最新刊期
卷 42 ， 期 6 ， 2026
- 摘要：With the evolution of 6G networks, malicious network behaviors are characterized by long-term lurking and persistent penetration. Accordingly, network behavior security analysis faces challenges such as parallel multi-path activities, fragmented evidence, and stealthy behaviors. To address the behavioral security protection requirements in 6G networks, a “three-layer, three-domain” multi-agent collaborative network architecture was proposed. This architecture was divided into three layers: the perception layer, the orchestration layer, and the execution layer. By leveraging cross-domain collaboration among the intelligence domain, knowledge domain, and tool domain, the implementation of network behavioral security analysis was supported. On this basis, a domain communication protocol for the interaction between agents and both the knowledge domain and the tool domain was specified. Together with the A2A protocol, a standardized workflow for multi-agent collaborative network behavioral security analysis was established. Finally, using the attack chain construction task under an advanced persistent threat (APT) attack scenario as an instance of network behavioral analysis, the effectiveness of the proposed architecture in supporting network behavioral security analysis in 6G agent networks was validated based on the Clearscope-e3 and Clearscope-e5 datasets.关键词：agent network;6G network;network behavior;agent communication protocol;multi-agent collaboration0|0|0更新时间：2026-07-21
- 摘要：The evolution of the Internet toward the AI agent Internet is driven by generative AI and large language models. As a key device for broadband user access to the network, the broadband network gateway (BNG) is difficult to adapt to the new requirements of ultra-large-scale management and edge computing power posed by massive heterogeneous AI agents. As an intelligent upgrade of traditional BNG, the intelligent BNG (iBNG) still faces a series of challenges. An innovative computing iBNG architecture for the AI agent Internet was proposed. Its technical architecture and key technology system were systematically elaborated, and its application paths were discussed in three typical scenarios: smart home, unmanned commerce, and small-scale industry. By deeply integrating the capabilities of “connection, computing power, and intelligence”, this proposed architecture provides a systematic access gateway solution for the AI agent Internet through the integrated design of endogenous computing power and intelligent management and control.关键词：Computing iBNG;AI agent Internet;Broadband Network Gateway;Computing Gateway47|20|0更新时间：2026-07-21
- 摘要：With the deep integration of artificial intelligence and network technologies, the Internet of agents has become an important direction for the evolution of the next-generation Internet. The connection paradigm of the traditional Internet centered on “location addressing” can hardly meet the demands of the Internet of agents for “business-driven and service-as-connection”. The service semantic routing technology system, paradigm evolution, core collaboration requirements in the Internet of agents, and the limitations of traditional routing technologies were systematically elucidated. It was clarified that semantic routing was the key to realize the transformation from “service demand” to “resource matching”. The solutions for four core semantic routing mechanisms in the Internet of agents, namely fuzzy search based on tag matching and flexible adaptation, precision-named DNS (domain name system) for “service-domain name” mapping, Anycast identification and routing based on IP of the business plane network, and URL (uniform resource locator) semantic routing with application-layer semantic encapsulation were analyzed in depth. Finally, the technical characteristics of each solution were summarized and compared, providing an important reference for subsequent research.关键词：Internet of agents;service identification;semantic routing;agent collaboration;resource scheduling0|0|0更新时间：2026-07-21
- 摘要：To realize the ubiquitous intelligence vision of IMT-2030 (6G), an endogenous intelligent core network architecture with multi-agent collaboration was proposed. Based on the service-based architecture (SBA), the AI-enhanced intent function (AIEF), agent control function (ACF), and agent registration and management function (ARMF) were defined to support endogenous intelligence in 6G networks. An intent-driven subnet generation task was selected to verify the proposed architecture, and a system platform was developed to validate registration, orchestration, and user equipment (UE) access.For 6G logical subnet construction, a two-stage multi-agent collaborative mechanism coordinated by the ACF was designed and evaluated on a single NVIDIA RTX 4090D GPU. The results demonstrate the feasibility of AI-agent-enabled endogenous intelligence for future networks.关键词：6G;mobile communication network;intent-driven;subnet generation;AI agent100|47|0更新时间：2026-07-21
- 摘要：The engineering deployment of large language model (LLM) has given rise to agent systems with autonomous decision-making and tool-calling capabilities. However, transitioning from controlled prototypes to enterprise production still faces six structural challenges: execution isolation, identity governance, tool ecosystem governance, multi-agent coordination, observability, and ecosystem distribution. Existing research primarily focuses on agent reasoning architectures or individual platform capabilities, lacking a systematic infrastructure framework for full-lifecycle enterprise deployment. A six-layer reference framework was proposed covering agent runtime, identity and security, tool integration, multi-agent orchestration, observability and evaluation, and marketplace distribution. The design logic of each layer was validated through dual-path verification against both enterprise cloud platforms and open-source community projects. An evolutionary convergence tendency across agent identity, tool invocation, multi-agent coordination, and governance security was revealed by a systematic comparison with 3GPP SA2#173 research. The framework provides an analytical lens from cloud computing practice for understanding agent-native 6G network architecture evolution.关键词：agent infrastructure;large language model;multi-agent collaboration;6G44|17|0更新时间：2026-07-21
