Theory-guided machine learning for licensee event reports of u.S. nuclear power plants to quantify organizational factors in probabilistic risk assessment

Four key questions in this line of research are: (I) how significant is the contribution of organizational factors to accidents and incidents? (II) how critical, with respect to improving risk assessment, is the explicit incorporation of organizational factors into Probabilistic Risk Assessment (PRA)? (III) what theoretical bases are needed for explicit incorporation of organizational factors into PRA? and (IV) what methodological bases are needed for the explicit incorporation of organizational factors into PRA? [1]. SoTeRiA (Socio-Technical Risk Analysis [2]) is used as a theoretical framework for integrating social aspects (e.g., safety culture) and structural features (e.g., safety practices) of an organization with a technical system PRA. This research develops a Data-Theoretic (DT) approach [3] where “data analytics” are guided by “theory." Theory enhances the completeness of “causality” being analyzed from data and helps avoid potentially misleading results from solely data-oriented approaches. DT includes two submodules: (a) DT-BASE that focuses on the development of detailed causal relationships in SoTeRiA, based on a theory-building process equipped with a software-supported BASEline quantification utilizing analyst interpretation of information extracted from articles and standards; and (b) DT-SITE [4,5] that relates to conducting data analytics (text mining) to quantify SoTeRiA causal elements based on industry event databases and by Bayesian updating of the baseline quantification from DT-BASE. A case study is conducted targeting the “training system” in Nuclear Power Plants (NPPs) using Licensee Event Reports (LERs) of U.S. NPPs.