Foreign Object Detection for Transmission Lines Based on Multi-Site Federated Learning

Foreign object detection plays a crucial role in power grid operation and maintenance, directly affecting the safety, stability, and efficiency of grid operations. However, existing foreign object detection systems for transmission lines face heterogeneous data distribution across sites, leading to challenges in federated learning. To address this issue, we propose a federated learning-based detection algorithm that includes a Vision Transformer (ViT)-based foreign object recognition model and a dynamic client gradient aggregation strategy. The proposed model leverages self-attention mechanisms to process global image features, eliminating reliance on local convolutional operations, thereby improving performance in foreign object recognition tasks. The federated learning method employs a dynamic gradient aggregation strategy to effectively resolve client heterogeneity, enhancing detection efficiency and ensuring data security. Experimental results demonstrate that compared to local client training, our federated learning approach mitigates overfitting caused by insufficient data, improving model accuracy by 42.5%. Compared to the gradient averaging strategy, the dynamic gradient aggregation strategy further increases model accuracy by 19.1%.