Hybrid data augmentation for uncertainty mitigation in machine learning based fault diagnosis of pressurized water reactor nuclear power plants

Uncertainty in machine learning-based fault diagnosis for nuclear power plants (NPPs) poses a critical challenge due to limited and noisy operational data. This study proposes two novel hybrid data augmentation techniques, spatiotemporal variation augmentation (STVA) and dynamic noise-augmented warping (DNAW) designed to enhance training data diversity and realism by integrating standard data augmentation techniques such as amplitude scaling, jittering, time warping, and controlled noise injection in structured sequences. The proposed methods significantly improve model robustness to inherent data noise and reduce overall predictive uncertainty in noisy environments. An attention-based long short-term memory (AT-LSTM) network, combined with Monte Carlo dropout at inference, serves as the classification model and enables entropy-based decomposition of predictive uncertainty into aleatoric (data-inherent) and epistemic (model-related) components. Results show that AT-LSTM models trained with the proposed hybrid data augmentations substantially outperform standard data augmentation approaches and the non-augmented baseline, achieving near-perfect classification accuracy (¿99.5%) together with reduction of estimated aleatoric and epistemic uncertainty (on the order of 1 0 − 5 ) and exceptionally high prediction confidence (.9999) on unseen data with up to 10% Gaussian noise. These findings highlight the vital role of hybrid data augmentation in mitigating the effects of irreducible data noise and model limitations, thereby enabling the development of more trustworthy diagnostic systems for safety-critical nuclear applications.