Table of Content

10 August 2026, Volume 46 Issue 8

Previous Issue   

Artificial intelligence

Survey on action representation and generation strategies in vision-language-action models for embodied intelligence

Wentao ZHANG, Aolan SUN, Xiaoyang QU, Xulong ZHANG, Jianzong WANG

2026, 46(8):  2383-2393.  DOI: 10.11772/j.issn.1001-9081.2025070900

Asbtract ( )   HTML ( )   PDF (1186KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Vision-Language-Action （VLA） models are critical pathway to realize embodied intelligence， with their core of seamless transformation of multimodal perception and understanding into concrete actions in the physical world. However， action representation and generation strategies of VLA models， serving as the pivotal bridge between “perception” and “execution”， face complex challenges among the high-dimensional continuous spaces， the diversity of action modalities， and the stringent demands of real-time robotic control. Therefore， this survey sorted and summarized a systematic review of the evolution， key methodologies， and future directions of action representation and generation strategies in VLA models. At the same time， we analyzed discrete and continuous action representations in detail， as well as autoregressive， non-autoregressive， and hybrid action generation strategies， highlighting their inherent trade-offs in terms of action precision， generation diversity， and reasoning efficiency. In addition， we summed up emerging high-efficiency strategies for real-time control， such as hybrid generation architecture. Finally， we presented a summary of the current technological landscape and prospected frontier challenges and research opportunities of VLA models integrating with world models and the development of cross-robot generalized representations， aiming to provide a reference for constructing more general and efficient embodied agents.

Survey of knowledge graph-based question answering methods

Xinliang LIU, Yushi XU, Dubai LI, Yanzhao REN

2026, 46(8):  2394-2410.  DOI: 10.11772/j.issn.1001-9081.2025070838

Asbtract ( )   HTML ( )   PDF (1891KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Significant advancements in Knowledge Graph Question Answering （KGQA） technology have been driven with the rapid evolution of artificial intelligence technologies and deepening research and application of Knowledge Graph （KG）. Owing to lacking systematic classification framework for the existing question-answering methods， current users have limited awareness of various KGQA models. Therefore， it is difficult to meet the reference needs of users across different domains. To address this issue， a systematic review of research progress in KGQA domain over past 15 years was conducted. First， four core question-answering strategies were identified including template matching， semantic parsing， deep learning and Large Language Model （LLM） augmentation. Second， the performance of eight typical question-answering methods under four categories of strategies under complex question， multi-hop reasoning， low-resource environments， and multi-lingual environments was compared， and applicable scenarios for these methods were provided. Third， common datasets and performance evaluation methods were organized. Finally， combining the current state of domain development， targeted suggestions and prospects for future research directions in KGQA technology were proposed， aiming to provide references for subsequent related studies and applications.

Diffusion-guided evolutionary neural architecture search algorithm

Zeyu TANG, Xiaoyi WEN, Zaipeng XIE

2026, 46(8):  2411-2420.  DOI: 10.11772/j.issn.1001-9081.2026010044

Asbtract ( )   HTML ( )   PDF (1457KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Evolutionary Algorithms （EAs） are widely adopted in Neural Architecture Search （NAS） due to their strong global exploration capability. However， traditional mutation ways often lack performance information guidance and have limited local exploitation capability， thereby constraining search efficiency and architecture quality. Therefore， this paper proposed a Diffusion-Guided Evolutionary NAS （DGEA） algorithm to enhance mutation quality and search capability jointly via distribution-aware and bias control mechanisms. The core of DGEA is to use a discrete diffusion model to generate structural perturbations， aiding evolutionary search process. By learning structural priors from the distribution of high-performance architectures， DGEA performed controlled noise addition and conditional denoising to create new high-performance architecture. At the same time， DGEA introduced an adaptive mutation strategy to balance exploration and exploitation dynamically. Besides， DGEA integrated a lightweight Harmony Search （HS） operator as local enhancement module for fine-tuning neighborhoods， thereby suppressing premature convergence. Experimental results on NAS-Bench-101， NAS-Bench-201， and ImageNet100 transfer tasks show that DGEA boosts search efficiency and effectiveness significantly. Under similar computational budgets， compared to EG-NAS （Neural Architecture Search with Fast Evolutionary Exploration）， GEA （Guided Evolutionary Neural Architecture Search with Efficient Performance Estimation） and REA （Regularized Evolution for Image Classifier Architecture Search）， DGEA reduces the architecture evaluations needed to reach the same accuracy level by 10%-84% with the overall accuracy gains of 0.28 to 0.42 percentage points. Ablation studies verify the contributions and synergistic effect of Diffusion-Guided Mutation （DGM） mechanism and HS module. It can be seen that by integrating distribution-aware generative mutation with adaptive HS local optimization， DGEA improves evolutionary search’s dynamic performance and convergence stability.

Graph representation learning enhanced by Wasserstein regularization

Lining YUAN

2026, 46(8):  2421-2431.  DOI: 10.11772/j.issn.1001-9081.2026010069

Asbtract ( )   HTML ( )   PDF (1529KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Probabilistic Graph Representation Learning （GRL） models typically rely on KL （Kullback-Leibler） divergence to constrain latent distribution of node representations， and suffer from posterior collapse， support set sensitivity， and unstable training gradient. Therefore， we proposed a GRL model based on Wasserstein distance — Wasserstein Graph AutoEncoder （WGAE） to use the closed-form solution of second-order Wasserstein distance as a regularization term to constrain latent distribution from the perspective of optimal transport theory. Compared to KL divergence， the advantage of Wasserstein distance lies in the robustness of support set and gradient continuity， which enables probabilistic graph representation learning models to avoid representation degradation caused by posterior collapse， encourages models to learn potential representations with clearer geometric structure and more disentangled dimensions， and optimize training dynamics. Experimental results demonstrate that WGAE not only improves the prediction performance of graph representation learning models in downstream tasks， but also enhances representation quality and training dynamics. At the same time， through multiple experiments on benchmark graph datasets， the proposed model’s advantages in prediction accuracy， training stability， information encoding efficiency， resistance to data perturbations， and embedding space entanglement resolution are verified comprehensively. It can be seen that Wasserstein regularization can complement and enhance various advanced GRL mechanisms， bringing stable performance gains， and providing a practical basis for building more robust and efficient GRL optimization components.

Aspect sentiment triplet extraction based on graph convolutional network and cross-domain data augmentation

Musheng CHEN, Wenqing FU, Xiaohong QIU, Junhua WU, Qiang WEN

2026, 46(8):  2432-2439.  DOI: 10.11772/j.issn.1001-9081.2025070808

Asbtract ( )   HTML ( )   PDF (1490KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aspect Sentiment Triplet Extraction （ASTE） identifies sentiment entities， attributes， and polarities， enabling precise analysis of sentiment associations in user opinions. This supports product and service optimization， public opinion monitoring， and consumer decision-making. To address limitations of traditional methods in modeling long-range semantic dependencies and insufficient target-domain data， a cross-domain data augmentation graph convolutional network model was proposed for ASTE. During cross-domain data augmentation stage， pseudo-label generation technology based on Maximum Mean Discrepancy （MMD） was introduced to label unlabeled target-domain data initially to mitigate domain shift， a domain-adaptive language model was trained to capture target-domain-specific semantics， and autoregressive data generation was performed by this domain-adaptive language model， so as to further expand the target-domain labeled data volume and diversity. During sentiment triplet extraction stage， the data after first-stage augmentation were used to fine-tune the graph convolutional network-based language model in the pseudo-label generation process， then the fine-tuned model was employed for sentiment triplet extraction. Experimental results on the ASTE-DATA-V2 dataset show that the proposed method outperforms baseline methods such as BGCA （Bidirectional Generative Cross-domain ABSA）， FOAL （Fine-grained cOntrAstive Learning）， and HiPM-hard （Hybrid Prompts Mixture）. Compared with the baseline BGCA， the proposed method has the average F1 score increased by 0.81 percentage points， improving the accuracy and stability of cross-domain sentiment triplet extraction.

