Privacy-Preserving Cross-Silo Federated Learning for Remaining Useful Life Estimation of Distribution Transformers

Due to the growing necessity for a reliable power supply and the aging of distribution transformers (DTs), estimating the remaining useful life (RUL) of these transformers has gained wide attention in the power industry. However, there is an inevitable data shortage that restricts the performance of the estimation. In addition, data sharing sometimes tends to result in the leakage of sensitive information. These are the main issues that need to be resolved currently. For this purpose, a privacy-preserving cross-silo federated learning (FL) method is developed for the RUL estimation of DTs, namely ADP-FedRUL. Specifically, a cross-silo FL architecture enables different data silos to train a global model without the need to share raw data. In this way, it extracts composite health index from multi-source sensor data and integrates the advantages of data to facilitate accurate RUL estimation while preserving privacy. Moreover, a user-level adaptive differential privacy (ADP) strategy is adopted to address potential problems of excessive and insufficient perturbations. Case studies show that the proposed method achieves a favorable balance between privacy budget and estimation performance, with estimation accuracy as high as 98.7 %.