最新刊期
- 摘要：To address the problem that significant Doppler frequency offsets caused by high-speed satellite movement and low signal-to-noise ratio (SNR) environments severely degrade the reliability of traditional orthogonal frequency division multiplexing (OFDM) signal detection methods in Starlink low Earth orbit (LEO) satellite downlink communications, this paper proposes a novel Starlink downlink signal detection algorithm integrating segmented cross-correlation spectrograms with the Starlink-Detect YOLOv9 (SD-YOLOv9) network. The proposed algorithm first utilizes the repetitive structure of the Starlink Primary Synchronization Signal (PSS) to generate two-dimensional (2D) feature maps with distinct spectral line structures via segmented cross-correlation and non-coherent accumulation, effectively converting the traditional one-dimensional (1D) signal peak detection problem into a 2D image feature recognition task. Furthermore, an enhanced object detection network named SD-YOLOv9 is designed. Specifically, an STCSPA (Swin Transformer-Enhanced CSP with Attention) module is introduced at the end of the Backbone to embed the local window attention mechanism of the Swin Transformer into the Cross Stage Partial（CSP）framework, thereby enhancing feature modeling and context representation capabilities. Simultaneously, the conventional downsampling convolution modules in the Backbone and Neck structures are replaced with SD-ADown (Adaptive Downsampling) modules to achieve adaptive multi-branch feature fusion, minimizing feature loss and improving detection performance for weak signals. Experimental results demonstrate that in the AWGN channel at an extremely low SNR of -13 dB, the detection accuracy of the proposed method reaches 100%. This represents an improvement of 7% and 48% compared to the baseline YOLOv9 model and the cross-correlation and CapNN-based algorithm, respectively. Additionally, it provides performance gains of 99%, 63%, and 41% compared to traditional singular differential cross-correlation, direct cross-correlation, and segmented cross-correlation algorithms, respectively. Ablation studies further validate the effectiveness of the STCSPA and SD-ADown modules. The effectiveness and superiority of this study in resolving detection challenges in high-dynamic wireless communications provide reliable technical insights for next-generation satellite communication systems.关键词：signal detection;LEO satellite;Starlink Signal;SD-YOLOv9;attention mechanism16|0|0更新时间：2026-07-23
- 摘要：To meet the quality of service (QoS) requirements of emerging intelligent services such as large language model inference and multi-agent collaboration, constructing an efficient resource scheduling mechanism has become a core issue in AI computing networks. For the resource scheduling scenario under QoS multi-objective optimization, a scheduling framework that integrates differentiated intelligent service modeling, heterogeneous computing-network resource awareness, and high-throughput remote direct memory access (RDMA) transmission optimization was proposed. Within this framework, a reinforcement learning (RL)-based intelligent scheduling algorithm was developed to adaptively generate scheduling policies by considering service priorities, computing-network resource loads, and data transmission characteristics, thereby improving QoS. Finally, a simulation platform was constructed using real-world network measurement data and actual topology parameters from an AI computing center to verify the performance of the proposed algorithm. Experimental results demonstrate that the proposed method achieves significant improvements over existing schemes in key QoS metrics such as task completion rate, average latency, and resource utilization.关键词：AI computing network;QoS assurance;resource scheduling;reinforcement learning;RDMA network28|10|0更新时间：2026-07-22
- 摘要：To strike a balance between point prediction accuracy and probabilistic calibration performance in traffic flow forecasting, a Graph-Coupled Residual Flow Matching (ST-RFM) framework was proposed, which integrated continuous flow matching into graph-structured spatiotemporal modeling. First, a lightweight multi-scale temporal condition module was designed to decouple and extract trend, periodicity, and volatility features. Second, a graph-coupled reaction-diffusion vector field was constructed to explicitly separate node-level evolution and spatial correction within the generative dynamics, thereby enhancing the capability to capture the dynamic evolution of road networks. Finally, a residual flow matching objective was adopted to learn future variations by anchoring to recent observations, which effectively preserved the continuity prior of traffic evolution. Experimental results on PEMS03, 04, and 08 datasets demonstrated that the proposed algorithm significantly improved probabilistic calibration performance while maintaining high point prediction accuracy. Specifically, the CRPS and MIS95 metrics on PEMS08 were reduced by 3.8% and 43.9%, respectively. This research provides an effective technical solution for uncertainty quantification in generative spatiotemporal traffic forecasting.关键词：Traffic flow prediction;Flow matching;Probabilistic forecasting;Graph-coupled vector fields37|6|0更新时间：2026-07-21
