Network telemetry is critical for low earth orbit (LEO) mega-constellation networks, where frequent topology changes and link disruptions make timely state collection challenging. Existing telemetry methods often suffer from limited coverage, high collection overhead, and poor observability resilience to link and node failures. This paper presents HyTel, a self-healing hybrid telemetry system designed for LEO mega-constellation networks. HyTel combines the controllability of active out-of-band probing with the fine-grained measurement capabilities of in-network telemetry and enables self-healing telemetry. In particular, HyTel proposes ring-based and tree-based telemetry mechanisms to balance probing cost, measurement coverage, and failure resilience. The ring-based mechanism provides low-overhead telemetry along orbital paths, while the tree-based mechanism improves observability resilience and failure localization through branch-based probing. We further design a lightweight telemetry protocol with a temporary-monitor-based recovery strategy for restoring observability after failures. Experiments on multi-plane, multi-shell dynamic constellations and analyses of representative failure scenarios compare HyTel with two-way active measurement protocol (TWAMP) and in-band network telemetry (INT) path-balance. Results show that HyTel-Tree achieves the lowest link-delay error and telemetry time per round while restoring telemetry availability through temporary-monitor election. HyTel-Ring incurs lower packet cost but provides reduced coverage when ring continuity is disrupted. Overall, HyTel offers a practical trade-off among accuracy, efficiency, overhead, and resilience.
This paper studies a surrogate-assisted two-base station (BS)-site deployment problem for low-altitude unmanned aerial vehicle (UAV) communication in dense urban environments. The study focuses on a fixed representative UAV altitude of 120 m and a fixed sectorized BS antenna configuration, so that the resulting signal-to-interference-plus-noise ratio (SINR) field is mainly determined by large-scale air-to-ground propagation, urban geometry, BS-site placement, and the prescribed antenna pattern. To avoid repeated ray-tracing simulations during deployment search, we synthesize dual-BS SINR maps from WinProp-generated path-loss (PL) maps and train RadioUNet++, an interference-aware cascaded U-Net surrogate, to predict dense SINR maps from the urban layout and two-BS-site location masks. Based on the trained surrogate, the two-BS-site placement problem is formulated as a surrogate-based continuous optimization problem and solved using proximal policy optimization (PPO) with a coverage-based reward. Simulation results over four urban scenarios in Nanchang, China show that RadioUNet++ achieves Pearson correlation above 0.95 and root mean square error (RMSE) below 3.3 dB on BS-pair-level held-out test sets, supporting its use as a fast SINR predictor under the considered setting. For deployment optimization, PPO reaches coverage ratios within 1.67–8.63 percentage points of exhaustive brute-force (BF) while using only 8 × 104 surrogate evaluations, roughly one order of magnitude fewer than BF (∼7 × 105), and achieves competitive performance compared with the genetic algorithm (GA) baseline. These results indicate that the proposed framework can serve as an efficient planning tool for two-BS low-altitude coverage optimization under the considered fixed-altitude and fixed-sector-antenna urban setting.
In underground emergency rescues, the inherent narrow bandwidth of magnetic induction (MI) communication severely conflicts with the delay-sensitive nature of critical data transmission. While existing works assume infinite-length coding to maximize throughput, they inevitably incur unacceptable transmission latency. To directly address this issue, we introduce finite blocklength coding (FBC) into MI communication to achieve ultra-low latency, while employing magnetic beamforming to compensate for the short-packet penalty and maintain high throughput. Based on perfect channel state information (CSI), we propose a joint beamforming and blocklength configuration algorithm that minimizes the transmission latency while satisfying the constraints on total energy consumption and throughput. For the case of imperfect CSI, a robust joint beamforming and blocklength configuration algorithm is proposed. This paper provides a novel framework integrating FBC with magnetic beamforming to address the trade-off between throughput and transmission latency. Numerical results show that the proposed methods can effectively reduce the transmission latency while satisfying the constraints on total energy consumption and throughput.
Underwater acoustic networks (UANs) are essential for long-term ocean monitoring, offshore exploration, and submerged target sensing, but their resource management is challenged by limited bandwidth, long propagation delay, time-varying noise, multi-path spreading, and shared-channel interference. Existing flat reinforcement learning (RL)-based schemes usually encode link scheduling, transmission mode selection, and power control into a single composite action, resulting in enlarged action spaces and ambiguous credit assignment among heterogeneous resource decisions. This paper proposes H2MARL-UAN, a heterogeneous hierarchical multi-agent reinforcement learning (MARL) framework for adaptive resource management in UANs. H2MARL-UAN decomposes each resource control decision into three physically meaningful layers: Q-mixing (QMIX)-based link scheduling, value decomposition network (VDN)-based transmission mode selection, and independent Q-learning (IQL)-based power control, thereby matching the cooperation mechanism to the coupling strength of each layer. Experiments in a physically consistent UAN simulator show that H2MARL-UAN improves throughput and packet delivery performance under random and bursty traffic while maintaining competitive energy efficiency. The results also reveal the trade-off between service capability and information freshness.
Age of information (AoI) is a freshness metric for time-critical wireless systems. We study periodic deterministic all-to-all status broadcasting under protocol interference and generate-at-will sampling. In each period, every source samples a fresh status at its first transmission, and each destination updates its timestamp at the first reception of that packet. This convention decomposes the network-average AoI into a half-period term and a mean first-reception delay term. The problem captures a tradeoff: spatial reuse shortens the update cycle, while rebroadcast waiting and interference shape dissemination delay. We develop three complementary solution methods. Complete-state packet-knowledge A∗ search certifies global optima on 992 connected nonisomorphic graphs, and an independent compact mixed-integer linear programming (MILP) reproduces their objective values. Bounded-state A∗ search constructs medium-scale schedules. For large graphs, a time-indexed MILP and its Dantzig-Wolfe reformulation motivate three completion-window policies: coverage-first, age-first, and dual-guided. Local greedy scheduling (LGS) implements each policy through distributed request/winner contention. Experiments on connected small graphs, medium-scale parameter regimes, and sparse random graphs with hundreds to thousands of nodes quantify the cycle-delay tradeoff and scheduling scalability.
