Conceptual Framework for Data Integrity Assessment for Artificial IntelligenceBased Abnormal Event Diagnosis in Nuclear Power Plants

Research on nuclear power plant state diagnosis has progressed across multiple directions, including diagnostic model development, performance optimization, uncertainty quantification, and explainable artificial intelligence. While these studies aim to improve diagnostic reliability, they commonly assume that the training data used to develop diagnostic models are inherently appropriate. In practice, training data for state diagnosis are generated according to different abnormal causes and their associated severity levels, which often leads to uneven coverage across scenarios, abnormal intensities, and event progression stages. Although these characteristics are frequently treated as secondary issues, they can fundamentally influence diagnostic behavior. Therefore, for artificial intelligence to be credibly deployed in safety-critical domains such as nuclear power plants, it is necessary to explicitly verify whether the training data themselves are appropriate for the intended diagnostic task, in addition to model-level validation. This study proposes a model-agnostic framework to quantitatively assess data integrity prior to diagnostic model development, specifically for nuclear power plant state diagnosis. Data integrity is quantified along three complementary dimensions: scenario integrity, which evaluates coverage and imbalance across target abnormal scenarios; severity integrity, which assesses the distribution of abnormal intensity within each scenario using a normalized multivariate deviation-based severity proxy; and phase integrity, which evaluates temporal coverage relative to event onset. These dimensions are integrated into a unified representation that quantifies discrepancies between the empirical data distribution and a diagnostically justified target distribution. The quantified integrity results are then directly used to guide under-sampling and data augmentation. The proposed framework is intended to serve as a principled basis for integrity-aware data preparation, with detailed validation and performance assessment to be addressed in future studies.

Keywords: Nuclear power plant, safety-critical systems, artificial intelligence, data integrity, abnormal event diagnosis, data augmentation, under-sampling.