Syntactically enhanced aspect-based sentiment analysis via mutual information maximization and contrastive learning

Huihui LYU, Zhaoman ZHONG, Yu ZHANG, Jidong FAN

2026, 46(8):  2440-2446.  DOI: 10.11772/j.issn.1001-9081.2025070843

Asbtract ( )   HTML ( )   PDF (893KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aspect-Based Sentiment Analysis （ABSA） aims to infer the sentiment polarity towards specific aspects. The existing methods often adopt an equal-weight strategy when aggregating neighbor node information， failing to distinguish the differences in node importance， which leads to insufficient semantic-syntactic coordination. Graph ATtention network （GAT） relies on indirect syntactic associations to capture the relationships between aspect words and sentiment words， resulting in a lack of global features and low distinguishability in the sentiment representation space， thereby causing ambiguous sentiment space. To address the above issues， a graph convolutional network model integrating Mutual Information （MI） alignment and supervised contrastive learning， MIC-GCN （Mutual Information and Contrastive learning-based Graph Convolutional Network）， was proposed. First， text words were converted into vector representations through an embedding layer. Second， a dual-attention dynamic routing module was designed to combine the dynamic interaction of self-attention and aspect-aware attention， so as to generate an attention score matrix， and a dynamic masking matrix was constructed on the basis of syntactic distance， so as to optimize the utilization of structural information. Third， the MIC-GCN module was used to construct the original graph and pseudo-graph. After extracting local features through convolution， local and global features were aligned by MI maximization. Finally， the supervised contrastive learning was introduced to optimize the sentiment representation space and enhance the ability to distinguish sentiment polarity， so as to complete sentiment polarity judgment through pooling and classification layers. Experimental results demonstrate that on three public datasets — Twitter， Lap， and Rest， the proposed model has the accuracy 0.10， 1.71， 0.30 percentage points higher than the best baseline models SSEGCN （Syntactic and Semantic Enhanced Graph Convolutional Network）， T-GCN（Type-aware Graph Convolutional Network）， and DGEDT （Dependency Graph Enhanced Dual-Transformer network）， respectively. It can be seen that the proposed model enhances semantic-syntactic coordination effectively， improves the efficiency of capturing global features and the distinguishability of sentiment polarity， demonstrating excellent performance in ABSA tasks.

Similarity-aware link prediction model for ethnic culture knowledge graph

Yan WANG, Xi CHEN, Zhongkai ZHAO, Huan ZHOU, Tao WU, Mengge AI, Xuesong XIAO

2026, 46(8):  2447-2456.  DOI: 10.11772/j.issn.1001-9081.2025070818

Asbtract ( )   HTML ( )   PDF (4787KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Mining potential associations among entities in ethnic culture knowledge graphs is valuable for revealing evolutionary relationships among cultural elements and provides new technical paths for systematic modeling and intelligent reasoning of ethnic culture. However， because such graphs have characteristics of many-to-one attribute affiliations， highly homogeneous structures， and numerous noisy edges， the existing link prediction models are limited in capturing fine-grained feature associations among cultural entities simultaneously and resisting noisy edge interference in the graphs effectively， which restricts prediction performance and model robustness. To address these problems， a Graph Neural Network （GNN） model with similarity-aware features， SRGCN （Similarity-aware Relational Graph Convolutional Network） was proposed. In SRGCN， a dynamic aggregation mechanism was constructed on the basis of node feature similarity， so as to capture feature associations among entities more accurately. At the same time， a dual-level contrastive learning framework was introduced to suppress noise interference effectively， and a linearly weighted multi-objective loss function was designed， where the weights of primary and auxiliary tasks were adjusted dynamically to further enhance the model robustness. Experimental results on the HeritEdge ethnic culture knowledge graph dataset show that SRGCN outperforms the optimal baseline model， LTRGN （Linear self-attention with multi-Relational Graph Network）， achieving improvements of 28.4% and 32.5% on Mean Reciprocal Rank （MRR） and Hits@10， respectively， demonstrating better performance.

Soft whitening inspired non-contrastive SSL framework for time-series forecasting

Dong LI, Yiji ZHAO, Haiyan DING, Hao WU

2026, 46(8):  2457-2466.  DOI: 10.11772/j.issn.1001-9081.2025070823

Asbtract ( )   HTML ( )   PDF (1332KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To overcome limitations in the existing Self-Supervised Learning （SSL） methods for time-series， such as reliance on negative sample pairs and the potential disruption of temporal structure by data augmentation， a Soft Whitening inspired non-contrastive SSL framework for Time-series Forecasting （SWTF） was proposed. In the framework， series representations with dense information and prediction ability were learned without negative samples. Its core method consists of two collaboratively optimized objectives. The first is a Temporal Soft Whitening （TSW） loss， which means that compact representations with rich information were learned by reducing the redundancy between representation dimensions and maintaining the variance of representations， so as to avoid the reliance on negative sample pairs and representation invariance constraints. The second is a denoising reconstruction task， which means that the model was guided to focus on and encode predictive time-series features. To optimize these two objectives jointly， a Siamese encoder was constructed to use parallel multi-scale convolutional networks to capture both micro-level local periodicities and macro-level long-range dependencies efficiently. Experimental results on several benchmark datasets （electricity， weather， and traffic） show that SWTF achieves a Mean Absolute Error （MAE） reduction of approximately 19.6% （13.4% to 26.4%） compared to mainstream baselines such as CoST and PDF（Periodicity Decoupling Framework）， verifying the effectiveness of SWTF in learning robust and highly predictive time-series representations.

Ride-hailing spatio-temporal demand forecasting method via large language model-driven semantic quantification

Yue ZHANG, Yuhan GUO

2026, 46(8):  2467-2476.  DOI: 10.11772/j.issn.1001-9081.2025070885

Asbtract ( )   HTML ( )   PDF (1468KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The accuracy of demand forecasting for ride-hailing services will affect the efficiency of urban transportation resource allocation and traffic operation directly. The existing ride-hailing demand forecasting methods suffer significant accuracy degradation during non-recurrent traffic events due to insufficient global multimodal pattern capturing capability and inadequate refined abnormal flow modeling mechanism. To address this issue， a data augmentation method based on Llama3 was proposed， integrating domain-adaptive fine-tuning with semantic event information fusion， and a quantitative representation system was constructed， incorporating event importance and congestion impact coefficients， thereby converting unstructured event descriptions into structured numerical features. At the same time， a dual-path collaboration architecture named InformerGRU （Informer with Gated Recurrent Unit） was designed， combining the Informer encoder’s long-sequence global modeling capability with the dynamic local-sequence capturing advantages of bidirectional Gated Recurrent Units （GRU）， so as to realize intelligent aggregation of multi-scale features through an adaptive feature fusion layer. Real-world experimental results demonstrate that the proposed data augmentation method reduces the model Mean Squared Error （MSE） by 71.44% in abnormal traffic flow scenarios， with enhanced peak forecasting capability. Compared to Informer， InformerGRU model achieves 4.33% lower Mean Absolute Error （MAE） and improves the direction accuracy by 2.68%.

