The deep integration of intelligence and communications has become the development trend of the sixth generation mobile communication (6G) and future communication systems. This study focuses on addressing three core issues in this deep integration：1) The inherent contradiction between increased communication bandwidth and a sharp rise in resource consumption；2) Improving the compression capability of the source to achieve information entropy reduction；and 3）Enhancing the adaptability of information systems to optimize communication system gain. Aiming at native artificial intelligence（AI）evolution，we propose，for the first time globally，a theoretical and technological system for modern semantic communication and 6G Intellicise networks. This system breaks through the limitations of classical information theory by establishing semantic information theory and expanding the boundaries of classical communication theory. It introduces a new pathway of “computation first，communication later”through modern semantic communication，ushering in a new communication paradigm. Furthermore，it constructs a 6G Intellicise network theory and technology system characterized by “intelligent endogenous and native simplicity”，providing core theoretical support and key technological approaches for overcoming future communication bottlenecks. Moreover，we have successfully established the world's first field trial network for 6G communication and intelligent integration，completing outfield trials in typical scenarios such as large-data-volume video transmission，unmanned vehicle communication，industrial Internet，and satellite-ground communication. In the realm of standardization，we have led the establishment of two core standards organizations—the China Communications Standards Association Technical Committee 630（CCSA TC630）Semantic Communication Promotion Committee and the International Mobile Telecommunications 2030（IMT-2030）6G Promotion Group Semantic Communication Task Force—to promote the integration and leadership of the independently intellectual property-rights-protected “Chinese solution”within the international standards system. These achievements have pioneered a new Intellicise evolution path globally in both industry and academia，achieving a full-chain progression from original theory and technological innovation to practical verification and international standards leadership. This work provides systematic technological support and a solid foundation for China to secure a leading position in 6G development，capture the strategic high ground in future communication technologies，and build a secure and controllable industrial ecosystem.

With the growing demand for understanding complex three dimensions (3D) environments in fields like Internet of things (IoT)-driven autonomous navigation and related applications, the automatic reconstruction of structured, editable computer-aided design (CAD) models from point cloud data has become a critical task. However, current research mainly focuses on reconstructing CAD models from CAD command sequences, sketches, and extrusion operations, and commonly faces challenges such as excessive reconstruction steps and strong platform dependence. To this end, a high-precision geometric and topological CAD reconstruction method is proposed, specifically for IoT 3D perception. First, the boundary representation of CAD is decomposed into parametric information of geometric primitives and their topological structure. Then, a primitive variational autoencoder (PVAE) and a topological variational autoencoder (TVAE) are designed to model their geometric features and topological relationships, respectively. Furthermore, PointNet++ is used to extract multi-scale local features from point cloud data and fuse them into global features. A topological decoder and primitive decoders are then used to predict topological tree sequences and primitive parameters, achieving high-precision CAD model reconstruction with clearer boundary details. To validate the method's effectiveness, metrics such as chamfer distance, edge chamfer distance, and structural consistency are used to evaluate its performance on two datasets. The experimental results show that the proposed model outperforms existing methods in terms of topological integrity and geometric reconstruction accuracy，enabling higher-precision CAD model reconstruction.

With the exponential growth of telecommunications business scale, customer complaint responsibility determination has become a key link in the compliance management of communication operators. Traditional methods face challenges such as semantic decoupling difficulties and low efficiency of knowledge reuse when dealing with complex complaints and unstructured historical cases. This study, based on the existing business process and large language model technology, proposes a linked retrieval enhanced generation framework. Through hierarchical semantic decoupling and dynamic retrieval of historical knowledge, it achieves the precision and efficiency of complaint responsibility determination. The framework proposes a two-level complaint point splitting mechanism：The first level uses a hybrid semantic parsing to strip off compound demands, and the second level extracts core complaint points from the first-level complaint list through large model prompt technology, while dynamically processing historical documents to adapt to the second-level complaint matching, completing the structural transformation of the database. Finally, experiments based on 423 actual complaint data from a provincial operator show that compared with several traditional retrieval augmented generation（RAG）models，this method improves the retrieval accuracy by 18.87% and the retrieval recall rate by 14.13% . This research provides an efficient and reusable technical path for intelligent responsibility determination under a strong regulatory background and has important practical significance.

