With the rapid advancement of artificial intelligence technology, fault diagnosis methods for nuclear power plants (NPP) based on deep learning have made significant research progress. However, the robustness and interpretability of most existing methods are insufficient, and when there are distribution shifts or noise interference between the test and training data, the diagnostic accuracy of the model is significantly reduced. Moreover, the “black-box” nature of deep learning methods makes it difficult for experts to explain the diagnostic results, thereby limiting their practical applicability. To address these challenges, this paper proposes a highly robust fault diagnosis method for NPP based on post hoc interpretability analysis. First, a multi-scale stacked Transformer is designed as a temporal feature extractor. Second, the Shapley Additive Explanations (SHAP) method is employed for post hoc interpretability analysis, identifying key parameters that significantly contribute to the classification results of the model. Finally, the identified key parameters are used to retrain the model, resulting in a highly robust temporal fault diagnosis model. Numerical experiments on complex NPP simulation data demonstrate that the retrained model achieves significantly improved testing accuracy across different power levels and exhibits enhanced resilience to various types of noise interference. Overall, the proposed method offers substantial performance advantages and application potential, laying the groundwork for further model refinement and optimization based on post hoc interpretability analysis results.