With the large-scale deployment of low earth orbit satellite network (LEO-SN), end-to-end (E2E) delay analysis under multi-traffic faces significant challenges from time-varying service capabilities, propagation delay-induced queue coupling, dynamic resource contention, etc. In this paper, we develop a martingale-based analytical framework to E2E delay violation probability in multi-hop, multi-traffic LEO-SN scenarios. First, bursty traffic arrivals are modeled through a Markov chain Monte Carlo process, while satellite-to-ground and inter-satellite link dynamics are captured through finite-state Markov channels, forming the basis for single-hop queuing analysis. We then extend stochastic network calculus to quantify queue coupling effects due to propagation delay and incorporate dynamic residual service allocation under a priority scheduling mechanism. To address the compounded complexity of multi-hop, multi-traffic interactions, we apply martingale theory and the stopping time theorem to derive closed-form upper bounds on backlog and delay violation probability. Furthermore, we introduce global and hierarchical stability conditions that characterize priority scheduling constraints and a scaling factor to reflect service heterogeneity across links. Extensive simulations validate the tightness of our bounds and demonstrate that, although priority scheduling significantly improves delay performance for higher-priority traffic, imbalanced link loads and transient service interruptions can amplify delay variability.

Unmanned Aerial Vehicle networks, also known as Flying Ad Hoc Networks (FANETs), offer significant advantages in applications such as disaster response and emergency communications, owing to their rapid deployment, wide-area coverage, and strong survivability. Routing protocols are critical to FANETs performance, ensuring reliable communication through dynamic path maintenance and three-dimensional (3D) topology optimization. However, most existing protocols are designed primarily based on node mobility or network topology, often overlooking traffic distribution. This can lead to severe congestion in high-traffic areas, resulting in increased end-to-end delays, delayed route updates, and packet loss. While some existing methods address congestion through flow prediction, they frequently disregard topological relationships among nodes, reducing the accuracy of predictions. To address these challenges, this paper proposes a Multi-Metric Joint Prediction (MMJP) routing protocol based on a Spatio-Temporal Graph Convolutional Network (STGCN). MMJP jointly models traffic dynamics and topological structure to enhance prediction accuracy and congestion control. The proposed STGCN model exchanges information with neighboring nodes, aggregates their features, and learns spatio-temporal representations to forecast future node resource availability. The resulting predicted network state is used to select routes that avoid congestion-prone areas. Simulation results demonstrate that MMJP outperforms benchmark protocols, including Destination-Sequenced Distance Vector (DSDV), 3D Geographic Greedy Routing (GREEDY), and the predictive protocol OPAR. Specifically, MMJP achieves over 30$\%$ higher Packet Delivery Ratio and a 35$\%$ improvement in throughput compared to DSDV and GREEDY, and outperforms OPAR by approximately 10$\%$ in both metrics.

In order to research the influence of joint beamforming on system performance within RIS aided systems, a double-RIS aided multiple-input multiple-output (MIMO) system is considered in this study, and the optimal energy efficiency (EE) within this system is investigated. To be specific, by optimizing the transmit beamforming at the BS, the phase shifts of the RISs and the receive beamforming at user terminals jointly, the EE of the double-RIS aided system is maximized. In order to address the complex problem, an alternating optimization (AO) algorithm is designed. This algorithm utilizes fractional programming (FP) and semidefinite relaxation (SDR) methods. In addition, a theoretical analysis is conducted on the double-RIS aided system’s EE performance. Simulations verify our analytical results and show that, compared with the benchmark scheme that ignores the receive beamforming, the proposed optimization scheme achieves at least 10% increase in energy efficiency performance.