- 摘要：Due to the reliability under high-mobility scenarios, Orthogonal Time Frequency Space (OTFS) has become one of the candidate waveforms for sixth-generation (6G) communication systems. However, as a multi-carrier modulation technology, OTFS faces the issue of high peak-to-average power ratio (PAPR). Iterative clipping and filtering (ICF) can effectively suppress PAPR with low complexities. However, the filtering operation induces peak regrowth. To improve the PAPR performance of ICF, this paper proposes a precoding-based ICF scheme aiming at suppressing peak regrowth. Building upon the use of ZC precoding to improve the amplitude distribution of OTFS signals, the scheme employs ICF to effectively suppress peak amplitudes. Simulation results demonstrate that, under the same number of iterations, the proposed precoding-based ICF method effectively improves PAPR suppression performance while improving bit error rate (BER) performance in high-mobility scenarios.关键词：OTFS;PAPR;ICF;precoding13|0|0更新时间：2026-07-21
- 摘要：随着物联网与人工智能技术的深度融合，由海量智能体协同互联构成的智能体互联网（Internet of Agents）正逐步迈向深度协同。然而，智能体在互操作过程中需频繁交换状态信息与决策知识，面临严峻的隐私泄露风险。现有研究大多将隐私保护视为性能优化的附属目标，缺乏系统性的内生安全机制。为此，本文提出一种隐私增强的互操作框架。该框架首先通过数字孪生技术为每个物理智能体构建虚拟映射，在孪生空间中形成隔离的交互沙盘；进而设计基于本地差分隐私的条件生成式对抗网络（LDP-CGAN），将加噪后的本地观测在孪生空间中复原为富含语义的虚拟数据，从源头切断原始数据流；最后提出联邦多智能体深度确定性策略梯度算法（FC-MADDPG），利用虚拟数据进行协同策略训练，并引入互操作性得分指导联邦聚合，实现通信、计算与隐私资源的联合优化。在MedMNIST医疗推理任务场景中的实验表明，与传统多智能体强化学习算法相比，本框架在引入强隐私保护的同时，任务平均时延与能耗仅分别增加5.1%与6.8%；模型逆向攻击的隐私泄露率从84.5%降至13.2%。实验结果验证了本框架在保障隐私安全的同时能够有效维持系统协同效能。关键词：智能体互联网;隐私增强;数字孪生;生成式对抗网络;多智能体强化学习;联邦学习13|0|0更新时间：2026-07-21
- 摘要：Aiming at the stringent delay and energy consumption requirements of computation-intensive tasks in the Space-Air-Ground Integrated Network (SAGIN), this paper explores how to rigorously map the complex heterogeneous physical characteristics of SAGIN into a Markov Decision Process (MDP) and introduces a multi-head attention mechanism to optimize the Multi-Agent Deep Deterministic Policy Gradient (AM-MADDPG) algorithm. Considering the curse of dimensionality in highly dynamic environments, a Centralized Training with Decentralized Execution (CTDE) architecture is introduced. Each edge agent utilizes multi-head attention during the feature extraction phase to accurately assess collision risks and achieves coordination with ultra-low delay. Simulation results show that the proposed mechanism effectively alleviates task delay constraints. Compared with mainstream benchmarks like MAPPO and distributed Graph Reinforcement Learning based on local neighborhood aggregation (DGN), the proposed algorithm demonstrates superior cost reduction under the premise of equivalent signaling overhead, enhancing system resilience in extreme heavy-load scenarios.关键词：space-air-ground integrated network;mobile edge computing;computation offloading;multi-agent deep reinforcement learning;multi-head attention mechanism43|1|0更新时间：2026-07-13
- 摘要：To address the insufficient accuracy and stability of network configuration translation under low-resource scenarios in cross-vendor environments, a Transformer-based network configuration translation model (NetConfigTM) was proposed. The model integrated domain characteristics of network configurations and constructed a configuration-semantic-oriented training mechanism. By expanding and modeling limited configuration corpora, effective feature learning was realized under low-label conditions. Furthermore, a multi-stage training strategy combining masked language modeling, denoising autoencoding, and supervised fine-tuning was introduced to enhance the modeling capability of cross-vendor configuration syntax structures and semantic mapping relationships. Experiments conducted on desensitized network configuration data demonstrated that the proposed method maintained stable and reliable translation accuracy in multiple typical configuration scenarios and reduced the configuration translation time from days of manual operation to minutes, verifying its engineering feasibility for automated configuration translation in cross-vendor network environments.关键词：network configuration translation;low-resource scenarios;transformer architecture;data augmentation28|1|0更新时间：2026-07-08
- 摘要：In active reconfigurable intelligent surface (Active RIS)-assisted millimeter-wave multiple-input multiple-output (MIMO) systems, mutual coupling (MC) induces electromagnetic interactions among RIS elements and renders the reflection matrix non-diagonal, which in turn leads to high dictionary coherence, limited estimation accuracy of conventional greedy algorithms, and high computational complexity of full-dimensional Bayesian methods. To address these issues, this paper establishes an MC-aware cascaded channel model and proposes a two-stage sparse Bayesian learning (SBL) channel estimation algorithm. In the proposed method, a low-complexity candidate support set is first obtained based on an approximate MC model, and dictionary reduction (DR) is then adopted to project the estimation problem into a low-dimensional subspace. Subsequently, refined channel estimation is achieved via SBL inference under the expectation-maximization (EM) framework. Simulation results verify that the proposed method significantly improves channel estimation performance in MC scenarios with controllable complexity, yielding an average normalized mean squared error (NMSE) gain of about 2 dB compared with the conventional two-stage orthogonal matching pursuit (OMP) algorithm.关键词：Active Reconfigurable Intelligent Surface;Mutual Coupling;channel estimation;sparse Bayesian learning;Dictionary Reduction34|2|0更新时间：2026-07-03
