Capturing cognitive causal paths in human reliability analysis with Bayesian network models
Introduction
A comprehensive probabilistic risk assessment (PRA) is an essential element of safety and reliability assurance for many complex engineering systems. The aim of the PRA is to understand the possible failure scenarios, the corresponding adverse consequences, and the failure scenarios’ probabilities. Most engineering systems can be characterized as human-machine systems, in which the human operator and the technical system are interacting. For that reason it is essential for a PRA to consider not only failures of technical components but also the effect of human actions and human inaction. Human reliability analysis (HRA) models human elements as part of PRAs; in general through identification and quantification of human failure events (HFEs) in PRA models. A variety of methods have been developed and applied in this field to determine human error probabilities (HEPs) corresponding to HFEs. Among the most important representatives are THERP [40], SPAR-H [15] and ATHEANA [7].
The limitations of existing HRA methods have been widely discussed previously [14], [18], [20], [28], [37], [4], [44]. Two interrelated shortcomings in existing HRA methods are the limited scientific basis used to develop those methods and the use of simplified modeling techniques, which lack causal structure and quantitative traceability.
Ongoing research into human performance is addressing the first shortcoming. The scientific foundations for human reliability have been explored and documented in the work by Whaley et al. [43] on the psychological basis of HRA. In particular, they introduce a set of psychological failure mechanisms and proximate causes, which can lead to human failure events. Furthermore, they provide detailed insight into the factors that affect human performance (Performance Influencing Factors, PIFs), the dependency between those factors, and the causal pathways from those factors to human errors. International data collection activities offer insight into human performance in complex engineered systems [31], [6], [8], which provide new opportunities to improve the quantitative basis of HRA.
The second shortcoming, the lack of causal structure and quantitative traceability, is being addressed through advanced modeling efforts. Bayesian Network (BN) models (also called Bayesian Belief Networks), have becoming increasingly popular within HRA as a means for addressing these shortcomings because of their ability to explicitly model cause and effect combined with the ability to incorporate information from different sources [1], [27]. Ongoing international research has demonstrated the ability of BNs both to capture the causal relationships among PIFs and to facilitate quantification of those relationships [16], [29], [33], [39].
The psychological foundation has been leveraged in the development of two new HRA Methods, the IDHEAS (Integrated Decision-Tree Human Event Analysis System) method [45] and the PHOENIX method [10], [11]. Both IDHEAS and PHOENIX introduce the concept of crew failure modes (CFMs), a characterization of ways that a human failure event can occur during a crew interaction with the system. Both methods include a quantitative model relating PIFs to CFMs. However, the quantitative models in IDHEAS fall short of both causal and quantitative traceability; e.g. the motivation for the exclusion of cognitive mechanisms and PIFs from the method remains unclear [36]. The PHOENIX method uses a BN model for quantification, but there are no directed arcs from one PIF to another, and thus the causal paths from the cognitive literature are not fully captured.
In this paper we propose a methodology to expand the scientific basis and traceability of HRA by capturing causal paths from cognitive literature in BN models. Furthermore we present a method for quantifying the BN model using Bayesian parameter updating to combine human performance data with expert elicitation results.
We introduce the methodology by developing a Bayesian network (BN) model for a single CFM from the IDHEAS method. We illustrate the procedure step by step, starting from the corresponding IDHEAS decision tree model, expanding the CFM model to a level where its cognitive foundation is modeled explicitly, and finally reducing the expanded model to a level where its quantification becomes straightforward. This process enhances the traceability between the HRA quantification models and the underlying cognitive literature basis. In addition we provide a method to quantify the new model based on expert elicitation and then show how a database can be used to update these expert elicited distributions, such that the final model is based on both expert knowledge and observed data.
Section snippets
Modeling and quantification tools
This section introduces Bayesian networks (BNs) and Bayesian updating, which provides the foundation for using a combination of experts’ estimates and data for quantification.
Crew failure modes in HRA
Two new HRA methods incorporate the concept of crew failure modes: the IDHEAS method developed by the U.S. NRC, and the PHOENIX method developed by the University of Maryland [10], [11]. PHOENIX and IDHEAS follow a similar modeling approach combining both qualitative and quantitative steps:
- –
Performing a qualitative task analysis and documenting crew failure paths in a crew response tree (CRT).
- –
Selecting applicable crew failure modes (CFMs) for each event in a CRT.
