Interpretable Bayesian-optimized Autoencoder for fault detection and diagnosis with application in nuclear power plants

Nuclear energy, as an essential clean energy source for combating climate change, is directly tied to sustainable energy development and societal safety through its safe and stable operation. However, fault diagnosis in nuclear power plants faces challenges such as large data volumes, scarcity of fault data, and complex multi-parameter correlations. To address these challenges, this paper proposes an interpretable Bayesian-optimized Autoencoder fault diagnosis model. The model relies solely on normal operating data and accurately diagnoses anomalies through abrupt changes in reconstruction error, overcoming the limitations of traditional methods that depend on fault data. By dynamically adjusting the hyperparameters of the Autoencoder using Bayesian optimization, the model enhances its adaptability to complex time-series data and improves diagnostic precision. Additionally, interpretable machine learning techniques are introduced to quantify the contributions of key parameters to faults, providing clear insights into fault causes and offering scientific support for subsequent engineering interventions. Experimental results demonstrate that the proposed method achieves fault diagnosis F1 scores exceeding 95% on both real-world operational data and simulation datasets from nuclear power plants, significantly outperforming traditional methods. In addition, the model can effectively evaluate the severity of different fault conditions and analyze the cause of the failure, demonstrating strong practical application potential.