Addressing challenges in complex structural health monitoring arising from heterogeneous sensor node sampling, feature drift, and limited model generalization ability in distributed vibration signals, this study constructs a distributed vibration signal with augmented generation (DVSAG) dataset. It utilizes cross-diffusion for adaptive sampling while preserving the spatiotemporal correlation of the original signal, combines the frequency domain to unify input dimensions, and enhances inputs by calculating residuals using fault-free reference signals. A fault diagnosis network with a convolutional block attention module (CBAM) is designed to extract multi-scale features from distributed vibration signals. These features are converted into word embeddings, combined with user questions, and input into a distributed vibration signal large language model (DVSLLM). Finally, a feature alignment and semantic mapping framework is used to achieve fine-grained interaction from vibration signals to natural language. Experiments show that the proposed method effectively improves fault diagnosis accuracy and model generalization ability under multiple operating conditions, providing reliable support for multi-task decision-making in complex structural health monitoring.

In the field of Internet of things (IoT) intrusion detection, federated learning has become an effective solution for implementing model weight integration updates. This distributed learning method allows devices to train models locally and transmit updated parameters to a central server for aggregation. However, existing intrusion detection methods based on federated learning still have limitations. In scenarios with non-independent and identically distributed data and heterogeneous client models, the intrusion detection performance of the global model will be severely affected. The significant communication overhead caused by simultaneously transmitting model parameters also hinders the actual deployment of federated learning schemes. To address the aforementioned issues, an efficient IoT intrusion detection method based on semi supervised federated learning is proposed. By utilizing unlabeled public data to enhance the model's understanding of the data, the performance of the client classifier is continuously improved. At the same time, a discriminator module is added to improve the quality of the client's predicted labels, and the combination of hard label strategy and voting mechanism effectively reduces communication overhead. The experimental results show that an accuracy of 86.97% is achieved in non-independent and identically distributed data and heterogeneous client model scenarios, which is superior to typical federated learning methods and achieves lower communication overhead.

To overcome the influence of beam split in terahertz communication, for the problems of a large number of large-range delay devices, high hardware complexity and high power consumption in traditional fully connected antenna structure schemes based on single-layer and double-layer delay devices, a multi-user hybrid analog/digital precoding scheme based on multi-layer delayers was considered. In this scheme, the time delays were divided into multi-layer arrangement at the base station, and hybrid precoding design was carried out based on the multi-layer delayers. Specifically, firstly the bit number of time delays and the set of discrete time delay based on the number of time delays in each layer and antennas were deduced. Secondly according to the array response vector at the center frequency, the phase shift of phase shifter was designed. Then the propagation delay of antenna array aperture corresponding to the subarray where the delayer is located was used as the delay of time delays, and the delayer delay was discretized based on the discrete delay set. Afterwards, analog precoding was designed based on the phase shift of phase shifter and the delay of time delays. Finally, low complexity zero forcing precoding technology was used to design digital precoding. Simulation and analysis results show that compared to traditional scheme, the proposed scheme can greatly reduce the number of large range delay devices and hardware complexity while sacrificing a small amount of rate performance.

A smart contract is a piece of computer program that runs on the blockchain, which has the characteristics of automatic execution, non-tampering, and publicity. Smart contracts control the flow of large amounts of high-value data, and attackers can exploit vulnerabilities in smart contracts to steal funds or resources. Existing detection methods, such as symbol execution, have problems such as path explosion and high false positive rate, while machine learning methods are black-box and uninterpretable. In order to solve the above problems, an expert mode based on expert mode and explainable machine learning was proposed to detect vulnerabilities in smart contract code, an expert mode for vulnerabilities was designed, and shapley additive explanations (SHAP) was used to explain the weights of multiple features, and the average detection accuracy of four vulnerabilities (re-entrancy vulnerability, timestamp vulnerability, integer overflow vulnerability, and permission control vulnerability) reached 90.36% , which achieved better detection results compared with classic tools such as Oyente and Mythril.

