One of the reasons that Human Reliability Analysis methods had been created is because when the first method was created there was not enough data available to estimate Human Error Probabilities. Such methods provide methodologies to adjust the probabilities according to the specific industrial context being assessed (organisational, technological and individual factors). Some examples are THERP, SPAR-H, HEART, CREAM and ATHEANA.
The availability and quality of human error data have improved consistently, thanks to the use of simulators and incident reports. However, the implementation of Bayesian networks models based on the available datasets faces several issues, mainly connected to the definition of adequate conditional probability tables. Indeed, the lack of data regarding the specific error events precludes the possibility of computing the frequency of their occurrence and hence to adopt suitable conditional probability distributions for the related set of event outcomes.
Several strategies have been proposed to tackle this issue in the literature, although none of them has obtained unanimously agreement. This paper proposes a Credal network model for the inference of human error. The main aim of the study is to investigate the capability of the proposed approach to fully capture the uncertainty of the input and to rigorously quantify the accuracy model outputs, improving the robustness of human error inference models. Finally, a numerical model is presented to test the feasibility and efficiency of the approach.