Computer Science
Started in January,1974（Monthly）
Supervised and Sponsored by Chongqing Southwest Information Co., Ltd.
ISSN 1002-137X
CN 50-1075/TP
CODEN JKIEBK
Authors
Reviewers
Editors
Current Issue
Volume 52 Issue 11A, 15 November 2025
  
Artificial Intelligence
Review of Artificial Intelligence Generated Content Applications in Natural Language Processing
YUAN Tianhao, WANG Yongjun, WANG Baoshan, WANG Zhongyuan
Computer Science. 2025, 52 (11A): 241200156-12.  doi:10.11896/jsjkx.241200156
Abstract PDF(2991KB) ( 14 )   
References | Related Articles | Metrics
With the explosive development of large language models in recent years,the applications of artificial intelligence ge-nerated content in natural language processing has become a research hotspot in the field of artificial intelligence.Unlike traditional analysis and prediction models,generative models have made significant progress in the field of natural language generation in recent years,including recurrent neural networks,long short-term memory networks,generative adversarial networks,Transformer models,variational autoencoders,and diffusion models.These models have found wide applications in various generation tasks within the natural language field.Owing to the rapid development of large language models,artificial intelligence generated content has achieved remarkable results in tasks such as question answering systems,text summarization,machine translation,information extraction,and other related tasks.However,despite the tremendous progress artificial intelligence generated content has made in natural language processing,many challenges still remain.In the future,it is necessary to further optimize the training process of related models,improve their generalization ability in multi-task and interdisciplinary applications,and address issues related to the quality and safety of generated content to meet the evolving demands of emerging tasks.
Research Progress on Application of Causal Machine Learning in Medical Decision-making
ZHOU Chan, WEI Zhengxi, LIU Jiang, CHEN Yuwen
Computer Science. 2025, 52 (11A): 240800160-8.  doi:10.11896/jsjkx.240800160
Abstract PDF(1784KB) ( 11 )   
References | Related Articles | Metrics
This paper summarises the core concepts and fundamentals of causal machine learning,as well as the research progress of its application in healthcare,providing an important reference for medical researchers,doctors and policy makers.It introduces the basic concepts of causal learning,the main causal models,and causal machine learning models,systematically sorts out the progress and challenges of the application of causal machine learning in medical decision-making.This paper points out that causal machine learning-related technologies can be effectively applied to the process of medical diagnosis,treatment,and prediction to enhance the ability to control and identify the disease,thus helping doctors and decision makers better understand and predict the treatment effect,and provide more effective medical solutions for patients.Therefore,causal machine learning has a broad application prospect in medical decision-making,but it still faces challenges in data quality and model interpretability at the moment.Future research should focus on how to overcome the existing challenges,provide more precise and personalised medical decision support to maintain patients’ health to a greater extent.
Review of Application of Information Extraction Technology in Digital Humanities
WEI Hao, ZHANG Zongyu, DIAO Hongyue, DENG Yaochen
Computer Science. 2025, 52 (11A): 250600198-10.  doi:10.11896/jsjkx.250600198
Abstract PDF(2196KB) ( 14 )   
References | Related Articles | Metrics
Digital humanities,as an emerging interdisciplinary field integrating computer science and humanities,aims to address research challenges in humanities through digital technologies,thereby advancing disciplinary development,cultural heritage preservation,and cultural dissemination.Information extraction,a core task in natural language processing,enables the automatic extraction of structured knowledge from unstructured texts,providing valuable data support for digital humanities research.This review systematically examines the applications of information extraction technologies in digital humanities,focusing on three key subtasks:named entity recognition,relation extraction,and event extraction.The study traces the evolution of these tasks from early rule-based and dictionary methods to traditional machine learning approaches,and further to current mainstream techniques based on deep learning and pre-trained language models,analyzing the trajectory of technological advancements.Furthermore,the review discusses the unique challenges of information extraction in digital humanities,including data scarcity,complex text structures,ambiguous entity boundaries,and implicit relationship expressions,while critically evaluating the applicability and limitations of existing methods.Finally,future research directions are outlined,such as multimodal information extraction,cross-lingual processing,optimization for low-resource scenarios,knowledge graph construction,and language generation technologies.The review offers theoretical insights and practical guidance for further research and applications of information extraction in digital humanities.
Human-Machine Trust Prediction Using Behavior Measures and Trust Relationships
ZHU Renze, YANG Ning, WANG Baohui
Computer Science. 2025, 52 (11A): 250300110-8.  doi:10.11896/jsjkx.250300110
Abstract PDF(2428KB) ( 13 )   
References | Related Articles | Metrics
With the rapid development of AI technologies in the aviation industry,understanding and quantifying human-machine trust have become especially important for flight safety,handling emergencies,and improving flight efficiency.Research has shown that physiological and behavioral features are closely related to trust,but current trust prediction studies seldom consider these features.To address this gap,this research proposes a deep learning-based human-machine trust prediction model,which integrates TCN and GAT,using gaze and behavioral features as inputs.To more effectively capture the complex relationships between trust states and their influencing factors(such as system performance and various behavioral features),a causal graph mo-del is constructed.The study conducts experiments based on the MATB task and collects trust-related data from 13 participants.Data analysis results indicate that the experimental design is sound and successfully identifies features that are highly correlated with trust,which are used for trust prediction.The experimental results demonstrate that,compared to existing trust models and traditional methods,the proposed model improves prediction accuracy by at least 16%.This result not only validates the effectiveness of the proposed approach but also highlights the significant potential of combining deep learning techniques with trust-rela-ted features for prediction tasks.Furthermore,since the model is not limited to a specific domain,this research can provide valuable references for trust prediction in other fields.
Japanese Text Clustering Based on Multi-attribute Word Embedding
YU Juan, LI Weiting, ZENG Xinyi, ZHAO Huiyun
Computer Science. 2025, 52 (11A): 241100087-9.  doi:10.11896/jsjkx.241100087
Abstract PDF(3172KB) ( 12 )   
References | Related Articles | Metrics
To address the problems of information loss in traditional Japanese text representation and the difficulty in processing high-dimensional sparse vectors,we study Japanese text word extraction and clustering methods.Firstly,the words are extracted using the improved atomic-word-step method based on Japanese linguistic characteristics.The Multi-attribute Fusion Weight (MFW) of the words is calculated combining their statistical features,positions,word lengths and semantic features so as to obtain a set of text feature words for retaining text information while reducing feature dimensionality.Then,Japanese texts are represented as the BERT-weighted MFWs of feature words,which is fused into the deep embedding model framework improved by the K-means++ algorithm to realize the clustering of Japanese texts.Experimental results on two Japanese text datasets with different topics show that the approach proposed in this paper improves both the NMI and Purity index values by more than 5% compared with the existing methods,which demonstrates a good clustering performance.
Multi-language Embedding Graph Convolutional Network for Hate Speech Detection
ZHAO Hongyi, LI Zhiyuan, BU Fanliang
Computer Science. 2025, 52 (11A): 241200023-8.  doi:10.11896/jsjkx.241200023
Abstract PDF(2314KB) ( 10 )   
References | Related Articles | Metrics
With the widespread use of social media,the issue of the spread of online hate speech has become increasingly severe,especially under the cover of anonymity on the Internet,allowing hate speech to spread rapidly,posing a serious challenge to the detection of hate speech.In order to effectively address this issue,this paper proposes a cross-lingual hate speech detection me-thod based on Multi-language Embedding Graph Convolutional Network(MEGCN).This method fully integrates the advantages of sequence modeling and graph modeling,and uses multi-language pre-trained models for feature extraction,thus being able to handle complex relationships between different languages.At the same time,this paper proposes a joint training method based on interpolation prediction to improve the accuracy and robustness of the model.Experiments on four public datasets show that MEGCN achieves better performance than all existing comparative models in the task of cross-lingual hate speech detection.This method not only maintains a high sequence modeling accuracy,but also effectively captures the structural relationships between texts,thereby improving the performance of the model in multi-language environments,especially in terms of semantic correspondence between different languages.
Bidding Document Named Entity Recognition Algorithm Based on Multi-head Attention Mechanism and Dictionary Feature Fusion
YANG Hua, WANG Baohui
Computer Science. 2025, 52 (11A): 241000154-6.  doi:10.11896/jsjkx.241000154
Abstract PDF(2371KB) ( 16 )   
References | Related Articles | Metrics
The preparation and review of bidding documents play a crucial role in ensuring the smooth operation of the bidding process.Entity recognition technology can notably enhance the accuracy and efficiency of information extraction,thereby improving the readability and retrievability of information during the review of bidding documents.However,due to the complexity of the content and the presence of numerous specialized terms,recognizing long entities poses a significant challenge.Traditional methods for named entity recognition(NER) perform poorly in addressing these issues.This paper proposes an NER approach named Roberta-DFF-BiLSTM-MHA-CRF,which integrates a multi-head attention mechanism,dictionary feature fusion,and the Roberta-BiLSTM-CRF model.Utilizing Roberta as the input layer,this method enhances the capability to capture long-range dependencies.The introduction of the multi-head self-attention mechanism improves the recognition of long entities.Meanwhile,incorporating domain-specific dictionary features addresses the issue of unclear term boundaries.Experimental results demonstrate that the proposed model significantly boosts the accuracy and efficiency of information extraction in the context of NER for bidding documents.When compared to the Bert-BiLSTM-CRF model,it achieves a 2.49 percentage point improvement in precision,a 4.28 percentage point increase in recall,and a 3.37 percentage point enhancement in F1 score.These improvements effectively reduce time and labor costs,offering an efficient new solution for information extraction from bidding documents.
Calculation of Police Incident Address Similarity Based on Fusion Model
ZHANG Shuo, JI Duo
Computer Science. 2025, 52 (11A): 241200035-8.  doi:10.11896/jsjkx.241200035
Abstract PDF(3579KB) ( 11 )   
References | Related Articles | Metrics
With the widespread application of big data technology in the field of public security,the improvement of police response speed has become one of the core goals to promote the modernization and efficient operation of public security.The rapid response system for police incidents replaces traditional manual dispatch with an automatic dispatch mechanism,and its core relies on the model’s accurate identification of police addresses.However,there are significant differences in feature representation between police addresses and regular addresses,and existing commercial address matching models often suffer from insufficient adaptability when dealing with police addresses.To address this issue,this paper proposes an improved method that combines address grading and pinyin information,aiming to replace traditional deep learning algorithms and address the limitations of commercial address calculation models in police address recognition.This method is optimized for the special phrases,multi-level address structure,homophones,and misspellings in Chinese police addresses.By using techniques such as pre-training models,data augmentation,address grading,and Pinyin information encoding,this paper aims to develop and train an efficient model specifically designed for calculating the similarity of police addresses,significantly improving the recognition accuracy and adaptability of Chinese police addresses.
Translation Quality Estimation Based on Cross-lingual Term Attention Mechanism
WANG Xueni, YE Na, ZHANG Guiping
Computer Science. 2025, 52 (11A): 250200007-9.  doi:10.11896/jsjkx.250200007
Abstract PDF(2375KB) ( 12 )   
References | Related Articles | Metrics
Translation quality estimation(QE) refers to the assessment of the quality of machine translations in the absence of a reference translation.Existing QE systems perform well in general domains but poorly in specific domains(e.g.,engineering,medicine,law) that contain a large number of specialised terms because they focus on evaluating the semantic similarity between the original text and the translated text,and lack sensitivity to translation bias of specialised terms.In order to solve this pro-blem,this paper proposes a translation quality estimation method based on cross-lingual term attention mechanism.Firstly,a prompt template is designed to guide GPT to complete the recognition of bilingual terms.Secondly,the sentence representation is obtained using the sentence encoding module,and then the enhanced sentence representation is obtained by explicitly fusing the bilingual term information.Then,the bilingual cross-lingual representation is generated using the cross attention mechanism and the semantic similarity value is computed as a term feature.Finally,a Knowledge Enhancement Layer(KEL) is introduced into the QE model to fuse the term features with the neural features output from the model,which is processed by a feed-forward neural network to obtain the predicted scores of the model.Experimental results on English-Chinese engineering literature dataset show that the proposed method improves the main metric of Spearman correlation coefficient by 3.77 percentage points,the auxiliary metrics of Pearson correlation coefficient and Kendall correlation coefficient by 3.07 percentage points and 4.45 percentage points,when compared to the state-of-the-art methods.
