Performance analysis and improvement of data-driven fault diagnosis models under domain discrepancy base on a small modular reactor

Data-driven fault diagnosis has attracted much attention in the field of nuclear energy. The complex operating environment of nuclear power plants tends to cause data to deviate from its own source domain, resulting in domain discrepancy thus performance degradation. So it is crucial to study the performance of the model under domain discrepancy. In this study, two evaluation metrics are proposed for domain discrepancy namely the generalization ability to fault severity and the robustness to data failures. Commonly used fault diagnosis models are built based on ACP100 reactor model, including simple Recurrent Neural Network (RNN), Long Short-Term Memory(LSTM), and Convolutional Neural Networks(CNN). The performance variations of these models are analyzed under the above domain discrepancy. The effect of noise is also analyzed in order to simulate real-world operations. In addition, the model performance is improved by Artificial Disturbance Methods, and the effects of different levels of model noise are analyzed. The LSTM model performs better than the other two models, according to the results under domain discrepancy. The maximum performance gain occurs when 10 dB of noise is introduced. These results provide a reference for the establishment and improvement of future fault diagnosis models for nuclear power plants.