- 摘要：As critical infrastructure，IP networks are facing an urgent need for automation and intelligent transformation in change management and security protection. An intelligent agent O&M platform is developed in this paper. Addressing the issues of IP network change monitoring and router APT security defense, two high-value L4 scenarios of the autonomous network are selected: configuration change and network security. Two major intelligent agents are focused, where configuration generation, configuration verification, configuration simulation, and service comparison is built by the change agent, based on digital twins. Security configuration verification, exposure surface management, and system intrusion detection functions are constructed by the security agent. Through pilot tests on real networks, review time is compressed to 30 minutes and efficiency is improved by 90%. Weak configuration exposure periods are reduced to the daily level with 90% efficiency improvement. Post-event auditing of device intrusion threats can be achieved by minute-level awareness. The O&M risks and security incident rates can be effectively reduced, while network stability and security protection levels can be significantly enhanced.关键词：Intelligent O&M;Change Agent;Security Agent27|4|0更新时间：2026-07-03
- 摘要：In the context of enterprise intelligent transformation, large model training and inference require massive intelligent computing resources. Self-built computing power often suffers from high costs and lengthy construction cycles, while rented computing power services face issues such as data exfiltration security risks. This paper proposes a Cloud-Edge Collaborative Distributed Training and Inference Solution. Through hierarchical model deployment, the enterprise edge side processes sensitive data, while the cloud-based intelligent computing center undertakes large-scale computing tasks. Between the cloud and edge ends, non-sensitive intermediate variables are transmitted via an RDMA (Remote Direct Memory Access)-supported cross-domain lossless transmission network. On the basis of ensuring enterprise data privacy, the solution achieves elastic scaling of computing power based on a secure network architecture, and finally forms a high-performance networking solution that is adaptable to diverse industry scenarios.关键词：cloud-edge collaboration;Distributed Training and Inference;Hierarchical Model Deployment;Cross-Domain Lossless Transmission41|22|0更新时间：2026-06-24
- 摘要：For OFDM/FBMC systems over doubly selective channels, the conventional interference removal (IR) method is susceptible to erroneous data feedback, which leads to error propagation and a performance floor in the high-SNR region. To address this issue, a posterior-weighted prediction-basis coefficient-domain update method is proposed. Based on the current channel estimate, the posterior mean and posterior variance of the data symbols are obtained, and a weighted correction model is constructed in the prediction-basis coefficient domain to reconstruct the full-matrix channel estimate. On this basis, two implementation forms, namely SoftMean and SoftVar, are further developed. Simulation results show that the proposed method outperforms the original IR method in the medium- and high-SNR regions, and that the performance advantage becomes more pronounced as the SNR increases. At an SNR of 35 dB, compared with the original IR method, the proposed method achieves an NMSE improvement of approximately 4 dB for full-matrix channel estimation and approximately 2 dB for pilot-position channel estimation. The computational complexity of the proposed method is also analyzed, and the results indicate that the additional computational cost mainly arises from the calculation of posterior statistics and the weighted solution in the prediction-basis coefficient domain.关键词：Doubly selective channels;Full-matrix channel estimation;MIR-based correlation prediction;Prediction-basis coefficient-domain updating;Posterior confidence weighting34|40|0更新时间：2026-06-24