Knowledge Graph Completion Model Based on Multi-semantic Extraction
LI Pengyan, WANG Baohui
Computer Science. 2025, 52 (11A): 241200012-7.  doi:10.11896/jsjkx.241200012
Abstract PDF(2120KB) ( 9 )   
References | Related Articles | Metrics
In the field of knowledge graph completion,the rich multi-semantic information between entities and relationships is of great significance for improving the accuracy of completion tasks.However,existing models often struggle to fully capture and integrate these multi-semantic features,which limits the effectiveness of completion.To address this challenge,this paper proposes a Knowledge Graph Completion Model Based on Multi-Semantic Extraction(MSE).Firstly,a multi-semantic aggregation encoder is designed to dimensionally split entity and relationship embeddings,integrating the multi-semantic information of neighboring entities and relationships.Secondly,a decoder based on multi-scale convolution is proposed,using convolutional kernels of different sizes to extract the deep semantic features of entities.Lastly,a loss function with independence constraints is designed,introducing a regularization term based on Pearson correlation coefficients to enhance the model’s multi-semantic expression capability.The experimental results show that on the FB15k-237 and WN18RR datasets,the MRR values of the MSE model are improved by 1.7% and 2.3%,respectively,compared with the optimal models of other baselines,which verifies its effectiveness on the knowledge graph complementation task.
Research on Application of Deep Learning-based Natural Language Processing Technology inIntelligent Translation Systems
FU Juan
Computer Science. 2025, 52 (11A): 241000037-6.  doi:10.11896/jsjkx.241000037
Abstract PDF(2209KB) ( 8 )   
References | Related Articles | Metrics
With the acceleration of globalization,the demand for translation is increasing,and the importance of intelligent translation systems is becoming increasingly prominent.This paper deeply studies the application of natural language processing technology based on deep learning in intelligent translation systems.Firstly,the intelligent translation system based on deep learning mainly relies on the architecture of recurrent neural networks,long short-term memory networks,and convolutional neural networks,and achieves high-quality translation through word vector representation and semantic understanding technology.In terms of system architecture,the encoder-decoder framework combined with attention mechanism significantly improves the quality of translation,while the Transformer-based model has made breakthroughs in handling long-distance dependencies.In practical applications,systems such as Google Neural Machine Translation and CUBBITT have achieved near-human translation performance through innovative data enhancement techniques and multilingual model training methods.However,current intelligent translation systems still face significant challenges in dealing with semantic ambiguity,adapting to linguistic diversity,and cross-cultural understanding.To address these issues,researchers have proposed solutions such as multi-source information fusion,cross-language pre-training,and knowledge enhancement,and have made significant progress in evaluation metrics such as accuracy and fluency.The future development of intelligent translation systems will be towards multi-modal fusion,knowledge-driven and lightweight deployment,while also needing to further improve capabilities in low-resource language translation and model interpretability.
ZHA_TGCN:A Topic Classification Method for Low-resource Sawcuengh Language
ZHAO Zhuoyang, QIN Donghong, BAI Fengbo, LIANG Xianye, XU Chen, ZHENG Yuehua, LIANG Yufeng, LAN Sheng, ZHOU Guoping
Computer Science. 2025, 52 (11A): 250100059-8.  doi:10.11896/jsjkx.250100059
Abstract PDF(3614KB) ( 11 )   
References | Related Articles | Metrics
Traditional graph convolutional network methods can effectively model graph structures under data-limited conditions.However,due to their reliance on sparse one-hot encoding,they face limitations in capturing the contextual relationships between words.This issue is particularly pronounced in low-resource language environments.Taking the Sawcuengh language text topic classification task as an example,this task faces not only data scarcity but also the challenge of complex linguistic structures.To address these challenges,this paper proposes a Sawcuengh language topic classification method suitable for low-resource settings－ZHA_TGCN.This method leverages the Sawcuengh pre-trained model,ZHA_BERT,to extract textual features,and combines these features with Sawcuengh tone features.These combined features are then input into a BiGRU to learn deep semantic representations.The learned representation vectors are used as node features for the GCN,which propagates labels to learn the feature representations of both the training data and the unlabeled test data.Finally,a Softmax layer is used to output the classification results.Experimental results show that the proposed method achieves an accuracy of 82.12%,precision of 90.08%,recall of 92.46%,and an F1 score of 90.18% in the low-resource Sawcuengh language topic classification task,demonstrating the effectiveness of the method.
Biased Retrieval-augmented Ensembling Translation Model for Aviation Manuals
YANG Chen, YE Na, ZHANG Guiping
Computer Science. 2025, 52 (11A): 241100022-10.  doi:10.11896/jsjkx.241100022
Abstract PDF(3489KB) ( 11 )   
References | Related Articles | Metrics
The aviation manuals refer to publications related to the design of large civil aircrafts,including flight manuals,maintenance manuals,and safety manuals.As a type of technical documentation that demands a high level of clarity and precision in language expression,the translation of aviation manuals requires adherence to the Simplified Technical English Specification(STE).STE is a controlled natural language that imposes explicit and stringent rules on the use of grammar and vocabulary in documentation.This paper proposes a biased retrieval-augmented ensembling translation model(BRAETM) for aviation manuals guided by STE.Within the model,biased target language sequences with the same sentence type and with lengths that meet the specification are cross-lingually retrieved to guide the translation generation at the decoder end,and a biased decoding strategy guided by the STE dictionary is adopted to correct the words in the translation.Outside the model,a non-passive translation model is selectively ensembled according to the estimation results of a prediction module,in order to generate more standardized translations in terms of sentence structure,voice and vocabulary.Experimental results show that the proposed model can generate translations that better adhere to the STE rules.Compared to the state-of-the-art baseline models,the BLEU scores of this model on two aviation manual test corpora are improved by 3.60 and 2.67,respectively.
Multimodal Entity-Relation Joint Extraction Method Based on Quantum Transformer
LI Daiyi, KONG Delong, WU Huaiguang, ZHANG Jiahui, HAN Yucan
Computer Science. 2025, 52 (11A): 241100071-8.  doi:10.11896/jsjkx.241100071
Abstract PDF(3151KB) ( 13 )   
References | Related Articles | Metrics
Multimodal Name Entity Recognition(MNER) and Multimodal Relation Extraction(MRE) are two key technologies in the construction of multimodal knowledge graphs.However,the existing MNER and MRE methods still have certain limitations in feature extraction and fusion of high-dimensional data.To address these issues,this paper proposes a multimodal entity relation joint extraction method based on quantum Transformer.Firstly,a parameterized quantum circuit for text data processing is design,which utilizes the superposition and entanglement characteristics in quantum mechanics,and combines with the Transformer model to extract deep features from text;Secondly,the pyramid visual feature extraction model is designed to obtain hierarchical features from high to low,which fully considers the multi-scale information of the image.Finally,by designing a hierarchical visual prefix network,the hierarchical multi-scale image features are aligned and fused with the text features to obtain a highly robust text representation.This study provides a new research approach for multimodal entity relation joint extraction.Experimental results on three public benchmark datasets show that the multimodal entity relation extraction method based on quantum Transformer proposed in this paper is effective and stable.
Research on Retrieval-augmented Generation Technology Combining Graph Retrieval and Contextual Ranking
XUE Xiaonan
Computer Science. 2025, 52 (11A): 250100011-4.  doi:10.11896/jsjkx.250100011
Abstract PDF(2130KB) ( 11 )   
References | Related Articles | Metrics
Complex question-answering tasks require models to efficiently retrieve relevant information from large-scale heterogeneous knowledge sources while supporting the generation of high-quality answers.However,existing retrieval-augmented generation methods face numerous challenges in knowledge retrieval,semantic relevance,and generation consistency:(1) the granularity and structured information of the knowledge retrieval module are insufficient;(2) there is a lack of contextual relevance in retrieval,limited ranking capability,and constrained generation quality;(3) generative models struggle to accurately integrate retrieved knowledge and produce contextually consistent answers.This paper proposes a novel framework,GraphRank-RAG,which combines graph-based retrieval-augmented generation with contextual ranking to address the issues mentioned.By introducing a graph-based retrieval mechanism,the framework captures deep semantic relationships within contexts,optimizing both the ranking process and answer generation.Experimental results demonstrate that the proposed method outperforms existing approaches on multiple open-domain question-answering datasets,achieving significant improvements in retrieval accuracy and generation quality.
Optimal Scheduling Algorithm for Electric Vehicle Charging and Discharging in Q-Learning Based Consortium Blockchain Framework
CAO Yongsheng
Computer Science. 2025, 52 (11A): 241200015-5.  doi:10.11896/jsjkx.241200015
Abstract PDF(2335KB) ( 10 )   
References | Related Articles | Metrics
The increasing number of grid-connected electric vehicles poses new challenges to the power system,particularly in terms of efficient energy management and secure trading.This paper introduces an optimization scheduling algorithm that integrates Q-learning with a consortium blockchain framework for EV charging and discharging.Initially,a secure and reliable power trading platform is established using consortium blockchain technology,ensuring the immutability and traceability of transactions.A comprehensive EV charging and discharging model is then developed,taking into account various physical constraints such as battery degradation and user waiting time.Based on this,a Q-learning-based intelligent scheduling algorithm is designed to identify optimal charging and discharging strategies,aiming at minimizing the overall system cost while enhancing operational efficiency.Simulation results demonstrate that the proposed method not only ensures the security of transactions but also significantly reduces system costs,validating its effectiveness and practicality.
Adaptive Red-billed Blue Magpie Optimization Algorithm Based on Mixed Strategy
DUAN Bowen, YIN Jibin, ZHANG Hang
Computer Science. 2025, 52 (11A): 241100005-10.  doi:10.11896/jsjkx.241100005
Abstract PDF(3436KB) ( 20 )   
References | Related Articles | Metrics
Aiming at the problems of rapid degradation of diversity,poor optimization accuracy,and susceptibility to local optima in the Red billed Blue Magpie Optimization Algorithm(RBMO),a hybrid strategy based adaptive Red billed Blue Magpie Optimization Algorithm(JRBMO) is proposed.Firstly,the Hammersley sequence is introduced to initialize the population,making the initial solution distribution more uniform and providing a foundation for optimization.Secondly,during the exploration phase,an adaptive spiral capture strategy is proposed to improve the search capability of RBMO by dynamically controlling the exploration range and direction of individuals.In the exploitation phase,the Levy flight strategy is introduced to locally perturb the current optimal solution and enhance the algorithm’s local development capability.Finally,an adaptive dimension mutation strategy is proposed to perform dimension mutation on individuals based on changes in population fitness distribution,avoiding the algorithm from getting stuck in local optima.The algorithm performance was evaluated on the CEC2017 and CEC2019 test sets,and the results showed that JRBMO had average win rates of 88.9% and 70%,respectively,verifying the effectiveness of JRBMO.In addition,applying JRBMO to the tension(compression) spring design problem and the three-dimensional wireless sensor network(WSN) node coverage problem,JRBMO achieves the optimal results,in which the WSN node mean coverage is 6.3% higher than that of the original algorithm,which demonstrates the universality of JRBMO in practical applications.
UAV Path Planning Method Based on Ant Colony Mixed Potential Field Method
YU Haonan, XI Wanqiang, QI Fei
Computer Science. 2025, 52 (11A): 241100179-6.  doi:10.11896/jsjkx.241100179
Abstract PDF(3049KB) ( 8 )   
References | Related Articles | Metrics
This paper proposes a path planning method based on ant colony hybrid potential field method for the path problem of UAV in motion.This method divides path planning into global path and local path,uses the improved ant colony optimization algorithm to plan the global path,and adds the dynamic improves potential field method for local path optimization.The improved ant colony algorithm improves the rapidity and safety by improving the heuristic function and safety rule,and the dynamic improved potential field algorithm improves the analysis ability of dynamic targets by adding the velocity potential field.Finally,the performance of the proposed algorithm,the traditional potential field method and the current classical optimization potential field method in different scenarios is compared in the simulation.The results show that the proposed algorithm performs well in the success rate of obstacle avoidance and the path length.
Construction and Optimization of Logic Circuit Models Based on DNA Switching Circuits
WEI Maoduan, LYU Hui
Computer Science. 2025, 52 (11A): 241200169-7.  doi:10.11896/jsjkx.241200169
Abstract PDF(3378KB) ( 11 )   
References | Related Articles | Metrics
As the complexity of DNA computational requirements increases,the corresponding DNA logic circuit models have become more intricate.To address the low applicability of current DNA switching circuits(DSC) modeling methods,the high time cost for network stabilization,and the limited number of output signals,This paper proposed the “0-1” network.This network aimed to construct molecular models of logic circuits using DSC,allowing flexible configuration of the number of output signals in multi-output logic circuits,thus expanding the modeling scope.By leveraging the programmability of DNA strands,a “transit station” molecular structure was designed to ensure smooth current flow within the circuit and reduce the stabilization time of the response network.Additionally,based on the DNA strand displacement principle,a DNA inert circuit was constructed,utilizing the mutual exclusivity of three output signals to ensure their independent expression while reducing the overall circuit size.Finally,combining the proposed method,DSC models for consistency discriminative circuits,binary classification networks,and feature discriminative networks were constructed and verified through Visual DSD simulations.The simulation results show that the proposed method not only simplifies the circuit structure but also accelerates the stabilization of the response network.These DSC-based logic circuit models demonstrate the potential of using biomolecules for signal processing.
Balanced Quantization Strategy for Efficient Post-training Quantization of BEVFormer
ZHANG Xiaoxuan, TANG Xiaoyong
Computer Science. 2025, 52 (11A): 241100059-7.  doi:10.11896/jsjkx.241100059
Abstract PDF(2872KB) ( 8 )   
References | Related Articles | Metrics
BEVFormer’s bird’s-eye view(BEV) representation achieves strong results in autonomous driving applications.However,its high memory use and computational demands make real-time deployment difficult on resource-constrained devices.BEVFormer’s ReLU activation values vary widely,creating an uneven distribution that traditional quantization metrics,such as cosine similarity and mean square error(MSE),struggle to address effectively.To overcome these limitations,this paper introduces a new post-training quantization(PTQ) method,the Balanced Quantization Strategy.This method is specifically optimized for BEVFormer,focusing on quantizing linear layers and ReLU activations.For linear layers,it uses predefined quantization ranges,while ReLU activations are quantized with customized ranges to retain key value accuracy.Further,Hessian matrix optimization dyna-mically adjusts scaling factors,reducing quantization errors and stabilizing the quantization process.Results show that the Balanced Quan-tization Strategy improves computational efficiency with minimal accuracy loss.In testing on the nuScenes dataset,the proposed 8-bit quantization method achieves less than a 1% drop in NDS,maintaining BEVFormer’s high performance.
Research on Structured Pruning Algorithm Based on Information Fusion
HUANG Haixin, XU Chenglong, FU Yao
Computer Science. 2025, 52 (11A): 241000041-6.  doi:10.11896/jsjkx.241000041
Abstract PDF(2383KB) ( 12 )   
References | Related Articles | Metrics
Aiming at the problems of high PPL( perplexity ),low text generation accuracy and slow model reasoning speed in Zero-shot Performance after the existing large-scale language model is processed by pruning algorithm,this paper proposes a pruning metric algorithm LAM based on the joint magnitude of loss.In the process of estimating the weight importance,the loss function information and the weight activation information are fused.By using the LAM algorithm,the limitations caused by the omission of the second derivative in the Taylor expansion of the gradient information in the process of weight importance evaluation are eliminated,and the accuracy and robustness of the model pruning process are improved.Enhance the versatility of the pruning algorithm.When establishing the coupling structure,a single coupling structure is proposed,and the neurons in the multi-layer perceptron( MLP ) in the Transformer block are selected as the initial trigger.Only the attention layer,the query vector,the key vector,and the value vector layer are considered to activate the neurons to establish the coupling structure.Thus,the number of parameters required to identify the coupling structure group is reduced,and the pruning speed and throughput are improved.The Zero-shot Performance experiments on WikiText2 dataset and PTB dataset show that when the pruning rate is 25 %,the PPL scores of LLaMA-7B are 20.24 and 36.05,respectively,which are lower than other pruning algorithms.The PPL scores of Vicuna-7B after pruning are 21.24 and 85.81,which are also better than other pruning algorithms,showing that the algorithm has higher universality and accuracy.
Path Planning for AGV Integrating Improved A* Algorithm and TEB Algorithm
PENG Ke, LIU Hongsheng, ZHANG Zhicheng, ZHU Liang, HE Maiqing, ZHANG Xuhui, ZENG Qijin, ZHANG Siyuan
Computer Science. 2025, 52 (11A): 240900148-7.  doi:10.11896/jsjkx.240900148
Abstract PDF(3609KB) ( 11 )   
References | Related Articles | Metrics
To enhance the autonomous navigation and obstacle avoidance capabilities of Automated Guided Vehicles(AGV),this study addresses the issues of poor path smoothness,non-optimal path length,and collision susceptibility inherent in the A* algorithm.We propose an AGV path planning method that integrates an improved A* algorithm with the Timed Elastic Band(TEB) algorithm.Initially,the search domain is expanded to 12 directions based on certain rules,broadening the AGV’s search horizon and making the search more directional.Nextly,by incorporating an obstacle factor into the heuristic function,the function can adaptively change according to the distribution of obstacles on the map,effectively reducing estimation errors.Finally,the globally optimal path planned by the improved A* algorithm is decomposed into global waypoints.Between these waypoints,the TEB algorithm is used for local path planning,ensuring that the AGV can dynamically avoid obstacles in real-time while following the globally optimal path.Simulations demonstrate that the improved A* algorithm significantly reduces the number of turns,path length,and nodes.An AGV experimental platform with an omnidirectional Mecanum wheel chassis was then constructed to test the integrated algorithm’s performance in autonomous navigation and obstacle avoidance.The results show that the proposed algorithm can effectively reduce the path length and travel time of AGV,ensuring safe arrival at the target point,thereby validating its superiority.
Hybrid Reinforcement Learning Algorithm Combined with 2-opt for Solving Traveling Salesman Problem
PENG Junlong, FAN Jing
Computer Science. 2025, 52 (11A): 250200121-8.  doi:10.11896/jsjkx.250200121
Abstract PDF(3449KB) ( 8 )   
References | Related Articles | Metrics
TSP is a classic NP-hard combinatorial optimization problem in operations research,aimed at the shortest possible cycle that visits each city exactly once from the origin point and returns back.For solving TSP,this paper presents a hybrid deep reinforcement learning approach(2+HRL) based on a pointer network,integrating graph attention mechanisms and the 2-opt heuristic.Specifically,graph attention networks(GATs) capture both local and global structural information of cities,while bidirectional LSTMs dynamically encode path dependencies for context-aware state representation.During training stage,the 2-opt strategy iteratively improves paths by local edge swaps to enhance solution quality.The policy gradient optimization via the REINFORCE algorithm is combined with an entropy reward function to avoid local optima,while a value network enhances parameter estimation accuracy.Experimental results show that 2+HRL algorithm performs better than traditional heuristics and exact algorithms,and it has faster computational speed and higher precision if limited with fewer training iterations,and its performance exceeds other deep reinforcement learning approaches as training progresses increment.
Novel Multi-modal Multi-objective Algorithm Based on Growing Neural Gas Network
XUAN Hejun, KOU Libo, LIU Ruyi
Computer Science. 2025, 52 (11A): 250100055-7.  doi:10.11896/jsjkx.250100055
Abstract PDF(2409KB) ( 12 )   
References | Related Articles | Metrics
Multi-modal multi-objective optimization is a complex multi-objective optimization problem with multiple Pareto solutions on the same Pareto front.It has become an important research direction in the field of multi-objective optimization.Existing algorithms can solve this problem well,but they have certain limitations in terms of solution diversity,convergence and handling of target conflicts,such as difficulty in effectively covering all solution sets or premature convergence during the optimization process.To solve these problems,a new multi-modal multi-objective optimization algorithm based on the environment selection strategy of the growing neural gas(GNG) network is proposed.This method introduces an adaptive topological structure to dynamically adjust the population distribution,and uses weighted Euclidean distance to calculate the crowding degree for environment selection,thereby improving the diversity and uniformity of the population.In addition,the knowledge transfer mechanism is introduced to enhance the algorithm’s search ability and further improve the diversity and convergence of solutions.To verify the effectiveness of the algorithm,experiments are carried out on the HYL and MMF test function sets.The experimental results show that the proposed algorithm performs better than the five comparison algorithms in terms of solution distribution uniformity,Pareto front convergence and target space coverage.
PPIS-MFH:Predicting Protein-Protein Interaction Sites Based on Multi-feature HybridNetwork Integrating ViT
HU Zhaolong, HU Chunling, HU Ruijie, GUO Longju
Computer Science. 2025, 52 (11A): 241000145-9.  doi:10.11896/jsjkx.241000145
Abstract PDF(3193KB) ( 7 )   
References | Related Articles | Metrics
The deeper principles of molecular life can be revealed through an in-depth study of protein-protein interaction sites(PPIS).However,existing methods for identifying PPIS are complex and time-consuming,and more accurate models are needed for PPIS prediction.Although deep learning techniques based on attention mechanisms and convolutional neural networks(CNNs) have made progress in PPIS prediction,they still face limitations in capturing amino acid features.To effectively capture long-range dependencies in protein sequences and accurately characterize amino acid properties,this paper proposes a multi-feature hybrid network(MFH),PPIS-MFH,for predicting protein-protein interaction sites.Protein-protein interaction sites are predicted by combining both global and local sequence features.For local sequence features,the PPIS-MFH model incorporates a Vision Transformer(ViT) module,which captures long-range dependencies and extracts local features from protein sequences.For global sequence features,the model employs a bidirectional gated recurrent neural network to discern intrinsic connections between amino acids in protein sequences.This is achieved through a feature crossover network that combines a text convolutional neural network(TextCNN) with an attention mechanism,specifically a text recurrent neural network(TextRNN-Attention).In this study,the PPIS-MFH model was evaluated on four datasets and compared with eight similar methods.The experimental results show that,on most metrics,the proposed method outperforms other similar methods.
Lightweight Deep Learning Network Algorithm Optimized Based on Pruning Algorithm
QIU Dandan
Computer Science. 2025, 52 (11A): 241000134-7.  doi:10.11896/jsjkx.241000134
Abstract PDF(4768KB) ( 30 )   
References | Related Articles | Metrics
With the continuous development of computer technology,many computer technologies have used intelligent algorithms to improve their intelligence level.Among them,lightweight deep learning network algorithms are one of the most frequently used,and many fields have used this algorithm to improve their production efficiency.However,current lightweight deep learning network algorithms still have drawbacks such as large algorithm scale and poor feature extraction performance.In order to solve the above problems,this study focuses on the One-dimensional convolutional neural network algorithm in deep network learning algorithms,and uses pruning algorithms to design lightweight convolutional neural network algorithms in order to optimize their performance.The study first compared the lightweight convolutional neural network algorithm with traditional algorithms,and the results showed that the speed of the lightweight algorithm was improved by nearly three times,reaching 3.7bps.At the same time,the storage requirements and energy consumption of the algorithm were significantly reduced,with energy consumption only 12.3%.Comparing the lightweight convolutional neural network learning algorithm of pruning algorithm with other lightweight algorithms,the results show that the average detection accuracy of this algorithm for different data is over 95%,far higher than other algorithms.The feature extraction effect of this algorithm is also significantly better than other algorithms,and the running time of this algorithm is only 4.98ms,far lower than other algorithms.From the above results,it can be concluded that the proposed pruning algorithm lightweight design method can improve the performance of deep learning network algorithms.
Lithium Battery State of Charge Estimation Based on Newton-Raphson Optimized UnscentedKalman Filter Algorithm
ZHANG Haonan, ZHANG Ancai, PAN Guangyuan, ZHENG Wenbo
Computer Science. 2025, 52 (11A): 241000150-6.  doi:10.11896/jsjkx.241000150
Abstract PDF(2925KB) ( 8 )   
References | Related Articles | Metrics
With the increasing precision requirements of battery management systems(BMS) in electric vehicles and energy sto-rage systems,the accurate estimation for the state of charge(SOC) of lithium-ion batteries becomes critical.To enhance the SOC estimation accuracy,this paper develops a new method using the Newton-Raphson optimized unscented Kalman filter(UKF) algorithm.Firstly,the mathematical model of lithium-ion battery is constructed based on a second-order RC equivalent circuit.In order to reduce the influence of noise’s initial value on SOC estimation accuracy,the Newton-Raphson algorithm is used to optimize the initial covariance matrices of process noise and observation noise in the UKF algorithm.This enhances the adaptability of the algorithm to the impact of noise.Finally,the incremental current experimental data are utilized to identify the parameters of the battery model.And the SOC estimation performance is validated through experiments conducted under constant current-rest and dynamic pressure test conditions.The experimental results show that the presented estimating algorithm has high precision and stability in both mean absolute error and root mean square error indices compared to the traditional UKF algorithm.This provides important technical support for optimizing battery management and ensuring the safe operation of lithium-ion batteries.
Research on Substation Camera Inspection Task Scheduling Method Based on Improved Discrete Black-winged Kite Algorithm
LI Haifeng, CHEN Qing, HUANG Yuehua, CHEN Xi, WEN Bin, WU Xichun
Computer Science. 2025, 52 (11A): 250900008-10.  doi:10.11896/jsjkx.250900008
Abstract PDF(4133KB) ( 8 )   
References | Related Articles | Metrics
Aiming at the problem of uneven task allocation and lack of flexibility in substation camera inspection,which leads to low efficiency of camera work,a camera inspection task scheduling method based on improved discrete black-winged kite algorithm is proposed.Firstly,considering the complex mapping relationship between cameras,substation equipment and inspection tasks,an optimal scheduling model for camera inspection tasks is constructed with inspection completion time,deflection angle and load balancing as the objectives.Then,heuristic joint rules based on the actual inspection-specific information are designed to generate the initial population of the optimization solution,which effectively solves the problem of random initialization uncertainty.Furthermore,the introduction of the discrete difference mutation operation and spiral search migration mechanism are introduced to improve the black-winged kite algorithm with hybrid multi-strategy search to increase the algorithm adaptability and search capability.Scenario test results show that the proposed method effectively improves the efficiency of substation camera inspection,and the method can make the camera have better stability in large-scale and long-cycle inspection tasks.
2QAN Quantum Circuit Scheduling Optimization Based on Quantum Firefly Algorithm
LI Hui, WANG Jiepeng, JI Yingsong, CHEN Yutong
Computer Science. 2025, 52 (11A): 250200097-10.  doi:10.11896/jsjkx.250200097
Abstract PDF(4503KB) ( 17 )   
References | Related Articles | Metrics
Aiming at the underconsideration of line structure characteristics and hierarchical demands of the traditional quantum scheduling strategy,and the execution will be occurred to reduce parallelism and increase the circuit depth during optimization execution process,this paper proposes the Quantum Firefly Algorithm(QFA) to apply it to 2QAN quantum circuit scheduling optimization.Quantum information is introduced to explore multiple locations simultaneously,and it increases the coverage of the search space.A balance between the exploration of the new solutions and the development of known solutions through the wave function evolution and collapse mechanism,meanwhile,the random perturbations is imported to enhance the search diversity,and the solutions will be jump out of the local optimum with quantum tunneling effect.The algorithm optimizes the order of quantum gate operations by evaluating the fitness values of different scheduling schemes to reduce the circuit depth and move operations,which in turn improves the circuit parallelism.Tests are conducted on four benchmark functions.The test results show that,compared with the firefly algorithm,the convergence speed of the quantum firefly algorithm is improved by approximately 40%,the quality of the solutions is enhanced by about 67%,and the search efficiency is increased by 45%.In the optimization of quantum circuit scheduling,compared with the traditional algorithm,the 2QAN circuit,the 2HQAA algorithm,and the combined algorithm of LCRA and LTSA,the number of SWAP gates of the quantum firefly algorithm is on average reduced by 42%,6.7%,10.4%,and 3% respectively,and the number of CNOT gates is on average decreased by 15.6%,10.8%,11%,and 2.2% respectively.
Research on Cross-Evaluation Method of Large Model
LIANG Binghao, ZHANG Chuangang, YUAN Mingming
Computer Science. 2025, 52 (11A): 241000129-7.  doi:10.11896/jsjkx.241000129
Abstract PDF(4305KB) ( 11 )   
References | Related Articles | Metrics
With the emergence of ChatGPT,large model have become a new track for global technology competition,and have begun to be widely used in all aspects of production and life.Many domestic technology companies have invested in large model research and development and open source work.As the application scenarios of large model continue to expand,there are more and more types and quantities of pre-trained large model that can be downloaded or invoked,and users’ demand for large model eva-luation is gradually increasing.At present,there is no standardized method for the evaluation of large model,and the industry mainly compares the capability of large models through the evaluation lists provided by third-party institutions.There is still a lack of effective measurement methods for the actual effect of large models in specific application scenarios.In this paper,a cross evaluation method is proposed to evaluate the application effect of the pre-trained large model in the vertical industry scenario,especially the answering ability of open questions,and its reliability and robustness are verified by experiments.The cross-evaluation method proposed in this paper has a high consistency with the official results,indicating that the method has a strong reliability.This method effectively improves the objectivity and convenience of large model evaluation,and helps users to quickly complete the horizontal comparison and selection of large models in personalized scenes.
Research on Multi-agent Joint Navigation Strategy Based on Improved Deep ReinforcementLearning
XIA Weihao, WANG Jinlong
Computer Science. 2025, 52 (11A): 250200095-7.  doi:10.11896/jsjkx.250200095
Abstract PDF(3191KB) ( 13 )   
References | Related Articles | Metrics
Driven by the rapid progress of artificial intelligence technology,multi-agent systems have shown their potential for cooperative navigation in many practical applications,such as environmental monitoring,disaster relief,and autonomous driving.These tasks can generally be summarized as the multi-agent cooperative navigation problem.However,with the increase of the number of agents involved in the task,the expansion of reinforcement learning in multi-agent systems faces problems such as inefficiency and learning inertia,which seriously restrict the performance of task execution.This paper proposes an innovative multi-agent reinforcement learning framework.The framework speeds up the learning process by building a two-tier strategy network that enables agents to consider their peers’ strategies in a partially observable environment.In addition,a dynamic reward mechanism is introduced to solve the problem of poor cooperative navigation.The experimental results show that this deep reinforcement learning model based on two-layer strategy network can significantly improve the cooperation efficiency in multi-agent cooperative navigation tasks,especially in the case of a large number of agents,its advantages are more obvious.
Image Processing & Multimedia Technology
Review of Applications of Artificial Intelligence Generated Content in Video Processing
WANG Zhongyuan, WANG Baoshan, WANG Yongjun, YUAN Tianhao
Computer Science. 2025, 52 (11A): 241200164-10.  doi:10.11896/jsjkx.241200164
Abstract PDF(4873KB) ( 12 )   
References | Related Articles | Metrics
Artificial intelligence generated content has become a key research focus in recent years,particularly in the field of video processing.With the emergence of new technologies such as Sora,a new wave of research enthusiasm has been sparked.This paper introduces the development and applications of artificial intelligence generated content in video processing and discusses future research directions and challenges.There are three parts in this paper.Firstly,it introduces the early foundational models of artificial intelligence generated content in the field of video processing,including generative adversarial networks,variational autoencoders,diffusion models and other models,summarizing the models that have made significant innovations or achieved excellent results in video generation tasks.Secondly,it compares the advantages and disadvantages of new video generation models before and after the introduction of Sora in 2023－2024 from three dimensions:basic properties,video generation quality and human subjective perspective.Finally,based on data analysis,this paper outlines the future development directions and challenges in the field of video generation,offering valuable insights for researchers in related fields and promoting the widespread adoption of generative artificial intelligence in video processing.
Music Emotion Transformation Model Based on CVAE-WGAN
XU Bei, ZHAO Dan
Computer Science. 2025, 52 (11A): 241100014-13.  doi:10.11896/jsjkx.241100014
Abstract PDF(3913KB) ( 11 )   
References | Related Articles | Metrics
Music is an important means of emotional expression for people,serving as a powerful tool for conveying feelings.Music emotion transformation technology allows for the conversion of original music into music with a target emotion,thereby mee-ting users’ demands for diverse emotional music and significantly improving creative efficiency.The existing music emotion transformation technologies achieve end-to-end emotion transformation by constructing sophisticated deep learning models.However,in current methods,the correspondence between the emotional vector representing music and the actual musical features is insufficient,resulting in a lack of interpretability in the intermediate network layers,which limits the accuracy of emotion transformation to a certain extent and may contribute to the problem of gradient vanishing.To address these issues,a new music emotion transformation model based on the CVAE-WGAN(Conditional Variational Autoencoder Wasserstein Generative Adversarial Network) architecture is proposed.This model uses WGAN-GP Network to replace the traditional GAN module and introduces Wasserstein distance and gradient penalty mechanism,which effectively avoids mode collapse and gradient vanishing,thereby further enhancing the stability of training and the quality of generated music outputs.Meanwhile,in order to address the lack of interpretability in the intermediate process of the generative model,64 kinds of intermediate perceptual features with clear interpretability are introduced,covering aspects such as music melody,harmony,rhythm,dynamics,timbre,expressiveness and form.These features are incorporated into the model as latent space variables to ensure that each dimension of the latent space corresponds to a specific and meaningful musical feature.In addition,a Gaussian mixture model is employed in place of the single Gaussian model tradi-tionally used in the variational autoencoder to capture and represent the nuanced distribution of musical features across different emotional categories.The experimental results show that this model performs excellently in transformations among six distinct emotions－happiness,sadness,tenderness,anger,fear,and surprise.Moreover,the proposed model outperforms comparative mo-dels in terms of emotional accuracy,reconstruction error,generation coherence,and generation diversity.
Drug Name Recognition Method Based on CRAFT and OCR Technology
XU Ying, LI Xiaoming, YU Fenghao
Computer Science. 2025, 52 (11A): 241200160-7.  doi:10.11896/jsjkx.241200160
Abstract PDF(3865KB) ( 12 )   
References | Related Articles | Metrics
In the operation of intelligent pharmacies,it is crucial for robots to accurately identify and retrieve drugs in order to achieve efficient and precise drug selection tasks.This study focuses on drug name recognition methods and proposes a CRAFT-OCR algorithm that integrates CRAFT algorithm and OCR technology to achieve efficient recognition of drug names.Among them,the CRAFT algorithm is used to detect the text area of the medicine box.To improve recognition accuracy,a drug name area localization method based on sorting rules is designed to determine the drug name area,and advanced OCR technology is finally used to complete text recognition.The drug name recognition experiments conduct on the collected dataset of medicine box images show that the accuracy of the CRAFT-OCR method in detecting drug name areas is 96.43%,and the accuracy of text re-cognition is 96.00%.The performance is better than existing algorithms in the literature,providing an effective solution for intelligent drug name recognition.
Multimodal Sentiment Analysis Based on Dominant Attention and Multi-space Domain Information Collaboration
FENG Guang, LIN Yibao, ZHONG Ting, ZHENG Runting, HUANG Junhui, LIU Tianxiang, YANG Yanru
Computer Science. 2025, 52 (11A): 250200022-9.  doi:10.11896/jsjkx.250200022
Abstract PDF(2497KB) ( 8 )   
References | Related Articles | Metrics
Multimodal sentiment analysis has significant applications in smart education,such as assessing students’ engagement and emotional states through speech,facial expressions,and tone to help teachers adjust teaching strategies in real time.How-ever,existing cross-modal attention mechanisms struggle to capture associations between heterogeneous modalities effectively,and the collaboration between shared and private spaces remains underexplored,limiting multimodal fusion learning.To address these issues,this paper proposes a multimodal sentiment analysis model that integrates heterogeneous modalities across multiple space domains using dominant attention.This mechanism enables effective fusion of heterogeneous modalities in both domains,enhancing cross-modal learning.Additionally,a gating mechanism preserves the modality independence of shared-space fusion vectors,ensuring complementary interactions between private and shared spaces.Experimental results on the MOSI and MOSEI datasets demonstrate that the proposed model achieves overall performance improvements,validating its ability to capture and integrate heterogeneous multimodal information effectively.
Defect Detection of Engine Engraved Surface Based on Generative Data Augmentation andImproved Faster-RCNN
TAN Jianhui, ZHANG Feng
Computer Science. 2025, 52 (11A): 241200025-7.  doi:10.11896/jsjkx.241200025
Abstract PDF(3033KB) ( 10 )   
References | Related Articles | Metrics
The engraved surface of automotive engine has the functions of carrying engine information,searching for lost,and preventing unauthorized disassembly and modification of the engine.The quality of the engraved surface will directly determine whether the vehicle can be registered and driven normally.However,in the field of automobile manufacturing,manual visual inspection is mainly used for defect detection of engraved surfaces at present,which poses a risk of missed detection.Although there have been some studies on surface defect detection in the industry,they cannot fully adapt to the defect detection of engine engraved surface,which can easily lead to false positives and false negatives.In order to innovate the method of detecting engine engraved surface defects,this paper proposes a defect detection method based on generative data augmentation and improved Faster-RCNN.Firstly,a method for generating engraved surface defect images based on stable diffusion model is proposed to address the problem of limited samples of engine engraved surface defects.This method controls the location of defect generation and restores the character features of the image through a dual mask image,thereby completing the generation of engraved surface defect images and achieving data augmentation of the dataset.Secondly,a synchronous bidirectional fusion feature pyramid network(SBFFPN) is proposed to replace the feature pyramid network(FPN) used in the original algorithm,enhancing the multi-scale feature fusion capability and solving the problem of wide target scale range of engraved surface defects.The experimental results show that the proposed method achieves mAP of 97.52% in detecting engine engraved surface defects,which is 34.73% higher than the original Faster RCNN model and can meet the detection requirements of engine engraved surface defects.
Lightweight Image Super-resolution Reconstruction Based on Feature Similarity Analysis
LIU Xingpeng, XUE Yiming, LIN Yuyang, LI Yan, PENG Wanli
Computer Science. 2025, 52 (11A): 250100057-8.  doi:10.11896/jsjkx.250100057
Abstract PDF(4552KB) ( 11 )   
References | Related Articles | Metrics
Lightweight image super-resolution(SR) networks based on Transformer have achieved promising results.However,most research efforts have focused on designing lightweight architectures while neglecting the analysis of structural redundancy within the SR networks.To address this,the feature similarity-based model design approach is proposed,which compresses attention groups with high feature similarity while retaining those with low similarity,effectively reducing redundancy within SR network.Furthermore,a novel feature extraction module integrating the frequency and spatial domains is proposed.By separately performing localized frequency domain and spatial domain feature extraction,the model can leverage a broader range of input pixels with positive influences,effectively enhancing its capability to restore fine texture details.By applying the proposed method to baseline models,comparative results on multiple benchmark datasets demonstrate that the proposed approach achieves superior visual perceptual quality and reconstruction performance while maintaining low computational complexity.
DEFM-YOLOv8-based Detection Algorithm for High-speed Rail Contact Network Wire State
GAO Yuli, WANG Baohui
Computer Science. 2025, 52 (11A): 241000155-9.  doi:10.11896/jsjkx.241000155
Abstract PDF(5028KB) ( 9 )   
References | Related Articles | Metrics
The high-speed rail contact network is a critical conductor in the electrified railway system,and ensuring the proper functioning of its wires is crucial for maintaining the stable operation of the railway.Traditional manual inspection methods are inefficient and prone to oversight.With the rapid development of deep learning technologies,the use of computer vision techniques for automated detection has become an urgent necessity.In response to the challenges of detecting the state of wires in high-speed rail contact networks under various complex outdoor backgrounds and diverse environmental conditions(such as night and day),this paper proposes a wire state detection algorithm based on the combination of a Detail Enhancement Fusion Module(DEFM) and YOLOv8.By incorporating spatial and channel attention mechanisms,the algorithm fuses infrared and visible light images,introducing multimodal fusion and the Shuffle Attention mechanism.Experiments conducted on a real dataset demonstrate the mo-del’s significant improvement in performance metrics such as detection accuracy and recall rate.The results show that the improved algorithm increases the recall rate by 0.94% and mAP by 2.09% compared to the original algorithms.Practical tests indicate that the DEFM-YOLOv8-based detection model performs effectively in detecting wires in the high-speed rail contact network,regardless of whether the environment is nighttime or daytime,even under complex backgrounds.
Cotton Disease Detection Based on Feature Enhancement and Group Mix Attention
WANG Hongqiang, ZHAO Hui, JIA Zhenhong
Computer Science. 2025, 52 (11A): 250200043-7.  doi:10.11896/jsjkx.250200043
Abstract PDF(3246KB) ( 9 )   
References | Related Articles | Metrics
In order to achieve rapid and accurate detection of cotton diseases in real field environments,this paper proposes a cotton disease target detection model based on feature enhancement and attention mechanism.To ensure the accuracy of the model’sdetection in real field environments,an improved feature enhancement module is used in the Neck module to weight the feature maps and reduce the interference of background or other objects on the targets in the image.After the feature enhancement mo-dule,Group Mix Attention is used to connect contextual information and enrich the feature map information.The proposed model can effectively improve the detection accuracy of models in real field environments,effectively reducing the occurrence of model false positives and false negatives using SIoU loss function.The experimental results show that the proposed model performs well on the self built real field environment cotton disease target detection dataset,effectively improving the detection accuracy of the model in real field environments.Compared with the baseline model,the mAP and Precision have increased by 2 percentage points and 4.5 percentage points.
C2P-YOLO:A Lightweight Crack Detection Algorithm for Wind Turbine Towers
DUAN Pengsong, GAO Yang, ZHANG Dalong, CAO Yangjie, ZHAO Jie
Computer Science. 2025, 52 (11A): 250100126-6.  doi:10.11896/jsjkx.250100126
Abstract PDF(2583KB) ( 10 )   
References | Related Articles | Metrics
The safety of wind turbine tower,as the support structure of the whole wind turbine,is crucial.As one of the main diseases of wind turbine tower,it is necessary to detect cracks accurately.Due to the lack of feature extraction capability,the existing crack detection algorithms have low accuracy and high model complexity,which cannot well meet the needs of end-side equipment on-site detection.For this reason,this paper proposes a YOLO-based wind tower safety detection algorithm C2P-YOLO.In the backbone network part,the algorithm utilizes the lightweight feature extraction module C2P instead of the redundant network structure,in order to extract richer feature information in the feature map.In the neck network part,the algorithm adds the lightweight up-sampling CARFE and attention mechanism modules to complement the information loss in the feature fusion process.Experimental results show that the algorithm achieves a mAP score of 84.9% on the publicly available dataset NEU-DET,which is 3%~8% higher than similar algorithms,and it can maintain a better lightweight property.
Multi-criteria Quality Assessment Method for Low-illumination Enhanced Images Based on Visual Loss
CHEN Qi, SUN Jin, WANG Jigang, HUANG Changcheng
Computer Science. 2025, 52 (11A): 241100114-7.  doi:10.11896/jsjkx.241100114
Abstract PDF(4199KB) ( 9 )   
References | Related Articles | Metrics
Low-light image enhancement improves the perception and interpretability of the images,and the assessment of the enhanced images impacts the image’s reliability and playing a guiding role in parameter selection and model adjustment of the enhancement algorithm.However,the existing image quality assessments are not completely for low-light enhanced images,which lead to discrepancies between the assessment results and subjective perceptions.In this paper,a multi-criterion based Low-light Enhanced Image Quality Assessment(MC-LEIQA) is proposedby analyzing of the visual loss factors based on human visual perception.According to the visual artifacts such as insufficient brightness gain,artifacts,false contours,and color shifts that occur during the process of enhancing low-light images,MC-LEIQA designs an assessment criterion based on the fusion of adaptive brightness gain degree using Kullback-Leibler divergence,structural recovery degree based on variance and gradient,and color recovery degree.Additionally,it introduces a correction coefficient for positive offset that incorporates automatic brightness perception to achieve accurate quality assessment of low-light enhanced images.Ablation experiments demonstrate the rationality and necessity of the selected assessment metrics in this study.Furthermore,comparative experiments with the classical image quality assessment methods on public datasets further validate that the proposed method exhibits higher assessment accuracy and effectiveness for low-light enhanced images.
Multi-modal Fusion Based Object Detection for All-day and Multi-scenario Environments
ZHANG Fan, LI Ang
Computer Science. 2025, 52 (11A): 241100137-10.  doi:10.11896/jsjkx.241100137
Abstract PDF(6405KB) ( 12 )   
References | Related Articles | Metrics
Traditional object detection methods have limitations in handling complex scenes,especially in low light conditions at night and shaded environments during the day,making it difficult to achieve ideal results.Existing multimodal image fusion techniques tend to emphasize the importance of infrared images in low light scenes,while neglecting the need for a balance between infrared and visible light fusion in complex daytime environments.Therefore,in response to the object detection demand for all-day and multi-scenario environments,this paper proposes a multi-modal fusion object detection method based on feature map classification and GAN.Unlike previous fusion methods that emphasize visual quality of images,this paper focuses on improving the object detection performance of fused images.By using a multi-scale attention mechanism to classify feature maps into saliency and detail feature maps,and optimizing the fusion effect through a generator and saliency and detail discriminators in a cross adversa-rial training network,key information of each modality is captured to meet the detection needs of different scenarios.The experimental results show that the proposed method performs well on TNO,RoadScene,and M3FD datasets,significantly improving the performance of multimodal fusion object detection.
Remote Sensing Image Restoration Based on DIL
JIANG Yujia, LI Hangqi, SUN Baodan, ZHANG Xinyi, JIANG Junhui, GONG Jianguang
Computer Science. 2025, 52 (11A): 250200079-5.  doi:10.11896/jsjkx.250200079
Abstract PDF(2270KB) ( 10 )   
References | Related Articles | Metrics
Currently,remote sensing images are widely used in environmental monitoring,disaster management and other fields.However,sensor failure or external environment results in the degradation of image quality in image acquisition influencing the application of remote sensing images.The DIL algorithm models different distortion degrees and distortion types,and uses the “backdoor” criterion in causal inference to derive the causal network for image restoration with fairly strong generalization ability.Therefore,this paper applies the DIL algorithm to remote sensing image restoration to use the collected remote sensing image data improving the restoration quality of remote sensing images.In this way,this paper uses DIL algorithm to improve the application of remote sensing images in environmental monitoring,disaster management and other fields.In this paper,the training data is normalized on the basis of the DIL algorithm to ensure that the variables of the training data are unique,so that they can better deal with the repair problems of remote sensing images.In the experiments,the DIL algorithm is used to denoise,rain and deblur the remote sensing images.The experimental results show that the DIL algorithm is better than the Noise2Noise,FFDNet,DnCNN and Restormer algorithms in remote sensing image restoration,and the image restoration quality is significantly improved.
RFI Suppression of the Yunnan 40-meter Radio Telescope Based on Deep Learning
LUO Xin, LIANG Bo
Computer Science. 2025, 52 (11A): 250300044-7.  doi:10.11896/jsjkx.250300044
Abstract PDF(5380KB) ( 11 )   
References | Related Articles | Metrics
In Radio astronomy,RFI refers to all phenomena that interfere with the weak astronomical signals received by radio telescopes and have a significant impact on the observation and study of astronomical signals.The existing suppression methods at present,such as manual labeling methods,signal processing methods,machine learning and deep learning methods.This paper proposes an improved deep learning neural network model based on DeepLabV3+,aiming to detect RFI in astronomical data.It trains the model using the real RFI observation data observed by the Yunnan Observatory and fully utilizes the ability of the convolutional neural network to extract image features,so that the model can better learn the features of RFI and thereby achieve more accurate RFI detection.It adopts the model to recognize the image.By estimating the probability that each data point belongs to RFI and using the trained model to determine whether RFI exists or not.When a certain data point is predicted as RFI by the model,we mark it as the interfering part;otherwise,we mark it as the non-interfering part,thereby achieving the marking and suppression of the part with RFI in the image.The experimental results show that the proposed method demonstrates a satisfactory level in terms of F1 score,Accuracy and MIoU.Meanwhile,the proposed model is compared with the traditional deep lear-ning model,and ablation experiments are conducted to further verify its performance advantages.
Research on Public Nuisance Website Identification Method Based on Multi-modal Data Fusion
ZHAO Chunlei, YU Jie, WANG Pengxiang, YOU Wei
Computer Science. 2025, 52 (11A): 241100171-10.  doi:10.11896/jsjkx.241100171
Abstract PDF(3613KB) ( 10 )   
References | Related Articles | Metrics
Currently,methods for identifying public nuisance websites,suffer from insufficient feature utilization and poor feature integration.Therefore,this paper proposes a multi-modal fusion model for identifying public nuisance websites,named RBI-RA.This model uses the ResNet50+Attention model to extract visual features from website screenshots,while utilizing OCR techno-logy to extract text from screenshots to enrich the website’s text features subsequently.The model employs the RoBERTa+Bi-LSTM+interactive attention mechanism model to extract features from HTML text and screenshot text separately,and integrates them through an interactive attention mechanism to enrich and expand the website text features.The model uses a self-attention mechanism to merge the website’s visual and text features,resulting in a multi-modal fusion classifier that leverages the complementary features across different modalities.Finally,to prove the effectiveness of the proposed model,experiments are conducted on a self-developed dataset.Experimental results show that the proposed model based on multi-modal data fusion effectively improves the performance of identifying public nuisance websites,with good precision,recall,and F1 scores.
RMSFF-SSD:Remote Sensing Image Object Detection Model Based on Reparameterization andMulti-scale Feature Fusion
CHEN Haiyan, MA Shuhao, ZHANG Zhenxiao
Computer Science. 2025, 52 (11A): 241000184-7.  doi:10.11896/jsjkx.241000184
Abstract PDF(3563KB) ( 11 )   
References | Related Articles | Metrics
Remote sensing image target detection has a wide range of applications in fields such as land resource survey,disaster monitoring,and military reconnaissance.In response to the difficulty of SSD(Single Shot MultiBox Detector) models in effectively extracting features of small targets during remote sensing image target detection,which is detrimental to the detection of small targets,this paper proposes a remote sensing image target detection model based on reparameterization and multi-scale feature fusion,named RMSFF-SSD(Reparameterization Multi-Scale Feature Fusion SSD).This model is an improvement based on the SSD model.Firstly,the convolutional layers in the backbone feature extraction network of SSD are replaced with convolutions that have reparameterization properties to extract features,and at the same time,the SE attention mechanism is introduced into the reparameterized convolutions to capture the dependencies between channels and suppress useless features.Secondly,the features extracted by the feature extraction network are fused through multi-level feature fusion to integrate global information and local detail information,further enhancing the target features.Finally,the six different scales of feature maps obtained after fusion are used for target detection.The experimental results of target detection on the NWPU VHR-10 dataset show that the average precision of the proposed RMSFF-SSD512 target detection model is 89.7%,which is significantly higher than the DSSD(78.7%) model,FSSD(86.7%) model,FPN(68.9%) model,Faster R-CNN(44.2%) model,and YOLOv5(83.7%) model.
Real-time Transformer Small Target Detection Model Based on Feature Extraction Enhancement and Pyramid Structure
ZHANG Wei, CAI Yufan, YE Lintao, LIU Dazhi
Computer Science. 2025, 52 (11A): 250100139-11.  doi:10.11896/jsjkx.250100139
Abstract PDF(5644KB) ( 19 )   
References | Related Articles | Metrics
To address the challenges in small target detection in outdoor environment,such as complex background,insufficient light,dense target and severe occlusion,an improved LDSD-DETR model based on real-time detection Transformer is proposed to enhance feature extraction and small target detection capability in complex background.In order to improve the efficiency of feature extraction,linear deformable convolution(LDConv) is used to improve the pooling layer and the subsampling part to extract features more effectively.Deformable attention mechanism is introduced into the attention-based feature interaction part of the scale to optimize the feature capture of the relevant regions of the target.For small target detection,a small target enhancement pyramid is designed in the cross-scale feature fusion part to enhance the sensitivity of small target.To further improve perfor-mance,the reconstructed structure combines DGCST modules to effectively capture both local and global features of the image.The experimental results show that the average detection accuracy of LDSD-DETR on Roboflow100 and its extended data set is better than other test models.Compared with the original model,all indexes are effectively improved,among which mAP50 is increased to 90%,an increase of 1.8 percentage points.In addition,the model is optimized in terms of computation amount,parameter number and weight file size,which provides a more accurate and efficient solution for real-time detection of small targets.
Unmanned Driving Scene Object Detection Method Based on Local Features and Feature Fusion
JI Tao, YANG Yifang, FENG Yachun, WU Lingfan, LI Xuliang, LI Yawei
Computer Science. 2025, 52 (11A): 250200051-7.  doi:10.11896/jsjkx.250200051
Abstract PDF(3581KB) ( 9 )   
References | Related Articles | Metrics
In the context of unmanned driving,the accuracy and robustness of object detection are of vital importance to the performance of the system.Aiming at the false detection and missed detection phenomena that occur when existing deep learning-based network models deal with small objects and occluded objects inunmanned driving scenarios,an LSDA-YOLO network model is proposed.Firstly,the LocalSimAM attention mechanism is proposed to address the issue of information loss,and it is applied to the Backbone.Meanwhile,the SHSA attention mechanism is introduced,and an information aggregation network is designed to enhance the detection ability for occluded objects.In the Neck part,by dynamically adjusting the upsampling ratio,the adaptability of the model to multi-scale features is enhanced,reducing the missed detection rate of small objects.In the Head part,the ASFF strategy is introduced to enhance the model’s multi-scale detection ability.Experimental results show that the LSDA-YOLO network model improves the mAP0.5 and mAP0.5:0.95 by 3.1 percentage points and 3.9 percentage points respectively on the KITTI dataset,outperforming the YOLOv11n baseline network model,and is suitable for high-precision real-time detection in unmanned driving scenarios.
Three-dimensional Object Detection Algorithm of Road Scene Based on Attention Mechanism
CAO Wenbo, WEI Mingyang, DUAN Xiaoyong, LIU Xueyuan
Computer Science. 2025, 52 (11A): 241100112-7.  doi:10.11896/jsjkx.241100112
Abstract PDF(3675KB) ( 11 )   
References | Related Articles | Metrics
With the development of deep learning and on-board LiDAR,driverless cars have increasingly high requirements for detection,which not only need to accurately detect obstacles on the road,but also have high requirements on detection speed.In the complex road scene,there are always obstacles and small volume of some targets,which make it difficult to accurately detect some targets.To solve this problem,this paper proposes an improved 3D target detection method of Pointpillars algorithm model to make it have higher accuracy while guaranteeing the detection speed.Firstly,a variety of data-enhancing operations are introduced to increase the diversity and magnitude of the dataset and reduce the overfitting phenomenon.Then,an attention matrix is added to the point column feature extraction,and the importance of each voxel is dynamically adjusted according to different voxel positions and semantic information,so that the model can focus on more useful features for target detection tasks.Finally,the channel attention mechanism(CA) and spatial attention mechanism(SA) modules are added to the backbone network of the model successively,which enhance the response of the model to useful information,suppresse the interference of unimportant features to the detection results,and thus improve the representation of target features.The experimental results show that the detection accuracy of the improved algorithm model is improved in each category and detection difficulty.
Target Recognition Algorithm in Urban Traffic Field of View Based on Improved YOLOv8
CHEN Junjie, ZHAO Hong, LUO Yong, DING Xiaoyun
Computer Science. 2025, 52 (11A): 241200131-8.  doi:10.11896/jsjkx.241200131
Abstract PDF(5003KB) ( 10 )   
References | Related Articles | Metrics
To reduce the issues of false detection and missed detection in target detection algorithms within urban environments,the YOLOv8 target detection model is used as the foundation,and a small-object detection layer is introduced to enable the network to better capture and recognize small-sized objects in the field of view,thereby improving its focus on target recognition.A novel remote sensing target detection model is integrated to reconstruct the C2f module,enhancing its perception of rich gradient flow information and its ability to dynamically adjust the receptive field.By applying topological optimization concepts to improve the CBAM attention mechanism,the GSAM attention mechanism is proposed and embedded at appropriate positions in the network to enhance the utilization of semantic information.To address the problem of missed detections,the performance of multiple IoU methods is compared,and the optimal EIoU is selected to accelerate the convergence speed of the algorithm and improve regression accuracy.Testing and ablation experiments conducted on the Cityscapes public dataset show that,compared to the baseline algorithm,the improved algorithm achieves increases of 2.5,5.8,and 6.1 percentage points in precision,recall,and mean average precision(mAP),respectively.These results effectively enhance the accuracy of vehicle target detection in urban traffic scenarios,providing reliable support for applications such as road video surveillance.
Optimization Study of Segmentation Algorithms for Lower Limb Bone on 3D CT Slices
SONG Lei, WANG Baohui, DU Hui
Computer Science. 2025, 52 (11A): 240900072-7.  doi:10.11896/jsjkx.240900072
Abstract PDF(3564KB) ( 12 )   
References | Related Articles | Metrics
Robot-assisted lower limb osteotomy requires precise bone models for accurate pin placement and osteotomy planning.Accurate segmentation of bone tissue in CT images is essential for creating these models.This paper proposes an enhanced U-Net convolutional neural network model,which incorporates a dynamic sliding window mechanism.This mechanism adjusts the window size dynamically during the processing of sequential CT slices,improving the model’s adaptability to different cross-sectional variations and enhancing segmentation accuracy.Validation with a CT image dataset of lower limb bones from Beijing Jishuitan Hospital shows that the improved model achieves a Dice coefficient of 84.948%.This represents a significant improvement over the U-Net model(80.353%) and the Attention U-Net model(83.580%).These results highlight the effectiveness of the proposed method in achieving more accurate bone tissue segmentation.
Detection of Brain Tumor Lesion Areas Based on Improved YOLO Model
RONG Changda, YIN Jibin
Computer Science. 2025, 52 (11A): 241000166-8.  doi:10.11896/jsjkx.241000166
Abstract PDF(4824KB) ( 7 )   
References | Related Articles | Metrics
Aiming at the problem that traditional manual detection is easily affected by subjective factors leading to misdiagnosis or omission in brain tumour reading,an improved YOLO model is proposed for intelligent detection of brain tumour foci region by combining the characteristics of brain tumour images.Aiming at the irregular shape of brain tumour lesion regions,deformable convolution is introduced to make the network adaptive to complex lesion morphology and improve the feature extraction ability of irregular lesions.Meanwhile,by embedding a global attention mechanism that combines global multi-attention,local attention and channel attention,the network focuses on the subtle features of the lesion region while reducing the negative impact of the complex background of the image on the feature extraction of the lesion region in order to obtain a higher recognition accuracy.In addition,for the actual situation that the prediction frames in the brain tumour dataset annotation are not necessarily accurate,the improved Wise-IoU is used instead of the original CIoU loss function to adapt to the problem of inaccurate manual annotation.The results of comparison experiments on the brain tumour dataset Brain Tumor Detection show that the proposed model improves the accuracy by 5.9％compared to the original model.
Optimization of 3D Reconstruction Algorithm Based on X-ray of Lower Limb Bone
WANG Baohui, DU Hui, ZHANG Yuan
Computer Science. 2025, 52 (11A): 241100152-7.  doi:10.11896/jsjkx.241100152
Abstract PDF(4116KB) ( 13 )   
References | Related Articles | Metrics
In clinical practice,bone malformations of the lower limb are common and difficult to treat in orthopedic medicine.Doctors usually rely on anteroposterior and lateral X-rays to judge the degree of malformations in the weight-bearing position,but this process is highly dependent on the professional level and experience level of doctors.Although three-dimensional CT imaging technology exists,it cannot meet the diagnostic requirements well due to the difference between patients who need to lie flat and standing weight-bearing position during CT imaging.Therefore,it is essential to create a more intuitive and accurate display of the lower limb bone malformation model.This will not only simplify the work of doctors,but also improve the accuracy of diagnoses and help them develop more effective treatment plans.This paper proposes a model based on PSSobel-X2CTGAN model,and on the basis of this model,Transformer mechanism is added to the reshape module.In addition,CycleGAN is used for data enhancement in the preparation of data sets.After validation of the data sets in the original paper,the experimental results clearly show that the structural similarity values of the model reaches 79.51% and 56.32% on the CT-PELVIC and SKI10 data sets,respectively,while the values of the original model are only 77.49% and 49.53%,indicating a significant improvement.
Retinal Vessel Segmentation Based on Multi-scale Attention
ZHU Sifan, ZHU Guosheng
Computer Science. 2025, 52 (11A): 241200112-10.  doi:10.11896/jsjkx.241200112
Abstract PDF(3664KB) ( 12 )   
References | Related Articles | Metrics
In medical image segmentation,retinal vessel segmentation is very important for the early diagnosis and treatment of ophthalmic diseases.Retinal vessel segmentation is not only helpful for the diagnosis of diseases such as diabetic retinopathy,glaucoma,and arteriosclerosis,but also has wide applications in analyzing ocular vascular morphology and hemodynamics.How-ever,existing methods cannot accurately segment small retinal blood vessels and blood vessel edges,and are still limited in terms of class imbalance,complexity of blood vessel morphology,and limited training samples.In order to improve the accuracy of blood vessel segmentation and reduce the false positive rate,this paper proposes a retinal vessel segmentation model based on multi-scale attention(MDAF-Net).The model introduces multi-scale dynamic convolution to adaptively adjust the attention to blood vessels of different scales,alleviates the problem of insufficient extraction of small blood vessels,combines channel and spatial attention mechanisms to optimize feature fusion,enhances the model’s ability to extract detailed features,and adopts a multi-scale feature fusion strategy to improve the segmentation effect under the complexity of blood vessel morphology.MDAF-Net verifies the model effect on the DRIVE and CHASE_DB1 datasets,and obtains a Dice coefficient of 0.764 and an MIoU of 78.3%(DRIVE) and a Dice coefficient of 0.820 and an MIoU of 82.5%(CHASE_DB1).The experimental results show that MDAF-Net has significant advantages in segmentation accuracy and false positive rate control,and solves the limitations of traditional methods in small blood vessel segmentation,category imbalance and false positives.
CINN:A High-speed and JPEG-resistant Medical Image Watermarking Network
ZHANG Xiaorui, XU Yanan, SUN Wei
Computer Science. 2025, 52 (11A): 241100037-7.  doi:10.11896/jsjkx.241100037
Abstract PDF(4101KB) ( 9 )   
References | Related Articles | Metrics
This paper proposes a watermark recovery algorithm for medical images against JPEG compression to address the problems of privacy protection and transmission efficiency of medical images in telemedicine.Traditional methods such as parity-check codes and Hemming codes have limitations in watermark error correction,while Reed-Solomon codes can effectively recover multi-bit error,but their recovery ability is limited when facing block processing attacks such as JPEG compression.With the development of deep learning,although INN-based watermarking technology realizes high-capacity information embedding,the computational burden is large,which affects the efficiency of information transfer.To solve these problems,this paper firstly applies Reed-Solomon code to preprocess the watermark information to improve its stability and recovery ability,and embeds the processed watermark into the DCT low-frequency coefficients of the carrier image.Secondly,in order to reduce the computation time,this paper is inspired by the structure of CSPNet,divides the features into two parts,optimizes the network structure of INN through cross-stage connection,reduces the number of model parameter,and accelerates the training process.The experimental results show that the algorithm achieves nearly 100% correct watermark recovery rate under JPEG compression with QF=50,and reduces the training time by about 40%,which significantly improves the computational efficiency and training speed of the proposed model.
