Hierarchical Federated Learning for Power Transformer Fault Diagnosis
Jun Lin, Jin Ma, Jianguo Zhu
Abstract
Accurate diagnosis of power transformer fault type is critical to maintaining its safe and stable operation. Existing methods require a large number of labeled data and are implemented in a centralized manner. However, the labeled samples are owned by different electricity companies, and they are unwilling to share their private data due to commercial competition and personal data privacy. In this paper, a federated learning (FL) model, called FL-CA-CNN, is proposed to infer the transformer fault type. Specifically, a channel attention-based convolutional neural network (CA-CNN) is designed to bridge the dissolved gases and the fault types. Its model layers are divided into shallow and deep layers. A hierarchical parameter aggregation strategy is adopted to update the model in the FL framework, through which only the model parameters are shared, and the data privacy is protected. Experiments on the actual dataset demonstrate the effectiveness of the proposed method.