To tackle the formidable challenges from future AI applications in communication and computing power, an urgent need arises for a scalable distributed learning framework, which should adeptly orchestrate the varied resources of widely dispersed nodes within 6G networks, offering flexible connectivity and computing services. For distributed AI training in 6G computing-power network (CPN), this paper proposes an adaptive split federated learning (SFL) framework. Considering the terminal computing power heterogeneity, it introduces three train modes: local-only, single base station (BS) collaboration, and dual BSs collaboration. Given the varying channel qualities, two transmission modes are used in the model aggregation phase: direct uploading to the BS and uploading via Device-to-Device (D2D) relays. Accordingly, we formulate a joint optimization problem to minimize overall task latency, which involves model splitting method, cooperative node selection, and multi-domain resource allocation, and then decompose it into two subproblems. First, a shortest path search algorithm is devised to solve the optimal model splitting method and cooperative node selection. Second, convex optimization is employed to derive the optimal multi-domain resource allocation. Simulation results show that the proposed framework attains lower total training latency while preserving high model accuracy.

In 6G-enabled Industrial Internet of Things (IIoT) systems, edge intelligence faces heterogeneous resource availability and highly uncertain operating conditions, which undermine semantic fault tolerance and low-latency inference. To this end, we propose a resilient edge intelligence framework with training–inference co-design, based on a system-level Non-Parametric Mixture of Experts (NP-MoE) in which each expert is instantiated as an execution-capable inference pipeline deployed on an edge node, and a latency-aware gating policy activates only a subset of experts per request for conditional computation. To avoid gradient-heavy parametric MoE training and gating updates, the framework uses lightweight modality-specific backbones together with a non-parametric semantic-evidence aggregation mechanism based on feature memory banks and statistical semantic-score alignment, where “non-parametric” refers to the absence of additional trainable gating or fusion parameters rather than parameter-free feature backbones, enabling retraining-free and robust multimodal inference under modality missingness. Across the lifecycle, split learning serves as a resource mapper to enable feasible model partitioning in resource-constrained edge environments, while the deployed experts form a distributed expert pool for elastic inference scheduling. Then the joint cooptimization of model partitioning and multi-domain resource allocation is formulated as a MINLP and solved via a hybrid graph-theoretic and convex optimization method to minimize lifecycle latency. Experiments on industrial datasets show that the proposed framework achieves a favorable accuracy–latency trade-off and remains robust under extreme modality missingness, outperforming representative multimodal fusion baselines.

Sudden disasters frequently disrupt terrestrial cellular infrastructure, severely degrading network capacity in affected regions and necessitating the rapid restoration of uplink connectivity for rescue and monitoring operations. To address this challenge, we propose a systematic capacity analysis framework for a three-tier emergency wireless network comprising a device access layer, an uncrewed aerial vehicle (UAV) relay layer, and a backhaul layer. First, within the access layer, heterogeneous ground devices are modeled using a Poisson point process (PPP). We analyze the uplink transmissions under non-orthogonal multiple access (NOMA), deriving the coverage probability and ergodic capacity while accounting for imperfect successive interference cancellation (SIC). Next, in the relay layer, multiple concurrent flows are scheduled via time-frequency resource allocation across end-to-end UAV links, incorporating practical communication constraints and multihop packet-loss effects. To balance throughput and fairness, we introduce a joint resource allocation scheme that maximizes a novel satisfaction function. In the backhaul layer, the capacity between UAVs and base stations (BSs) is characterized under space-division multiple access (SDMA) with directional beamforming, explicitly modeling the impact of antenna-orientation jitter at both the UAV and BS sides. Finally, the overall end-to-end system capacity is formulated as the bottleneck capacity among the three tiers. Analytical and simulation results validate the accuracy of the proposed theoretical model. Furthermore, comparative evaluations demonstrate that the proposed resource allocation scheme for the UAV relay layer significantly improves allocation fairness among concurrent flows relative to conventional baselines.

To address the limitation of traditional federated learning in fully exploiting topological structures and complex inter-node features within 6G networks, which leads to restricted learning efficiency, this paper proposes a federated graph neural network learning architecture and its corresponding algorithm for 6G networks. This method first models the 6G network as a graph structure, with network devices as nodes and communication links as edges, comprehensively capturing the connectivity relationships within the network. On this basis, a graph partitioning algorithm is employed to divide the overall network into multiple subgraphs, and a Graph Neural Network (GNN) is utilized to extract topological features within subgraphs and interaction information between nodes, thereby effectively learning node embedding representations to support network performance optimization. Furthermore, a federated learning mechanism is introduced to collaboratively aggregate the learning results from each subgraph. This approach not only fully utilizes the distributed computing resources of edge nodes but also avoids the centralized transmission of raw data, significantly enhancing privacy protection capabilities.