Data science and technology

Text-ID sequential recommendation model based on multi-strategy contrastive learning and adaptive label smoothing

Jiaming LIU, Fangju AI

2026, 46(8):  2477-2484.  DOI: 10.11772/j.issn.1001-9081.2025070928

Asbtract ( )   HTML ( )   PDF (1014KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Significant progress has been made in sequential recommendation systems for capturing the evolution of user interests， where the effective use of side information plays a crucial role. For example， item-related text data is fused with item ID in sequential recommendation effectively in the transformed domain space， and this text-ID semantic fusion is then used for sequential recommendation， so as to enhance recommendation performance. The existing models suffer from limitations such as the complex user behaviors， limited generalization capability after heterogeneous information fusion， and data sparsity. Therefore， a text-ID sequential recommendation model based on multi-strategy contrastive learning and adaptive label smoothing named MCLALS （Multi-strategy Contrastive Learning and Adaptive Label Smoothing） was proposed. First， textual representations were processed through a domain-expert fusion network and knowledge from different feature subspaces were integrated dynamically. Second， the processed text and ID representations were fused semantically at sequence level through frequency-domain transformation to capture deep interactions among heterogeneous information. Third， on the fused high-level representation， multi-strategy contrastive learning was applied to generate diverse augmented representations using multiple complementary data augmentation strategies， thereby alleviating data sparsity effectively. Finally， adaptive label smoothing was adopted to optimize the training objective， thereby enhancing the model’s robustness to noise and long-tail distribution. MCLALS was evaluated systematically on three public benchmark datasets： Food， Office， and OR. Experimental results show that the proposed model outperforms several baseline models， including SASRec （Self-Attentive Sequential Recommendation） and TedRec （Text-ID Semantic Fusion for Sequential Recommendation）， in overall performance. In particular， on the OR dataset， MCLALS achieves at least 6.69% improvement in NDCG@10 compared to the best baseline model， verifying its effectiveness.

Learning path recommendation model integrating multi-behavior modeling and reinforcement learning

Pengyu CHEN, Baojun TIAN, Lichang ZHAO, Jiandong FANG

2026, 46(8):  2485-2493.  DOI: 10.11772/j.issn.1001-9081.2025070864

Asbtract ( )   HTML ( )   PDF (1160KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address issues of sparse interaction data and unreasonable learning resource planning in learning path recommendation， a Learning Path Recommendation Model integrating Multi-behavior Modeling and Reinforcement Learning （RL）， LPRMMMRL， was proposed. First， user behavior data were classified and fused into a course knowledge graph as entity nodes， and a Graph Convolutional Network （GCN） with attention mechanism was used to capture multi-source heterogeneous features. Second， three data augmentation strategies were designed from both behavioral and learning resource perspectives， and contrastive learning was applied to carry out representation learning to the augmented data， which were integrated into RL as supplementary information. Finally， after fully considering individual differences of the users， a personalized learning path recommendation model was built using a Deep Q-Network （DQN）， and a dual reward function of both knowledge points and paths was designed to guide model convergence. Experiments were conducted for comparing LPRMMMRL with learning path recommendation models such as LPG （Learning Path Generator）， KTKDM （Knowledge Tracing based Knowledge Demand Model） on four real-world courses from MOOPer and MOOCCubeX datasets. The results show that LPRMMMRL achieves at least 5.53% and 5.74% improvements in precision， recall， and F1-score on the MOOPer dataset， respectively， and at least 4.56% and 3.41% improvements on the MOOCCubeX dataset， respectively. Besides， parameter sensitivity analysis and ablation studies validate the effectiveness of the proposed model.

Multi-view attention fusion for academic community identification in citation networks

Yuan QIU, Hailong PENG, Rong FEI, Qingzheng XU, Qianxi LI, Cheng XUE

2026, 46(8):  2494-2504.  DOI: 10.11772/j.issn.1001-9081.2025070851

Asbtract ( )   HTML ( )   PDF (1049KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Academic community identification in citation networks aims at uncovering internal correlations among scholars. However， heterogeneity of node attributes （content features） and topological structures （association patterns） often cause feature misalignment and information dilution during fusion， and the existing methods often ignore stable association patterns in core subnetworks formed by highly-connected nodes， which limit identification accuracy. To address these issues， a Multi-view Attention Fusion for academic Community Identification （MAFCI） method was proposed. In the method， a dual-view design was adopted： original node attributes and the adjacency matrix were encoded by one view， while core-node information and its topology extracted by k?core were incorporated by the other one view. A sparse graph attention mechanism was applied to extract low-dimensional embeddings from both views， and a dynamic weighting strategy was employed to fuse multi-source information adaptively， thereby balancing local structure and global semantics. In the decoder， the adjacency matrix was reconstructed via inner product and node attributes as well as core information were restored using attention mechanism. At the same time， a self-supervised clustering loss was introduced to guide embedding representations in converging to discriminative cluster centers and enhance community boundary separability. Experimental results on four citation network datasets （Cora， Citeseer， ACM， and DBLP） demonstrate that MAFCI method outperforms nine advanced baseline methods， including DDGAE （Deep Dual Graph Attention auto-Encoder） and BCDAN （Balanced method for Community Detection in Attribute Networks）， on most core metrics. Specifically， on the DBLP dataset， MAFCI method achieves a Normalized Mutual Information （NMI） of 0.513 and an Accuracy （ACC） of 0.809， which are improved by 2.3 and 1.3 percentage points， respectively， compared to the optimal baseline method； on the ACM dataset， MAFCI method achieves an NMI of 0.674 and an F1-score of 0.903， representing improvements of 1.2 and 0.8 percentage points， respectively， compared to the optimal baseline method； on the Citeseer dataset， MAFCI method has the NMI achieved 0.448， which is improved by 0.7 percentage points； on the Cora dataset， MAFCI method has the NMI reached 0.543， also surpassing all comparison methods. It can be seen that MAFCI method enhances the accuracy and robustness of academic community identification and has strong generalization across citation networks of varying scales and structures effectively through dual-view sparse attention and dynamic fusion mechanism.

Cyber security

Zero-trust authentication scheme for non-person entities in cloud environments

You JIA, Yaling WU, Qingpeng MIAO

2026, 46(8):  2505-2514.  DOI: 10.11772/j.issn.1001-9081.2025070901

Asbtract ( )   HTML ( )   PDF (1060KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Current cloud environments commonly use perimeter-based security mechanisms， only relying on firewall port policies or simple authentication for access control， so there are several limitations： first， they may lead to identity theft and repudiation among internal entities， lateral movement， traffic-based attacks， or sensitive data leakage； second， the existing mechanisms such as JWT （JSON Web Token） and TLS （Transport Layer Security）， though partially adopted， still suffer from replay attacks， port exposure， and management complexity. To address these risks， we proposed a Non-Person Entity （NPE） identity registration and authentication scheme based on the zero-trust principle. In the scheme， we designed a multi-level identity registration mechanism and employed the Chinese national cryptographic algorithm SM9 to achieve NPE key generation， identity binding， and automatic distribution. By improving the existing Single Packet Authorization （SPA） technique and integrating JWT， we realized bidirectional authentication and fine-grained authorization among NPEs. Security analysis shows that the proposed scheme can mitigate Distributed Denial-of-Service （DDoS） attacks and resist various network threats such as replay and man-in-the-middle attacks， offering security advantages over authentication schemes such as JWT， TLS and UDP （User Datagram Protocol）-based SPA. Experimental results demonstrate that compared with the Software-Defined Perimeter （SDP） controller-based authentication scheme， the proposed scheme reduces the communication latency by 50% and decreases the authentication overhead by 62.03%.

Deepfake detection method based on fusion of multi-modal physical prior features

Renkun LYU, Peng SUN, Yubo LANG, Hong GUO, Zhe SHEN, Di TIAN

2026, 46(8):  2515-2523.  DOI: 10.11772/j.issn.1001-9081.2025070826

Asbtract ( )   HTML ( )   PDF (2963KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The existing deepfake detection methods mainly model on the basis of pixel-level clues of images， and seldom consider the impact of the synthesis process on the forged images. Although good detection results are achieved， it is difficult to explain the detection process. Therefore， a multi-modal physical prior feature fusion-based explainable detection method for deepfakes was proposed. First， optical flow features， illumination features， edge features and DCT（Discrete Cosine Transform） features were used to describe the inter-frame motion differences in temporal videos， the illumination inconsistency in single-frame videos， and edge artifact information， respectively， so as to obtain multi-modal physical prior features with explainability. Second， a multi-modal mixture of experts network was proposed to construct expert sub-networks for different modalities， and after cross-modal attention weighting， the sub-networks were fused through a gated unit and input into the discriminative network for classification. Third， the SIAM （Spatial Intersection Attention Module） was introduced into the discriminative network， and the fully connected structure was replaced by the KAN （Kolmogorov-Arnold Network） structure. Finally， the multi-modal physical prior features were used to train different expert sub-networks， respectively， and the Shapley value analysis of different input features was given， thereby constructing a pre-feature-post-explanation explainable analysis framework to provide pixel-level explanations for model inference and prediction. Experimental results show that compared with algorithms such as CORE（COnsistent REpresentation learning）， SRM（Rich Models for Steganalysis）， and UCF（Uncovering Common Features）， the proposed method achieves the best performance on AUC and accuracy， with an accuracy range of 97.35% to 98.75% and an average accuracy of 98.22% on FaceForensics++ dataset， and the model’s interpretability also is improved significantly.

