Federated learning-based ISAC network in cohesive clustered satellite: resource optimization in heterogeneous datasets and systems

With the continuous advancement of space technology and the rapid increase in the number of low Earth orbit satellites, cohesive clustered satellites (CCS) are thriving. Additionally, the growing computational capabilities of onboard satellite equipment have enhanced constellations’ data processing power. These developments have made federated learning (FL)-based CCS (CCSFL), a feasible and promising approach. Therefore, this paper proposes a CCSFL-based integrated sensing and communication (ISAC) network, where FL convergence, transmission latency, and energy consumption are optimized using a deep reinforcement learning (DRL) approach under heterogeneous datasets and system conditions. To further enhance communication performance, we adopt intra-orbit inter-satellite link (ISL) multi-hop routing, inter-orbit ISL neighbor forwarding, and sparse gradient compression techniques. Specifically, we introduce a utility function based on the sensing signal-to-noise ratio (SNR) as a reward for the double deep Q-network (DDQN) algorithm, addressing the optimal client selection problem under heterogeneous datasets and systems. Additionally, we employ the deep deterministic policy gradient (DDPG) algorithm to optimize system-wide latency and energy consumption. Simulation results show that the proposed algorithm outperforms the benchmark in both FL accuracy and resource utilization.