Abnormal Object Detection of the Transmission Lines with YOLOv5 and Federated Learning
Kaihong Zhang, Yimin Zhou · 2023
Detecting abnormal object intrusions of the transmission lines in a timely and effective manner is of great significance for the safe operation of the electrical power grid. However, due to the particularity and privacy policy of the power grid companies, there is a lack of available and effective dataset with open source. Besides, individuals tend to gather data in their commonly own regions, which would lead to data silos. In this paper, a dataset of the abnormal object intrusions of the transmission lines is first configured to meet the model training requirements. An improved abnormal object detection (AOD) method is proposed utilizing the federated learning. The FasterNet backbone and partial convolution are incorporated to improve the YOLOv5 based model detection speed. Further, the loss function Wise-IoU (WIoU) and Squeeze & Excitation Networks (SE) are applied to enhance the detection precision. Finally, the distributed training strategy of the federated learning is applied to overcome the data silos. Through the ablation experiments, practicality together with merits of the suggested approach are validated on the federated learning platform, achieving a delicate balance between the detection accuracy and speed with comparison analysis for the abnormal object detection in the transmission lines.