Autonomous Fault Diagnosis in Reactor Coolant Pump with a Mixed Deep Learning Model

Abstract

Fault diagnosis in nuclear reactor coolant pump is of great significance to improve the safety of nuclear power plant. Data-driven fault diagnosis methods which may greatly enhance the ability to analyze large amount of monitoring data, have become a trending topic in fault diagnosis. However, data-driven models need the experience of experts to optimize the model hyper-parameters. It consumes a lot of time without any guarantees to find the optimal model. This paper proposes a CNN-based mixed model with automatic neural architecture search (NAS) for fault diagnosis in reactor coolant pump. The CNN-based mixed model is proposed to fuse heterogeneous-structured time-series data to extract informative features for characterizing the faults. By properly determining the search space, search strategy and model evaluation measure, NAS automatically outputs the optimal model for fault diagnosis. The performance of the proposed model is verified with the seal leakage data of the reactor coolant pump.

Keywords: CNN, Fault diagnosis, Mixed model, Neural architecture search, Reactor coolant pump