Semantic communications have emerged as a key paradigm for intelligent sixth-generation (6G) wireless networks, which aim to convey the meaning of information rather than accurate bit sequences. However, in open-space low Earth orbit (LEO) satellite links, the broadcast nature and wide beam coverage expose semantic transmissions to severe eavesdropping risks. This paper establishes a unified theoretical and algorithmic framework for secure semantic downlink transmission in satellite networks. In particular, we first develop an integrated mathematical model that couples the semantic representation process, physical-layer satellite propagation characteristics, and information-theoretic secrecy into a single analytical formulation. By defining a joint semantic security cost function, the antagonistic trade-off between semantic fidelity and secrecy capacity is quantitatively characterized under realistic power, beamforming, and propagation constraints. To balance semantic fidelity and information secrecy, a reinforcement-learning-based optimization framework is proposed, wherein an actor-critic agent learns optimal power allocation and semantic weighting strategies through continuous interaction with the environment. This learning-based optimization approach enables autonomous control without requiring explicit channel distribution knowledge or offline parameter tuning. Extended simulation results show that the proposed approach consistently enhances both semantic fidelity and secrecy performance compared with conventional power-control schemes and demonstrate its potential as a foundational architecture for secure and intelligent semantic communications in next-generation satellite networks.

When deploying Reconfigurable Intelligent Surface (RIS) to improve System Sum-Rate (SSR), the timeliness and accuracy of SSR optimization methods are difficult to achieve simultaneously through a single algorithm. Some algorithms focus on timeliness, while some focus on accuracy. In this paper, in order to take into account the timeliness and accuracy of the system comprehensively, we construct SSR analysis model of RIS-assisted multi-user downlink communication system and propose several new optimization methods. The goal is to maximize SSR by using the proposed algorithms to jointly optimize power allocation and reflection coefficients. To solve this comprehensive problem, two sets of Alternating Optimization (AO)-based timeliness algorithms and one set of Monotonic Optimization (MO)-based accuracy algorithms are proposed separately to jointly optimize system performance. First, the Water-Filling (WF)-based and penalty-based low complexity algorithms are developed to optimize power allocation and reflection coefficients respectively. To improve the reality of the calculation, penalty-based algorithm cleverly considers residual noise that is difficult to calculate. Then, for further improve the timeliness, a new Successive Convex Approximation (SCA)-based low complexity algorithm is designed to further optimize reflection coefficients and its convergence is proved. Third, in order to verify the effectiveness of the proposed timeliness algorithms, we further propose MO-based accuracy algorithms, in which, the Polyblock Outer Approximation (POA) algorithm, the Semidefinite Relaxation (SDR) method, and the bisection search algorithm are combined in a novel way. Numerical results confirm the timeliness of AO-based algorithms and the accuracy of MO-based algorithms. They supervise and complement each other.

low-Earth orbit (LEO) satellite networks, with their advantages of low latency, wide coverage, and high data rates, have become a core component of 6G networks. However, due to the complexity and interactivity of satellite networks, they face higher uncertainty and vulnerability, and there is currently a lack of effective frameworks for survivability and recovery in harsh environments. This article proposes a resilience evaluation and optimization model for large-scale LEO satellite networks, based on entropy theory and the minimum-cut theorem. Considering the time-varying characteristics of satellite networks, a spatiotemporal extended graph is used to model the network topology, along with a network communication model that incorporates node state transitions. By analyzing various factors, such as topological structure, node functionality, and link quality in real time, entropy theory is leveraged to provide a dynamic and comprehensive evaluation of LEO satellite network resilience. In response to network failures, multiple attributes, including node load and computational capacity, link signal-to-noise ratio (SNR), and link duration, are integrated into an optimization decision-making framework, formulating a resilient recovery optimization problem. To address this, a network resilience bottleneck identification and global load optimization-based edge augmentation strategy is proposed to enhance the network's adaptive and rapid recovery capabilities. Simulation results demonstrate that the proposed resilience evaluation model effectively reflects the resilience performance of LEO satellite networks in complex and dynamic environments, and the proposed optimization method significantly enhances network resilience.

