Table of Content

10 July 2026, Volume 46 Issue 7

Previous Issue   

Artificial intelligence

Large language model-driven method for qualifier enhancement in hyper-relational knowledge graphs

Jiahang LI, Qilong HAN, Lijie LI, Hui ZHANG

2026, 46(7):  2043-2053.  DOI: 10.11772/j.issn.1001-9081.2025060753

Asbtract ( )   HTML ( )   PDF (1197KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aiming at the problems of incomplete semantic representation of hyper-relational facts as well as insufficient task accuracy and generalization ability caused by the qualifier sparsity in Hyper-relational Knowledge Graphs （HKGs）， a Large language model-driven method for Qualifier Enhancement in HKGs （L-EQs） was proposed. First， semantically rich description information was obtained from the associated knowledge bases to mitigate semantic ambiguity caused by the same labels of entities and relations. Second， prompt templates were used to guide the LLM in generating the matched qualifiers and their semantic explanations as external knowledge. To ensure the quality of the qualifiers， an iterative prompting strategy was employed to filter the generated qualifiers， with multiple iterations conducted until the predetermined quantity was reached. Third， embeddings of external knowledge were constructed through the semantic explanations， and an external knowledge aggregation module was introduced to suppress noise， thereby providing high-quality semantic support for the downstream task by integrating external knowledge and original knowledge embeddings. Finally， the task results were predicted through a decoder， and extensive experiments were conducted on the open-source datasets WikiPeople and WD50K to validate the effectiveness of the L-EQs. Experimental results show that compared to the optimal results of the baseline models， the Mean Reciprocal Rank （MRR）， Hit@1， Hit@5， and Hit@10 of the L-EQs are increased by 8.4， 4.7， 5.3， and 6.8 percentage points， respectively， on the WikiPeople dataset， and are improved by 4.8， 2.3， 2.7， and 3.0 percentage points， respectively， on the WD50K. L-EQs addresses the qualifier sparsity in HKG and the issues caused by the incomplete semantic representations.

Dual-role interaction mechanism-based large language model for psychological health

Yadian CHENG, Yingying LI, Ping ZHANG, Fangbing QIU, Xiaonan CHAI, Yuqiao SHU

2026, 46(7):  2054-2063.  DOI: 10.11772/j.issn.1001-9081.2025060783

Asbtract ( )   HTML ( )   PDF (1485KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Existing Artificial Intelligence （AI） psychological counseling systems are generally limited to the counselor role， making it difficult to address users' dual needs for both professional advice and emotional support dynamically. To overcome this limitation， a psychological health Large Language Model （LLM） based on dual-role interaction mechanism， named STAR （Supportive Therapeutic Adaptive Responder）， was proposed. First， a dynamic role switching mechanism based on system prompts was designed to enable stylistic shifts between the psychological counselor and friend roles， thereby addressing users' dual needs for both professional advice and emotional support in psychological counseling scenarios. Then， a training process incorporating data quality control was implemented， where model feedback was continuously collected to dynamically optimize both data generation strategies and system prompt design， thereby forming a closed loop between data generation and model training and enhancing dataset quality. Experimental results on the Chinese psychology benchmark testing framework CPsyCoun （Chinese Psychological Counseling） show that the STAR model achieves improvements of 9.80% in comprehensive score of comprehensiveness， professionalism， and authenticity over the base model Qwen2.5-7B-Instruct， and 26.81% over the single-role psychological LLM EmoLLM （Emotional Large Language Model）. STAR model also outperforms models fine-tuned on open-source datasets such as SoulChat， as well as general-purpose LLMs like GPT-4o-mini. The STAR model achieves dual-role adaptive responding effectively， significantly enhancing both effect and experience of psychological counseling.

Cybersecurity named entity recognition method based on large language model linking

Didi HOU, Shaodong HONG, Yujie FU, Yunhe CUI, Guowei SHEN

2026, 46(7):  2064-2073.  DOI: 10.11772/j.issn.1001-9081.2025070867

Asbtract ( )   HTML ( )   PDF (1589KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Cybersecurity knowledge graph is an important means to conduct deep correlation analysis on unstructured Cyber Threat Intelligence （CTI） data. The Use of Large Language Model （LLM） can overcome the limitations of traditional cybersecurity named entity recognition methods， such as lack of understanding of text context， difficulty in capturing long-distance dependencies， and lack of robustness when facing noise and redundancy. However， the extraction of cybersecurity named entities by a single LLM will lead to poor adaptability of the model capability outside the domain boundary， insufficient generalization ability on Out-Of-Distribution （OOD） data， and cognitive limitations of a single model. To solve these problems， an LLM collaborative linking method that integrates fine-tuning was proposed. In the method， Low-Rank Adaptation （LoRA） fine-tuning of LLMs was performed on a base LLM using a CTI corpus， and the model's understanding of the cybersecurity field was enhanced through parameter optimization. At the same time， static voting and dynamic meta-learner linking were adopted by LLM collaboration. In the static mechanism， prediction results of multiple models were integrated through voting； and in the dynamic mechanism， the feature space adaptation function was designed to quantify the model capability boundary. Simultaneously， dynamic mapping from input data features to model performance was achieved through meta-learners such as random forests. Experimental results show that the F1-score of the proposed method reaches 91.1%， which is 6.8% higher than that of the semantic enhancement network method based on multiple features， providing an efficient and scalable solution for named entity recognition in cybersecurity.

Smart contract generation framework driven by LLM and SWRL for cross-border data compliance across heterogeneous blockchains

Linkai ZHU, Lu YANG, Hongjie ZHANG, Zijin WU, Lu LIU

2026, 46(7):  2074-2084.  DOI: 10.11772/j.issn.1001-9081.2026010107

Asbtract ( )   HTML ( )   PDF (2934KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Enterprises operating across borders often face conflicting data protection regulations and must address coordinated execution across heterogeneous blockchains， making compliant cross-border data transfer complex and risky. Therefore， a smart contract generation framework driven by Large Language Model （LLM） and Semantic Web Rule Language （SWRL） was proposed for cross-border data compliance across heterogeneous blockchains. First， centering on formalized rule definition-driven access control， cross-chain interoperability between two heterogeneous blockchains was achieved by cooperating with notarization services. Additionally， the data subject's latest authorization was injected into access decisions in real-time by utilizing an integrated consent module. Second， a rule matching framework based on finite-state decision process was constructed to decompose Semantic Web Rule Language （SWRL） compliance rules into reusable and modular logical units. Finally， a fine-tuned LLM was introduced to complete the SWRL rules semantically， which were automatically converted into executable smart contract code on the blockchain. Experimental results show that in the smart contract code generation task， the fine-tuned LLM achieves a comprehensive code generation success rate （FullRate） of 76.1%， which is improved by 18.6， 3.2， and 1.2 percentage points compared to that of the baseline models LLaMA-3-8B-Instruct， DeepSeek-V3， and DeepSeek-R1， respectively. Furthermore， the performance of the prototype system was evaluated under varying network node counts and adjustable channel configurations， demonstrating that it meets enterprise-level scalability requirements. The proposed framework significantly enhances the efficiency of smart contract generation while maintaining high security and logical correctness， demonstrating the practical feasibility and engineering application value of the framework in efficient and compliant cross-border data sharing scenarios.