High-performance electromagnetic shielding materials can effectively protect against the increasingly prominent electromagnetic interference and damage, in order to overcome the shortcomings of traditional rigid electromagnetic shielding materials such as poor flexibility and controllability in practical applications, a flexible, mouldable and conductive silver nanosheets composite hydrogel shielding material was developed, and its three-dimensional structure and effective component composition were determined by scanning electron microscope (SEM) , X-ray diffraction (XRD) and X-ray photoelectron spectroscopy (XPS) . A vector network analyser was further used to confirm the excellent electromagnetic shielding performance of the hydrogel in the frequency band of 8.0～12.4GHz (consistently stabilized above 35dB). Based on the controllable water content of the hydrogel, a dehydration-induced spatial interconnection of silver nanolayers to form a macroscopic conductive pathway is designed and proposed, which greatly improves the electrical conductivity of the composite hydrogel and thus achieves a high level of electromagnetic shielding, and to a certain extent, provides a practical and effective solution for the controllable design and preparation of high-performance electromagnetic shielding materials.

To effectively mitigate Sinkhole attacks in wireless sensor networks (WSN) , this paper proposes a novel Sinkhole attack detection and defense strategy (IF-GBO) that integrates isolation forest (IF) and gradient-based optimizer (GBO) . First, a detection threshold is established to trigger the IF-GBO intrusion detection mechanism, thereby reducing network overhead and improving detection efficiency. Second, considering the characteristics of Sinkhole attacks and the dynamic real-time nature of WSN data, a multidimensional feature dataset is designed, incorporating node hop count, energy consumption, packet reception/forwarding rate, and time delay. The model is trained using a sliding window sampling approach, which not only enhances the algorithm's operational efficiency but also improves the accuracy of malicious node identification. Finally, a multi-objective path selection function is developed，leveraging the GBO algorithm to assist nodes in rapidly identifying alternative transmission paths to counter Sinkhole attacks. This approach effectively ensures reliable data transmission，extends network lifetime，and resolves the issue of delayed delivery of anomaly detection results to the legitimate sink node. Experimental results demonstrate that compared to conventional anomaly detection models such as support vector machine（SVM），k-nearest neighbors（KNN）and local outlier factor（LOF），IF-GBO achieves higher accuracy in identifying malicious nodes with lower false positive rates and superior generalization capability. Furthermore，when compared to dedicated Sinkhole attack detection algorithms like hop count-based detection scheme for Sinkhole attack（HCODESSA）and a Sinkhole detection algorithm based on the random routes selected by minimum hop（RMHSD），the GBO-based defense strategy significantly mitigates the disruptive effects of Sinkhole attacks on the network，ensuring routing security and reliability.

Backscatter communication has received widespread attention in the field of the Internet of things due to its advantages of low cost and low energy consumption. However, with the gradual expansion of data services, new challenges have been posed to the transmission performance of backscatter communication. Multiple-input multiple-output (MIMO) technology has been introduced into backscatter communication due to its advantages in system reliability and transmission rate. One of them, a dual ended joint coding technology called block-level unitary query space-time block code (BUTQ-STBC) , can fully exploit the diversity potential of backscatter channels, but at the cost of sacrificing transmission rate. To combat this issue, a new signal transmission scheme that selects the optimal communication link for data transmission based on the idea of spatial matching is proposed. The simulation results show that the proposed transmission scheme is superior to the BUTQ-STBC scheme. The proposed scheme provides a solution for efficient and reliable transmission of MIMO backscatter communication.