To enable the rapid deployment of mobile base stations and optimize operations in urban environments, a network coverage planning and optimization method based on multi-agent reinforcement learning was proposed. This method was designed to address the issue of reducing network coverage due to user mobility and the interference caused by densely deployed base stations. During the deployment phase, a hybrid optimization algorithm combining particle swarm and fruit fly optimization was employed to determine the optimal base station locations while minimizing construction costs. In the operational phase, a joint optimization algorithm featuring multi-agent deep deterministic policy gradient and lightweight gradient boosting algorithms was designed to optimize base station locations based on terminal signal strength. Additionally, when performance indicators failed to meet requirements, new base stations were automatically added in suitable locations. Simulation results demonstrate that the proposed algorithm outperforms traditional heuristic algorithms in terms of coverage and service rates, while the designed joint operational optimization algorithm shows superior recovery capability in network coverage compared to the traditional k-means clustering algorithm, adapting to a wider range of scenarios.

The future 6G (sixth-generation) mobile communication technology is required to support advanced network services capabilities such as holographic communication, autonomous driving, and the industrial internet, which demand higher data rates, lower latency, and greater reliability. Furthermore, future service classifications will become more fine-grained. To meet the requirements of these low-latency services with varying granularities, this work investigates fine-grained network slicing for low-latency services in 6G networks. A fine-grained network slicing algorithm for low-latency services in 6G based on GCNs (graph convolutional networks) is proposed. The goal is to minimize the end-to-end delay of network slicing while meeting the constraints of computational resources, communication resources, and the deployment of SFCs (service function chains). This algorithm focuses on the construction and deployment of network slices. First, due to the complexity and diversity of 6G networks, DAGs (Directed Acyclic Graphs) are used to represent network service requests. Then, based on the depth-first search algorithm, three types of SFCs of latency-type network slices are constructed according to the available computing and communication resources. Finally, the GCN-based low-latency service fine-grained network slicing algorithm is used to deploy SFCs. The simulation results show that the latency performance of the proposed algorithm outperforms that of the Double DQN and DQN algorithms across various scenarios, including changes in the number of underlying network nodes and variations in service sizes.

In the cloud–edge–end communication architecture of the new power system, heterogeneous perception services face a fundamental and long-standing demand–supply mismatch with multi-dimensional resources (computing, storage, spectrum/bandwidth, and power) under QoS constraints such as delay, reliability, and accuracy. To uniformly measure and minimize this mismatch under resource-limited and time-varying network conditions—thereby enabling precise and efficient perception—this paper proposes an intelligent perception-service efficiency evaluation and optimization method for electric power information and communication networks based on fit entropy. First, based on the theory of information entropy, the fit entropy is defined for the degree of matching between the requirements of perception services such as delay and reliability and the provision of resources. Then, based on the fit entropy, a three-layer matching model of business domain- logical domain- physical domain is constructed, and then a many-to-many matching optimization problem between the business, service function chain and physical device is formed. Furthermore, a dynamic hypergraph neural network based on the gated attention mechanism is designed to solve this problem, where the multi-type aware service requests are dynamically mapped to cross-domain hyperedges, and the fit entropy is used as the weight of the hyperedges to quantify the global fit among the three domains. The fit entropy is optimized by adaptively adjusting the hypergraph structure and the weight of the hyperedges. The simulation results show that this method can significantly improve the quality of service of perceptive services and effectively balance the utilization of network resources and service adaptability.