- 摘要：To address the “stability-plasticity dilemma” faced by multi-agent systems based on large language models and the risk of catastrophic forgetting caused by fine-tuning, a frozen main branch dual-loop evolution framework was proposed. General reasoning was decoupled from policy preferences by this framework. Discrete value iteration was utilized to optimize immediate actions by the behavioral loop. A prediction error-modulated semantic reflection mechanism was introduced by the semantic loop that transformed natural language feedback into high-dimensional semantic gradients, driving the evolution of external latent vectors in a continuous semantic space. In experiments conducted in the GridWorld game environment, this method achieves a 85.6% task success rate under fully frozen parameters, with the highest path-weighted success rate, while exhibiting extremely low inference latency (2.1 s/step) and token consumption (3.2 k/session). The experiments also observe socially cooperative behavior based on implicit causal reasoning, validating the significant potential of this paradigm in terms of low resource consumption and high interpretability.关键词：large language model;multi-agent system;continuous learning;dual-loop evolution0|0|0更新时间：2026-07-21
- 摘要：The large-scale deployment of 5G and industrial digital transformation are driving network operations and maintenance (O&M) toward collective intelligence. The fragmentation of AI applications has resulted in data silos, capability fragmentation, and cross-domain coordination challenges, hindering the advancement of autonomous intelligent networks toward level 4 (highly autonomous). A three-layer agent architecture for O&M of autonomous networks was proposed: the agent capability layer, the operations management layer, and the security protection layer. Through innovative layered decoupling and multi-agent collaborative mechanisms, a dual-dimensional evaluation framework (maturity-generality) and a full-lifecycle security protection system were constructed, enabling the evolution of O&M paradigms from single-point to collective intelligence. A practical paradigm was provided by the proposed architecture, technical pathways, and security operation methodologies for the communications industry to accelerate progress toward L4 autonomous intelligent networks.关键词：autonomous intelligent network;network operations and maintenance;multi-agent coordination0|0|0更新时间：2026-07-21
- 摘要：Compared with 5G, 6G will go beyond the traditional communication function and deeply integrate cross-domain technologies such as communication, sensing, computing, artificial intelligence (AI), big data, and security to build a new generation of mobile information network featuring everything as a service (XaaS). Based on the vision requirements and design objectives of 6G, a technical framework integrating communication, sensing, computing, and intelligence with ubiquitous space-air-ground integrated coverage was systematically articulated. On this basis, the system design of key 6G technologies was elaborated, aiming to provide a reference for 6G technology research and standardization.关键词：6G;integration of communication-sensing-computing and intelligence;space-air-ground integration;technical system0|0|0更新时间：2026-07-21
- 摘要：Terahertz (THz) non-orthogonal multiple access (NOMA) is regarded as a candidate in 6G and beyond systems. By exploiting the ultrabroad bandwidth and power domain, THz-NOMA can realize massive connectivity by assigning each sub-band to different users. To unleash the potential of the THz-NOMA system, it is significant to allocate power quickly under quality of service (QoS) requirements. Focusing on the instant power allocation, a novel Transformer-based double deep Q-network solution adaptive to general user distributions was proposed. Transformer was used to learn the relationships among allocation strategies for different users, and a double deep Q-network was adopted to achieve a more stable decision optimization process. The simulation results validate that the proposed algorithm realized the throughput close to the optimum given by exhaustive search method within the millisecond level. The proposed method demonstrates high real-time performance and robustness, which suggests its high practicability.关键词：terahertz communication;non-orthogonal multiple access;resource allocation24|5|0更新时间：2026-07-21
- 摘要：A gridless channel estimation scheme combining simultaneous orthogonal matching pursuit (SOMP) and variational Bayesian inference (VBI) was proposed to address the near-field effect and spatial non-stationary effect in XL-MIMO systems. Firstly, based on group time block code (GTBC), the received signals of each subarray were extracted, and the spatial non-stationary channel estimation problem over the entire antenna array was transformed into a spatial stationary channel estimation problem for each subarray. Subsequently, the SOMP algorithm was employed to obtain the sparse channel support set for each subarray. Finally, a simplified codebook was constructed using the obtained support set, and the codebook matrix and the complex path gain expectation matrix were iteratively updated by means of VBI, enabling channel estimation for each subarray and thereby obtaining the estimated channel for the entire array. Simulation results demonstrate that, compared with the GP-SOMP and GP-SIGW algorithms, the proposed channel estimation scheme achieves significant performance improvement in terms of normalized mean square error (NMSE).关键词：XL-MIMO;near-field;spatial non-stationary;channel estimation;simultaneous orthogonal matching pursuit;variational Bayesian inference0|0|0更新时间：2026-07-21