- 摘要：With the rapid advancement of generative artificial intelligence, AI agents, as the core of next-generation human-computer interaction and automated decision-making, are gradually being integrated into the core processes of real-time communication. However, a fundamental conflict exists between the flexible control flow of agents and the fixed data processing patterns of the media plane, which severely restricts the in-depth application of AI technologies in real-time communication scenarios. To address this challenge, this paper proposes a plug-in and agent collaborative scheduling architecture for the converged media plane oriented towards real-time communication networks. A General Media Framework is constructed to support the efficient embedding of agents into the media pipelines in the form of plug-ins. For high-concurrency media stream processing scenarios, a zero-copy memory mechanism is proposed to significantly reduce latency during stream processing. Furthermore, a two-level scheduling optimization model based on a universal agent is presented to address the redundant computation problem caused by multiple agents running in parallel. Experimental results demonstrate that various AI plug-ins can be stably integrated into the GMF pipeline, and the proposed approach satisfies the end-to-end latency requirements of real-time communication, providing both a theoretical foundation and engineering reference for the AI-native evolution of real-time communication networks.关键词：real-time communication;AI Agents;Plug-in Architecture;Pipelines24|21|0更新时间：2026-06-24
- 摘要：In distribution and metering field operations, traditional operation and maintenance approaches that rely on manual device identification and information recording suffer from low efficiency and high error rates, making them inadequate for large-scale and refined management of power distribution and utilization terminals. To address these issues, this paper proposes an intelligent field operation and maintenance method for distribution and utilization terminals based on near-field sensing. Bluetooth broadcasting and QR code scanning are employed to achieve automatic terminal identification, while essential operation tools, including device rebooting, time synchronization, process detection, and access point switching, are integrated into mobile terminals to establish a portable and visualized field operation framework. Furthermore, a near-field container management mechanism and SIM card information identification method are designed to unify terminal identification, work order association, and operation execution, thereby improving operational accuracy and closed-loop capability. Field deployment and experimental results demonstrate that the proposed method can reduce device identification and registration time by approximately 60% and improve on-site work order processing efficiency by about 35%, effectively reducing reliance on manual experience and enhancing the intelligence level of field operation and maintenance for distribution and utilization terminals.This method addresses key aspects such as target identification, work order association, anomaly resolution, and hierarchical transformation at the implementation level. Furthermore, it analyzes the cost-benefit analysis, application value, and future research directions from an engineering application perspective.关键词：Near-Field Sensing;Distribution and Utilization Terminals;Field Operation and Maintenance;Intelligent Operation and Maintenance;Work Order Management24|21|0更新时间：2026-06-24
- 摘要：To address the large-scale variation of defect targets and severe background interference in substation equipment inspection images, this paper proposes a substation equipment defect detection model, termed the Substation Equipment Defect Detection Transformer (SED-DETR), based on multi-scale feature fusion. In the encoder, a Grouped Cascaded Self-Attention (GCSA) module is introduced to enhance focus on critical regions and improve cross-region contextual modeling. In the neck, a Cross-Scale Multi-Branch Fusion Feed-Forward Network (CSMB-FFN) is constructed, where an improved Unit Step Fusion Convolution (US-Conv) is employed for precise multi-scale feature alignment, and a Cross-Stage Partial Omni-Kernel Block (CSP-OKB) module is incorporated to strengthen the fusion of global context and local texture information. During training, the Generalized Intersection over Union (GIoU) term in the bounding-box regression loss is replaced with the Minimum Pairwise Distance Intersection over Union (MPDIoU) to improve the sensitivity of regression supervision under highly overlapped samples. Experimental results show that SED-DETR achieves mAP@0.5 values of 94.1% and 95.4%, and mAP@0.5:0.95 values of 77.6% and 85.2%, on the self-collected dataset and the public IGV Dataset, respectively, outperforming current mainstream detection models. Edge-side deployment results further demonstrate that the proposed model maintains favorable real-time inference capability on the Jetson Orin Nano 4GB platform; after FP16 acceleration, the inference time is reduced to 23.9 ms while the mAP@0.5 remains 92.5%. These results indicate that SED-DETR achieves a favorable balance between detection accuracy and computational efficiency, providing a practical solution for substation equipment defect detection.关键词：Substation Equipment Fault Detection;deep learning;Multi-scale Feature Fusion;self-attention mechanism;RT-DETR;Cross-stage Feature Fusion;Feature Alignment25|35|0更新时间：2026-06-24
