Nuclear power plant fault diagnosis is often complicated by the presence of unknown faults and frequent variations in operating conditions, which pose significant challenges to the reliability and trustworthiness of conventional data-driven diagnosis models. These challenges are especially critical in safety-sensitive nuclear power plants, where incorrect or overconfident decisions can lead to serious consequences. To address the above issues, a novel fault diagnosis framework based on Bayesian Neural Network (BNN) is proposed in this study, which incorporates principled uncertainty analysis to enhance both diagnostic credibility and model self-awareness. Unlike traditional deterministic models, the BNN-based approach outputs probabilistic diagnoses that capture the model's confidence level, enabling not only the accurate identification of known faults but also the effective recognition of unseen faults or domain discrepancy scenarios. Numerical experiments are conducted on simulated high-dimensional and strong-nonlinear nuclear power plant data, on which the framework is systematically evaluated in terms of diagnostic accuracy, confidence distribution, sensitivity to domain discrepancy, and uncertainty quantification and decomposition. Results show that although BNN and CNN achieve comparable accuracy in in-distribution tasks, BNN is significantly superior to CNN in uncertainty expression, domain sensitivity, and unknown fault detection, enabling more trustworthy and interpretable fault diagnosis for real-world nuclear power plants.
扫码关注我们
求助内容:
应助结果提醒方式:
