Adaptive Personalized Federated Reinforcement Learning for Satellite Formation Fault Diagnosis

This paper introduces an approach for diagnosing faults in satellite formations based on adaptive personalized federated reinforcement learning . Firstly, Markov decision model of fault diagnosis problem is established, and multi-agent reinforcement learning method is used for local training. Secondly, the adaptive weights α
i
are added to the federated learning framework to mix local and global models. This method can generate a personalized model for each satellite, reduce the drift of the local model under the influence of non-iid distribution of different satellite data, and improve the generalization ability of satellite fault diagnosis. Finally, a simulation environment was created using unity to embed the real satellite attitude model into the environment. The results show that the algorithm has high diagnostic accuracy and fast convergence speed.