Advanced computing

User-centric satellite edge computing architecture for task offloading optimization

Juan CHEN, Yujie CHEN, Zongling WU, Di TIAN, Jie ZHONG

2026, 46(8):  2524-2532.  DOI: 10.11772/j.issn.1001-9081.2025070842

Asbtract ( )   HTML ( )   PDF (1356KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Satellite Edge Computing （SEC） improves computational efficiency for remote users by offloading tasks to satellite servers. However， SEC system performance may be affected by wireless transmission quality， since there is signal interference among User Equipments （UEs）. To this end， we proposed a User-Centric SEC （UCSEC） architecture to allocate Access Point （AP） clusters dynamically according to the real-time location of users and network conditions. To handle dynamic task offloading in UCSEC， we designed a Master-Slave multi-agent Proximal Policy Optimization with Gated Recurrent Unit （GRU） and Multi-head Attention mechanism （MSPPO_GMA） algorithm. The algorithm adopted a centralized control-distributed execution framework. The master agent， deployed on the SEC server， monitored global resource constraints and coordinated slave agents， while slave agents deployed on the UE executed task offloading and resource allocation according to network states. In MSPPO_GMA algorithm， GRU captured temporal correlations of network states， while multi-head attention focused on critical information of task offloading and resource allocation. Experimental results demonstrate that compared with MAPPO， MATD3 and other algorithms， MSPPO_GMA algorithm reduces task completion latency by 3.7% to 13.6% and UE energy consumption by 6.5% to 13.2%.

Multi-agent path planning with hierarchical adaptive implicit quantiles

Jingying XUE, Liqun KUANG, Zhixun WANG, Zhengtao GUO, Huiyan HAN

2026, 46(8):  2533-2540.  DOI: 10.11772/j.issn.1001-9081.2025070804

Asbtract ( )   HTML ( )   PDF (1397KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address limitations of real-time and local conflict coordination for multi-objective path planning in Virtual Reality （VR） scenes， a multi?agent path planning algorithm based on hierarchical adaptive implicit quantile was proposed. First， collaborative optimization of high- and low-level strategies was introduced to unify global planning and local execution， and an implicit quantile network was employed for value distribution modeling. The high-level strategy was used to generate sub-goals considering both global environmental distribution and target positions， while the low-level strategy was used to perceive local environmental dynamics according to sub-goals in real time. Then， an adaptive Conditional Value at Risk （CVaR） mechanism was used to dynamically adjust decisions. Finally， a goal-smooth reward function was designed to guide agents toward rapid and stable target achievement by optimizing feedback structure of the strategy. Experimental results show that the proposed algorithm outperforms the suboptimal algorithm， adaptive_IQN （Adaptive Implicit Quantile Network）， by 2， 3， 3， and 4 percentage points under scenarios with 3 agents and 5， 10， 15， and 20 targets， respectively， optimizing path planning efficiency and enhancing fluency and immersion of multi-user interactions in VR scenes.

Two-stage energy consumption feature selection method based on tree-based models

Boran TIAN, Jiantao ZHOU, Daming ZHAO

2026, 46(8):  2541-2547.  DOI: 10.11772/j.issn.1001-9081.2025070854

Asbtract ( )   HTML ( )   PDF (741KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In cloud computing platforms， multidimensional resource features of application services influence data center energy consumption significantly. Extracting key energy consumption indicators through feature selection techniques serves as an effective means to build accurate prediction models. The existing studies fail to adequately integrate feature interpretability with global optimal search capability， resulting in issues such as high redundancy， low prediction accuracy， or lack of decision interpretability in selected features. To address these problems， a two-stage energy consumption feature selection method based on tree-based models was proposed. In the first stage， SHapley Additive exPlanations for Tree-based models （TreeSHAP） was used to quantify marginal contribution of features and realize highly transparent elimination of redundant features. In the second stage， rapid convergence of Ant Colony Optimization （ACO） algorithm was combined with global search capability of Gravitational Search Algorithm （GSA） to localize the optimal feature combination in the reduced feature space. Through the two stages working collaboratively： the low-dimensional interpretable space filtered by TreeSHAP was used to provide a foundation for ACO-GSA， with computational complexity reduced significantly， while key feature combinations were identified by ACO-GSA efficiently through a hybrid search strategy. By comparing different feature selection methods and validating the performance of the feature subsets on multiple prediction models， a comprehensive evaluation was conducted in terms of feature dimensionality， prediction accuracy， and generalization performance. Experimental results demonstrate that the proposed method reduces the feature subset dimensionality by 58.3% compared to Least Absolute Shrinkage and Selection Operator （Lasso） and improves the prediction accuracy of the feature subset by 9.1% compared to ACO on the University of Melbourne cloud dataset. It can be seen that the proposed method outperforms comparative methods in feature compactness， prediction accuracy， and generalization reliability， verifying its effectiveness.

Network and communications

Channel estimation of STAR-RIS-assisted hybrid-field communication system based on deep learning

Xiaojin GUO, Xuyang SUI, Kenan ZHOU

2026, 46(8):  2548-2554.  DOI: 10.11772/j.issn.1001-9081.2025070859

Asbtract ( )   HTML ( )   PDF (791KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems of high pilot overhead and low accuracy in channel estimation of Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surface （STAR-RIS）-assisted hybrid-field communication systems， a Multi-scale Feature-aware Fusion Super Resolution Network （MFSRN） was proposed. In the network， channel estimation problem was modeled as a channel matrix reconstruction task. First， common features were extracted from an upsampled channel matrix； then， an attention gate was utilized to separate user-specific channel features； finally， reconstruction of multi-user channel matrices was completed in parallel. To enhance network performance， an Enhanced Convolution Block （ECB） and a Dual-Stream Feature-Aware （DSFA） block were designed to extract both local and global features of hybrid-field channel effectively， and a Feature Pyramid Attention （FPA） block was introduced to strengthen model’s capability for multi-scale feature aggregation. Experimental results demonstrate that compared with classic DRSN and U-MLP schemes， the proposed scheme achieves a lower pilot overhead and has the Normalized Mean Square Error （NMSE） reduced by an average of 3.62 dB and 1.28 dB respectively at different Signal-to-Noise Ratio （SNR）， showing better estimation performance.

Personalized federated learning intelligent edge aggregation method for space-air-ground integrated networks

Yadong JIN, Jing FAN, Enkang XI, Hua DONG, Hao YU, Yihang SUN

2026, 46(8):  2555-2566.  DOI: 10.11772/j.issn.1001-9081.2025070920

Asbtract ( )   HTML ( )   PDF (1906KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Space-Air-Ground Integrated Network （SAGIN） serves as a comprehensive network supporting global data transmission and plays a key role. However， its highly dynamic and heterogeneous network characteristics lead to an issue that data at each node exhibits statistical heterogeneity， specifically manifests Non-Independent and Identically Distributed （Non-IID） characteristics. Federated Learning （FL）， as a distributed machine learning approach， uses distributed data for model training. When FL trains models on Non-IID data in SAGIN， the generalization performance of global model at each node is damaged. To address this issue， a Personalized FL-based Intelligent Edge Aggregation （FedIEA） method was proposed， which captured necessary information from global models of node models through Personalized Federated Learning （PFL）. The core of FedIEA method is an Intelligent Edge Aggregation （IEA） module， which was able to aggregate downloaded global model and the edge models into the edge-side targets at each node intelligently to train the initialized local models in each iteration. To evaluate the effectiveness of FedIEA method， comparative experiments were conducted on three datasets， namely MNIST， FashionMNIST and CIFAR-10， with typical algorithms such as Federated Averaging （FedAvg）， Federated Proximal （FedProx）， and Federated learning via Logits Calibration （FedLC）. The results show that the test accuracies of FedIEA under different Dirichlet heterogeneous settings are higher than those of benchmark algorithms like FedAvg； and based on Convolutional Neural Network （CNN） model， the test accuracies of FedIEA surpass FedAvg by 25.30 and 41.28 percentage points on FashionMNIST and CIFAR-10 datasets， respectively； on these three datasets， FedIEA method has significant advantages in the test accuracy based on CNN and Deep Neural Network （DNN） models over 9 state-of-the-art PFL algorithms.