Knowledge-aware recommendation performance enhancement method based on multi-heterogeneous topology collaboration and hierarchical contrastive learning

Haiwei FAN, Wenshuai DING, Hongzheng XING

2026, 46(7):  2085-2095.  DOI: 10.11772/j.issn.1001-9081.2025060795

Asbtract ( )   HTML ( )   PDF (1064KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address knowledge-aware recommendation limitations， including insufficient high-order semantic modeling， weak exploitation of item deep correlations， and weak exploitation of knowledge graph structure feature， a knowledge-aware recommendation performance enhancement method based on Multi-heterogeneous Topology collaboration and Hierarchical Contrastive Learning （MTHCL） was proposed. First， three heterogeneous graphs were constructed： a Collaborative Filtering-based Item-item Graph （CFIG） was designed to capture collaborative filtering signals， a Knowledge Graph-based Item-entity Graph （KGIG） was established to mine semantic associations， and a User Preference Graph （UPG） was developed to model multi-hop interest evolution via dynamic attention mechanism. Then， a hierarchical contrastive learning strategy was designed to establish a complementary mechanism across heterogeneous topological views. In the local layer， semantic accuracy was ensured by focusing on the direct interaction features between users and items； in the dynamic layer， scene adaptability was enhanced by using attention weighted aggregation of multi-hop relationships； in the global layer， semantic consistency was guaranteed through fusing multi-view information. Finally， regularization term was introduced to suppress overfitting， and an integrated loss function was used to optimize the method. Experimental results demonstrate that MTHCL achieves the highest Area Under Curve （AUC） and F1-score （F1）， outperforming the best baseline method — Multi-level Cross-view Contrastive Learning for Knowledge-aware Recommender System （MCCLK） by 2.23 and 2.19 percentage points on the Book-Crossing dataset， 1.24 and 1.78 percentage points on the MovieLens-1M dataset， and 1.86 and 1.97 percentage points on the Last.FM dataset. Ablation study results confirm the critical contributions of multi-view collaboration and hierarchical contrastive learning to performance enhancement， and also verify the superiority of multi-view collaboration and hierarchical contrastive learning in effectively enhancing recommendation performance.

High school English reading comprehension model integrating knowledge augmentation and contrastive learning

Yuqi WANG, Yangsen ZHANG, Pu WANG

2026, 46(7):  2096-2103.  DOI: 10.11772/j.issn.1001-9081.2025070871

Asbtract ( )   HTML ( )   PDF (783KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of information overload， lack of content matching， and insufficient interpretability of the educational exam's models of English Multi-Choice Reading Comprehension （MCRC） in high school， the MCRC model integrating Knowledge augmentation and Contrastive Learning （KCL-MCRC） was proposed. First， the Large Language Model （LLM） was employed to generate a summary and extract clue sentences from the article for knowledge augmentation to reduce the input of irrelevant information. Second， the embedded representations of the article， questions and options were obtained by using the Bidirectional Encoder Representations from Transformers （BERT） model. Third， the information difference vectors of the article， questions and options were obtained through the contrastive learning strategy combining average pooling and absolute difference， and were subsequently concatenated with the original embeddings to jointly influence the answer selection. Finally， each question type was labeled by the LLM， and the performance of the KCL-MCRC model was evaluated to analyze the reasons for the model’s performance gains through accuracy in answering various types of questions. Experimental results on the RACE-H dataset show that KCL-MCRC achieves the optimal overall accuracies of 65.75% on validation set and 64.12% on test set， surpassing the models MDT （Multi-Decision-Transformer） and MMA （Multi-stage Maximization Attention） by 1.02， 0.35 and 0.88， 0.92 percentage points， respectively， achieving the best performance； both knowledge enhancement and contrastive learning are crucial for model's performance improvement， with significant the accuracy gains in three question types： logical reasoning， main idea summarization， and author's attitude. This validates effectiveness of designing modules for different question types.

Scalable dynamic graph representation learning with neighborhood memory

Yue ZHANG, Ying WANG

2026, 46(7):  2104-2110.  DOI: 10.11772/j.issn.1001-9081.2025070907

Asbtract ( )   HTML ( )   PDF (790KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Dynamic graph learning addresses graph data in which nodes and edges evolve over time. Traditional Graph Neural Networks （GNNs） typically rely on information from nodes and their neighbors， but in dynamic graph scenarios， models must capture real-time dynamic evolution information of nodes and edges while balancing global temporal dependencies and local structural details when facing large-scale data. To address these challenges， MTGN （Memory-based Temporal Graph Network）， a scalable dynamic graph representation learning model with neighborhood memory was proposed to realize efficient modeling via subgraph sampling and neighborhood memory mechanism. First， a parallel temporal sampler was used for efficient subgraph sampling. Second， a cache-based node neighborhood memory mechanism was introduced to maintain node historical status dynamically. Third， temporal relationships among dynamic neighbors were modeled through an attention mechanism adaptively， thereby generating more stable and robust dynamic graph representations. Finally， experiments were conducted on four real-world datasets. The results demonstrate that MTGN has excellent prediction accuracy in the link prediction tasks with the prediction accuracies of 99.38% and 74.46% on the MOOC and LASTFM datasets， respectively. In the dynamic node classification tasks， MTGN achieves prediction accuracies of 90.50% and 79.08% on the Wikipedia and Reddit datasets， respectively， showing significant performance improvements compared to the baseline methods. Additionally， MTGN has outstanding training efficiency， reducing training time by at least 86.9% and 93.6% on the Wikipedia and Reddit datasets， respectively， compared to the baseline methods， validating its scalability for large-scale dynamic graph tasks.

Interpretation method based on selective Softmax gradient for layer-wise relevance propagation

Chong CHEN, Hongyang ZOU, Jing CAO, Wenjing JIANG, Jie CHEN, Fumin GAO

2026, 46(7):  2111-2118.  DOI: 10.11772/j.issn.1001-9081.2025070906

Asbtract ( )   HTML ( )   PDF (2128KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The inherent “black-box” nature of neural network models leads to opaque internal logic and behavioral decision， restricting their application in critical areas. To address this issue， based on the Layer-wise Relevance Propagation （LRP） method， the interpretability of image classification models such as VGG16 and ResNet50 was researched， and an interpretation method for image classification models named SSGLRP （Selective Softmax Gradient for Layer-wise Relevance Propagation） was proposed. By introducing activation values for the positive gradients of output neurons and modifying the initial relevance values for non-target classes， the problem of LRP heatmaps containing noise and lacking class discrimination was effectively solved. The SSGLRP method was quantitatively evaluated using maximum patch masking and fixed-point game experiments， with classic LRP， Selective Layer-wise Relevance Propagation （SLRP）， and Softmax Gradient Layer-wise Relevance Propagation （SGLRP） as baseline methods. The results of the maximum patch masking experiments show that on VGG16， the prediction change obtained by SSGLRP is on average 91.8%， 60.3%， and 34.2% higher than those obtained by LRP， SLRP， and SGLRP， respectively； on ResNet50， the improvements are 61.9%， 64.9%， and 33.2%， respectively. The SSGLRP method has higher class discrimination ability and less noise with superior performance in interpreting VGG16 and ResNet50 models.