- 摘要：High-precision close-range measurement is identified as a critical technology in millimeter-wave radar systems. However, existing high-precision ranging algorithms are hampered by high computational complexity, which made them difficult to deploy effectively on resource-constrained embedded platforms. To address this, a Macleod-CZT(MCZT) range estimation algorithm integrating Macleod interpolation and the Chirp-Z transform (CZT) was proposed. A three-stage cascaded structure consisting of "coarse estimation–refinement–fine interpolation" was adopted by the algorithm. By combining the interpolation correction of the Macleod method with the frequency-domain refinement capability of the CZT, the picket fence effect caused by discrete sampling was effectively suppressed. It was demonstrated by experimental results on the ADT6101P radar system that under a low transform-point configuration (M=32), superior ranging accuracy was exhibited by the MCZT algorithm under complex conditions such as low signal-to-noise ratios (SNR) and non-integer frequency bins. Without introducing significant additional computational load, the measured error was found to be stably controlled within 20 mm. This scheme provides a high-efficiency and practical ranging solution for resource-constrained millimeter-wave radar systems.关键词：millimeter-wave radar;Macleod algorithm;Chirp-Z transform;frequency estimation2|0|0更新时间：2026-07-21
- 摘要：As optical transport network (OTN) evolve toward a higher-level autonomous and intelligent ecosystem, the prediction accuracy of OTN health models was directly related to the reliability of network optimization decisions. To overcome the limitation of conventional general time-series augmentation methods lacking physical constraints, an evaluation method for health model accuracy was proposed by integrating data augmentation technology with OTN physical transmission characteristics. In this method, basis functions were designed based on the physical variation patterns of key OTN performance parameters, and a data augmentation mechanism of “physics-guided mathematical transformation followed by optical-domain verification” was constructed by combining a function generator with an optical network simulator. High-fidelity synthetic datasets were generated to simulate various performance variation scenarios such as gradual changes and abrupt changes, thereby forming benchmark datasets for model training and validation. Furthermore, a complete workflow encompassing data loading, augmentation, processing, model training, and evaluation was detailed, and the effectiveness of the method was validated through experiments.关键词：optical transport network;data augmentation technology;healthy model;performance prediction;long short-term memory network2|0|0更新时间：2026-07-21
- 摘要：With the rapid development of online education, how to accurately model learners' knowledge status and provide personalized support has become a key part of realizing efficient teaching systems. Knowledge tracing aims to model students' potential knowledge status by using their historical answer data. In recent years, the Transformer model has been introduced into the field of knowledge tracing for its excellent sequence data processing capabilities, but its computational complexity is high when processing long sequences, and there is still room for further improvement in prediction accuracy. To address this, a Transformer-based knowledge tracing model integrating sparse attention mechanism, named SATKT (sparse attention Transformer knowledge tracing), was proposed. In this model, a question‑knowledge embedding module was employed to comprehensively capture the evolution of students’ knowledge states from both the question level and the knowledge concept level. A sparse attention mechanism was introduced to selectively focus on key interaction information. Furthermore, a hybrid loss function was incorporated to achieve multi‑objective joint optimization, thereby further improving the convergence stability and prediction accuracy of the model. Experimental results on four knowledge tracing datasets show that SATKT achieves an average improvement of approximately 1.33% in AUC compared with existing mainstream models, and also performes better in terms of ACC and RMSE, demonstrating higher prediction accuracy and stronger generalization ability. This study provides an effective and scalable new approach for personalized learning analysis in smart education scenarios.关键词：knowledge tracking;Transformer;smart education;sparse attention mechanism0|0|0更新时间：2026-07-21
- 摘要：With the rapid development of the Internet and communication technology, segment routing over IPv6(SRv6), as an innovative network technology, simplifies network architecture and enhances the flexibility and programmability of the network topology, thereby meeting the diverse demands of various new services. In contrast, traditional static policies are unable to respond in real-time to the vast and complex changes in the network states, while network traffic is difficult to monitor and predict in real-time, and network optimization often requires trade-offs among multiple conflicting objectives. To address these problems, an SRv6 dynamic optimization framework that combined spatio-temporal mixed neural networks with enhanced reinforcement learning was proposed. It utilized long short-term memory (LSTM) and graph convolutional network (GCN) to capture the spatio-temporal features of network traffic, subsequently merging these features to derive network performance metrics. The enhanced Q-learning algorithm was then employed to interact with the environment, obtaining the optimal strategy. Experimental results show that compared with the basic Q-learning, this approach shows improved learning speed while maintaining a high bandwidth utilization rate, providing a technical reference for enhancing the efficiency of SRv6.关键词：SRv6;reinforcement learning;dynamic optimization;network traffic0|0|0更新时间：2026-07-21