Multimedia computing and computer simulation

Dynamic dilated convolution and hierarchical local-global attention model based on improved Lite-Mono architecture

Guanghui LI, Licheng QU

2026, 46(8):  2567-2576.  DOI: 10.11772/j.issn.1001-9081.2025070832

Asbtract ( )   HTML ( )   PDF (1665KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Monocular depth estimation is a core technology for 3D environmental perception and is of significant value in applications such as autonomous driving and robot navigation. As a representative lightweight monocular depth estimation model in these fields， Lite-Mono achieves an excellent balance between performance and efficiency through dilated convolutions and a local-global attention mechanism. However， there are two major bottlenecks in current Lite-Mono architecture. The first one is that fixed dilation rate in the dilated convolutions leads to mismatch between receptive field and target scale， so that features of small objects are lost due to excessively large detailed rate， while contextual information of large objects is lacked due to insufficient receptive field. The second one is that high computational complexity of O（H2W2C） in local-global attention module hinders processing of high-resolution input， thus limiting real-time application. To address these issues， an enhanced model based on Lite-Mono architecture， DDHL （Dynamic Dilated convolution and Hierarchical Local-global attention）， was proposed. First， a Dynamic Dilated Convolution Module （DDCM） was introduced to adjust the receptive field in real time via a fusion weight rate predictor and generate adaptive weights by combining channel attention. Second， a hierarchical local-global attention module was designed to reduce the computational complexity to O（（HW/M）2C）. Finally， a learnable global token was incorporated to establish cross-window dependencies， so as to assist models in understanding and analyzing visual content more accurately in complex scenes， thereby enhancing perception and modeling capabilities of the overall scene. Experimental results on Make3D dataset demonstrate that DDHL model achieves significant improvements on key metrics： Absolute Relative Error （Abs Rel） is reduced from 0.462 to 0.290 with a decrease of 37.2%， and Squared Relative Error （Sq Rel） is decreased by 49.7%. It can be seen that this model achieves good balance between accuracy and efficiency， demonstrating practical application value.

Zero-shot composed image retrieval algorithm for remote sensing images based on image edit proxy

Jie ZHANG, Zhiyong HUANG, Ruijin WANG, Fengli ZHANG

2026, 46(8):  2577-2583.  DOI: 10.11772/j.issn.1001-9081.2025070813

Asbtract ( )   HTML ( )   PDF (1036KB) ( )  

Figures and Tables | References | Related Articles | Metrics

With the rapid development of Composed Image Retrieval （CIR） technology， its application is explored in the field of Remote Sensing （RS） image retrieval to improve the accuracy of retrieving target images from RS image libraries. However， the existing algorithms fail to bridge the semantic gap between image and text modalities effectively and are limited by the lack of high-quality annotated datasets for training CIR models in the RS field. To address these challenges， a zero-shot composed image retrieval for Remote Sensing images based on Image Edit Proxy （IEP4RS） algorithm was proposed to use image editing techniques to generate proxy images aligned with the query image and text description， thereby enhancing the query representation. In IEP4RS algorithm， image editing instructions were generated based on the query image and the text description of the target image， these instructions， along with the query image， were then fed into an image editing model to produce a proxy image， and composite query image features were constructed by fusing the features of the proxy image and the original query image. In this algorithm， the semantic gap between image and text modalities was bridged effectively through direct image feature matching， and a zero-shot learning paradigm was used to avoid the dependency on annotated datasets required by traditional algorithms. Experimental results on the public RS CIR benchmark dataset， PatternCom， demonstrate that IEP4RS algorithm improves retrieval performance significantly. With RemoteCLIP as the feature encoder， compared to the baseline algorithm WEICOM （WEIghted COMposed image retrieval）， IEP4RS algorithm has the mean Average Precision （mAP） improved by 9.74 percentage points. Furthermore， IEP4RS algorithm outperforms the mainstream zero-shot composed image retrieval （ZS-CIR） algorithms， achieving improvements of 11.79， 7.81， and 3.99 percentage points in mAP compared to Pic2Word （mapping Pictures to Words for zero-shot composed image retrieval）， SEARLE （zero-Shot composEd imAge Retrieval with textuaL invErsion）， and FREEDOM （composed image retrieval for training-FREE DOMain conversion）， respectively.

Point cloud registration network with dual-branch multi-level feature fusion

Ming LIU, Dongqi SHEN, Ziyang MENG

2026, 46(8):  2584-2593.  DOI: 10.11772/j.issn.1001-9081.2025070887

Asbtract ( )   HTML ( )   PDF (1581KB) ( )  

Figures and Tables | References | Related Articles | Metrics

A point cloud registration network with dual-branch multi-level feature fusion， DMFNet （Dual-branch Multi-level Feature Fusion Network）， was proposed to address poor registration accuracy and insufficient robustness problems in point cloud registration caused by partial overlapping， occlusion， and noise interference. In the network， during the encoding stage， a rotation branch and a translation branch were configured in parallel， and self-attention fusion as well as cross-attention fusion modules were inserted at shallow， middle， and deep layers， so as to achieve multi-scale feature interaction and deep fusion between source and reference point clouds. At the same time， a rotation-translation feature fusion module was designed to strengthen pose estimation capability. During the regression stage， based on the lightweight Set Transformer regressor， multi-layer induced-attention blocks and an attention-pooling module were used to regress the quaternion and the translation vector directly. DMFNet does not rely on overlap-region detection or explicit mask estimation， and it shows strong adaptability and generalization. Comparative experiments with six point cloud registration methods were conducted on the ModelNet40 dataset， and a generalization study was carried out on the Stanford 3D scanning dataset. The results show that on the ModelNet40 dataset under noisy conditions and two-time sampling， DMFNet reduces RMSE（t） and Error（R） by 21.32% and 14.47% compared with MAximum Clique-based registration （MAC） method， respectively， and demonstrates superior robustness and registration accuracy.

Representation learning method of hierarchical rotation-invariant geometric structure for point cloud classification and segmentation

Wei LIU, Weigang LI, Zhiqiang TIAN

2026, 46(8):  2594-2602.  DOI: 10.11772/j.issn.1001-9081.2025070897

Asbtract ( )   HTML ( )   PDF (1331KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The existing point cloud deep learning methods can process point cloud data from fixed perspectives. However， in practical applications， changes in object directions affect the point cloud description by rotational transformations， thereby reducing recognition accuracy of the deep learning network. To address this issue， a representation learning method of hierarchical rotation-invariant geometric structure for point cloud classification and segmentation was proposed. First， point cloud samples were modeled using triangulated local geometric structure， so that a triangular surface was constructed within each neighborhood， and rotation-invariant features describing the geometric relationship between Euclidean space and the tangent plane were extracted. Then， the extracted rotation-invariant features were expressed through convolutional operators， and the self-attention enhanced convolutions were employed to aggregate local neighborhood structure， so as to achieve adaptive fusion of local and global information， thereby further extracting the refined rotation-invariant features as well as improving their expressiveness and global consistency. Finally， a hierarchical Inverted Residual Block （IRBlock） was introduced to fuse features from shallow geometric features to deep semantic features hierarchically through multi-level nonlinear mapping and progressive channel expansion， thereby enhancing the high-level expressiveness and discriminative ability of rotation-invariant features， and improving the ability to express and distinguish complex spatial structures and diverse rotations. Experimental results show that the proposed method achieves the Overall Accuracy （OA） of 93.9% on the ModelNet40 dataset， the OA of 87.8% on the ScanObjectNN dataset， and the mean Intersection over Union （mIoU） of 82.3% in the segmentation tasks of the ShapeNet dataset. It can be seen that the proposed method has strong classification and segmentation performance with rotation-invariance， and exhibits excellent robustness and generalization capability.