Sample augmentation method for fault diagnosis with high-frequency guided diffusion model

Xiaohong QIU, Zixiang YU, Yujie HU

2026, 46(7):  2119-2129.  DOI: 10.11772/j.issn.1001-9081.2025060774

Asbtract ( )   HTML ( )   PDF (4013KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issue of scarce fault samples in industrial fault diagnosis that limits the performance of deep learning models， as well as the training instability and mode collapse problems in the existing Generative Adversarial Network （GAN） -based data augmentation methods， a sample augmentation method for fault diagnosis with a high-frequency guided diffusion model was proposed. Firstly， the one-dimensional vibration signals were converted into two-dimensional gray-scale spectrograms using Short-Time Fourier Transform （STFT）， and a U-Net structure-based denoising diffusion probabilistic model was constructed. Then， a high-frequency guidance module was designed to enhance the extraction and fusion of high-frequency features via high-pass filtering and cross-attention mechanism， thereby preserving fine-grained texture information in fault signals. Meanwhile， an auxiliary classifier was introduced at each denoising step of the diffusion process to improve intra-class diversity and inter-class distinction of samples via class-conditional constraints. Finally， a multi-objective optimization strategy， combined with noise prediction and classifier guidance， was adopted to stabilize the generation process and enhance sample quality. Experimental results on two public industrial fault datasets showed that the classification model trained on samples generated by the proposed method achieved maximum diagnostic accuracies of 97.41% and 96.66%， respectively. Compared with the sub-optimal method on the two datasets， WGAN-GP （Wasserstein Generative Adversarial Network with Gradient Penalty） and Conditional Denoising Diffusion Probabilistic Model （CDDPM）， the improvements were 1.92 and 2.16 percentage points， respectively. The proposed method enhances fault diagnosis performance under few-shot conditions significantly and provides a stable and reliable solution to the class imbalance problem in industrial scenarios.

Data science and technology

Dual-channel recommendation algorithm based on FFT denoising and reverse path collaborative enhancement

Liu WANG, Xuebin CHEN, Yuan GAO, Kaiguang MA, Tong ZHAO

2026, 46(7):  2130-2138.  DOI: 10.11772/j.issn.1001-9081.2025070869

Asbtract ( )   HTML ( )   PDF (2944KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of rating fluctuation and limited linear prediction ability in traditional collaborative filtering algorithms， a dual-channel recommendation algorithm based on Fast Fourier Transform （FFT） denoising and reverse-path collaborative enhancement named FRP-DCR （Frequency-domain denoising and Reverse-Path enhanced Dual-Channel Recommendation） was proposed， which integrates FFT-based denoising and reverse-path collaborative enhancement. First， user rating sequences were denoised in the frequency domain using FFT， so as to extract long-term preference trends， thereby reducing the impact of short-term emotional fluctuations on recommendation. Furthermore， a time decay mechanism was introduced to increase the algorithm's ability to represent temporal dynamics. Second， a reverse-path collaborative enhancement structure was designed during the construction of collaborative feature vectors， to simulate indirect user-item-user collaborative relationships， thereby addressing the inability of traditional methods to capture weak connections between users due to their reliance on direct user interactions. Finally， a regularized Broad Learning System （BLS） was applied to model the main rating structure， and a gradient boosting framework based efficient decision tree algorithm was adopted to model the nonlinear residuals， thereby improving prediction accuracy. The proposed algorithm， FRP-DCR， was compared with five baseline methods， including QP-DNN-TF-IDF （Quadratic Polynomial-Deep Neural Networks-Term Frequency-Inverse Document Frequency）， Light Graph Convolutional Network （LightGCN）， and Neural Collaborative Filtering （NCF）， to verify its performance. Experimental results on the MovieLens dataset show that FRP-DCR reduces the Mean Absolute Error （MAE） by 0.029 5 to 0.667 5， and reduces the Root Mean Square Error （RMSE） by 0.036 6 to 0.798 3. The algorithm achieves good performance in both metrics.

Parameter tuning method based on Monte Carlo tree search

Jintao GAO, Zhiyuan HU, Lulu JIANG

2026, 46(7):  2139-2151.  DOI: 10.11772/j.issn.1001-9081.2025070884

Asbtract ( )   HTML ( )   PDF (1840KB) ( )  

Figures and Tables | References | Related Articles | Metrics

DataBase Management System （DBMS） configuration and optimization are directly related to overall performance， but default settings of DBMS make it difficult to achieve optimal performance. The evaluation of experience-driven manual tuning shows that it is constrained by the inability to continuously capture the system state and data distribution， often resulting in suboptimal results. Bayesian Optimization （BO）， as the mainstream scheme， tends to fall into local optimum on complex objective functions， limiting the ability to discover the global optimal configuration. Therefore， a Monte Carlo Tree Search （MCTS） -based method for DBMS parameter tuning （MTune） was proposed. In MTune， the strategy tree is adopted to divide the multi-dimensional knob space into multiple regions， where each tree node is defined as an independent subspace. The objective function of Monte Carlo Tree Search （MCTS） is constructed by sending knob configurations to Bayesian Optimization （BO） and receiving corresponding evaluation metrics. With this objective function， each tree node is scored based on the Upper Confidence Bound （UCB） criterion. Besides， the root node is iteratively partitioned via the k-means algorithm to realize the growth of the strategy tree. In this way， the search space is gradually narrowed down， and a balance between exploration and exploitation can be maintained under the guidance of spatial partitioning. By refining search regions step by step， the risk of falling into local optima is effectively alleviated， and the capability of finding the global optimum is improved significantly. Experimental results show that MTune outperforms advanced baselines under YCSB-A and YCSB-B workloads. On PostgreSQL v9.6， MTune with optimal HeSBO （Hashing-enhanced Subspace Bayesian Optimization） has reduced transaction latency rate by 97.13%-97.91%， increased throughput by 24.83%-48.56%， and decreased system overhead by 1.62%-16.26%. On PostgreSQL v13.6， MTune has decreased latency by about 95% and increased throughput by 10%-25%. The system overhead is equal to or slightly better than Deep Deterministic Policy Gradient （DDPG）， and is the best compared to SMAC （Sequential Model-based Algorithm Configuration） under HeSBO-16. The proposed method can identify high-quality regions and generate approximately optimal knob configurations， and is robust and effective in practice.

Cyber security

Multi-channel covert communication model based on blockchain and social network

Wei SHE, Kebing XIA, Jiawei MA, Wei LIU, Zhao TIAN

2026, 46(7):  2152-2163.  DOI: 10.11772/j.issn.1001-9081.2025060787

Asbtract ( )   HTML ( )   PDF (2846KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems of single-channel vulnerability， low transmission efficiency， and security risks when combined with the InterPlanetary File System （IPFS） in the existing blockchain covert communication models， a multi-channel covert communication model based on blockchain and social network was proposed. First， the Content IDentifier （CID） returned after uploading secret files to IPFS was split into multiple shares using the Shamir secret sharing algorithm. Second， an embedding algorithm based on Ethereum smart contracts and another based on zero-width characters in social networks were designed to achieve cross-channel collaborative transmission of shares. Third， a redundant share mechanism was introduced， so that the CID could still be reconstructed under unstable network or partial share loss conditions， and the share parameters were supported to be adjusted dynamically to meet different security， latency， and cost requirements. Finally， the receiver was required to extract a sufficient number of shares reaching the recovery threshold from Ethereum and social networks to reconstruct the CID， thereby retrieving the secret information from IPFS. Experimental results show that with baseline parameters， the proposed model requires 21.02 s and only 1.01 USD to transmit 100 MB of data. Compared with comparison models， the transmission efficiency of the proposed model is reduced from the minute level to the second level with the cost reduced by at least 40%， and the model demonstrates superior security and robustness， making it suitable for efficient， secure， and low-cost covert transmission of large-scale data.

Distributed graph publishing method based on dual-granularity privacy protection

Yuan GAO, Xuebin CHEN, Liu WANG, Kaiguang MA

2026, 46(7):  2164-2173.  DOI: 10.11772/j.issn.1001-9081.2025070872

Asbtract ( )   HTML ( )   PDF (1671KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the challenges of limited feature extraction from sensitive subgraph data and the insufficient utility of synthetic graphs under strong privacy constraints in distributed graph scenarios， a Distributed graph publishing method based on dual-granularity privacy protection （DPP-DGP） was proposed. During the data collection phase， both coarse-grained and fine-grained information were acquired. Specifically， degree information was perturbed by the random jump algorithm， and neighbor lists were perturbed through randomized response. The neighbor lists were then sampled based on the perturbed degree information. During the synthetic graph generation phase， after receiving the privacy subgraph information from users， the data publisher employed a community detection algorithm to obtain community information，thereby retaining more topological structures. At the same time， edges were generated randomly based on edge count outside the community， and edges were generated within the community using a degree-sequence-based random graph generation method. Experimental results on four real-world datasets （email-Eu-core， Chameleon， Facebook， Wiki-Vote） show that the proposed method preserves more topological information on published synthetic graphs while guaranteeing differential privacy， compared with current distributed graph publishing methods such as Randomized Neighbor List and LDPGen. For instance， the Kullback-Leibler（KL）divergence of degree distribution of the proposed method is at least 8.0 lower than those of the baseline methods， and all four metrics（KL divergence of degree distribution， relative error of modularity， relative error of eigenvector centrality， and average path length） perform well under strong privacy protection.

Dynamic targeted recovery method for backdoor model purification

Xinming CHENG, Rong HUANG, Hao LIU, Xueqin JIANG

2026, 46(7):  2174-2183.  DOI: 10.11772/j.issn.1001-9081.2025060683

Asbtract ( )   HTML ( )   PDF (3430KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Backdoor attacks on Deep Neural Networks （DNNs） severely compromise the trustworthiness of model decisions. However， the existing defense methods relying on one-time pruning or global fine-tuning often lead to significant degradation in ACCuracy （ACC）. To address this problem， a Dynamic Targeted Recovery method for backdoor model purification （DTR） was proposed. First， pre-activation was utilized to characterize neuron behaviors， enabling the localization of poisoned neurons with abnormal behaviors. Second， during model purification， targeted recovery was implemented by fine-tuning only the poisoned neurons， thereby effectively avoiding disturbances to clean neurons and maintaining model ACC. Third， in the process of model purification， neuron behaviors were monitored to obtain feedback of the neurons to the purification， thereby realizing dynamic localization of poisoned neurons. Finally， a tabu search strategy was introduced to exclude interference from stubborn neurons with minimal contribution to purification， thereby accelerating the convergence of model purification. Experimental results on 3 benchmark datasets against 6 backdoor attacks including BadNets （Backdoored neural Network） show that the proposed method reduced the average Attack Success Rate （ASR） to at most 0.21% and improved the ACC by 0.10-2.84 percentage points， outperforming 5 other defense methods such as ABL （Anti-Backdoor Learning）. It can be seen that the proposed method effectively overcomes the ACC degradation of traditional methods caused by one-time pruning or global fine-tuning， providing a more reliable method for enhancing DNN security.

Federated learning backdoor defense algorithm based on gradient features

Qi ZHONG, Shufen ZHANG, Zhenbo ZHANG, Tao LI

2026, 46(7):  2184-2195.  DOI: 10.11772/j.issn.1001-9081.2025070873

Asbtract ( )   HTML ( )   PDF (2100KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aiming at the problem that the existing defense algorithms can not effectively distinguish the differences of client gradients in federated learning backdoor attacks and the large computational overhead， a federated learning backdoor defense algorithm based on gradient features — GradGuard was proposed. First， normalization and scaling were applied to each client gradient， and the problem of masking the backdoor dimension by benign dimensions with large absolute changes was addressed by scaling the relative changes in local model gradients. Second， a dominant gradient clustering algorithm based on cosine similarity was designed to search for small client clusters with the lowest risk probability and eliminate malicious gradients that might contain backdoors. Finally， an adaptive gradient clipping strategy was proposed， in which the clipping threshold was adjusted adaptively based on the ratio of L2 norm of the benign client dimension and the attacker， thereby ensuring the stability of the global model. Experimental results show that compared to the Datadefense algorithm， GradGuard reduces the Backdoor Accuracy （BA） of backdoor attacks by 1.91， 1.36， 1.28， and 0.49 percentage points， respectively， under four different attack frequencies， and reduces the BA of edge backdoor attacks by 8.93， 3.34， 9.51， and 0.79 percentage points， respectively. Additionally， compared to the Scope algorithm， GradGuard reduces the computational burden on servers， with training time reduced by 42.58% and 27.82% on the EMNIST and CIFAR-10 datasets， respectively， thereby enhancing the efficiency of federated learning systems.

Active defense framework against Byzantine attacks in federated learning

Zhiqiang REN, Xuebin CHEN, Changsheng QU

2026, 46(7):  2196-2207.  DOI: 10.11772/j.issn.1001-9081.2025070875

Asbtract ( )   HTML ( )   PDF (2959KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Federated Learning （FL）， an emerging distributed machine learning method， enables multiple participants to train models collaboratively while protecting data privacy. However， the existing studies have shown that FL systems face the threat of Byzantine attacks， which may lead to significant degradation of model performance or hinder model convergence. Aiming at Byzantine attacks， an active defense framework for detecting malicious updates and weakening the impact of malicious updates was proposed. The framework consists of two core modules： malicious update detection module and malicious update weakening module. The malicious update detection module identifies malicious updates by constructing a malicious sample library and combining a scoring mechanism and a clustering algorithm； while the malicious update weakening module reduces the negative impact of malicious updates on the global model effectively by distributing a specific “test model” to clients， and adjusting aggregation weights according to model updates from clients using the scoring and weight allocation mechanisms. Experimental results show that the proposed framework has the accuracy of malicious update detection close to 100% for some attack types， while ensures that aggregation weights of benign updates account for more than 90% through malicious update weakening in most attack scenarios.

Differential-linear analysis method for data encryption algorithm LiCi

Yanjun LI, Dingyun HUANG, Zhong WANG, Xin CHEN, Yuan GAO

2026, 46(7):  2208-2215.  DOI: 10.11772/j.issn.1001-9081.2025060713

Asbtract ( )   HTML ( )   PDF (710KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Currently， differential-linear security analysis of the LiCi block cipher is still missing. To fill this gap， this paper proposed a differential-linear analysis method for LiCi based on constraint programming （CP） modeling and key recovery. First， the LiCi algorithm was decomposed into three layers： a differential layer， a middle layer， and a linear layer. For each layer， constraints on the propagation of differential and linear masks were established， including inequality constraints for the S-box， branch propagation， and XOR operations of the differential and linear masks， as well as constraints for the objective function. Then， the Gurobi solver was used to perform the search. Experimental results show that a 13-round differential-linear distinguisher for LiCi with a probability of 2^-27.4 exists. By appending 4 rounds before and extending 3 rounds after this distinguisher， a 20-round key-recovery attack can be mounted， recovering 91 bits of key information with a time complexity of 2^114.3 20-round encryptions and a data complexity of 2⁴⁸ plaintexts. This result advances the number of breakable rounds of LiCi from 17 to 20， and for the first time provides a remaining security margin of 11 rounds out of the full 31 rounds.

Advanced computing

Location optimization model for ride-hailing vehicle charging stations considering temporal variability

Zichen TIAN, Qinming LIU, Chunming YE, Yujie WANG

2026, 46(7):  2216-2228.  DOI: 10.11772/j.issn.1001-9081.2025060760

Asbtract ( )   HTML ( )   PDF (941KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Significant temporal variability exists in the charging demand of ride-hailing vehicles， but the traditional static location methods tend to cause problems such as insufficient coverage during peak peirod and resource waste during off-peak period. To address these issues， a location optimization model for ride-hailing vehicle charging stations that considers temporal variability was proposed. First， an in-depth analysis of the temporal distribution of charging demand was conducted based on Nanjing's ride-hailing vehicle operation data. The whole day was divided into three periods： peak， shoulder and off-peak periods. Differentiated weights and service time thresholds were set according to the proportions of demand and user sensitivity. Second， a dual-objective model was constructed with the goals of "minimizing total deadhead time cost" and “minimizing uncovered penalty in high-difficulty areas”， and the difficulty of ride-hailing was introduced as a priority adjustment factor. Constraints such as station service capacity and regional balance were established. Finally， by integrating the local fine exploitation ability of Artificial Bee Colony （ABC） and the global exploration ability of Grey Wolf Optimizer （GWO）， an Improved artificial Bee colony-grey Wolf hybrid Optimization algorithm （IBWO） was proposed to solve the proposed model， and adaptive parameter adjustment and population restart mechanismes were also added to avoid the local optimum. Using the core area of Nanjing as a case study and optimally selecting 100 stations from 1 465 candidate locations， the proposed model achieved 100% demand coverage across all periods， with uncovered demand in high-difficulty areas at 0. After 100 iterations， compared with the optimal results of Genetic Algorithm （GA）， Particle Swarm Optimization （PSO）， GWO and ABC for solving the proposed model， the optimal fitness value of IBWO is 16 180.1， which is 35.1% lower than that of the best-performing ABC （24 929.8）； the number of stations of IBWO is reduced by 50.7% compared with ABC （203）； the calculation time of IBWO is 334.87 s， which is 14.5% shorter than that of ABC （391.75 s）. The proposed model can accurately adapt to temporal fluctuations in charging demand， and the proposed IBWO offers fast calculation speed and avoids local optimum. This paper establishes a quantitative correlation framework between charging demand temporal dynamics and location decision-making， improves the infrastructure location model under multi-period constraints. They can provide accurate decision-making basis for the planning of urban electric ride-hailing vehicle charging facilities， reduce construction costs by more than 50% while ensuring all-weather service quality， and contribute to the efficient coordination of transportation and energy systems under “carbon peaking and carbon neutrality” goal.

Quality-aware incentive mechanism based on Stackelberg game for federated learning in internet of vehicles

Tianle SUN, Tengfei CAO

2026, 46(7):  2229-2238.  DOI: 10.11772/j.issn.1001-9081.2025060784

Asbtract ( )   HTML ( )   PDF (1484KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Focusing on the issues of Non-Independent and Identically Distributed （Non-IID） data， low-quality data contributed by vehicle participants， and insufficient incentives in Internet of Vehicles （IoV）' Federated Learning （FL）， a quality-aware incentive mechanism based on Stackelberg game for federated learning in IoV was proposed. First， a clustering method based on the similarity of gradient optimization directions of vehicle participants was designed to reduce the impact of Non-IID data on model convergence and improve training efficiency. Second， a dual-dimensional quality evaluation metric was developed that combined gradient direction consistency and loss reduction to measure data quality， and a joint incentive was implemented according to data contribution. Furthermore， a Stackelberg game model was constructed to describe the interaction among cloud servers， edge servers， and vehicle clients， and the optimal strategy combination was derived using backward induction， with the existence and uniqueness of the Stackelberg equilibrium analyzed. Simulation results show that the proposed quality-aware incentive mechanism outperforms baseline mechanisms， achieving significant improvements in both social utility and cloud server utility. Compared with FedAS （Federated parameter-Alignment and client-Synchronization）， the proposed mechanism improves the accuracy by 0.28%， 1.77%， and 5.08% on the MNIST， CIFAR-10， and BelgiumTSC datasets， respectively， verifying the superiority and effectiveness of this mechanism.

Master-apprentice evolutionary algorithm based on improved local search for minimum weakly connected dominating set problem

Yu LI, Xuegang CHEN

2026, 46(7):  2239-2249.  DOI: 10.11772/j.issn.1001-9081.2025060694

Asbtract ( )   HTML ( )   PDF (716KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the limited solving accuracy of the existing heuristic algorithms for Minimum Weakly Connected Dominating Set Problem （MWCDSP）， a Master-Apprentice Evolutionary algorithm based on improved Local Search （MAE-LS） was proposed. First， double simplification rules were used in the initial solution construction stage to identify vertices that must be included in the optimal solution， thereby narrowing the searching space and improving quality of the initial solution. Second， a feasible solution construction mechanism with priority of weak connectivity was introduced in the local search stage. Compared to traditional methods， this mechanism maintained the weak connectivity and dominance of solutions more effectively. At the same time， combined with an efficient vertex selection strategy， the addition to the candidate solutions or removal from the solutions of vertices was guided dynamically， thereby improving diversity and feasibility of the solutions. Third， a dual cycle suppression mechanism combining frequency and tabu strategy was designed to effectively reduce the probability of repeated searches. Finally， a path-breaking strategy and a perturbation strategy were proposed to help the algorithm escape from the local optimum. The proposed algorithm was compared with the CPLEX （C Programming Language for EXpressions） exact solver and 6 state-of-the-art algorithms on 98 instances of the 4 benchmark test sets. Experimental results show that MAE-LS is the algorithm obtaining the largest number of optimal solutions， especially on the NDR （Network Data Repository） instance， where the number of optimal solutions obtained by MAE-LS is 61.5% higher than that of the algorithm with suboptimal performance， FPLS （Local Search algorithm based on the Feedback mechanism and the Perturbation strategy）， significantly， fully proving the significant advantages of this algorithm in terms of solving accuracy and algorithm stability. MAE-LS provides an effective solution to NP （Nondeterministic Polynomial） -hard network optimization problems， and is of great value to practical applications such as wireless sensor network construction.

Network and communications

Channel estimation algorithm for orthogonal frequency division multiplexing system based on deep neural network

Hailin XIAO, Qingyang CAO, Hailong JIANG, Zhongshan ZHANG, Zhiqun HU

2026, 46(7):  2250-2258.  DOI: 10.11772/j.issn.1001-9081.2025060782

Asbtract ( )   HTML ( )   PDF (1459KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the limited accuracy and high complexity of channel estimation in Orthogonal Frequency Division Multiplexing （OFDM） systems， a Deep Neural Network （DNN） based channel estimation algorithm for OFDM systems was proposed. First， the channel estimation and equalization processes used in traditional communication systems were replaced by the proposed algorithm， and the pilot information and data blocks at the receiver were used as input features， enabling rapid learning and estimation of channel characteristics. Second， during the model training phase， OFDM frames containing pilot blocks and data blocks were constructed， and combined with a wireless channel propagation model， a large number of simulated data pairs were generated， thereby ensuring that the training data covered a variety of complex channel environments and noise conditions. Finally， the OFDM system was used to perform a nonlinear mapping of the input-output relationship， transforming this mapping relationship into a task suitable for supervised DNN learning， thereby enabling the trained model to output accurate channel estimation results at the receiver efficiently， as well as recovering the original transmit data reliably and improving overall system performance. At the same time， a key computational module for DNN model training was designed on Field Programmable Gate Array （FPGA） to monitor and optimize the trained model in real time. Experimental results showed that， compared with the Least Squares （LS）， Minimum Mean Square Error （MMSE）， and Discrete Fourier Transform-Least Squares-Wiener （DFT-LS-WIENER） algorithms， at a Signal-to-Noise Ratio （SNR） of 10 dB， the proposed algorithm reduced the Bit Error Rate （BER） by 92.82%， 80.32%， and 72.27%， respectively， and outperformed the above comparative algorithms in Normalized Mean Squared Error （NMSE）. The proposed algorithm has good anti-interference performance and estimation accuracy under complex channel conditions； furthermore， it has moderate resource consumption on an FPGA hardware platform， and controllable operating power consumption， making it suitable for deployment on FPGA.

