An interpretable deep transfer learning method for fault diagnosis of nuclear power plants under multiple power level conditions

Abstract

Nuclear power plants (NPPs) operations under different power level conditions (i.e., different operating modes) often exhibit non-independent and identically distributed (non-IID) characteristics in their fault-related parameters, posing significant challenges to traditional data-driven fault diagnosis methods. To address this issue, the study proposes a fault diagnosis approach that combines deep transfer learning with an interpretable multi-variable gated recurrent unit (IMV-GRU) model. The proposed approach incorporates a hybrid loss strategy integrating adaptive focal loss (AFL) and maximum mean discrepancy (MMD) to improve cross-power-level feature transfer capability. The interpretability of IMV-GRU is demonstrated through its autonomous quantification of multi-variable contribution degrees, enabling feature selection to optimize computational efficiency and mitigate interference from non-key variables. Experimental results demonstrate that the proposed method is effective in cross-power-level fault diagnosis, with particularly significant accuracy improvements under sparse data conditions. Furthermore, the effectiveness of extracting multi-variable contribution degrees is validated, highlighting its value in fault diagnosis.