Robust data-driven human reliability analysis using credal networks

Despite increasing collection efforts of empirical human reliability data, the available databases are still insufficient for understanding the relationships between human errors and their influencing factors. Currently, probabilistic tools such as Bayesian network are used to model data uncertainty requiring the estimation of conditional probability tables from data that is often not available. The most common solution relies on the adoption of assumptions and expert elicitation to fill the gaps. This gives an unjustified sense of confidence on the analysis. This paper proposes a novel methodology for dealing with missing data using intervals comprising the lowest and highest possible probability values. Its implementation requires a shift from Bayesian to credal networks. This allows to keep track of the associated uncertainty on the available data. The methodology has been applied to the quantification of the risks associated to a storage tank depressurisation of offshore oil & gas installations known as FPSOs and FSOs. The critical task analysis is converted to a cause-consequence structure and used to build a credal network, which extracts human factors combinations from major accidents database defined with CREAM classification scheme. Prediction analysis shows results with interval probabilities rather than point values measuring the effect of missing-data variables. Novel decision-making strategies for diagnostic analysis are suggested to unveil the most relevant variables for risk reduction in presence of imprecision. Realistic uncertainty depiction implies less conservative human reliability analysis and improve risk communication between assessors and decision-makers.