Multimedia computing and computer simulation

Point-MLPBLS： point cloud semantic segmentation network based on MLP cascaded broad learning system

Guoyou ZHANG, Hongyu NIE, Lihu PAN, Rundong LEI

2026, 46(7):  2259-2266.  DOI: 10.11772/j.issn.1001-9081.2025060788

Asbtract ( )   HTML ( )   PDF (1278KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Semantic segmentation of 3D point clouds is a key technology in the fields of autonomous driving and computer vision. Deep learning-based semantic segmentation models suffer from inefficient operation due to their complex network structures. To address the inefficiency of deep neural networks in 3D point cloud semantic segmentation， Point-MLPBLS， a point cloud semantic segmentation network based on Multi-Layer Perceptron （MLP） cascaded Broad Learning System （BLS） was proposed. First， the overall network adopted a dual network collaborative architecture： a feature extraction network and a point cloud segmentation network. Second， the feature extraction network performed a sampling-grouping-aggregation operation to the input point cloud data， and an inverted residual module was added to enhance the local feature representation capability. Third， a feature propagation module was used to achieve upsampling， thereby restoring the original point cloud resolution and enabling multi-scale feature fusion， thereby constructing a refined point cloud representation. Finally， the point cloud segmentation network replaced deep iterative training with a flattened MLP cascade structure， and spatially-aware MLPs were embedded in the feature mapping layer， so as to achieve high-precision inference. Experimental results show that Point-MLPBLS improves the mean Intersection-over-Union （mIoU） by 12.9 percentage points and reduces the segmentation time by 41.1% compared with PointNet++ on the S3DIS dataset， improving the efficiency of point cloud segmentation， and providing an efficient solution for 3D point cloud semantic segmentation.

Point cloud completion method based on smooth-sharpen graph convolution

Xiangyi WU, Hailiang YE, Feilong CAO

2026, 46(7):  2267-2276.  DOI: 10.11772/j.issn.1001-9081.2025060668

Asbtract ( )   HTML ( )   PDF (1412KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aiming at the problem that the graph learning-based methods in point cloud completion are easy to lead to the smoothing phenomenon of point features when the number of network layers is deepened， thereby causing local geometric detail loss， a point cloud completion method based on smooth-sharpen graph convolution was proposed， so as to improve the completion quality of irregular point clouds. Firstly， the SA （Set Abstraction） layer in PointNet++ was cross-stacked with the point Transformer module to extract multi-level features and global shape features of the incomplete point cloud. Secondly， the graph structure was constructed by generating the sub-point features through point-wise splitting， and the local similarity and coordinate difference information between point clouds were captured by using smooth graph convolution and sharpen graph convolution， respectively， so as to retain key geometric details and achieve coarse completion. Finally， global shape features and the results of coarse completion were fused in the refinement module to generate high-resolution point clouds gradually. Experimental results show that the proposed method outperforms comparison methods ODGNet （Orthogonal Dictionary Guided shape completion Network） and CRA-PCN （Cross-Resolution Aggregation Point Completion Network） on the ShapeNet and PCN （Point Completion Network） datasets. Specifically， the proposed method has the CD-T （Chamfer Distance-TopNet）， CD-P （Chamfer Distance-Point completion network） and F1-Score increased by at least 0.220， 0.285 and 0.009 on the PCN dataset， respectively， and the CD-P and F1-Score increased by at least 0.021 and 0.003 on the ShapeNet extended dataset， indicating that the proposed method has more advantages in geometric structure recovery. In addition， the completion results on the real-world KITTI （Karlsruhe Institute of Technology Toyota Technological Institute at Chicago） dataset further show that the proposed method can reconstruct the vehicle contour more completely， avoid blurring and artifacts effectively， and improve the effect of point cloud completion.

Zero-shot human-object interaction detection method via multimodal collaborative prompt optimization

Yue MA, Huicheng LAI, Di JIANG, Liejun WANG

2026, 46(7):  2277-2287.  DOI: 10.11772/j.issn.1001-9081.2025060786

Asbtract ( )   HTML ( )   PDF (4059KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the challenge of identifying unseen categories in zero-shot Human-Object Interaction （HOI） detection， a zero-shot HOI detection method via multimodal collaborative prompt optimization， named MCPNet （Multimodal Cooperative Prompt Network）， was proposed to enhance the model's generalization ability under zero-shot conditions. The method leverages text-driven prompt learning and visual feature generation mechanisms to improve the recognition performance of unseen interaction categories and the overall generalization performance from both semantic and visual modeling perspectives. First， in terms of semantic modeling， a prompt-guided module was developed by integrating human-designed prompt templates with learnable prompts， which were encoded by the Contrastive Language-Image Pre-training（CLIP）text encoder to produce fine-grained semantic representations. Then， a generation module conditioned on learnable prompts was introduced， in which a Variational AutoEncoder （VAE） was employed to model the latent semantic space and to synthesize discriminative image features for unseen categories， thus alleviating the long-tailed distribution problem. Finally， for visual modeling， a collaborative mechanism between local and global features was proposed. Local interaction region features were integrated with global image semantics as inputs to the interaction detection head， enhancing the recognition performance of HOIs in complex scenes. Experimental results in zero-shot detection tasks on the HICO-DET （Humans Interacting with Common Objects DETection） dataset show that the proposed method achieves significant gains in mean Average Precision （mAP） of unseen categories， surpassing CLIP4HOI （towards adapting CLIP for practical zero-shot HOI detection） method by 8.59% under the Unseen Combination （UC） setting， by 9.94% under the Rare-First UC （RF-UC） setting， by 8.24% under the Non-rare-First UC （NF-UC） setting， and by 9.37% under the Unseen Object （UO） setting. The conclusion verifies the effectiveness of the proposed method.

Video snapshot compressive imaging reconstruction method based on dense spatio-temporal deformable attention

Xiuli DU, Xing GAO, Xiaoyu ZHANG, Chengsheng PAN, Qijie ZOU

2026, 46(7):  2288-2296.  DOI: 10.11772/j.issn.1001-9081.2025060699

Asbtract ( )   HTML ( )   PDF (4934KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Deep learning-based reconstruction methods for Video Snapshot Compressive Imaging （VSCI） have achieved promising results in many tasks. However， challenges such as insufficient detail recovery and high computational overhead remain in dynamic scene reconstruction. To address these issues， a VSCI reconstruction method based on dense spatio-temporal deformable attention was proposed. First， the compressed measurements and masks were input to obtain initial feature representations. Second， a deformable attention module was designed to enhance the model's ability to capture local deformations and global temporal dependencies effectively. Finally， the dense connectivity was improved by designing a channel splitting factor to enable dynamic group-wise progressive feature extraction and fusion， thereby improving feature representation ability and reducing reconstruction time. Experimental results on multiple simulated grayscale video benchmark datasets （e.g.， Kobe， Runner， Drop） and color video benchmark datasets （e.g.， Beauty， Bosphorus， Jockey） showed that compared with the suboptimal methods M2BA-SCI and SCT-SCI， the proposed method improves the average Peak Signal-to-Noise Ratio （PSNR） by 0.45 dB and 0.79 dB， with reconstruction times of 0.43 s and 3.37 s respectively， demonstrating that it significantly enhances the reconstruction quality and computational efficiency for complex motion scenes.