To address the challenges of excessive resource demands and difficulty in meeting low-power requirements in the embedded domain when field programmable gate array (FPGA) are utilized to accelerate convolution operations, a resource-efficient FPGA-based convolution acceleration method for binary neural networks (BNN) is proposed. First, the computational characteristics and resource consumption patterns of various parallelization schemes during the forward inference process of the convolution layer are systematically analyzed. Leveraging the lowbit-width feature of BNN, a high dimensional data splicing and dimensionality reduction storage scheme is introduced. Subsequently, a channel dimension reduction rotation cache (CDR) structure tailored for BNN is proposed, aiming to achieve the combined benefits of intra-convolution kernel parallelism and inter-feature map parallelism with moderate cache bandwidth expansion. Furthermore, to fully exploit the performance advantages of the CDR structure, a specialized CDR processing unit is designed, and the adder tree structure is optimized. The processing unit supports flexible adjustment of different pipeline levels and can achieve higher-level parallel computing capabilities through a repeated invocation mechanism, adapting to diverse system requirements. Experimental results demonstrate that the BNN accelerator based on the CDR architecture achieves significantly superior computational density and storage density compared to existing state-of-the-art solutions when deployed on the Xilinx XC7Z020 chip. It also exhibits low-power characteristics and enables faster inference speed，making it well-suited for resource-and power constrained embedded platforms.

A method is proposed to maximize the minimum user rate for full-duplex (FD) rate splitting multiple access (RSMA) systems. Firstly, the optimization problem of maximizing the minimum user rate in the system is constructed by taking the base station precoding matrix, power allocation, and common rate as optimization parameters, and the user power and rate requirements as constraint conditions. Then, the optimization objective function and the constraints are simplified by introducing slack variables respectively, and the first-order Taylor expansion is used to linearize constraints, so as to transform the non-convex optimization problem into a convex one. Finally, the optimization problem is solved based on iteration and convex (CVX) method. The simulation results show that the minimum user rate increases with the increase of the transmit power of the base station, and the minimum user rate of the proposed scheme is higher than that the existing scheme for the same scenario.

The luminescent properties of gallium oxide (Ga₂ O₃) doped with rare-earth erbium (Er³⁺) have gained significantattention due to their potential in optoelectronic and semiconductor applications. In this study, five Ga₂ O₃ bulk materials with varying erbium doping concentrations were prepared using solid-phase sintering. The crystal structure, micro-morphology, fluorescence characteristics, and the effect of doping concentration were systematically examined. The results show that Er³⁺ substitutes Ga³⁺and also forms a new erbium-gallium garnet (Er₃ Ga₅ O₁₂) phase. As the doping concentration increases, the content of the garnet phase rises, leading to a linear increase in luminescence intensity. However, the sample with 1.00at% erbium doping shows lower red-green luminescence intensity compared to the 0.75at% sample dominated by the garnet phase, indicating that the erbium-gallium garnet phase offers superior fluorescence performance in these regions compared to the substitutional doping phase.

Deep convolutional neural networks are widely used in structural magnetic resonance imaging (sMRI) analysis for the early diagnosis of Alzheimer's disease. To address the challenge of efficient representation learning in sMRI, this study proposes a two-pathway convolutional network that improves the computational efficiency of sMRI feature extraction by representation decoupling, and further strengthens the representation discriminability through adaptive feature fusion. The network consists of three parts：1) A high-channel-capacity slice path, which processes sparse slices to encode semantic information of slice images；2) A low-channel-capacity context path, which processes dense slices to capture inter-slice contextual information；3) An adaptive feature fusion module, which integrates the decoupled information from both paths to generate more effective sMRI representations. The proposed method was evaluated on two tasks—Alzheimer's disease classification and mild cognitive impairment conversion prediction—using the Alzheimer's disease neuroimaging initiative (ADNI) dataset. The results demonstrate that the proposed approach surpasses the baseline models in both computational efficiency and diagnostic performance, while achieving results comparable to those of current state-of-the-art methods.

