Aiming at the limitation of aerial insulators detection by Unmanned Aerial Vehicle(UAV) combined with deep learning technology, we propose a federated learning framework for aerial insulators orientation detection, dubbed FedIOD. It addresses the challenge of utilizing original insulator image data, which cannot be involved in centralized deep learning model training due to the reasons of policy and confidentiality, as well as horizontal anchor boxes select insulators and their defective regions inaccurately. In FedIOD, we improve the head structure and loss function of the YOLOv5 algorithm, so that it can detect insulators and their defective parts directionally, and add an attention mechanism module to its backbone network to enhance the feature extraction capability. In addition, FedIOD trains the global model collaboratively based on federated learning, and the parameters are aggregated by using the weighted average method. The experimental results show that the precision and recall of FedIOD can reach above 79.16% and 80.06%, respectively. FedIOD has a great practical significance for building an intelligent grid equipment detection system.