Unsupervised dehazing method for remote sensing images with non-uniform haze

Wenling LIU, Yong LI, Jiahui LI, Houkang ZHANG

2026, 46(7):  2297-2306.  DOI: 10.11772/j.issn.1001-9081.2025060755

Asbtract ( )   HTML ( )   PDF (7845KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problem that the existing dehazing methods are difficult to suppress artifact generation and insufficient texture detail recovery caused by non-uniform haze distribution in real remote sensing images effectively， an unsupervised dehazing method for remote sensing images with non-uniform haze was proposed. First， based on Cycle-consistent Generative Adversarial Network （CycleGAN）， a Residual Multi-scale Attention Mechanism （RMAM） was designed in the generator to expand the receptive field and enhance the extraction ability of multi-scale texture and structural information， thereby effectively restoring realistic texture details. Second， a Haze Distribution Enhancement Module （HDEM） was designed to explicitly enhance the expression of haze features， and combined with a dual-branch fusion strategy， the network was guided to accurately identify and process non-uniform haze， thereby alleviating the artifact issue caused by insufficient haze distribution perception. Finally， a Feature Attention （FA） mechanism was embedded in the discriminator to strengthen its ability to distinguish the authenticity of local structures and textures in the image， thereby improving its ability to restore dehazed images. Experimental results on the synthetic remote sensing datasets SateHaze 1k and RICE showed that the proposed method outperformed the best-performing baseline method DedustGAN by 4.20% and 2.44%， respectively， in Peak Signal-to-Noise Ratio （PSNR）， and by 0.96% and 0.51%， respectively， in Structural Similarity Index Measure （SSIM）. Experimental results on the real-world dataset RRSD300 showed that the proposed method outperformed the best-performing baseline method Cycle-SNSPGAN by 2.04% and 0.45% in Natural Image Quality Evaluator （NIQE） and Integrated Local NIQE （IL-NIQE）， respectively. The proposed method effectively removes non-uniform haze， suppresses artifact generation， and restores texture details in remote sensing images.

Robotic end-to-end dynamic grasping method based on curriculum reinforcement learning

Yanyang LIANG, Wenxuan XIE, Wei CUI, Hongfei LYU, Da LI, Dongzhou ZHONG

2026, 46(7):  2307-2317.  DOI: 10.11772/j.issn.1001-9081.2025060749

Asbtract ( )   HTML ( )   PDF (2184KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the low efficiency and the difficulty of balancing high success rate with motion smoothness in robotic end-to-end dynamic grasping tasks， a robotic end-to-end dynamic grasping method based on Curriculum Reinforcement Learning （CRL） was proposed. First， a multi-modal input network that fuses color images， depth maps， and robot proprioceptive states was constructed to map raw sensory data directly to continuous action commands for the end-effector. Second， a curriculum mechanism with synchronously increasing difficulty and smoothness constraints was designed， and combined with a staged reward function， the agent was guided to master grasping from static to dynamic ones progressively. Finally， Domain Randomization （DR） was employed to enhance the policy's transfer capability of Simulation-to-Reality （Sim-to-Real）. Simulation results show that the proposed method achieves a grasping success rate of nearly 100% at target speeds ranging from 0.15 to 0.40 m/s， elevating the upper speed limit for stable grasping from 0.25 m/s of the object detection-based baseline method to 0.40 m/s. Compared to Simple Curriculum Learning （SimpleCL） with only increasing difficulty， the proposed method increases the success rate by 3.6 percentage points and reduces the average joint acceleration and jerk norm by 58.34% and 69.25%， respectively， in the most difficult test. In physical experiments， the grasping success rates of the proposed method for static scene and two dynamic scenes are 95.0%， 90.0%， and 70.0%， respectively. It can be seen that this method effectively coordinates success rate and smoothness in robotic dynamic grasping tasks by jointly optimizing task difficulty and behavioral constraints.

Frontier and comprehensive applications

Survey of research on applications of explainable deep learning in tropical cyclone forecasting

Miaogen LING, Rui JING, Wei FANG

2026, 46(7):  2318-2326.  DOI: 10.11772/j.issn.1001-9081.2025070874

Asbtract ( )   HTML ( )   PDF (618KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Deep learning offers a new path for tropical cyclone forecasting， yet its black-box characteristic hinders its application in critical fields. Therefore， a review was conducted on the application research of explainable deep learning in tropical cyclone forecasting. First， the latest advances of deep learning research on tropical cyclone forecasting were classified systematically. Second， the advantages and disadvantages of traditional methods of forecasting were reviewed. Third， attention is focused on the applications of Convolutional Neural Network （CNN）， Recurrent Neural Network （RNN）， Long Short-Term Memory （LSTM） Network， and Transformer for forecasting tasks， and the frontier explorations of methods such as Generative Adversarial Network （GAN） were discussed. On this basis， the importance of explainable deep learning in tropical cyclone forecasting was analyzed in a deep way， the specific application of explainable deep learning to tropical cyclone forecasting was explored in three dimensions： feature importance interpretation， spatio-temporal mode interpretation， and uncertainty quantification， and its potential in enhancing model transparency and user confidence was pointed out. Finally， to address current challenges faced by explainable research， future research prospects of combining physical mechanisms， multi-granularity visualization， and causal inference were introduced.

Bi-directional diffusion dynamic graph convolutional network with multi-source information fusion and graph-level attention for traffic flow prediction

Jianqiang YAN, Beibei DONG, Boting QU, Chen PENG

2026, 46(7):  2327-2333.  DOI: 10.11772/j.issn.1001-9081.2025070876

Asbtract ( )   HTML ( )   PDF (819KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Traffic data often exhibits significant dynamic spatial correlations and sudden pattern changes. Traditional models struggle to fully capture the spatio-temporal dependencies of real-world transportation networks. To address this challenge， a Bi-directional Diffusion Dynamic Graph convolutional Network with multi-source information fusion and graph-level attention for traffic flow prediction （BDDGNet） was proposed. The network integrated spatio-temporal information and multi-source dynamic features， utilizing a multi-head attention mechanism to construct dynamic graph adjacency relationships between nodes， thereby capturing spatial connectivity strength evolving over time in the traffic network. For spatial modeling， a bi-directional diffusion dynamic graph convolution module was designed to learn the propagation dynamics of node states in both directions， thereby enabling collaborative modeling of spatial dependencies in traffic flow. At the same time， a graph-level attention mechanism was introduced to extract global semantic information of key nodes， thereby enhancing the perception capability for the overall topological structure of the traffic network. Experimental results on two real-world traffic flow datasets， PEMSD4 and PEMSD8， showed that for 15-minute， 30-minute， and 60-minute prediction tasks， on the PEMSD4 dataset， BDDGNet achieved the Mean Absolute Error （MAE）， Root Mean Square Error （RMSE）， and Mean Absolute Percentage Error （MAPE） of 18.26， 30.26， and 12.16%， representing reductions of 25.2%， 19.3%， and 26.9% compared with Diffusion Convolutional Recurrent Neural Network （DCRNN） based on static graph structure， respectively； on the PEMSD8 dataset， BDDGNet achieved the MAE， RMSE， and MAPE of 14.07， 23.44， and 9.24%， representing reductions of 16.3%， 11.1%， and 15.4% compared with DCRNN， respectively. The proposed network can capture spatio-temporal dependencies in dynamic traffic environments more accurately and significantly enhance prediction accuracy.

