Ensuring the safety and reliability of high-voltage electrical equipment requires robust and accurate online monitoring systems for detecting partial discharges (PDs) in switchgear. However, the transmission and storage of data in such systems raise significant concerns regarding data security and privacy. To address these issues, federated learning (FL) has gained popularity by enabling collaborative model training across clients without sharing raw data. Unlike previous approaches that compute model aggregation weights based solely on individual client performance, we propose a novel FL framework that evaluates client models using a strategy based on maximum diversity and minimum redundancy, resulting in a more informative and representative global model. Furthermore, a spatial-logic alignment module is introduced to enhance client model personalization for local data, with spatial alignment reinforced via a novel knowledge distillation mechanism. The proposed framework features a hybrid architecture composed of leaf clients, branch clients, and a central server. This structure allows branch clients to benefit from the aggregated global knowledge while preserving privacy among isolated clients. Leveraging edge computing at the client level ensures fast data processing and low-latency decision-making, thereby improving its monitoring capabilities. We validate our framework on a custom-built PD dataset and three public datasets. Experimental results demonstrate that our method outperforms the state-of-the-art approaches.
Ji et al. (Thu,) studied this question.