Boosted HyperNet neural Min-Sum decoder: LDPC error floor mitigation in Space-Air-Ground integrated networks

Abstract

Ensuring extremely high reliability and efficiency of communication in dynamic heterogeneous environments is critical for Space-Air-Ground Integrated Networks (SAGIN). To address the error floor phenomenon encountered in low-density parity check (LDPC) codes, we design an innovative decoder based on HyperNet and boosted learning: boosted HyperNet Neural Min-Sum (BHNMS) decoder. This decoder features: (1) Layered Activation Mechanism dynamically activates HyperNet to solve the explosion of training resources due to forward computation required for HyperNet, multi-stage training for boosted learning, and block-wise iterative chunk training; (2) Quantization-aware weight generation relies on differentiable quantization gates, dynamic precision-aware training, and quantization domain consistency constraints to solve the compatibility disconnect between HyperNet’s floating-point computation and quantization weight generation; (3) Cross-stage State Relay achieves continuous reconstruction of node evolution trajectories across stages through spatiotemporal compression transmission and adaptive accuracy control to solve the dynamic feedback imbalance issue; (4) Block Gradient Bridge reconstructs the global gradient propagation path through learnable bridge parameters and cross-block consistency constraints, ensuring the end-to-end trainability of dynamic weight generators to solve the problem of gradient propagation disruption. Simulation shows that BHNMS does not exhibit significant error floors, and its performance far exceeds that of traditional decoders, reaching ${10}^{-9}$ earlier