Unsupervised clustering research of nuclear power plants under unlabeled unknown fault diagnosis scenario

Nuclear power plants (NPP) must promptly and accurately identify all types of faults due to the radioactive hazards in the event of accidents. Most existing data-driven fault diagnosis methods are supervised learning techniques that require labeled data for known faults. However, the complex operational characteristics of NPP cause unknown faults that are not trained in practice. Therefore, studying the fault diagnosis algorithm for the automatic clustering of unlabeled unknown faults is of great significance. To address this issue, we propose a two-step method called Double Weighting of positive and negative samples for Clustering by Adopting Neighbors (DWCAN). By utilizing positive sample expansion and adaptive weighting of negative samples, the existing contrastive learning (CL) method is improved to more effectively capture the trends and local similarities in the evolution of system faults. Numerical experiments show that DWCAN can achieve superior unsupervised clustering results and realize automatic decision of unknown number of faults. In addition, this study reveals that using weak and strong data augmentation strategies separately can better leverage the superiority of CL. In general, this method can extend the application of NPP fault diagnosis to unlabeled unknown fault data and lay the foundation for the further development of a continuous fault diagnosis system.