Anomaly detection of rotating machinery in small modular reactors with accelerated life data augmentation

Lead–bismuth eutectic small modular reactors (LBE-SMRs) are considered among the most promising generation IV nuclear energy systems, offering significant advantages in terms of inherent safety and system simplicity because of their specialized coolants. However, early-stage LBE-SMRs often lack operational data, resulting in data shortage scenarios. This data shortage severely hinders the development of effective anomaly detection models for key rotating machinery, such as main pumps and bearings. To address these challenges, a novel anomaly detection framework based on the reconstruction of accelerated life test (ALT) data, named the accelerated life data augmentation network (ALDA-Net), is proposed in this paper. First, a multiscale transformer variational autoencoder (VAE) is employed to extract robust features from the degradation data of key components obtained via the ALT under high-stress conditions. These features are then mapped to actual operating conditions using a life–stress relationship model. Conditional score-based diffusion models for imputation (CSDI) are subsequently introduced to address the missing parts of the mapped data, thereby augmenting a complete dataset. Finally, an anomaly detection model is established using a Wasserstein generative adversarial network with gradient penalty (WGAN-GP). The results on the case study dataset show that the proposed method achieves an average AUC of 99.16%, effectively addressing the problem of anomaly detection under data shortage.