- 摘要：To address the strong coupling between computing and communication resources and the high scheduling complexity caused by heterogeneous node hardware and multi-network access in edge computing power networks, a task scheduling mechanism and algorithm were proposed. A hierarchical edge computing power network architecture for the coordinated management of heterogeneous computing resources and multiple network domains was constructed, and a heterogeneity-aware task scheduling algorithm based on proximal policy optimization was developed on this basis. By jointly optimizing computing-node selection and network-path allocation, system cost and end-to-end execution delay were reduced while task latency constraints and trusted execution requirements were satisfied. Simulation results showed that the proposed mechanism effectively reduced task execution delay and overall cost in dynamic edge environments and achieved better overall performance than conventional scheduling methods. The proposed mechanism provides an effective solution for task scheduling in heterogeneous multi-network edge computing power networks.关键词：heterogeneous multi-network environment;computing power network;task scheduling;deep reinforcement learning;collaborative resource scheduling38|37|0更新时间：2026-06-24
- 摘要：As the Open Radio Access Network (O-RAN) for Sixth Generation mobile communication technology (6G) evolves toward intelligent autonomy, the "black-box" nature of highly relied-upon Artificial Intelligence (AI) models conflicts with the requirement for transparent and trustworthy decision-making. The deep integration of eXplainable AI (XAI) is essential to demystify this black-box, achieve transparent decision-making, and construct a trustworthy autonomous network. This paper provides a comprehensive review of XAI research in 6G O-RAN and proposes a multi-layer integrated deployment framework and roadmap: we expound on the O-RAN architecture and XAI deployment logic; analyze implementation paths and typical applications across radio resource management, network slicing, network security enhancement, intelligent collaboration, and zero-touch operation; discuss technical challenges, including the trade-off between explanation computational overhead and real-time requirements, the lack of unified evaluation standards, interface-induced coordination constraints, and data privacy risks; and explore future trends, such as lightweight XAI algorithms, neuro-symbolic causal reasoning models, closed-loop autonomous mechanisms, and interdisciplinary integration, offering insights for constructing intelligent, transparent, and secure 6G trustworthy autonomous networks.关键词：6G;O-RAN;XAI;Network Intelligence;Transparent Decision-Making;Trustworthy Autonomous Networks24|78|0更新时间：2026-06-24
- 摘要：Three-dimensional point cloud models hold significant application value in fields such as digital copyright protection and content authentication. However, reversible data hiding for 3D point clouds still faces the challenge of balancing embedding capacity with geometric fidelity. A reversible data hiding algorithm for 3D point clouds was proposed by integrating multi-scale spherical coordinate transformation with adaptive hierarchical clustering. The point cloud was processed through a three-level hierarchical scheme based on local cluster point density and scale features. A Gaussian-weighted radial distance predictor was constructed within the local spherical coordinate system of each cluster, and reversible watermark embedding and accurate extraction were achieved based on Prediction Error Expansion (PEE) technique. Experimental results showed that the proposed algorithm achieved a significant improvement in embedding capacity, while the signal-to-noise ratio was maintained above 65 dB under the condition of threshold
, and the bit error rate was zero (recovery rate 100%) when parameter ; ablation experiments further validated the effectiveness of each module, including the spherical coordinate transformation, the adaptive hierarchical strategy, and the Gaussian-weighted predictor. The embedding density of the proposed method demonstrates a marked improvement over existing plaintext-domain point cloud methods. 关键词：3D point cloud;reversible data hiding;adaptive hierarchical clustering;multi-scale spherical coordinate transformation;prediction error expansion18|20|0更新时间：2026-06-24 -
Specific Emitter Identification Based on Inner–Outer Modulation Feature Fusion and SGE-LATransformer
