Foreign Object Detection of Electric Transmission Line with Dynamic Federated Learning

Abstract

The bird’s nests and other foreign objects on the transmission lines bring about huge threats to the electric power security. The aggregation of foreign objects data of different regions can boost the generalization of foreign objects with deep learning. However, due to policy constraints and other data security reasons, it is impossible to collect the data of different power companies to train a joint foreign object detection model. Therefore a dynamic federated learning based foreign object detection method for transmission lines is proposed. For the clients distributed in power companies, each model is trained and uploaded to the central server in an asynchronous way for dynamic fusion. For the model trained with pre-trained model, the pre-trained part is frozen and will not be trained. Thus the feature extraction backbone is not uploaded to central sever, which contributes to the decrease of the communication consumption by reducing the amount of uploaded data. The experimental results demonstrate that the dynamic federated learning based foreign object detection algorithm can maintain the same accuracy level compared with centralized training. It can train the joint model without uploading the original data of the edge nodes, which guarantees the security and privacy of local data. Compared with training individually, it has great improvement in accuracy. And it is also applicable with different network sizes. Therefore the dynamic federated learning based foreign object detection has strong privacy, accuracy, efficiency, flexibility and scalability.