Short-term power load forecasting model based on dynamic convolution decomposition and multi-scale graph

Li ZHU, Lugema MI, Chunqiang ZHU, Wanru XU, Jingkai GAO, Jinqi QU

2026, 46(7):  2334-2346.  DOI: 10.11772/j.issn.1001-9081.2025060676

Asbtract ( )   HTML ( )   PDF (2180KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problem that the existing short-term power load forecasting methods have difficulty in effectively modeling non-linear structures and lack cross-scale and cross-variable interaction capabilities， a short-term power load forecasting model based on Dynamic Convolution Decomposition and Multi-Scale Graph （DCDMSG） was proposed. First，to cope with the complex non-linear structure， a dynamic convolution decomposition method was adopted to decompose the trend and seasonal items in the load sequence. Second， the trend item was forecasted using a linear layer directly， and for the seasonal item， a multi-scale sequence was constructed for deep modeling using an adaptive multi-scale sequence construction method. Third， to model the complex dependencies within and outside the multi-scale sequence， a multi-scale fusion graph was used to capture dependencies within sequences at different scales， a multi-variable correlation graph was utilized to model correlations between variables， and a multi-scale mix-hop propagation was employed to aggregate features in the graph. Finally， the seasonal items were forecasted， and the obtained forecast results were weighted and fused with the trend items' forecast results to obtain the final forecast values. Experimental results show that， on the Australian dataset， the Mean Absolute Error （MAE） and Root Mean Square Error （RMSE） of DCDMSG reached 0.379 MW and 0.529 MW， respectively， which were reduced by 7.11%-21.37% and 9.88%-26.32% compared with all comparison models. On the Cele dataset， DCDMSG achieved an MAE of 0.437 MW and an RMSE of 0.708 MW， representing decreases of 5.61%-24.66% and 5.22%-24.12% over all comparison models. In the short-term power load forecasting task， DCDMSG effectively improves the forecast accuracy through seasonal-trend decomposition as well as cross-scale and cross-variable modeling.

Bearing life prediction method based on dynamic knowledge embedding

Jing LIU, Fengfeng LYU, Wei NIU, Haipeng JI, Jian WU, Xiao ZHANG

2026, 46(7):  2347-2354.  DOI: 10.11772/j.issn.1001-9081.2025070866

Asbtract ( )   HTML ( )   PDF (2815KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In practical industrial scenarios， the significant divergence in bearing state evolution caused by complex operating conditions imposes dual constraints on bearing life prediction： incomplete mechanism understanding and distribution shift in monitoring data. Although the existing data-driven methods perform well under stable conditions， they suffer from strong reliance on annotated data and limited generalization capability under rappidly changing conditions. To address these issues，a method for Bearing Life Prediction based on Dynamic Knowledge embedding （DK-BLP） was proposed in this paper. The method dynamically integrated data-driven features and domain prior knowledge. Firstly， a dynamic knowledge graph was constructed， the prior knowledge of bearing degradation was encoded into computable triplets， and a sliding window confidence mechanism was introduced， so as to realize adaptive updating of domain knowledge. Second， relational graph convolutional networks were used to extract physically meaningful embedding vectors， which were then fused cross-modally with time-frequency features of vibration signals extracted by hierarchical convolutional networks. Finally， dynamic interactions between features and knowledge were modeled by using a multi-head self-attention Transformer， thereby allowing the model to adaptively balance the contributions of data features and mechanistic knowledge. Experimental results on the PHM2012 and XJTU datasets demonstrate that the proposed method significantly improves the accuracy of cross-condition bearing life prediction.

Rock thin section lithology identification model based on improved GhostNet

Miaomiao LIU, Yuhong ZHANG, Qiang ZHANG, Ruishan DU

2026, 46(7):  2355-2363.  DOI: 10.11772/j.issn.1001-9081.2025060785

Asbtract ( )   HTML ( )   PDF (3387KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of low identification accuracy and poor efficiency in rock thin section lithology identification， a rock thin section lithology identification model based on improved GhostNet was proposed. First， an edge guidance module was constructed using multi-directional Sobel operator to extract edge and texture information from rock thin section images effectively. Second， CBAM （Convolutional Block Attention Module） was embedded at key positions within the GhostNet model to enhance the feature extraction capability and further improve the model's identification performance. Finally， the classifier structure was redesigned by replacing the original classifier with a global average pooling layer， so as to reduce the model's parameters， thereby improving the model training efficiency. A hierarchical classification method was adopted to divide the dataset into three major categories and 108 small categories for coarse-grained and fine-grained identification， respectively. Experimental results showed that the proposed model achieved a first-level classification accuracy of 98.15% and a second-level classification accuracy of 96.39% on the test set， with a model size of 16.3 MB. Compared with EfficientNetV2， the proposed model achieved a 1.31 percentage point improvement in first-level classification accuracy and an 81% reduction in parameters. The proposed model demonstrates superior classification performance and provides an efficient and accurate solution for lithology identification of rock thin sections.

Elevator multivariate time series anomaly detection method based on dynamic spatial-temporal encoder

Jianwei CHEN, Jiawei LU, Qibing WANG, Mengke ZHAO

2026, 46(7):  2364-2372.  DOI: 10.11772/j.issn.1001-9081.2025060800

Asbtract ( )   HTML ( )   PDF (1839KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The multivariate time series data generated during elevator operation are crucial for building an elevator safety monitoring system. To address the issue that existing methods ignore the correlation between series， an elevator multivariate time series anomaly detection method based on Dynamic Spatial-Temporal Encoder （DSTE） was proposed. First， a dynamic graph aggregation module was introduced to compute the global node self-attention through graph convolution and attention mechanism for spatial feature aggregation. Meanwhile， a graph pooling method was employed to dynamically update the graph structure and adjust the degree of correlation between series. Then， a temporal feature aggregation module was constructed by incorporating gated convolution modules into the traditional Temporal Convolutional Network （TCN）， and the temporal feature propagation was calculated utilizing neural ordinary differential equations， thereby realizing precise spatial-temporal feature modeling. Finally， multi-layer convolutions were used for decoding and prediction， so as to identify anomalies. Experimental results on a real-world elevator operation dataset show that compared to the suboptimal model DTAAD （Dual TCN-Attention networks for Anomaly Detection）， DSTE achieves a 0.34% improvement in precision， a 10.85% increase in recall， a 5.39% enhancement in F1-Score， and a 56.78% reduction in False Negative Rate （FNR）， while balancing the FNR and False Positive Rate （FPR）， verifying that the proposed method can identify elevator anomalies effectively and reduce the missed detection of abnormal events.

Cross-condition microstructure data generation method for titanium alloys based on improved CGAN

Jing LIU, Shaoze ZHAO, Xingang LIU, Haozhe NIU, Haipeng JI

2026, 46(7):  2373-2382.  DOI: 10.11772/j.issn.1001-9081.2025070865

Asbtract ( )   HTML ( )   PDF (1578KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Titanium alloys are widely used in aerospace and other fields， and optimizing their microstructure is the key to improving material properties. Therefore， exploring the correlation between thermal processing parameters and microstructure evolutionary patterns of titanium alloys has become a research hotspot. Traditional research methods heavily rely on experimental methods， obtaining data through physical sample preparation and characterization. However， the complex combinations， long experimental cycles， and high costs of thermal processing conditions make it difficult to obtain complete microstructural data across all operating conditions. To address the above issues， a cross-condition microstructure data generation method for titanium alloys based on improved Conditional Generative Adversarial Network （CGAN）， namely GL-CGAN-TL， was proposed. The proposed method consists of two modules： first， a microstructure data generation module based on global-local generator fusion， which efficiently generates microstructure data under new operating conditions through the global-local generator； second， a data quality transfer validation module based on adaptive fine-tuning， which verifies the quality of generated data by combining transfer learning technology. Experimental results on the microstructure sample data of TC18 titanium alloy show that the proposed method can maintain the generated data completing the real data effectively， while reducing the Mean Absolute Error （MAE） and Root Mean Square Error （RMSE） by more than 27.6% compared to the existing methods， which is significantly better than similar methods.