- –
Quantifying the individual CFMs
Development of a BN structure for each CFM
As explained in Section 2, the directed acyclic graph (DAG) part of a BN ideally represents the causal relationships between the random variables in the model. Furthermore, the structure also defines the information (i.e., the marginal and conditional probabilities) needed to quantify the BN. In this section, we illustrate the development of two BN structures for each CFM: a first BN that contains an expanded causal structure based on cognitive literature [43] and PIF specification nodes
CFM BN quantification
This section describes the quantification of the BN structures developed in the previous section. We first present the straightforward quantification of the BN model in Fig. 5 based on the IDHEAS DT and how this simple model can be augmented with expert elicited data about the PIFs. Thereafter we show how the BN of Fig. 7 can be quantified using expert estimates, and finally how information from the SACADA (Scenario Authoring, Characterization, and Debriefing Application) [6] or similar
Updating with data
In this section, we illustrate how the SACADA database [6] could be used to update the probabilities of the CFM node in the IDHEAS-BN. These HEPs in IDHEAS are conditional on the relevant PIFs. Since SACADA and IDHEAS are not completely consistent, it is not always possible to deterministically decide in which states the IDHEAS PIFs are for a given SACADA case. Nevertheless, SACADA still provides information, which can and should be used to improve the quantitative side of IDHEAS. To this end,
Example results with the “critical data misperceived” BN
With the established BN for critical data misperceived, HEPs conditional on different observations are investigated (Table 8). Case I gives the prior HEP before having knowledge about the states of the PIFs or the PIF specification nodes. The states of the PIF specification nodes occur in that case according to the probabilities elicited from the experts. The BN gives reasonable prior HEPs if the CPTs of the PIF specification nodes are elicited (either based on data, experts or similar sources)
Discussion
We present a comprehensive framework for the application of BNs to address shortcomings of HRA with respect to scientific basis and traceability (both causal and quantitative). A main advantage of BNs is that they allow for models that are causally traceable. To this end, unobservable PIFs and concepts from psychology can be included in the BN structure and removed in a later step. Furthermore the quantification of BNs can rely on different information sources, such as data and expert
Conclusion
We propose a framework for developing BN models for HRA directly from causal dependencies found in cognitive literature. The framework is illustrated through the causal paths that were identified during the development of IDHEAS. In order to develop the BN structure, a two-level approach is proposed. In a first step, identified causal paths for a crew failure mode are modeled in a qualitative BN structure. Since quantification of such a BN structure is difficult, the model is reduced in a
Acknowledgment
The authors would like to acknowledge the support from the survey respondents as well as April Whaley (INL), Stacey Hendrickson (SNL), Susan Stevens-Adams (SNL), and the NRC (especially Jing Xing).
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-AC04-94AL85000.
References (45)
- et al.
Comparing the treatment of uncertainty in Bayesian networks and fuzzy expert systems used for a human reliability analysis application
Reliab Eng Syst Saf
(2015) - et al.
Phoenix – a model-based human reliability analysis methodology: qualitativequalitative analysis procedure
Reliab Eng Syst Saf
(2016) - et al.
Human reliability analysis: a critique and review for managers
Saf Sci
(2011) - et al.
A Bayesian method for using simulator data to enhance human error probabilities assigned by existing HRA methods
Reliab Eng Syst Saf
(2014) - et al.
Bridging the gap between HRA research and HRA practice: a Bayesian network version of SPAR-H
Reliab Eng Syst Saf
(2013) Looking for errors of omission and commission or the hunting of the snark revisited
Reliab Eng Syst Saf
(2000)- et al.
Bayesian inference in probabilistic risk assessment—the current state of the art
Reliab Eng Syst Saf
(2009) - et al.
Model-based human reliability analysis: prospects and requirements
Reliab Eng Syst Saf
(2004) - et al.
A virtual experimental technique for data collection for a Bayesian network approach to human reliability analysis
Reliab Eng Syst Saf
(2014) - et al.
OPERA—a human performance database under simulated emergencies of nuclear power plants
Reliab Eng Syst Saf
(2007)
Considerations on the elements of quantifying human reliability
Reliab Eng Syst Saf
Human reliability modeling for the next generation system code
Ann Nucl Energy
On taking human performance seriously in risk analysis: comments on dougherty
Reliab Eng Syst Saf
Probabilistic risk analysis: foundations and methods
Recognition-by-components: a theory of human image understanding
Psychol Rev
The SACADA database for human reliability and human performance
Reliab Eng Syst Saf
Behavioral decision theory: processes of judgement and choice
Annu Rev Psychol
Cited by (60)
Identifying human failure events for human reliability analysis: A review of gaps and research opportunities
2024, Reliability Engineering and System SafetyCausal Graph Attention Network with Disentangled Representations for Complex Systems Fault Detection
2023, Reliability Engineering and System SafetyA novel integrated methodology for human reliability assessment in hydrogen fuelling stations
2023, International Journal of Hydrogen EnergyA traceable process to develop Bayesian networks from scarce data and expert judgment: A human reliability analysis application
2023, Reliability Engineering and System SafetyA framework for quantitative analysis of the causation of grounding accidents in arctic shipping
2022, Reliability Engineering and System Safety