Research on Heart Semi-supervised Segmentation Algorithm Based on Global-local Information Fusion LPV-Net and 3D-EDA
HU Huichen, LIU Ruixia, LIU Zhaoyang, GUO Zhenhua
Computer Science. 2025, 52 (11A): 241100077-7.  doi:10.11896/jsjkx.241100077
Abstract PDF(2820KB) ( 10 )   
References | Related Articles | Metrics
Heart disease is one of the main causes of death worldwide,which seriously threatens human life and health.As a non-invasive medical imaging technology,cardiac magnetic resonance imaging(MRI) is widely used in clinical diagnosis,helping doctors accurately and efficiently diagnose and treat heart diseases.However,cardiac MRI segmentation faces great challenges in practical applications,manual segmentation methods are time-consuming and subjective,while existing fully supervised and semi-supervised segmentation methods are not effective in dealing with complex cardiac structures and pathological changes,limited by the scarcity of data sets.This study aims to solve the challenges of cardiac MRI segmentation by proposing a 3D left atrial semi-supervised segmentation framework based on global-local information fusion to address the time-consuming and subjective pro-blems of manual segmentation.Although the current fully supervised heart segmentation methods are effective,they are limited by the scarcity of data sets.The semi-supervised methods come into being,but they are still limited by the small amount of data,especially when dealing with complex cardiac structure and pathological changes.To solve this problem,this study propose a new cardiac MRI segmentation method that combines Linformer and Performer merge V-Net(LPV-Net) and 3D Enhanced Discriminator with Attention(3D-EDA) technologies to achieve an effective fusion of global-local information.The LPV-Net module,created by LinPerBlock and improved V-Net,aims to standardize the training process of the model and achieve effective fusion of global and local information,thus improving the accuracy and robustness of segmentation.In addition,we also introduced a new discriminator 3D-EDA for the specification of unlabeled data.The most critical module in the model is CARELayer,which integrates a custom attention module to enhance the ability of the model to capture important information in the feature,and the auxiliary segmentation network improves the segmentation performance.By conducting a comprehensive experiment on the left atrial dataset,comparing the proposed method with several advanced semi-supervised methods.The experimental results show that the proposed method performs well on the baseline dataset,especially when training with limited label data.For example,when training with only 10% and 20% labeled data,the Dice coefficients of 88.50％ and 90.39％ were obtained.
Medical Image Segmentation Model Based on Frequency Texture Prior and Frequency Feature Enhancement Fusion
ZHONG Yanjie, JIAN Muwei, ZHANG Haoran, LING Yukun
Computer Science. 2025, 52 (11A): 241200125-8.  doi:10.11896/jsjkx.241200125
Abstract PDF(3708KB) ( 13 )   
References | Related Articles | Metrics
The proposed model leverages frequency-domain information extracted via Fourier transform as an optimization basis,enhancing the network’s ability to identify camouflaged lesion regions under highly similar backgrounds.By designing the Frequency Feature Enhancement Module (FFEM),the network can significantly strengthen lesion-related features across different frequencies,thereby enabling more precise capture of subtle camouflaged patterns in complex contexts.In addition,a novel strategy integrates frequency-domain prior maps into the loss function through weighted fusion,guiding the optimization process to focus on lesion features and improving the network’s sensitivity and adaptability during training.Furthermore,a Cross-Attention Fusion Module(CAFM) is designed to perform differentiated enhancement of multi-frequency features,further enhancing the network’s ability to regulate and balance frequency-specific information.The proposed method demonstrates outstanding segmentation performance across multiple medical imaging datasets(skin datasets ISIC 2016,ISIC 2017,ISIC 2018;colon polyp datasets CVC-Clinic,Kvasir,CVC-ColonDB,ETIS-LaribPolyDB;breast dataset BUSI).In quantitative evaluations,including Dice coefficient,Intersection over Union(IoU),and Accuracy(ACC),the method outperforms existing models,achieving superior accuracy and robustness.
Research and Application of Pipe Center-line Extraction Method for Fusion Reactor CoolingPipe Visualization
LUO Yuetong, DONG Ziqiu, PENG Jun, ZHAO Dongsheng
Computer Science. 2025, 52 (11A): 241000137-5.  doi:10.11896/jsjkx.241000137
Abstract PDF(2623KB) ( 10 )   
References | Related Articles | Metrics
Cooling pipes are crucial components distributed throughout the fusion reactor,whose impact on nuclear safety of the fusion reactor is significant.Therefore,the visualization of cooling pipes is of great importance for improving the safety of fusion nuclear processes.Because cooling pipes are distributed linearly,visualization based on an accurate pipe centerline is a commonly used method.However,extracting the centerline from complex cooling pipe surface models is highly tedious.To address this issue,a solution is proposed by this paper.First,the mean curvature flow algorithm is used to extract discrete points near the center-line.Then,a set of optimization methods are designed,based on the prior knowledge that the pipe segments are cylinders or rings and the connection relationship between the pipe segments,to construct accurate centerline segments from the discrete points,including the type,equation,and connection relationship of the centerline segments.The paper validates the effectiveness of the proposed method by using the cooling pipes of the International Thermonuclear Experimental Reactor(ITER),and the experimental results show that the centerline extracted from the pipes meets the requirements and can effectively support subsequent visualization tasks,proving that the proposed method is effective.
Construction and Research of Convolution Enhanced Adaptive Classification Model
CHEN Yizhuo, ZOU Wei, WANG Hongda
Computer Science. 2025, 52 (11A): 241200069-5.  doi:10.11896/jsjkx.241200069
Abstract PDF(3095KB) ( 9 )   
References | Related Articles | Metrics
Classical convolutional Neural Networks(CNNs) have been successfully widely used in the image application field.However,when images undergo rotations or scaling transformations,the relative positions and scales of features change,presenting challenges for traditional CNNs in extracting stable and invariant image features.To address this issue,this paper introduces a Con-volutional Enhanced Adaptive Classification Model(CEACM),which consists of two parts:feature extraction and classifier design.In the feature extraction stage,a feature invariant layer is used to enhance the CNN,applying rotational transformations to enhance the convolutional neural network features,allowing the model to extract stable and representative features from the input data.In the classifier part,an adaptive enhancement model based on Particle Swarm Optimization(PSO) is proposed,where the PSO algorithm is used to optimize the weights of the classifier to avoid local optima,thus improving the model’s generalization and classification performance.Finally,the model’s performance is evaluated using a series of image datasets.Experimental re-sults indicate that the proposed CEACM outperforms traditional machine learning models and a series of improved models in terms of classification effectiveness.
SAM-MR:SAM-based Mixed Region Matching Expert Adaptation Algorithm for FabricDetection
LUO Qifeng, XIAO Xing, WEN Chaofei, CHI Mingmin, PENG Bo
Computer Science. 2025, 52 (11A): 241200124-6.  doi:10.11896/jsjkx.241200124
Abstract PDF(2838KB) ( 12 )   
References | Related Articles | Metrics
Supervised anomaly detection has been widely applied to fabric quality inspection due to its high precision in industrial scenarios.However,existing unified-architecture methods often suffer from limited feature adaptation capabilities,making it difficult to distinguish diverse and highly similar fabric defects.This paper proposes a novel approach based on a Mixture of Region Experts(SAM-MR),which introduces a Mixture of Adapter Experts module to differentiate between various types of fabric defects.Additionally,an Align and Differencing module is employed to align features between template and defect images,further enhancing the localization of anomalous regions.The model is also extended to incorporate component detection,enabling semantic recognition of defect-related components on top of defect localization.Experimental results demonstrate that SAM-MR outperforms existing methods on fabric defect datasets,and qualitative,quantitative,and ablation studies validate the effectiveness of the proposed approach in multi-task prediction.
Obstacle Recognition Method for Grassland Inspection Robot Based on Improved YOLOv8
DOU Zhuolun, YU Chunzhan, ZHANG Jialin, LI Yulong
Computer Science. 2025, 52 (11A): 241100065-6.  doi:10.11896/jsjkx.241100065
Abstract PDF(3497KB) ( 9 )   
References | Related Articles | Metrics
In order to solve the problem of difficulty in balancing accuracy and real-time performance of obstacle recognition algorithms for grassland inspection robots due to complex external environments and insufficient computing power,a lightweight detection model for grassland obstacles based on YOLOv8 is proposed,which utilizes an efficient multi-scale attention module to enhance network feature extraction capabilities.At the same time,1X1 convolution is added to the neck structure of the network for dimensionality reduction mapping processing,reducing the number of parameters in the network.This paper also replaced the loss function of the original network with WIoU,reducing the impact of low-quality images on the model during training.Experiments conducted on self-built datasets have shown that the improved model has an F1 score of 93% and an average accuracy value(mAP) of 96.2%,which is 1 and 1.9 percentage points higher than the original model.The model parameter size is 1.96×106,which is 34.7% lower than the original model.Finally,the model was ported to an embedded platform and FP16 quantization was performed,resulting in a 35% increase in running frame rate.The proposed method can balance accuracy and real-time performance,and is a lightweight detection method suitable for embedded platforms,providing technical support for obstacle detection of grassland inspection robots.
Automatic Recognition of Irrelevant Individuals in Videos Based on Multi-object Tracking
MA Yixin, ZENG Junhao, YANG Xinyan, LIANG Gang
Computer Science. 2025, 52 (11A): 241100155-8.  doi:10.11896/jsjkx.241100155
Abstract PDF(3496KB) ( 10 )   
References | Related Articles | Metrics
Automatic identification of irrelevant individuals aims to detect and identify irrelevant persons in videos to solve their privacy protection issues.Existing privacy protection methods extract high-level visual features to identify individuals irrelevant to the subject.However,the extraction of high-level features significantly affects the processing efficiency of the video and makes it difficult to process massive video data.At the same time,the existing single-frame recognition method does not consider the temporal characteristics of the target,resulting in low accuracy.Therefore,this paper proposes an automatic recognition algorithm to efficiently identify irrelevant individuals,and introduces a multi-target tracking method to determine the correlation between people and videos.The method can extract five lightweight features from the time and space dimensions of the individual’s motion trajectory.In addition,in order to solve the challenges brought by occlusion and blur during video motion,an observation-based trajectory association algorithm is adopted to improve the accuracy of motion tracking.Extensive experiments conducted on various datasets demonstrate that the proposed method achieves significant improvements across multiple evaluation metrics compared to state-of-the-art approaches.Specifically,the MOTA metric shows a maximum improvement of 10.87 percentage points,the HOTA me-tric achieves a maximum increase of 10.95 percentage points,and the accuracy of irrelevant individuals recognition reaches 98.13%.
Real-time Helmet Detection Algorithm for Roadway Engineering Construction Based on UAV Visual Inspection
WEN Ming, WU Xingtang, SHANG Yuhao, ZHEN Jian, YU Fucai
Computer Science. 2025, 52 (11A): 250100047-7.  doi:10.11896/jsjkx.250100047
Abstract PDF(3776KB) ( 13 )   
References | Related Articles | Metrics
To ensure the safety of highway engineering construction personnel and reduce safety risks during the construction process,real-time detection of helmet usage has become a critical safety supervision method.Highway projects are characterized by numerous,long,and wide construction sites,often involving complex terrains such as mountain ranges and rivers.Traditional fixed-camera coverage has limitations and high costs.Drones,as flexible,low-cost,and highly visible image acquisition tools,can effectively address these challenges,especially in high-risk areas that are difficult to cover with traditional methods.This paper proposes a real-time helmet detection algorithm based on an improved eXtended Difference of Gaussians(XDOG) and YOLOv5,aiming to solve the issues of misdetection and missed detection under variable lighting conditions,scale,and shape changes in images captured by drones.In complex construction environments,the features of safety helmets are often hard to distinguish from backgrounds or other objects.The XDOG module is introduced to enhance edge information in images,thereby highlighting the structural and detailed features of helmets to be detected.The difference-of-Gaussians results are further normalized and non-linearly activated to eliminate the effects of lighting variation and noise interference in construction environments.To ensure compatibility with the YOLOv5 network,the algorithm uses a 1×1 convolution layer to adjust the number of channels in the enhanced feature maps,and a residual connection is used to fuse the enhanced feature maps with the input image,thereby improving the robustness and accuracy of the network.Experimental results show that compared to traditional YOLOv5 and YOLOx models,the XDOG-YOLOv5 significantly improves detection accuracy,with notable gains in mAP@50 and mAP@50-95,demonstrating its effectiveness in real-time helmet detection for construction personnel.
