A data-based comparison of BN-HRA models in assessing human error probability

Abstract

Bayesian Network (BN) has been increasingly exploited to improve different aspects of Human Reliability Analysis (HRA), resulting in a new generation of HRA techniques, known as BN-HRA models. However, validating and evaluating the accuracy of BN-HRA models is still a challenging task. In this study, we have assessed and compared the performance of some of well-known BN-HRA techniques using human performance data obtained from an offshore evacuation simulation. Based on the role of data in quantifying the BN-HRA models, three categories of BN-HRA models have been considered: (i) BN-CREAM and BN-SPARH, which are based on predefined rules (rule-based methods), (ii) Bayesian Parameter Learning (BPL), which is entirely based on the available data (data-based method), and (iii) BN-SLIM model which is based on both the available data and the predefined rules (hybrid method). The results of the present study show that the data-based methods, i.e., BN-SLIM and BPL, in general outperform the rule-based methods. Cross-validation analysis further demonstrates the superiority of BN-SLIM over BPL, particularly in case of data scarcity.