Accurate traffic forecasting in wireless mesh networks is critical for optimizing resource allocation and ensuring ultra-reliable low-latency communication in 6G-enabled scenarios. However, existing models often suffer from feature entanglement in sequential spatio-temporal architectures, limiting their ability to decouple multi-domain dependencies (e.g., periodic, topological, and transient dynamics). To address this, we propose MeshHSTGT, a novel hierarchical spatio-temporal framework that synergizes TimesNet for multi-periodic temporal-frequency modeling and a Channel Capacity-Weighted Graph Convolutional Network (CCW-GCN) with Temporal Encoding GRU (TE-GRU) for topology-aware spatial-temporal dependency learning. Unlike conventional serial architectures, MeshHSTGT employs a parallel feature re-extraction paradigm to independently capture domain-specific patterns, followed by a Transformer-based adaptive alignment module to dynamically fuse multi-domain features via self-attention. Experiments on real-world mesh network datasets and the Milan cellular traffic benchmark demonstrate that MeshHSTGT reduces MAE by 5.4-31.4% and RMSE by 13.3-19.5% over state-of-the-art baselines (e.g., TSGAN, STFGNN) across short- to long-term forecasting tasks. Ablation studies validate the necessity of parallelized multi-domain modeling, highlighting a 40% improvement in handling irregular traffic spikes compared to serial counterparts.

To tackle the problems of large data volume, multiple devices, and difficult data acquisition in large-scale sensor networks, we propose a multi-UAV (Uncrewed Aerial Vehicle) cooperative trajectory planning and data acquisition method based on the fusion of multi-traveling salesman problem (mTSP) and multi-agent reinforcement learning. Considering the factors such as the wide and uneven distribution of sensing nodes, limited transmission range, and limited energy of UAVs, firstly, large-scale sensing nodes perform intra-cluster data fusion processing by cluster heads, and multi-UAV cooperative trajectory planning acquires cluster head node data to balance the working time of each UAV, thereby improving energy efficiency. Then, the multi-UAV data acquisition problem is modeled as a mTSP, and an improved genetic algorithm with 2-opt optimization operator to optimize the sequence of data acquisition for multi-UAV, so that each UAV acquires the data of the cluster head nodes in its own area in the shortest time. To optimize the UAV's trajectory based on the data acquisition sequence, we model it as a partially observable Markov decision process, train the model with the counterfactual multi - agent policy gradient method, and solve the agents' belief allocation problem using the advantage function based on difference reward. Simulation results show that the proposed algorithm has good performance in terms of flight time, data acquisition and energy consumption.

The emergence of 6G applications such as artificial intelligence, augmented reality, and digital twins has imposed stringent requirements on the high reliability, low latency, and energy efficiency of computing networks. Therefore, in this paper, we propose a novel high-reliability resource scheduling optimization method for heterogeneous computing networks, leveraging hypergraph neural networks (HGNNs) and coded distributed computing (CDC). We first construct a multidimensional resource representation model for computing networks based on hyper-networks, effectively illustrating heterogeneous nodes and their interactions within computing networks. Then, targeting the need for collaborative optimization of task offloading, as well as computing, communication, and caching resources in cloud-edge-end computing networks, we propose the collaborative task offloading and resource allocation (CTOHRA) problem, which minimizes the total task processing delay. By incorporating CDC, we enhance resilience against edge node failures and unstable network links. To solve this problem, we utilize hypergraph neural networks to capture high-order correlations and improve the accuracy of dynamic resource scheduling, and combine particle swarm optimization (PSO) to handle discrete variables and find the global optimal solution. Extensive simulations show that the proposed method can significantly improve the task success rate, reduce the average system latency, and minimize energy consumption, especially under unfavorable network conditions.

