Fault Diagnosis for Power Distribution Network Line via Cloud-Edge Collaborative Federated Deep Learning

Abstract

Fault diagnosis methods for power distribution networks face efficiency and applicability challenges with large-scale data and complex patterns. A Cloud-edge Collaborative Federated Deep Learning (CCFDL) method is proposed for power distribution network line fault diagnosis. First, an effective cloud-edge collaborative federated learning framework is established, enabling distributed training where participants train locally and aggregate parameters to build a global model, ensuring data privacy while improving accuracy. Second, a novel pooling self-attention network and deep learning model is introduced. This model combines the pooling self-attention, which reduces computational complexity by integrating pooling and self-attention, and the deep star learning network, which utilizes depth-wise separable convolution and gating mechanisms for efficient feature representation. Experimental results show CCFDL achieves an average diagnostic accuracy of 98.6% with a running time of 10.70 seconds, outperforming most existing models.