A Bayesian two-stage approach to integrate simulator data and expert judgment in human error probability estimation. (March 2023)

• Record Type: Journal Article
• Authors: Greco, Salvatore F.; Podofillini, Luca; Dang, Vinh N.
• Is Part Of: Safety science. Volume 159(2023)
• Publication Date: 2023-03
• Subjects: Human reliability analysis -- Nuclear power plant -- Simulator data -- Expert judgment -- Bayesian models -- Population variability; Industrial accidents -- Periodicals; Accident Prevention -- Periodicals; Safety -- Periodicals; Travail -- Accidents -- Périodiques; 363.11
• DOI: 10.1016/j.ssci.2022.106009
• Languages: English
• ISSNs: 0925-7535

Abstract

Highlights: Empirical basis of HRA comes through traceable integration of data and judgment. A two-stage Bayesian model is proposed for data and judgment aggregation. The model can be used to enhance results of HRA methods. A way to link human failure probability to plant-specific evidence is presented. Abstract: With the ongoing efforts to collect new data for Human Reliability Analysis (HRA) (in particular, from nuclear power plant control room simulators), it becomes important that the coming data will be processed traceably, addressing its underlying variability, eventually in combination with expert judgment. In this direction, this work presents a two-stage Bayesian model to integrate expert-elicited probability estimates and empirical evidence from simulator data in the quantification of HEP values and of the associated variability distributions. The general aim is to provide a data aggregation framework able to mathematically combine diverse information sources throughout the HEP estimation process, in a systematic and reproducible way, contributing to strengthening the empirical basis of future HRA methods. The Bayesian model can be used to produce reference values and bounds for HRA methods as well as to improve the quality of plant-specific HEP estimates for use in Probabilistic Safety Assessment applications. The model is first verified with artificial data and then applied to quantify the HEP of human failure events from literature. Model sensitivity to biases in expert estimates is also investigated.