Deriving causal Bayesian networks from human reliability analysis data: A methodology and example model

Abstract

Within the probabilistic risk assessment community, there is a widely acknowledged need to improve the scientific basis of human reliability analysis (HRA). This has resulted in a number of independent research efforts to gather empirical data to validate HRA methods and a number of independent research efforts to improve theoretical models of human performance used in HRA. This paper introduces a methodology for carefully combining multiple sources of empirical data with validated theoretical models to enhance both qualitative and quantitative HRA applications. The methodology uses a comprehensive set of performance influencing factors to combine data from different sources. Further, the paper describes how to use data to gather insights into the relationships among performance influencing factors and to build a quantitative HRA causal model.  To illustrate how the methodology is applied, we introduce the Bayesian network model that resulted from applying the methodology to two sources of human performance data from nuclear power plant operations. The proposed model is introduced to demonstrate how to develop causal insights from HRA data and how to incorporate these insights into a quantitative HRA model. The methodology in this paper provides a path forward for carefully incorporating emerging sources of human performance data into an improved HRA method. The proposed model is a starting point for the next generation of data-informed, theoretically-validated HRA methods.