Fault diagnosis in reactor coolant pump with an automatic CNN-based mixed model

Fault diagnosis in nuclear reactor coolant pumps (RCPs) is crucial for enhancing the safety of nuclear power plants (NPPs). Data-driven fault diagnosis methods, which can significantly reduce reliance on experts' prior knowledge and enhance the analysis of large volumes of monitoring data, have emerged as a trending topic in fault diagnosis. However, data-driven models, particularly deep learning models, often require experts' experience to optimize hyper-parameters. This process can be time-consuming and offers no guarantees of finding the optimal model. This paper proposes a CNN-based mixed model with automatic neural architecture search (NAS) for fault diagnosis in RCPs. NAS enables the automatic discovery of optimal architectures and parameters based on monitoring data. The CNN-based mixed model is designed to integrate heterogeneous-structured time-series data to extract informative features for fault. The study adopts a space of sequentially structured neural networks as the search space and Bayesian optimization as the search strategy. Model accuracy serves as the evaluation criterion for constructing the NAS algorithm. NAS automatically generates the optimal model for fault diagnosis. The proposed model's performance is validated using seal leakage data from RCPs compared to benchmark methods.