Accounting for dependencies among performance shaping factors in SPAR-H using a regularized autoencoder and WINGS-AISM

Abstract

The standardized plant analysis risk human reliability analysis (SPAR-H) method is widely used for human reliability analysis to adjust the nominal human error probability (HEP) by assigning different multipliers to the performance shaping factors (PSFs). Nevertheless, SPAR-H suffers from assuming PSFs to be independent without considering any overlaps and dependencies. Therefore, this study introduces a new systematic method to analyze the relationships among the PSFs in SPAR-H qualitatively and quantitatively to obtain more reasonable HEP estimation results. The proposed method comprises three primary aspects: 1) a regularized autoencoder for the denoising and feature extraction of expert evaluation results, 2) the weighted influence non-linear gauge system-based adversarial interpretive structure modeling (WINGS-AISM) method to analyze the relationships among the PSFs and construct their causal hierarchy, and 3) a new relative weighting system to modify the PSF multipliers based on this hierarchy. The results of experiments comparing the proposed method with conventional methods highlight that our method effectively reduces the double counting of overlapping PSFs in SPAR-H, providing more reasonable and accurate HEP estimation results.