliab 2009b; 58: 69–87.
useful tool for improving process safety? J Loss Prev Process Simon C, Weber P, Evsukoff A. Bayesian networks inference
Ind 2009; 22: 769–777. algorithm to implement Dempster Shafer theory in reliability
Pearl J. Probabilistic reasoning in intelligent systems: networks analysis. Reliab Eng Syst Saf 2008; 93: 950–963.
of plausible inference. San Mateo, CA: Morgan Kaufmann Siu N, Apostolakis G. Modeling the detection rates of fires in
Publishers, 1988. nuclear power plants: development and application of a
Peng X, Yan Z. Estimation and application for a new extended methodology for treating imprecise evidence. Risk Anal 1986;
Weibull distribution. Reliab Eng Syst Saf 2014; 121: 34–42. 6: 43–59.
Peng W, Huang H, Xie M, Yang Y, Liu Y. A Bayesian approach for Siu NO, Kelly DL. Bayesian parameter estimation in probabilistic
system reliability analysis with multilevel pass-fail, lifetime and risk assessment. Reliab Eng Syst Saf 1998; 62: 89–116.
degradation data sets. IEEE Trans Reliab 2013; 62: 689–699. Soliman AA. Reliability estimation in a generalized life-model
Phimister JR, Oktem UG, Kleindorfer P, Kunrether H. Near-miss with application to the Burr-XII. IEEE Trans Reliab 2002; 51:
incident management in the chemical process industry. Risk 337–343.
Anal 2003; 23: 445–459. Swaminathan S, Smidts C. Identification of missing scenarios in
Pörn K. On empirical Bayesian inference applied to Poisson ESDs using probabilistic dynamics. Reliab Eng Syst Saf 1999;
probability models. Linköping Studies in Science and 66: 275–279.
Technology, Linköping. Dissertation no. 234, 1990. T-book reliability data of components in Nordic nuclear power plants.
Pulkkinen U, Simola K. Bayesian models and ageing indicators for Vattenfall AB, S-162, Vallingby, Sweden: ATV Office, 1987.
analysing random changes in failure occurrence. Reliab Eng Tan Z, Xi W. Bayesian analysis with consideration of data uncertainty
Syst Saf 2000; 68: 255–268. in a specific scenario. Reliab Eng Syst Saf 2003; 79: 17–31.
Quigley J, Walls L. Mixing Bayes and empirical Bayes inference Tan Q, Chen G, Zhang L, Fu J, Li Z. Dynamic accident modeling
to anticipate the realization of engineering concerns about for high-sulfur natural gas gathering station. Process Saf
variant system designs. Reliab Eng Syst Saf 2011; 96: Environ Protect. ISSN 0957-5820, http://dx.doi.org/10.1016/j.
933–941. psep.2013.03.004. Available online April 10, 2013.
Quigley J, Hardman G, Bedford T, Walls L. Merging expert and Targoutzidis A. The effectiveness of Bayesian updating in dynamic and
empirical data for rare event frequency estimation: pool complex systems. J Loss Prev Process Ind 2010; 23: 492–497.
homogenisation for empirical Bayes models. Reliab Eng Syst Torres-Toledano JG, Sucar LE. Bayesian networks for reliability analysis
Saf 2011; 96: 687–695. of complex systems. In: Proceedings of the 6th Ibero-American
Rathnayaka S, Khan F, Amyotte P. SHIPP methodology: predictive conference on AI (IBERAMIA 98). Lecture Notes in Artificial
accident modeling approach. Part I: methodology and model Intelligence. vol. 1484. Berlin, Germany: Springer, 1998: 195–206.
description. Process Saf Environ Protect 2011a; 89: 151–164. Troffaes MCM, Walter G, Kelly D. A robust Bayesian approach
Rathnayaka S, Khan F, Amyotte P. SHIPP methodology: predictive to modeling epistemic uncertainty in common-cause
accident modeling approach. Part II: validation with case failure models. Reliab Eng Syst Saf. 2014; 125: 13–21. ISSN
study. Process Saf Environ Protect 2011b; 89: 75–88. 0951-8320, http://dx.doi.org/10.1016/j.ress.2013.05.022.
Rathnayaka S, Khan F, Amyotte P. Accident modeling approach for Vaurio JK. On analytic empirical Bayes estimation of failure rates.
safety assessment in an LNG processing facility. J Loss Prev Risk Anal 1987; 7: 329–338.
Process Ind 2012; 25: 414–423. Vaurio JK. Extensions of the uncertainty quantification of common
Reese CS, Wilson AG, Guo JQ, Hamada MS, Johnson VE. A Bayesian cause failure rates. Reliab Eng Syst Saf 2002; 78: 63–69.
model for integrating multiple sources of lifetime information in Vaurio JK. Uncertainties and quantification of common cause failure
system-reliability assessments. J Qual Technol 2011; 43: 127–141. rates and probabilities for system analyses. Reliab Eng Syst
Robert CP. The Bayesian choice. New York: Springer, 2001. Saf 2005; 90: 186–195.
Røed W, Aven T. Bayesian approaches for detecting significant Vaurio JK, Jänkälä KE. Evaluation and comparison of estimation
deterioration. Reliab Eng Syst Saf 2009; 94: 604–610. methods for failure rates and probabilities. Reliab Eng Syst Saf
Saleh JH, Saltmarsh EA, Favarò FM, Brevault L. Accident precursors, 2006; 91: 209–221.
near misses, and warning signs: critical review and formal Vose, D. Risk analysis: a quantitative guide. Chichester: John Wiley
definitions within the framework of discrete event systems. & Sons, 2000.
Reliab Eng Syst Saf 2013; 114: 148–154. Weber P, Jouffe L. Complex system reliability modelling with
Sarhan A. Bayes estimation of the general hazard rate model. Reliab dynamic object oriented Bayesian networks (DOOBN). Reliab
Eng Syst Saf 1999; 66: 85–91. Eng Syst Saf 2006; 91: 149–162.
Sarhan AM. Reliability estimations of components from masked Weber P, Munteanu P, Jouffe L. Dynamic Bayesian networks
system life data. Reliab Eng Syst Saf 2001; 74: 107–113. modelling the dependability of systems with degradations

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
 A. Roy et al.: Risk assessment in process industries 499

and exogenous constraints. In: Proceedings of the 11th IFAC Wilson AG, Anderson-Cook CM, Huzurbazar AV. A case study for
Symposium on Information Control Problems in Manufacturing quantifying system reliability and uncertainty. Reliab Eng Syst
(INCOM’04). Salvador-Bahia, Brazil, April 5–7, 2004. Saf 2011; 96: 1076–1084.
Weber P, Medina-Oliva G, Simon C, Iung B. Overview on Bayesian Yang M, Khan FI, Lye L. Precursor-based hierarchical Bayesian
networks applications for dependability, risk analysis and approach for rare event frequency estimation: a case of
maintenance areas. Eng Appl Artif Intell 2012; 25: 671–682. oil spill accidents. Process Saf Environ Protect 2013; 91:
Wilson AG, Graves TL, Hamada MS, Reese CS. Advances in data 333–342.
combination, analysis and collection for system reliability Zheng X, Liu M. An overview of accident forecasting methodologies.
assessment. Stat Sci 2006; 21: 514–531. J Loss Prev Process Ind 2009; 22: 484–491.

Brought to you by | University of Utah

Authenticated
Download Date | 7/11/15 1:05 AM
View publication stats