FCMdepth： monocular depth estimation framework with multi-scale feature optimization

Fengchun LIU, Xinying SHAO, Chunying ZHANG, Liya WANG, Jing REN

2026, 46(8):  2603-2611.  DOI: 10.11772/j.issn.1001-9081.2025070890

Asbtract ( )   HTML ( )   PDF (1394KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of insufficient feature extraction and inadequate context modeling in monocular depth estimation， we proposed a multi-scale feature fused monocular depth estimation optimization framework， FCMdepth， to enhance prediction performance. FCMdepth adopted an Encoder-Decoder （ED） structure， where the encoder， FC-Net， consists of MobileNetV3-F and CDBlock， and optimized features through multi-scale features and dilated convolutions， while the decoder， LapMA-Net， combined the Laplacian pyramid with an Efficient Multi-scale Attention （EMA） module to enhance cross-scale feature fusion and output accurate depth maps. Experimental results on the KITTI datasets show that， compared to Lapdepth， FCMdepth achieves lower values for the three error metrics： Root Mean Square Error （RMSE）， Root Mean Square Logarithmic Error （Log_RMSE），and Square Relative error （Sq_Rel）， with reductions of 0.831， 0.009， and 0.145， respectively， and improves three accuracy metrics by of 0.4， 0.8， and 0.3 percentage points， respectively. It can be seen that FCMdepth has superior performance compared to other models in most metrics and provides an effective reference for monocular depth estimation and 3D reconstruction in complex scenes.

Cross-camera multi-target online tracking algorithm combining DINOv2 and ByteTracker

Shihao SHI, Qunshan SHI, Yang ZHOU, Xiaofei HU, Kai QI, Fengchao LI

2026, 46(8):  2612-2619.  DOI: 10.11772/j.issn.1001-9081.2025070846

Asbtract ( )   HTML ( )   PDF (1499KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Multi-Target Multi-Camera Tracking （MTMCT） holds significant application values in fields such as intelligent surveillance and intelligent transportation. However， due to differences in visual angles， lighting， and imaging quality of different cameras， there are high intra-target variability and inter-target similarity， posing severe challenges to cross-camera multi-target tracking. Although the existing offline methods have made progress， the limitations such as high latency and complex post-processing operations make them unsuitable for real-world scenarios requiring online processing. Therefore， an online cross-camera multi-target tracking algorithm combining DINOv2 （Data-efficient image network version 2） and ByteTracker was proposed. First， during feature extraction， DINOv2 was employed to design and train an appearance feature extraction network， thereby obtaining more robust and discriminative target appearance features. Then， based on the ByteTrack algorithm， by optimizing the data association strategy and trajectory management mechanism， the ByteTracker algorithm was proposed to generate candidate trajectories for each camera. Finally， through the Inter-Camera Association （ICA） module， hierarchical clustering was employed to associate these candidate trajectories across different cameras， thereby linking the targets to the global identities. Validation results on the CityFlowV2 dataset’s Scene 02 （S02） show that the IDentification F1 score （IDF1）， IDentification Precision （IDP）， and IDentification Recall （IDR） metrics of proposed algorithm are 81.24%， 84.59%， and 78.14%， respectively， which are improved by 2.84， 5.79， and 0.14 percentage points， respectively compared to those of the baseline model Shim. Moreover， the tracking time of the proposed algorithm under multi-camera online conditions is only 0.14 seconds.

Low-altitude dynamic navigation method based on visual AI and 3D grid computing

Shengwen HE, Shujun ZHANG, Yixuan LIU, Hui LI, Yufeng LIU

2026, 46(8):  2620-2629.  DOI: 10.11772/j.issn.1001-9081.2025070852

Asbtract ( )   HTML ( )   PDF (4366KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aiming at problems of slow update of 3D aeronautical chart， poor dynamic obstacle avoidance ability and lack of global perception in Unmanned Aerial Vehicle （UAV） navigation， a low-altitude dynamic navigation method based on visual Artificial Intelligence （AI） and 3D grid computing was proposed. In the method， the 3D spatial grid representation system was constructed， grid index was used to represent spatial information associations efficiently， combined with the airborne visual AI model， the environmental semantic information was perceived in real time， and the perceptual results were mapped into the 3D spatial grid in real time through the semantic feature embedding algorithm， so as to support the autonomous obstacle avoidance flight of UAVs. In local grid corresponding to the same sub-level airspace， a self-organizing communication network was constructed to share position， speed， height and other states of UAV in real time， thereby forming a low-altitude dynamic information pool， so that UAVs were able to grasp the neighboring UAV states in high-density areas such as take-off and landing areas accurately. At the same time， the 3D grid was subdivided recursively， and hierarchical search path was used to achieve overload grid intelligent flow distribution through the local grid interactive negotiation protocol. Taking the low-altitude scene of 50 square kilometers of city as the experimental object， three kinds of scenes， simple， medium and complex， were set up to compare the proposed method with Global Navigation Satellite System （GNSS） pure positioning navigation and monocular visual feature point navigation. The results show that in the three scenarios， the average obstacle avoidance efficiency of the proposed method is 96.7%±1.2%， 95.3%±1.1% and 93.8%±1.5%， respectively， which is 10.5， 10.8 and 11.7 percentage points higher than that of GNSS pure positioning navigation， and 26.9， 27.8 and 28.6 percentage points higher than that of monocular visual feature point navigation. It can be seen that the proposed method can solve the core pain points of UAV navigation effectively， and improve navigation efficiency and safety in complex low-altitude environment significantly.

Object detection algorithm based on improved YOLOv11 in foggy weather

Li TANG, Jianyu ZHANG, Rui YAO

2026, 46(8):  2630-2639.  DOI: 10.11772/j.issn.1001-9081.2025070921

Asbtract ( )   HTML ( )   PDF (1908KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address challenges of detecting multi-scale objects under adverse weather conditions such as fog weather， including severe occlusion， low image visibility， leading to frequent false positives and missed detections， a object detection algorithm based on improved YOLOv11 in foggy weather， named MFA-YOLO （Multi-scale frequency-aware Feature Alignment YOLO）， was proposed. First， the multi-scale edge feature enhancement module C3k2_MSEFE （C3k2_Multi-Scale Edge Feature Enhancement） was designed to replace the C3k2 module in YOLOv11 network， so as to enhance network’s ability to extract features from multi-scale objects. Second， the Lightweight Frequency-Aware Feature Pyramid Network （LFFPN） was designed to address the issue of inaccurate bounding box localization caused by low contrast in foggy images， so as to improve both accuracy and robustness of detection. Finally， the Dynamic Feature Alignment Detection Head （DFADHead） was designed to strengthen capability to detect overlapping and occluded objects in dense scenarios through mechanisms such as dynamic convolution and task decomposition. Experimental results demonstrate that on the RTTS dataset， the proposed algorithm achieves an mAP@50 of 76.5% and an mAP@50：95 of 52.6%， representing improvements of 2.7 and 1.9 percentage points compared to the baseline model， respectively. Meanwhile， the number of parameters and model size are reduced by 15.3% and 23.4%， respectively. Furthermore， verification results on two types of foggy images with different fog concentrations synthesized from Cityscapes dataset show that the proposed algorithm achieves the mAP@50 of 48.5% and 46.7%， respectively， which are 2.4 and 3.5 percentage points higher than those of YOLOv11s， respectively， further validating the adaptability and effectiveness of the proposed algorithm in various foggy weather scenarios.

