An Object Detection Model for Electric Power Operation Sites Based on Federated Self-supervised Learning

In the various edge cloud devices within the power system, a large amount of discrete data is generated, with the most significant amount being image data captured by each construction site. However, due to distance, privacy protection, transmission loss, and other issues, it is difficult to gather these data together for use, resulting in the waste of valuable data resources. Even if these data are collected, annotating them one by one is challenging, leading to limited training effectiveness for the model. As a result, supervision of electric power operation sites still depends on manual remote viewing of surveillance video, which is inefficient and error-prone. To address these issues, this paper proposes the use of Federated Self-supervised learning for object detection at electric power operation sites. This method effectively utilizes limited visual data in each terminal scene to jointly train a neural network model, improving the model's recognition accuracy under a distributed learning framework, increasing data utilization, and ensuring data security. The proposed approach has the potential to enhance the efficiency of electric power production.

I. Introduction

With the popularity of edge cloud technology in power systems, vast amounts of data are generated daily from various terminal devices, among which image data is the most common. They can significantly contribute to the safe production and efficient operation of electric power. However, these image data generated from terminal devices have the following characteristics: 1. They are incredibly discrete in distribution and challenging to aggregate. 2. The amount of data generated from individual devices is small, and the performance of the devices could be improved, making it impossible to train high-performance models locally. 3. The distribution of data is uneven, making it almost impossible to annotate all images.