摘要：To address the problems of insufficient feature discriminability, limited noise robustness, and degraded unknown-emitter identification performance in civil aviation communication specific emitter identification under complex electromagnetic environments, a specific emitter identification method based on inner–outer modulation feature fusion and an SGE-LATransformer was proposed. Considering that, in AM–MSK composite-modulated civil aviation signals, both outer modulation and inner modulation were jointly formed by different hardware stages along the transmitting chain while exhibiting distinct physical mechanisms and environmental sensitivities, the time–frequency representation of the original signal was constructed to characterize the overall modulation characteristics introduced by outer-modulation-related hardware components such as power amplifiers, amplitude control units, and envelope shaping circuits, and the time–frequency representation of the MSK inner modulation, obtained after removing the outer modulation, was constructed to highlight the intrinsic modulation characteristics determined by the non-idealities of oscillators, modulators, and filters; these representations were concatenated to form a joint time–frequency feature of inner and outer modulation. The fused representation enabled complementary modeling of outer-modulation hardware discrepancies and inner-modulation hardware fingerprint features in the time–frequency domain, thereby facilitating in-depth characterization of individual emitter signatures from the perspective of civil aviation signal modulation mechanisms. Furthermore, a local-aware Transformer network incorporating the Spatial Group Enhance (SGE) module, referred to as the SGE-LATransformer, was designed, in which global–local feature collaborative modeling and a group-level spatial attention mechanism were employed to emphasize stable emitter-discriminative regions while suppressing noise-induced irrelevant responses. Experimental results demonstrated that the proposed method outperformed traditional single time–frequency features, bispectrum and differential bispectrum methods, as well as conventional deep learning models in terms of identification accuracy, noise robustness, and unknown-emitter identification capability, validating its effectiveness and engineering practicality for civil aviation communication specific emitter identification.关键词：feature fusion;SGE-LATransformer;time–frequency feature;group-level spatial attention mechanism19|0|0更新时间：2026-06-23 - 摘要：As telecom networks continue to expand in scale, traditional manual operation and maintenance management models have become inadequate to meet the demands of efficient network operations. Domestic and international operators are accelerating the construction of L4 high-level autonomous networks. Network management agents (refer to “AI agents for network management” ), as the core enabling technology driving network intelligence upgrade, have become a hot topic in industrial research and application. However, the industry has not yet reached a consensus on the capability boundaries, functional architecture, interaction interfaces, and multi-agent collaboration of network management agents, which restricts large-scale deployment. This paper focuses on the concept connotation and key technologies of network management agents. It systematically reviews the generic functional architecture, workflow, deployment modes, and interaction interface schemes of agents, comparatively analyzes the characteristics and applicable boundaries of different multi-agent collaboration architectures and strategies, and clarifies and assesses the current status and development trends of standardization both domestically and internationally. The research results of this paper can provide theoretical support and practical reference for the technological R&D, scenario implementation, and industrial ecological improvement of network management agents, and facilitate the large-scale deployment of autonomous networks.关键词：Network management agent;autonomous network;AI Agent;AI;Network operation and maintenance management52|9|0更新时间：2026-06-23
- 摘要：Collaborative inference between low-Earth-orbit (LEO) satellites and unmanned aerial vehicles (UAVs) provides a new solution for real-time intelligent sensing tasks. Existing collaborative computing methods based on task offloading or non-pipelined model partitioning tend to cause mutual waiting between communication and computation, resulting in low resource efficiency. To address this problem, a satellite-UAV collaborative inference scheme combining model splitting and pipeline parallelism was proposed. In this scheme, the deep neural network was partitioned, with the head and tail subnetworks deployed on the UAVs and the computation-intensive intermediate subnetwork deployed on the satellite. Complete inference was achieved through intermediate feature transmission. On this basis, the micro-batch pipeline parallel mechanism and non-orthogonal transmission were employed, allowing the computation and transmission stages of different micro-batches to be executed in parallel. Furthermore, with the objective of minimizing the end-to-end collaborative inference delay, the neural network deployment structure, micro-batch partitioning granularity, and uplink and downlink transmission power were jointly optimized. Simulation results showed that the proposed scheme could effectively reduce the collaborative inference delay and improve system resource utilization.关键词：Model splitting;distributed inference;pipeline parallelism;non-orthogonal multiple access22|0|0更新时间：2026-06-23
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
- Total Visits：1237723 Visits Today：1479
0