The requirements of the emerging smart grid pose a challenge for power line carrier communication systems in low-voltage networks. However, a large-scale dual-mode communication mesh network, with its advantages, such as high reliability and throughput, holds promise for widespread application in power communication systems. Prior to constructing the new network, it is essential to perform a capacity pre-assessment. Yet, building an accurate analytical model for capacity analysis is complicated because of numerous interrelated factors. To address these challenges, the uplink capacity of dual-mode communication in large-scale mesh networks was investigated. Firstly, a throughput model for multi-hop transmission was established based on the device hierarchy structure within the sub-net mesh. Constraints such as frequency hopping and device half-duplex mechanism were then introduced into the model, and a satisfaction function was proposed to address the issue of unfair resource allocation for information flows, forming a convex optimization problem for throughput in one sub-net. Then, considering the constraints of sustained coefficients between multiple subnets, an integer-constrained optimization problem for throughput in multiple sub-nets was formulated, and the optimal solution represented the uplink capacity of the network. The simulation results show that the network throughput of dual-mode communication is at least 50% higher than that of the single mode communication. The proposed resource allocation algorithm improves throughput by 10% compared with traditional resource allocation algorithms. By improving the network concurrency ability and the receiving ability of the master node, a significant gain in network load capacity can be achieved.

The large-scale self-organizing networks (LSNs) for new power systems are the multi-mode mesh networks combining high speed power line communication and radio-frequency communication (HPLC/RF), which are a promising means to solve the "last mile" problem of ubiquitous access. Due to the complex network topology and communication environment, and rapid service iteration, it is challenging to improve transmission efficiency. In this paper, we proposed the routing and resource allocation algorithms based on graph attention network (GAT) and model for new power systems. First, according to historical data such as delay, reliability, SINR, load ratio, betweenness centrality and level, GAT can calculate different probabilities for nodes with higher level than the central node and select the optimal next hop with the largest probability. This process is executed multiple times on different central nodes to get the best end-to-end path. Secondly, considering the wired and wireless channel interference, packets priority and quality of service (QoS) comprehensively, a dynamic resource allocation model based on the optimal next hop generated by GAT is established, which is aimed to maximize the number of successfully transmitted packets between node pairs. And then a packet priority-based resource allocation algorithm is presented for LSNs for new power systems. Simulation results show that the proposed routing and resource allocation algorithms outperform baselines in terms of throughput, end-to-end delay, hop count and network reliability.

With the rapid expansion of data scale, compute-intensive tasks will become a core application of 6G networks. As Unmanned Aerial Vehicle (UAV) technology advances, UAVs can assist in task offloading for mobile edge computing by collaborating to overcome individual UAV limitations in battery life and computational capacity. Hence, in this paper, we propose a task offloading algorithm for multiple UAVs based on a temporal graph. We first formulate an optimization problem to minimize the total completion time of UAV swarm task offloading by classifying tasks and determining task priorities and subtask dependencies. To solve this problem, we introduce a temporal graph to simulate service nodes and task sequences in computing networks. It can reveal task execution priorities by calculating proximity indices, which indicate the ratio of physical distance to the sum of task weights, and determining timestamp offsets. In the following, to reduce unnecessary waiting and computation resource allocation risks, we transform the optimization problem into a directed acyclic graph connectivity problem, which identifies the fastest temporal paths for each UAV, forming a dedicated service network. Finally, we propose a two-stage matching algorithm that achieves optimal matching based on service node locations, statuses, task types, and offloading demands. Simulation results demonstrate that the algorithm performs exceptionally well, reducing task completion times and significantly outperforming other algorithms in terms of task utility.

Currently, federated learning (FL) is attracting increasing attention in the research of intelligent endogenous 6G networks. However, the increasing number of users and growing model sizes of FL cause considerable communication overhead. To reduce this communication overhead and improve the efficiency of FL, we design a hierarchical blockchain-based FL architecture, which divides a network into edge and center layers and further divides the edge layer into different shards. In addition, due to the massive scales of 6G networks, a main-shard blockchain architecture with IOTA and the practical byzantine fault tolerance (PBFT) consensus algorithm are used to ensure the safe sharing of the model and improve the transactions per second (TPS) of the system. Moreover, comprehensively considering the node computing resources, number of training samples, blockchain consensus delay, network communication delay, and FL training delay, we also establish a learning efficiency optimization model for FL and deduce the training round delay for the main and shard chains. Furthermore, the influences of different numbers of shards on the learning efficiency levels of different FL models are analyzed. The simulation results show that the proposed algorithm greatly improves the learning efficiency, increases the system TPS and reduces the communication overhead by more than 30%.