Answers are grounded in this paper and cite their sources.
The generalization ability of data–driven fault diagnosis models is critically challenged in Nuclear Power Plants (NPPs) due to domain shifts arising from varying operating conditions. In existing studies, model generalization ability is primarily enhanced through modifications to model architectures or learning algorithms, whereas the impact of data representations on generalization ability is rarely investigated. To address the above limitations, an integrated framework is proposed to analyze the impact of data representations on generalization ability and the potential mechanisms. By constructing different representations of the raw data through various feature scaling methods, the impact of data representations on model generalization ability is assessed via systematic training and evaluation. Furthermore, the mechanisms by which data representations influence model generalization ability are elucidated by SHAP (SHapley Additive exPlanations)-based interpretability analysis from both global and local perspectives. Distinct feature–dependence patterns across trained models are induced by disparities in the relative distributions of Domain–Invariant Features (DIFs) and Domain–Sensitive Features (DSFs), thereby resulting in variations in generalization ability. By systematically characterizing the influence of data representations on model generalization ability and elucidating the potential mechanisms that lead to performance variations, novel insights and practical guidance are provided for enhancing the generalization ability of fault diagnosis models in NPPs, which can be achieved by obtaining more generalizable data representations through optimized feature scaling methods.
No takes yet. Share an insight, caveat, or question.
Niu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: