Communication – Efficient Federated Learning for LEO Satellite

In the realm of low Earth orbit (LEO) satellite constellations, the burgeoning need for robust, real-time data processing and machine learning (ML) capabilities to address global challenges is palpable. Traditional centralized ML confronts significant challenges in satellite networks. The bandwidth constraints of download links have a substantial impact on how long data takes to transfer between satellite and ground servers. This delay limits the ability to handle data in real time, which is crucial for many satellite-based applications. The connection of download links to satellites is erratic and irregular, resulting in relatively brief communication periods with ground stations. This limitation restricts the quantity of data that may be transmitted inside each window, hindering the timely and effective processing of satellite data. Satellites collect massive volumes of data, most of which is redundant or useless to specific purposes. Downloading this large amount of raw data to ground stations for processing is not only inefficient, but also impracticable due to bandwidth limitations. The bandwidth constraints of download links have a substantial impact on how long data takes to transfer between satellite and ground servers. This delay limits the ability to handle data in real time, which is crucial for many satellite-based applications. The connection of download links to satellites is erratic and irregular, resulting in relatively brief communication periods with ground stations. This limitation restricts the quantity of data that may be transmitted inside each window, hindering the timely and effective processing of satellite data. Satellites collect massive volumes of data, most of which is redundant or useless to specific purposes. Downloading this large amount of raw data to ground stations for processing is not only inefficient, but also impracticable due to bandwidth limitations. Federated Learning (FL) has gained traction in academia and industry in recent years. FL allows ground stations and satellites to collaborate on training a global ML model without having to share the raw images captured by the satellites. This strategy dramatically decreases the need to download vast amounts of data, addressing the bandwidth and latency issues associated with classic ML methods. However, federated learning still suffers from inevitable high communication latency, especially in satellite networks. To overcome this challenge, we have proposed a revolutionary approach to improving communication efficiency for ML training in satellite networks by using Delayed Gradient Averaging (DGA) algorithm. This technology enhances the scalability and responsiveness of satellite networks, allowing for more robust and efficient federated learning applications in space-based systems. Moreover, our technique takes advantage of intra-plane inter-satellite links (ISLs) to build intra-orbit aggregations, which can further minimize communication costs. Numerical findings show that our proposed framework outperforms various baselines, demonstrating a significant progress in making satellite networks more efficient for machine learning applications. This novel technique, which combines DGA and intra-orbit aggregation via ISLs, opens new possibilities for using satellite networks to address global concerns by improving data processing and analytic capabilities.