- 摘要：In response to the rapid evolution of fraud call scripts, the reliance on a large number of labeled samples in traditional methods, and the potential hallucination risks in end-to-end recognition based on large models, a fraud call identification method based on large language models and event evolution graphs was proposed. By introducing the concept of one-class learning, the zero-shot capability of large language models was leveraged to abstract caller speech into standardized verb-object phrase event evolution chains without the need for negative samples. An incrementally expandable event evolution graph was constructed through semantic similarity fusion, and risk discrimination was performed using key event node matching and connectivity detection. Experimental results show that the proposed method achieves an F1 score of 84.84% on a sample scale of 2 000 instances, which is further improved to 91.46% on a scale of 8 000 instances, with performance exhibiting certain scaling law characteristics as the sample size increases. The proposed method provides a precise and scalable technical approach for identifying complex and continuously evolving fraudulent behaviors.关键词：large language model;event evolutionary graph;fraud call2|0|0更新时间：2026-07-21
Research and Development
- 摘要：The diversified service requirements of 5G networks drive the collaborative development of ultra-reliable low-latency communication (URLLC) and enhanced mobile broadband (eMBB), where the quality of service (QoS) assurance and service scheduling play the central roles in addressing heterogeneous resource competition. A comprehensive review of the technical background, core principles, and standard framework of 5G QoS and service scheduling was provided , with an in-depth analysis of both traditional scheduling algorithms and intelligent optimization approaches. By comparing representative scheduling schemes across different stages of evolution and linking them with 3GPP standards and real-world application scenarios, the challenges and opportunities of deep reinforcement learningdriven dynamic resource allocation in domains such as industrial automation and intelligent transportation were highlighted. Building upon this foundation, the taxonomy of radio resource scheduling from the perspective of the fundamental resource conflict between URLLC and eMBB was redefined. The evolutionary trajectory from conventional scheduling mechanisms and new radio(NR) coexistence schemes to intelligent scheduling was systematically reviewed and the pivotal impacts of 3GPP Releases 15 through 19 on service-oriented scheduling strategies were analyzed. Finally, the key research directions were outlined and the structured open problems for intelligent scheduling in the context of 6G were formulated.关键词：5G;QoS;Service Scheduling;priority45|24|0更新时间：2026-07-21
- 摘要：With the rapid adoption of artificial intelligence (AI) technologies in the telecommunications industry, the security of large-scale model software—serving as the “foundation of safety”—has become an increasingly pressing governance demand. As an extension of the software bill of material (SBOM), the artificial intelligence bill of material (AIBOM) systematically records information related to components, models, and datasets, and has emerged as a key instrument to ensure the trustworthiness and compliance of large-scale AI systems. Firstly, the definition and core elements of AIBOM were reviewed, as well as its distinctions from traditional SBOM, its strategic significance in large-model software security governance was highlighted. Then industry cases were analyzed to summarize technical and governance challenges in AIBOM adoption. Building on this, the unique challenges faced by telecom operators in large-model software security governance were explored, and the central role of AIBOM in addressing AI asset transparency, cross-departmental coordination, and regulatory compliance was elaborated. Finally, the practices of China Telecom were drawn on to demonstrate exploratory efforts in standards, toolchains, and risk management, and future research directions for AIBOM development were outlined.关键词：AIBOM;AI security;telecom operator;software security governance0|0|0更新时间：2026-07-21
- 摘要：In view of the development of computing-network convergence and computing-network infrastructure in China in recent years, the system positioning and functional architecture of heterogeneous computing power scheduling were expounded, and four key components for system implementation were extracted, including computing power scheduling adapters that shielded the differences of heterogeneous computing power, a job scheduling controller that performed two-level scheduling based on scheduling algorithms, a computing-network resource orchestration module that provided computing awareness and computing-network routing functions, and three center modules for computing power services that realized integrated computing power transaction services. A multi-objective scheduling algorithm model was proposed for the multi-dimensional constraints of scheduling effects. Combined with scenarios such as the East Data West Computing project, artificial intelligence training-inference collaboration, supercomputing interconnection, and live video streaming, the important role of heterogeneous computing power scheduling was discussed. A practical deployment case was given combined with engineering practice, and the innovation and application prospects of heterogeneous computing power scheduling in the evolution of computing-network convergence were summarized.关键词：computing-network convergence;heterogeneous computing power scheduling;cross-domain scheduling;computing-network collaboration0|0|0更新时间：2026-07-21
Engineering and Application
- 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：1199133 Visits Today：688
0