Foreign object detection method for transmission lines based on improved Hyper-YOLO model

Cheng WANG, Zhilong LIU, Junnan DU, Wen YANG, Tianyi WANG

2026, 46(8):  2640-2651.  DOI: 10.11772/j.issn.1001-9081.2025070839

Asbtract ( )   PDF (3006KB) ( )  

References | Related Articles | Metrics

Deep learning-based foreign object detection methods for transmission lines have become a critical way for continuous and stable operation of next-generation power grids. To enhance detection performance for foreign objects on power transmission lines， a hypergraph-based global-perspective position-aware foreign object detection model was proposed using Hyper-YOLO’s ability to model and propagate high-order semantic correlations across levels and positions. Firstly， a multi-path feature aggregation network with position awareness — PMFA （Position-aware Multi-branch Feature Aggregation） was introduced to position critical features needed to be represented. Additionally， a Global SPPF （Global Spatial Pyramid Pooling-Fast） module with a global receptive field was designed to enhance boundary features extracted by the backbone， so as to promote the modeling and learning abilities of the model to high-order semantic correlations. Finally， an attention-guided LAWD （Lightweight Attention-guided Weighted Downsampling） module was proposed to retain critical information as more as possible from low-quality data samples. The effectiveness and superiority of foreign object detection in transmission line scenarios of the proposed model were verified through ablation studies， comparative analyses， and data augmentation on the constructed dataset of foreign object images of transmission lines. Experimental results demonstrate that the proposed model achieves a precision of 90.6%， a recall of 87.6%， and a mAP@50 of 93.4%， and compared with the baseline model， it has the number of parameters reduced by 20.5%， and the floating-point operations reduced by 20.4%， providing technical support for the detection and management of foreign objects on transmission lines.

Remote sensing image object detection algorithm based on improved real-time detection Transformer

Lin HUANG, Yingshun LI, Weiyan TONG, Shuyuan ZHANG, Zihan WANG

2026, 46(8):  2652-2659.  DOI: 10.11772/j.issn.1001-9081.2025070856

Asbtract ( )   HTML ( )   PDF (2386KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address low detection precision of remote sensing image object detection caused by complex backgrounds， numerous small objects， and long-tail objects， a remote sensing image object detection algorithm named SLT-DETR （Small and Long-Tailed object-aware real-time DEtection TRansformer） was developed on the basis of improved Real-Time DEtection TRansformer （RT-DETR）. First， a Multi-scale Feature Enhancement Pyramid （MFEP） was employed to optimize the original continuous convolutional fusion， and an OmniKernel module was introduced to further enhance the ability of small object detection. Then， the content-guided attention was utilized to fuse features output by P3， P4， and P5， thereby the addressing information distribution imbalance effectively. Finally， a FasterBlock module was adopted for lightweighting model with improved inference speed. Experimental results show that on the self-constructed remote sensing dataset， SLT-DETR improves the mean Average Precision （mAP） by 5.0% compared to RT-DETR， with the number of model parameters reduced by 1.9×10⁶； compared with classic models YOLOv8m， YOLOv8l， and Swin Transformer， SLT-DETR achieves mAP improvements of 5.7%， 3.7%， and 10.9%， respectively； on the RSOD and HIT-UAV datasets， SLT-DETR achieves improvements in mAP of 1.1% and 1.3%， respectively. It can be seen that the proposed model enhances the detection of small and long-tailed objects effectively while maintaining lightweight model design.

Sparse CT artifact suppresion Transformer network based on multi-scale attention adaptive fusion

Jielong BAI, Chenyun FANG, Zhiwei QIAO

2026, 46(8):  2660-2667.  DOI: 10.11772/j.issn.1001-9081.2025070817

Asbtract ( )   HTML ( )   PDF (1852KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Sparse Computed Tomography （CT） reconstruction can reduce patient radiation dosage and is of great significance for clinical diagnosis. In deep learning-based image reconstruction tasks， classic networks such as Uformer （U-shape Transformer）， Restormer （Restoration Transformer）， and AST （Adaptive Sparse Transformer） fail to consider multi-scale and directional information in images， neglecting the balance between local details and global structure， resulting in limited artifact suppression. To address this issue， a Multi-scale Attention Adaptive Fusion Transformer （MAAF-Transformer） network was proposed. In this network， a parallel attention fusion strategy was adopted， a Multi-Scale Channel Direction-aware Attention （MSCDA） module was combined to capture artifact features， and a Convolutional Adaptive Spatial Channel Attention （CASCA） module was used to adjust the weights dynamically. Then， effective information was filtered through a Gated Deep-convolutional Feedforward Network （GDFN）. Experimental results show that at 60 sparse angles， MAAF-Transformer achieves 0.714 1 dB higher Peak Signal-to-Noise Ratio （PSNR）， 0.33% higher Structural SIMilarity （SSIM）， and 7.76% lower Root Mean Square Error （RMSE） compared with classic Uformer， with excellent performance in terms of visual effects as well. It can be seen that MAAF-Transformer has higher sparse reconstruction accuracy and stronger artifact suppression capability.

Frontier and comprehensive applications

Research and development of SWJTU-CAX： a self-developed integrated platform for design-simulation-optimization of composite materials and structures

Ming TANG, Ligang WANG, Yifan PENG, Jiantao LIU, Qingfeng ZENG, Meitong FU, Yunxiao HOU, Kang GUAN

2026, 46(8):  2668-2680.  DOI: 10.11772/j.issn.1001-9081.2025070923

Asbtract ( )   HTML ( )   PDF (4631KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Composite materials have wide application prospects in large equipment in multiple fields， such as aerospace， shipping， and high-speed rail， due to their advantages of lightweight， high strength， and performance designability. However， current commercial software mainly focuses on single-scale functions， such as digital design of composite materials， performance prediction， macroscopic response simulations， and structural optimization. And combined use of multiple software leads to version incompatibility， inconsistent data format and other problems， which makes it difficult to satisfy the current high-end equipment development requirements of on-demand design for the integration of composite material-component-system. Therefore， a cross-system common support platform SWJTU-CAX for design-simulation-optimization software research and development of composite materials and structure was designed and self-developed by combining object-oriented method with modular design idea. In the platform， cross-system development environment Qt and data visualization library VTK were used as basic tools， and data layer， logical processing layer， and presentation layer were used as overall architecture to deal with core difficulties of data management and circulation， functional organization and module expansion， efficient access and visualization of large-scale data in whole process of composite material and structure designs， cross-scale simulation， performance prediction， parallel solving， and optimization design. Besides， code organization and functional testing of SWJTU-CAX were carried out using modular approach and CMake tool， with the help of XML file for management， addition and removal of required functional modules on the platform， and rapid integration and expansion of composite material and structural design-simulation-optimization， as well as other functional modules were implemented. Finally， ceramic composite materials and components were taken as examples， and integrated optimization design and software development of digital modeling， cross-scale performance prediction， response simulation， and structural materials for ceramic composite materials with arbitrary weaving forms were realized on SWJTU-CAX platform， verifying the feasibility of this cross-system platform architecture， data organization， and project management method. It can be seen that the development of SWJTU-CAX platform provides a powerful tool for in-depth exploration of cross-scale failure mechanism of heterogeneous materials and on-demand design of integrated equipment-component-material as well as service safety assessment of equipment composite components， and provides technical ideas and references for domestic industrial software architecture and research and development.

