Fault diagnosis of oil immersed transformers based on federated learning and imbalanced sample data

This paper proposes a fault diagnosis mechanism for oil immersed transformers based on federated learning and unbalanced sample data to address the problem of data silos and difficult diagnosis in actual industrial environments. Unlike traditional deep learning centralized training methods, this paper adopts a multi participant distributed training approach. Each participant uses LeNet-5 deep neural network for local training, and the central server aggregates the local models. At the same time, an improved data sharing strategy is introduced to reduce the imbalance of non-independently identifiable distribution (Non IID) data. The experimental results demonstrate that this mechanism achieves collaborative training of various transformer diagnostic participants under independent and unbalanced data, and the diagnostic accuracy of transformer fault types can reach 97%.