Orbital-Edge-Assisted Adaptive Neural TOR Detection for LEO Satellite Massive MIMO Systems

This paper proposes an orbital-edge-assisted TOR-Net framework for uplink massive multiple-input multiple-output (MIMO) systems in low Earth orbit (LEO) satellite communications. Reliable detection in LEO massive MIMO is difficult because channels can be ill-conditioned and affected by phase uncertainty and atmospheric fluctuations, whereas conventional relaxation detectors require sensitive parameter tuning and many iterations. The proposed method unfolds the three-parameter over-relaxation (TOR) recursion into a trainable TOR-Net, where relaxation parameters are learned while the low-complexity triangular-update structure is preserved. To avoid costly onboard retraining, a condition-aware orbital-edge model-management scheme is developed. Condition-specific models are trained offline in a terrestrial cloud, stored at a nearby orbital edge node, and delivered to the receiver LEO satellite through an inter-satellite link. Simulation results show improved bit-error-rate performance, reduced complexity, lower onboard energy consumption, and compact model-transfer overhead for LEO massive MIMO detection.