Message consistency parsing framework for aviation domain

Mengting WANG, Shuanliang GAO, Wenrui LIAO, Geng LI, Jiaxing ZHANG, Yuanfang ZHU, Tiange ZHAO, Yonghui LIU, Hongru LIANG

2026, 46(8):  2681-2690.  DOI: 10.11772/j.issn.1001-9081.2025091103

Asbtract ( )   HTML ( )   PDF (935KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In aviation domain， stable communication between air units and ground units requires unstructured text messages be transformed into structured information in a fixed format for transmission. Although the existing methods use the instruction-following capabilities of Large Language Models （LLMs） to perform such transformation from text into structured data， they ignore the inherent hallucination and lack of interpretability of LLMs — which pose serious threats to information security. Therefore， a Message Consistency Parsing Framework for the Aviation domain （MCPFA） was proposed to integrate model self-reflection and human-machine collaboration to ensure semantic fidelity and operational safety during the parsing process systematically. First， a Consistency Verification Method based on Self-Reflection （CVMSR） was developed to establish a progressive validation mechanism across three dimensions： syntax， semantics， and pragmatics， so as to realize dynamic verification and correction of the structured output. Second， an Intelligent Aviation Message Human-Machine Collaboration Inspection System （IAMHCIS） was designed to use confidence awareness and cross-layer consistency analysis to trigger expert intervention precisely and incorporate human feedback into model iteration， so as to form a closed-loop optimization. Finally， the framework was evaluated comprehensively through systematic experiments. The results demonstrate that MCPFA outperforms the existing methods significantly in both consistency judgment and correction tasks， identifying and fixing critical errors caused by term confusion， unit misinterpretation， or regulation violation effectively. Furthermore， the framework exhibits strong generalization capability on unseen message types. Importantly， after integrating expert knowledge， performance gains continuously， confirming the proposed framework’s growability and practical value， and the proposed framework provides a new accurate， secure， and interpretable way for LLM application in aviation and other safety-critical domains.

Electricity theft detection method based on improved Wasserstein generative adversarial network with gradient penalty and time-frequency feature fusion

Jiangyu YAN, Shiqian WANG, Hong WANG

2026, 46(8):  2691-2698.  DOI: 10.11772/j.issn.1001-9081.2025070919

Asbtract ( )   HTML ( )   PDF (1033KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Electricity theft has become a growing cause of non-technical losses in power grids， posing significant challenges for the existing detection methods by class imbalance and inadequate feature representation. Therefore， an electricity theft detection method based on improved Wasserstein Generative Adversarial Network with Gradient Penalty （WGAN-GP） and time-frequency feature fusion was proposed. First， an adaptive gradient penalty was integrated with WGAN to model and augment original electricity theft data， thereby mitigating the problem of data sample imbalance. Then， a time-frequency feature fusion method for electricity theft detection was constructed to use a feature extraction module with attention mechanisms to capture global and periodic characteristics of user electricity usage behavior in time-domain dimensionality， while extracting frequency-domain features via Fourier transform in frequency-domain dimensionality， thereby obtaining dual representation of time and frequency domain features. Finally， high-dimensional features were fused and compressed to achieve accurate theft detection. This method was validated on a real electricity usage dataset from a power company. The results show that compared with traditional oversampling methods such as Synthetic Minority Over-sampling TEchnique （SMOTE） and ADAptive SYNthetic sampling （ADASYN）， the four comparative models， K-Nearest Neighbors （KNN） algorithm， Long Short-Term Memory （LSTM） network， Wide & Deep Convolutional Neural Network （WDCNN）， and Flowformer all achieve an accuracy of over 90.00% and an Area Under the Curve （AUC） exceeding 95.00% after applying the proposed data augmentation. On this basis， the time-frequency feature fusion electricity theft detection method attains an accuracy of 95.54% and an F1-score of 95.39%， representing respective improvements of 1.90 and 1.89 percentage points over the Flowformer model. It maintains a high detection rate of 92.71% while controlling the false alarm rate at a low level of 1.66%. The proposed method outperforms all comparative approaches in both the authenticity of generated electricity theft samples and the accuracy of electricity theft user identification， exhibiting strong practical applicability.

Physiological signal extraction and heart rate measurement based on spatial-temporal Mamba model

Yan GAO, Yudong YAO

2026, 46(8):  2699-2707.  DOI: 10.11772/j.issn.1001-9081.2025070889

Asbtract ( )   HTML ( )   PDF (1437KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Non-contact heart rate measurement methods based on remote PhotoPlethysmoGraphy （rPPG） technology face two main challenges： the accurate selection of facial Regions of Interest （RoI）， and the precise interpretation of periodic patterns in rPPG from long video sequences. To address the inaccurate localization of facial RoI in the existing works and the neglect of temporal lag among different signal sources， an end-to-end rPPG signal extraction model， named RppgMamba， was proposed for non-contact heart rate measurement. In the model， the hierarchical spatial-temporal Mamba module with cross-scan mechanism was designed to capture spatial perception relationships of weak signals from multi-scale spatial-temporal receptive fields， so as to localize facial RoI adaptively， thereby avoiding interference from background and motion noise. And the frequency-domain refinement module was designed to further learn the quasi-periodic patterns of rPPG signals in the frequency domain， which enhanced output signal quality and corrected phase shifts of the output signals combining with phase compensator， thereby reducing the impact of inherent temporal lag. Experimental results demonstrate that the proposed model achieves the best performance on three public datasets， confirming its high accuracy and applicability in heart rate measurement.

Model-free photovoltaic hosting capacity assessment method using attention mechanism and deep reinforcement learning

Yuanhang HUANG, Na RONG

2026, 46(8):  2708-2715.  DOI: 10.11772/j.issn.1001-9081.2025070844

Asbtract ( )   HTML ( )   PDF (3869KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aiming at limitations of traditional Photovoltaic Hosting Capacity （PHC） assessment methods， such as over-reliance on detailed physical models and high computational complexity， as well as challenge of improving accuracy and stability of deep reinforcement learning methods in policy optimization， a model-free PHC assessment method combining attention mechanism and deep reinforcement learning was proposed. In the method， a Deep Neural Network （DNN） was constructed to predict node voltages and was combined with Soft Actor-Critic （SAC） algorithm for PHC assessment， and intrinsic correlation between state-action pairs was focused on by the interactive attention mechanism adaptively， so as to improve Q-value estimation accuracy and training stability. Case study was conducted in a real power distribution network and the proposed method was compared with the DIgSILENT-based physical modeling method and the model-free method without attention mechanism. Experimental results demonstrate that the proposed method achieves Normalized Mean Absolute Error （NMAE） of 0.045 2， Normalized Root Mean Square Error （NRMSE） of 0.061 8， Mean Absolute Percentage Error （MAPE） of 5.93%， and coefficient of determination （R²） of 0.930 5， and the maximum voltage deviation of ±3 V， verifying the validity of this method.

Drug repositioning method based on Empirical meta-path and principal component analysis

Sixiu WANG, Xinzhou CHEN, Min LI, Xiaomin ZHAO

2026, 46(8):  2716-2724.  DOI: 10.11772/j.issn.1001-9081.2025070819

Asbtract ( )   HTML ( )   PDF (1286KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Most studies in drug repositioning field rely on the “similar drugs treat similar diseases” hypothesis， requiring similarity data between diseases and drugs. However， such data has acquisition difficulties， different methods have significant discrepancies in computational results， and the research has inability to conduct when data is missing. To address these issues， this study proposed a drug repositioning method， EMP-PCA， based on Empirical Meta-Path （EMP） and Principal Component Analysis （PCA） that achieved drug-disease association prediction without similarity data. First， five meta-paths corresponding to different interaction data were introduced to generate a commuting matrix for mining multi-source correlation information. Second， PCA was employed to identify the directions with the highest variances， dimensionality reduction was performed through data projection， and computations were simplified while retaining core information. Finally， Gradient Boosting Tree （GBT） method was used to construct base classifiers for various meta-paths， and the classifiers were combined into an ensemble classifier to integrate multi-source data effectively. Experimental results of comparing EMP-PCA with similarity data-based methods like Drug Repurposing via Heterogeneous Graph Convolutional Network （DRHGCN）， Additional Neural Matrix Factorization （ANMF） model， and Layer-wise Attention Graph Convolutional Network （LAGCN） demonstrate that EMP-PCA can fuse multi-source interaction data among drugs， proteins， and diseases without requiring similarity data effectively， and the method outperforms the comparison methods in key evaluation metrics including Area Under Curve （AUC）， precision， and recall， resolving data dependency and missing value issues inherent in similarity-based methods effectively， and has high practical application value.