Author: Camille Levine
Date/Time: March 27th, 2025 at 9:30am EST
Location: EGR-2164, Glenn L. Martin Hall | Zoom
Committee members:
Dr. Katrina Groth, Chair
Dr. Mohammad Modarres
Dr. Peter Sandborn
Dr. Yunfei Zhao
Dr. Michelle Bensi, Dean’s Representative
Title of dissertation: Expanding the Causal Logic Foundations of Human Reliability Analysis
Abstract:
Human reliability analysis (HRA) seeks to identify opportunities for operator error in complex engineering systems and quantify the probabilities of these errors occurring. Achieving accuracy and realism in HRA is a challenging process due to the lack of validated models and data available to analysts. First, it is necessary to design HRA methods capable of addressing the complexity of human cognition and the wide variety of operational scenarios where human errors are critical. HRA models should capture the multitudes of causal pathways to failure and their probabilistic nature. Second, error probabilities quantified through data-driven methods result in more accurate, traceable estimates; it is important that HRA methods completely and transparently incorporate human reliability data into the error quantification process.
This project develops robust causal logic models with a strong basis in cognitive psychological research that are quantified using human reliability data and engineering literature. This research has produced three literature-substantiated Bayesian network models that characterize human-machine error pathways and the mechanisms through which they occur. The network structures are parameterized with a variety of data to render them capable of quantifying human failure event (HFE) probabilities. Several data resources were synthesized, including nuclear power plant training simulator data, expert knowledge, HRA dependency idioms, and psychological literature. The contextual information in these quantitative resources allowed derivation of conditional probabilities for performance influencing factors (PIFs), human failure mechanisms, and crew failure modes (CFMs). The applicability of these models is demonstrated through a multi-stage validation. The model structures are validated against pre-analyzed nuclear event narratives from the ATHEANA HRA method, while the quantitative outputs are validated against German nuclear power plant operational experience. Then, the models are holistically evaluated against nuclear power plant simulator data and other established method outputs from the US HRA Empirical Study. Lastly, the models are applied to an external flooding hazard scenario as an extension of the cognitively focused Phoenix HRA method. This work presents a novel application of new and existing HRA data sources to create validated causal models via Bayesian network structures. The resulting models provide a robust technical basis suitable for scenarios beyond those covered in conventional HRA, and are intended for integration into the Phoenix HRA method.