A Two-layer Federated Learning for LEO Constellation in IoT

The rapid growth of low earth orbit (LEO) satellite communications has become pivotal in the realm of satellite Internet of Things (IoT), facilitating data collection and transmission. Due to the wide coverage and large volume of data collected in satellites, it can be used as training data for machine learning. Federated learning (FL) emerges as an apt paradigm for collaborative training across multiple satellites, mitigating the need for excessive communication resources. However, the high mobility of LEO satellites introduces communication asynchrony, posing challenges in FL, such as learning delays. To address this challenge, we propose a two-layer FL architecture with an innovative asynchronous learning design. This design not only mitigates learning delays but also reduces transmission scheduling complexity. Simulation results show that compared with the benchmarks (FedAvg and FedAsync), our proposed scheme has better performance in accuracy and learning speed.