A low complexity cyclic redundancy check (CRC)-headbiting convolutional code (HBCC)-Miller cascaded encoding scheme has been proposed to combat interference in long-distance wireless transmission environments and greatly improve the communication quality of cellular passive Internet of things systems. This scheme can utilize the combination of CRC and HBCC for encoding and decoding. Compared with convolutional CC encoding, theminimum number of registers can be reduced to 1/3 of the number of long term evolution (LTE) tail-biting CC (TBCC) registers, significantly reducing the computational cost of encoding and further reducing label encoding power consumption. The simulation results show that the proposed low complexity cascaded encoding scheme has a performance improvement of 4.2dB compared to Miller8 encoding, further enhancing anti-interference performance, expanding single station communication distance, and retaining rich clock information between codewords, which can better adapt to the label encoding requirements of ultra-low power cellular IoT systems.

The 5th generation of mobile communications system (5G) virtual private network carrying power business not only increase system capacity, but also bring new security challenges due to the openness. Conducting security analysis on wireless power systems can effectively identify security risks and guide the design of subsequent security protection strategies. However, existing methods lack targeted framework construction and fail to provide intuitive quantitative evaluations. In this paper, a security risk assessment method for power 5G virtual private network is proposed based on the characteristics of power business. On the basis of building the assessment indicators framework, the analysis of security risks is conducted from subjective and objective aspects through analytical hierarchy process and entropy value method. Finally, the effectiveness and applicability of the proposed safety risk assessment method is proved through example analysis.

The in-depth analysis of the semantic information contained in traditional Chinese medicine (TCM) prescriptions is of great significance for both clinical applications and the discovery of new formulas. Existing TCM prescription generation algorithms define the interactions between all symptom herb pairs solely based on co-occurrence, without considering the categorization of herbal properties. To address this issue, this paper proposes a prescription recommendation algorithm based on herbal property driven compatibility mechanism semantic modeling (HPDCM). First, the analysis of prescriptions takes into account the herbal property categories, which are defined as entities when constructing the knowledge graph (KG). Second, the algorithm integrates compatibility rules to model the interactions between symptoms and herbs with weighted connections. This is followed by aggregating higher-order heterogeneous path information of nodes through a graph convolutional network (GCN) model. Finally, an attention mechanism is employed to fuse information from symptom interaction graphs, symptom-herb interaction graphs, and herb interaction graphs, distinguishing the influence of different dimensions of TCM semantic information. Experimental results, compared with existing formula generation algorithms, demonstrate that HPDCM achieves higher accuracy and is more in line with the TCM diagnostic and therapeutic principles of syndrome differentiation and treatment.

In the sixth generation (6G) mobile communication system, intelligent reflecting surfaces (IRS) enhance wireless transmission efficiency by dynamically adjusting the wireless propagation environment. In distributed IRS-assisted multiple-input multiple-output (MIMO) systems, when the number of users significantly exceeds the number of antennas at the base station (BS) , the joint optimization of user selection and beamforming is crucial for reducing BS transmission power and promoting green communication. To address this, a model is established with the objective of minimizing BS transmission power by jointly optimizing user selection, BS beamforming vectors, and phase shift matrices of distributed IRSs while ensuring user quality of service (QoS). To simplify the model, it is decoupled into two subproblems that are iteratively optimized to approximate the optimal solution of the original problem. First，the artificial bee colony（ABC）algorithm and second-order cone programming（SOCP）are employed to determine the optimal user selection strategy and BS beamforming vectors. Then，the phase shift matrices of the IRSs are optimized using the semidefinite relaxation（SDR）method. Simulation results demonstrate that the proposed algorithm not only achieves good convergence but also effectively reduces the transmission power of the BS.

