Research on fault diagnosis method and interpretability of nuclear power plant based on hybrid transformer model

Abstract

In order to improve the accuracy of the fault diagnosis, a hybrid Transformer (HTransformer) model for fault diagnosis in nuclear power plants (NPPs) is proposed. By concatenating convolutional neural networks and gated recurrent unit networks in front of the Transformer encoder, the spatiotemporal feature information of the input information is effectively capture. However, the high reliability demand for NPPs requires researchers to be able to explain the decision-making behavior of deep learning models. The interpretability of the HTransformer model has been preliminarily studied based on the Shapley additive explanations (SHAP) method. The result shows that the HTransformer model has a higher fault diagnosis accuracy of 99.5% compared to other deep learning fault diagnosis models. The preliminary research result on the interpretability of fault diagnosis models based on the SHAP method provides a possibility for deep learning models to be applied in the practical engineering of NPPs.