Enhanced learning for nuclear power plant condition diagnoses using information metric based on Bayesian neural networks and UMAP

This study introduces an enhanced active learning framework utilizing Bayesian neural networks for nuclear power plant condition diagnoses. A novel multi-component information need metric combining uncertainty, density, entropy, and diversity with adaptive weighting is proposed to efficiently identify informative training samples. Validation using the International Atomic Energy Agency's integral pressurized water reactor simulator with 26 abnormal conditions across 10 independent runs demonstrates significant improvements. Our approach achieved 70.24 ± 2.18 % accuracy, outperforming both BALD (67.03 ± 2.12 %) and random selection (52.68 ± 5.62 %). Notably, the method required 50 % fewer samples than random selection to achieve equivalent performance, while maintaining superior stability. F1-Macro scores (0.726 vs BALD's 0.672) indicate balanced performance across all conditions. UMAP visualizations confirmed that the multi-component metric effectively identifies samples at decision boundaries and sparse regions. The framework demonstrates that integrating multiple information components provides superior accuracy, stability, and data efficiency compared to single uncertainty-based methods, making it valuable for safety-critical applications where labeled abnormal data is scarce and limited.