Bayesian inference in probabilistic risk assessment—The current state of the art
Introduction
The advent of Markov chain Monte Carlo (MCMC) sampling has proliferated Bayesian inference throughout the world, across a wide array of disciplines. The freely available software package known as Bayesian inference using Gibbs sampling (BUGS) has been in the vanguard of this proliferation since the mid-1990s [1]. However, more recent advances in this software, leading first to WinBUGS and now to an open-source version (OpenBUGS), including interfaces to the open-source statistical package R [2], have brought MCMC to a wider audience. Problems which would have been intractable a decade ago can now be solved in short order with these software packages. We will use this software to illustrate several applications from the field of risk assessment. While our examples are from applications of risk assessment to technology, it should be noted that there have also been many applications of Bayesian inference in behavioral science [3], finance [4], human health [5], process control [6], and ecological risk assessment [7]. We mention these other applications to indicate the degree to which Bayesian inference is being used in the wider community, including other branches of risk assessment; however, our principal focus is on applications of risk assessment to engineered systems (including non-nuclear technology). Within this focus, we will explore specific Bayesian topics and examples in greater detail, including
- •
Hierarchical modeling of variability
- •
Modeling of time-dependent reliability (with and without repair)
- •
Modeling of random durations, such as the time to suppress a fire or recover power
- •
Treatment of uncertain and missing data
- •
Regression models
- •
Model selection and validation
Siu and Kelly's earlier work [8] is the starting point for this paper, as it presented a tutorial on Bayesian inference for probabilistic risk assessment, and the elementary portions of that paper remain vital today. For the reader needing a more basic introduction to Bayesian inference, that paper is a good starting point, and it also provides a good point of entry to the vast (and vastly expanding) literature on Bayesian inference. However, advances have been made since the publication of that paper in 1998, and it is to these advances that the current paper is devoted. Additional relevant information on the Bayesian approach may be found in [9] and [10] has become a popular practical reference for the PRA community.
Section snippets
Background
First, we provide a brief overview of key concepts relevant to later discussion. In general, we have a goal of performing inference calculations, starting with data. “Data” are the observed values of a physical process and may be subject to uncertainties, such as imprecision in measurement, censoring, and interpretation errors. Next, at a higher level from data, we have “information,” which is the result of evaluating, manipulating, or organizing data and other information in a way that adds to
Hierarchical Bayesian modeling of variability
As discussed by Siu and Kelly [8], treatment of variability that can exist among sources of data is important if total uncertainty, including population variability, is to be properly represented by the resulting posterior distribution. Siu and Kelly [8] discusses hierarchical Bayesian analysis briefly, but (due to software and computer limitations that existed at that time) focuses on an older approach of Kaplan [14] called “two-stage Bayes,” and on (parametric) empirical Bayes. As pointed out
Modeling time trends
It is sometimes the case that the usual Poisson and binomial models are rendered invalid because the parameter of interest (λ or p, respectively) is not constant over time. In the US for example, recent data analysis in [18] has suggested decreasing values of λ and p for several important components. As an example, let us examine valve leakage data from [19]. These data are shown in Table 5.
We first carry out a qualitative check to see if there appears to be any systematic time trend in p. To
Modeling random durations
There are many instances in which time is the random variable of interest. For example, a facility may be threatened by fire, and the time of interest is the time needed to suppress the fire, which must be shorter than the time to damage vital equipment (which is also a random variable). Another important application for many modern facilities is inference on the time needed to restore ac power once it has been lost. And of course such models can be used for the renewal process described above.
Treatment of missing and uncertain data
It is not uncommon in risk applications to encounter situations in which the observed data, which would normally enter into Bayes’ theorem via the likelihood function, are either missing or the exact values are not known with certainty. For example, in the simple Poisson aleatory model used for initiating event frequencies and failure rates, one may not know the exposure time, t, accurately. Sometimes only an interval estimate is available. This is especially the case for failure rates, less
Bayesian regression models
Regression models are commonplace in statistics, and have been applied in risk analysis, as well. An example discussed earlier is the use of a logit or loglinear model for trend in a binomial or Poisson parameter. In these models, time was the predictor variable. In human reliability analysis (HRA), regression models have been used to estimate human error probabilities. The original SLIM (see [34]) and its derivatives are examples. We will examine a simple example from the published literature.
Bayesian model validation
The frequentist approach to model checking or validation typically involves comparing the observed value of a test statistic to percentiles of the sampling distribution for that statistic. Given that the null hypothesis is true, we would not expect to see “extreme” values of the test statistic. One Bayesian approach to model checking involves calculating the posterior probability of the various hypotheses and choosing the one that is most likely. One can also use summary statistics derived from
Other models
A variety of other “models of the world” are amenable to inference via Bayes’ Theorem, including Bayesian belief networks (BBN), influence diagrams, and fault trees. Graph models such as BBNs and influence diagrams are very similar to the directed acyclic graphs discussed earlier. In fact, an influence diagram is just a directed acyclic graph to which decision nodes have been added, as discussed in [47]. An example of these types of models is shown in Fig. 11, where we list six possible
Conclusions
In the scientific and engineering communities, we rely on mathematical models of reality, both deterministic and aleatory. These models contain parameters—whose values are estimated from information—of which data are a subset. Uncertain parameters (in the epistemic sense) are inputs to the models used to infer the values of future observables, leading to an increase in scientific knowledge. Further, these parameters may be known to high precision and thus have little associated epistemic
References (53)
- et al.
Bayesian parameter estimation in probabilistic risk assessment
Reliab Eng Syst Saf
(1998) - et al.
Bayesian models and ageing indicators for analysing random changes in failure occurrence
Reliab Eng Syst Saf
(2000) - et al.
Uncertainty in counts and operating times in estimating poisson occurrence rates
Reliab Eng Syst Saf
(2003) - et al.
Bayesian analysis with consideration of data uncertainty in a specific scenario
Reliab Eng Syst Saf
(2003) - et al.
Foundations of probabilistic inference with uncertain evidence
Int J Approx Reason
(2005) - et al.
Bayesian networks for multilevel system reliability
Reliab Eng Syst Saf
(2007) A fully bayesian approach for combining multilevel failure information in fault tree quantification and optimal follow-on resource allocation
Reliab Eng Syst Saf
(2004)- et al.
Pitfalls in risk calculations
Reliab Eng
(1981) WinBUGS—a Bayesian modelling framework: concepts, structure, and extensibility
Stat Comput
(2000)- R Development Core Team. R: a language and environment for statistical computing [Online] 2.6.1, R Foundation for...
Bayesian methods: a social and behavioral sciences approach
Contemporary Bayesian econometrics and statistics
Statistical methods in spatial epidemiology
Bayesian process monitoring, control, and optimization
Hierarchical modelling for the environmental sciences
Probability theory—the logic of science
Handbook of parameter estimation for probabilistic risk assessment
An introduction to Bayesian inference and decision
The concept of probability in safety assessments of technological systems
Science
On a two-stage Bayesian procedure for determining failure rates
IEEE Trans Power Apparatus Syst
Probabilistic reasoning in intelligent systems: networks of plausible inference
Inference and monitoring convergence
Industry-average performance for components and initiating events at US commercial nuclear power plants
Repairable systems reliability
Cited by (213)
Dynamic risk assessment model for third-party damage to buried gas pipelines in urban location class upgrading areas
2023, Engineering Failure AnalysisDeveloping a framework for generating production-dependent failure rate through discrete-event simulation
2023, International Journal of Production EconomicsTravel time variation and its impact on the elasticity of route choice model parameters
2023, Travel Behaviour and SocietyUsing experts’ opinion for Bayesian prior reliability distribution of on-demand equipment: A case study of a novel sliding sleeve valve for open-hole wells
2023, Reliability Engineering and System Safety