Federated-Transfer Learning for Scalable Condition-based Monitoring of Nuclear Power Plant Components

Condition-based monitoring (CBM) techniques are widely being used for maintenance activities in nuclear power plants (NPPs). As faults are rare events, it is highly unlikely that all potential fault modes are captured for a single component. In addition, fault signatures extracted from a single component cannot be robust enough to handle unseen fault patterns. On the other hand, privacy, security, legal, and commercial concerns restrict data-sharing across different plant systems/different systems within one plant. This research presents federated-transfer learning (FTL) to scale machine learning (ML) models for CBM across a component or plant system by combining federated learning (FL) and transfer learning (TL) approaches. FL enables a centralized server to develop an aggregated global CBM model, while the training data are safely and privately distributed on the devices of plant systems, and TL enables application of the developed aggregated model to different but related systems within the same plant site, or to the same system at different plant sites. FTL was demonstrated using circulating water system data for two plant sites-one two-unit plant and one single-unit plant-to predict the health condition of a circulating water pump. The FTL framework was verified using a multikernel adaptive support vector machine and an artificial neural network. The results show significant improvement in prediction performance, reducing overfitting issues and data heterogeneity.