Aiming at the problems of inadequate utilization of structural semantics and poor repair results of detailed features in the existing deep learning methods for repairing ancient murals, a structure-guided diffusion generative algorithm was proposed. Firstly, a mural structure reconstruction module composed of gated convolution and fast Fourier residual block is constructed, and the edge structure after reconstruction is used to guide the repair of damaged murals, so as to overcome the problem of insufficient utilization of structural semantic repair. Then, a generative diffusion module based on stochastic differential equation is proposed, which performs forward diffusion processing on the mural image to be repaired by stochastic differential equation. Next, a mask-enhanced backward iterative reconstruction module is designed to enhance the semantic consistency between the damaged area and the intact area of the mural, and improve the repair ability of the detailed features of the mural. Finally, the digital inpainting experiments and analysis are carried out on the Dunhuang mural data set. The experimental results show that the proposed algorithm can effectively complete the mural restoration, and the objective evaluation indicators are better than the comparison algorithms.

Brain tumor segmentation is a key task in medical image analysis due to the heterogeneous and irregular nature of tumor regions. To address the limitations of existing methods in modeling long-range dependencies and reducing resource consumption, we propose a lightweight segmentation model based on a hybrid convolutional neural network (CNN) and Transformer encoder. Depthwise separable convolutions are employed in shallow layers to reduce computation, while the proposed shuffle former block (SFB) integrates Transformer and ShuffleNet v2 to effectively capture both global and local context. Furthermore, lightweightattention modules are introduced to model long-range dependencies and enhance local perception. Experimental results on the BraTS 2019 dataset demonstrate that our model achieves Dice scores of 93.1% in whole tumor (WT) , 92.2% in tumor core (TC) , and 91.2% in enhancing tumor (ET) , with only 0.98M parameters and 54.60G floating point operations per second, achieving a superior balance between segmentation accuracy and computational efficiency for deployment in resource-constrained clinical settings.

To address the issue that the single red-green-blue (RGB) modality contains limited semantic information, is susceptible to noise interference, and exhibits suboptimal segmentation performance, this paper proposes an RGB-depth (RGB-D) semantic segmentation algorithm based on an asymmetric feature interaction method. First, a two-stream network is employed to extract features from the RGB and depth modalities separately. By incorporating an asymmetric feature correction module, features from one modality are used to correct those of the other, thereby suppressing intra-modal noise. Then, an asymmetric fusion module is applied to further enhance information interaction between the modalities. Additionally, multi-scale feature fusion is introduced in the decoder, and adversarial training is adopted as an auxiliary strategy during the training process to effectively leverage contextual information and improve overall accuracy. Experimental results demonstrate that the proposed algorithm effectively suppresses intra-modal noise and enhances the interaction of valid semantic information across modalities，achieving mean intersection over union（mIoU）scores of 57.4% and 52.1% on the New York University depth dataset v2（NYUDepthv2）and the Stanford University RGB-D dataset（SUN-RGBD），respectively.

Aimingat solving the problems of large amounts of computation and simplifying high complexity of M-ary low density parity check code (LDPC) decoding algorithm, a M-ary LDPC improved decoding algorithm is proposed based on the characteristics of BeiDou-3 (BD-3) B1C satellite signal. On the one hand, the proposed algorithm adopts the dimension reduction and edge complement mode to reduce the calculated amount of decoding algorithm by reducing the calculation dimension between nodes in the updating process of verification equation；On the other hand, to reduce the decoding complexity, the algorithm uses the way in which the nodes update simultaneously, which means the check node updates as the variable node updates. The simulation results show that the improved algorithm reduces the decoding computation and simplifies the decoding algorithm complexity without affecting the decoding performance in the dynamic range of receiving level of BD-3 B1C satellite signals. Compared with before optimization, the calculation amount is reduced to 25% of the original algorithm in one iteration, and 20000 times of data storaging and reading operations are saved when the calculation dimension is 32 by updating the node algorithm. Moreover, with the improvement of signal to noise ratio (SNR), the decoding computation amount and complexity can be more simplified by using smaller dimensions.