Temporal causal graph-based attention gated recurrent unit for interpretable fault diagnosis in nuclear power plants

Abstract

Intelligent fault diagnosis is crucial for safe operation of nuclear power plants (NPPs), yet challenges remain due to complex architectures and dynamic time-series data. Besides, the black-box nature of conventional data-driven models also hinders their practical deployment in fault diagnosis scenarios. To address these issues, this paper proposes a novel temporal causal graph-based gated recurrent unit with cell-level graph attention (GraphAttGRU-Cell) network for interpretable and efficient system-level fault diagnosis in NPPs. A temporal causal discovery method integrating domain expertise with data is first adopted to learn stable and credible temporal causal graph that incorporate both instantaneous and lagged dependencies from multivariate sensor data. The proposed GraphAttGRU-Cell Network is designed from the bottom up at the cell level, in which the attention and gated mechanisms can capture spatial and temporal features within the graph. The proposed method is validated by a simulated Gen-IV NPP simulation accident dataset. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art baselines in accuracy and interpretability, offering a robust and explainable tool for NPP maintenance decision-making.