Outdoor Self-supervised Monocular Depth Estimation Method Based on Gram Matrix Attention
JIA Hongjun, ZHANG Hailong, LI Jingguo, ZHANG Huimin, HAN Chenggong, JIANG He
Computer Science. 2025, 52 (11A): 250300040-9.  doi:10.11896/jsjkx.250300040
Abstract PDF(4236KB) ( 9 )   
References | Related Articles | Metrics
In the outdoor depth estimation task,traditional U-network based models often ignore the correlation and difference between different features in the feature extraction and fusion stage,and fail to fully utilize the interaction information between features.To address this problem,this study proposes an outdoor monocular depth estimation method based on Gram matrix attention.Specifically,firstly,the correlation matrix and difference matrix between features are designed by utilizing the properties of Gram matrix decomposition,so as to enhance the information interaction and characterization ability between features.On this basis,the mask generated by the Gram matrix attention mechanism is further deeply fused with the features extracted from the convolutional layer.By combining the important features concerned by the attention mechanism with the fine details captured by the convolutional layer,the diversity and completeness of feature representation is realized.Numerous experimental results show that the performance of the network is improved with the introduction of the Gram matrix attention mechanism on the outdoor scene dataset KITTI.The proposed method in this study achieves an improvement on the δ1 metric to 0.880,while the absolute error metric decreases to 0.112.In addition,the test results on the Make3D dataset further validate the superiority of the proposed model,which is shown by the fact that the absolute relative error,the root-mean-square relative error,and the root-mean-square error reach 0.318,respectively,3.174 and 7.163 excellent levels,respectively.
Expression Detection Algorithm Based on SSD Network Model Reconstruction
CHEN Ping’an, DENG Qi
Computer Science. 2025, 52 (11A): 250200066-6.  doi:10.11896/jsjkx.250200066
Abstract PDF(2237KB) ( 12 )   
References | Related Articles | Metrics
Facial expression recognition and detection are prominent research areas in computer vision and deep learning,widely applied across various scenarios.However,traditional expression detection methods perform poorly under unconstrained conditions,while deep learning approaches face challenges such as low feature discriminability and susceptibility to posture and expression variations.To address these issues,this paper proposes an IML-SSD algorithm based on SSD network model reconstruction and center loss optimization to enhance the accuracy and robustness of facial expression detection.Firstly,this paper introduces a fast SSD-based facial expression detection algorithm optimized through network reconstruction.By restructuring the base and auxiliary layers of the SSD model,the algorithm achieves improved recognition speed,accuracy,and robustness.Subsequently,the SSD algorithm is further optimized by incorporating a center loss function,which enhances the aggregation of features within the same category while increasing separation between different categories,thereby strengthening the discriminative capability of facial expression features.Test results demonstrate that the proposed algorithm outperforms comparative methods,achieving approximately a 6.5 percentage points increase in mAP values on the FERPlus dataset.
Point Cloud Registration Network Integrating Adaptive Optimization and Multi-dimensional Focusing
YUE Qianwen, WANG Dongqiang, ZHANG Qiang
Computer Science. 2025, 52 (11A): 250100019-7.  doi:10.11896/jsjkx.250100019
Abstract PDF(2800KB) ( 9 )   
References | Related Articles | Metrics
In the field of point cloud registration,effectively capturing detailed features and enhancing registration accuracy,particularly when handling point clouds with low overlap rates,are two primary challenges.Traditional feature extraction methods have achieved some success but remain insufficient in mining geometric information,leading to limited feature discrimination.Current techniques primarily rely on position encoding and geometric embedding strategies,which enhance geometric understanding but still struggle with accuracy in high-outlier scenes.To address these issues,this paper introduces ROPNet,a novel registration network that integrates adaptive optimization and multi-dimensional focusing.ROPNet’s design includes multi-dimensional focusing,adaptive modulation kernels,and a dynamic optimization selector.These components enable the network to capture both global features and local details,accurately identify spatial positions and correspondences,and better understand the intrinsic structure of point cloud data.Additionally,ROPNet’s robust design significantly improves the identification of inliers,thereby significantly improving registration accuracy.Experimental results demonstrate ROPNet’s superior performance.On the 3DMatch dataset,ROPNet achieves a 92.4% registration recall rate and a 71.3% inlier ratio.On the KITTI dataset,it attains 99.8% registration accuracy,with relative rotation and translation errors reduces to 0.24 degrees and 6.6 cm,respectively.
Research on High-robustness Encoding and Localization Methods Based on Damaged QR Dode
KANG Bohan, GAO Wanlin, JIA Jingdun
Computer Science. 2025, 52 (11A): 241000179-7.  doi:10.11896/jsjkx.241000179
Abstract PDF(3073KB) ( 12 )   
References | Related Articles | Metrics
With the popularization of mobile devices and the development of the IOT,QR code has become widely used as a convenient and efficient means of data transmission.However,QR code is susceptible to wear and corrosion during prolonged usage.In particular,damage such as corner loss can lead to the failure of the position detection module and the format information encoding module,making it difficult for users to decode QR code with traditional software.To cope with the problem,this paper proposes improved algorithms for the center position and edge corner detection area,made successful for localization by the decoder system when parts of the QR code’s position detection area are missing.Additionally,this paper introduces a novel structure for format version information to replace the functionality of the traditional structure of QR code.Experimental results demonstrate that the proposed methods can enhance more robustness in decoding than the conventional method with the corner loss of QR code,thereby possessing high significance in practical application.
Optimization and Absolute Scale Recovery of SFM Algorithm in GCP-assisted Colmap Framework
LI Pengfei, GUAN Xiancai, ZHU Youjian, LI Yuanqiao, WANG Jun
Computer Science. 2025, 52 (11A): 250100015-6.  doi:10.11896/jsjkx.250100015
Abstract PDF(3394KB) ( 12 )   
References | Related Articles | Metrics
With the rapid development of the digital economy,the demand for 3D reconstruction technology has significantly increased.However,existing commercial 3D reconstruction systems often rely on closed standalone or cluster architectures,which limit flexibility and efficiency,while open-source frameworks face deficiencies in absolute coordinate and scale recovery.This paper introduces an SFM algorithm based on a GCP assisted Colmap framework to address these issues.The algorithm precisely converts the free network results of SFM in Colmap to absolute coordinates through constructing residual equations,applying similarity transformation,and global bundle adjustment.Experimental results show that this method achieves computational accuracy comparable to commercial systems like Agisoft and DJI Terra,while maintaining high computational efficiency in scale reco-very.This study not only enhances the absolute scale recovery capabilities of open-source 3D reconstruction systems but also lays the theoretical and practical foundations for future cloud-based applications and large-scale data processing.Future efforts will focus on realizing a fully automated cloud architecture for 3D reconstruction and exploring its application prospects in 3D monitoring with IoT devices.
Few-shot Image Generative Adaptation for Power Defect Scenes
YANG Lan, ZHAO Jinxiong, LI Zhiru, ZHANG Xun, DI Lei, CAI Yunjie, ZHANG Hehui
Computer Science. 2025, 52 (11A): 241100149-8.  doi:10.11896/jsjkx.241100149
Abstract PDF(3968KB) ( 12 )   
References | Related Articles | Metrics
In the operation and maintenance of power systems,timely and accurate detection of power defects is crucial to ensure the safety and stability of the system.However,due to the difficulty in obtaining image data of power defect scenes,deep learning models often face the problem of insufficient training samples.To solve this problem,this paper applies the diffusion model to power defect image generation and proposes a few-shot generative adaptation method based on texture modulation and EMA parameter update to expand the power defect image dataset.Specifically,this paper introduces a texture modulation module into the diffusion model,and improves the image’s detail capture ability and spatial structure alignment ability through a two-stage injection mechanism.In addition,this paper designs a cross-domain adaptive training strategy for EMA parameter update,which combines style loss and diffusion loss to smooth the model training process and improve the quality and stability of generated images.Experimental results show that this method performs well on multiple few-shot datasets of power equipment defects,and the ge-nerated images have high spatial structure consistency and detail restoration capabilities,showing its application potential in power defect detection.
Big Data & Data Science
Review of Impact of Personalized Recommendation Algorithms on User Decision-makingBehavior
XU Fuping, ZHOU Xiaohang, ZHANG Ning
Computer Science. 2025, 52 (11A): 241100086-8.  doi:10.11896/jsjkx.241100086
Abstract PDF(2140KB) ( 8 )   
References | Related Articles | Metrics
The rapid development of the Internet has generated massive data,and the phenomenon of information overload has thus become increasingly prominent.In order to help users effectively filter and capture data and make high-quality use of the huge amount of data,personalized recommendation algorithms have been proposed,and have been continuously developed in different scenarios and applications,and have a guiding role in the perception and decision-making behaviors of users.This paper focuses on four typical personalized recommendation algorithms,namely,collaborative filtering-based recommendation,content-based recommendation,association rule-based recommendation,and hybrid recommendation,analyzes their characteristics and applicability in the big data environment and different scenarios,and explores the development of personalized recommendation algorithms in the introduction of relevant theories and integration of emerging technologies from the perspective of Internet content platforms,e-commerce platforms,and social scenarios.The study also explores the influence of personalized recommendation algorithms on users’ decision-making behavior from the perspective of willingness to use and purchase decision,and then explores the functional role of personalized recommendation algorithms in users’ decision-making and related research prospects.
Comprehensive Review of Hierarchical Time Series Forecasting Methods and Applications
XIANG Yi, CONG Lili, WANG Weipeng, ZHOU Xiaohang
Computer Science. 2025, 52 (11A): 241000139-7.  doi:10.11896/jsjkx.241000139
Abstract PDF(1877KB) ( 13 )   
References | Related Articles | Metrics
Hierarchical time series involve multiple time series with hierarchical constraints,where the data of the upper-level nodes are the cumulative sum of all their child node data.The main challenge in hierarchical time series forecasting is to accurately predict each series while ensuring consistency across different levels,i.e.,the forecast must satisfy the additive constraints within the hierarchical structure.With the emergence of large-scale data,this complex and challenging problem has demonstrated greater research value and a broad range of application prospects.This study reviews the literature on hierarchical time series forecasting methods,summarizing and generalizing from the aspects of classification methods and theoretical applications,while also discussing the challenges faced by this technology and the gaps in practical applications.The analysis indicates that hierarchical time series forecasting methods can mainly be divided into two stages,which are forecasting models and revision models,gradually introducing machine learning and deep learning methods,and evolving into end-to-end methods that integrate forecasting and revision models.These methods are widely applied in the fields of business operations and government governance.In terms of future research trends,the first area of focus should be the impact of massive data on the accuracy of two-stage method forecasts,followed by in-depth research on end-to-end hierarchical time series forecasting models to avoid the issue of non-coherent parameters between stages.Additionally,research in government management and business operations can focus on modeling the differences in attention levels at various hierarchical levels caused by specific issues.
Tensor Graph Diffusion Share Nearest Neighbor Density Peaks Clustering
LIU Qiaoming, WEI Qianran, LI Zhi, WANG Jian, LI Yuanfang
Computer Science. 2025, 52 (11A): 241200068-11.  doi:10.11896/jsjkx.241200068
Abstract PDF(6034KB) ( 8 )   
References | Related Articles | Metrics
Density Peak Clustering(DPC) is based on the idea of density clustering.When dealing with high-dimensional data,the DPC algorithm has the problem of the “clustering” effect and “domino” effect in the process of similarity calculation and cluster distribution respectively,which limits the analysis efficiency of DPC in practical application.To solve these problems,a Tensor Graph Diffusion Share Nearest Neighbor Density Peaks Clustering(TGD-SNN-DPC) is proposed.Firstly,an adaptive tensor graph construction module is designed based on tensor graph theory to mine diverse local neighborhood information among data points.On this basis,an efficient tensor graph diffusion learning module is proposed and an efficient update strategy is introduced.On the premise of not increasing the computing burden of the model,this module is used to mine the global high-level topological information of the data,and the adjacency similarity information between samples with stronger robustness is obtained by using the above two modules.Finally,the adaptive shared neighborhood clustering module is designed.Based on the high-order adjacency matrix generated by tensor graph diffusion,the local density and relative distance based on shared neighbor information are introduced,and the adaptive neighborhood non-center sample allocation strategy is designed to improve the accuracy of the model matrix.Experimental results on 6 synthetic datasets and 12 real UCI datasets show that TGD-SNN-DPC algorithm outperforms the benchmark algorithm in clustering.
Spatiotemporal Active-sampling and Joint Inference of Urban Air Quality Data
LANG Aoqi, HUANG Weijie, YU Zhiyong, HUANG Fangwan
Computer Science. 2025, 52 (11A): 241000116-9.  doi:10.11896/jsjkx.241000116
Abstract PDF(3195KB) ( 9 )   
References | Related Articles | Metrics
Currently,environmental data in cities are still sampled by fixed stations as the mainstream sampling method,but the high cost of full sampling makes it difficult to be scaled up on a large scale.In this context,the method of extrapolating the remaining unsampled data through local sampling and inference algorithm has become a hot topic in current research.Existing studies usually use two different models for active sampling and missing inference,respectively,which suffer from the shortcomings of high computational cost and easy accumulation of errors.Based on this,this paper proposes a spatiotemporal active-sampling and joint inference(SAJI) integration model.The model can not only select the sampling sites with high prediction accuracy,but also determine their own active sampling time.Finally,the missing values of all sites can be inferred jointly by using Multiple Mea-surement Vector(MMV) recovery algorithm.The experimental results show that compared with the baseline algorithms,SAJI can make full use of spatiotemporal correlation to obtain valuable prefilled values for the unsampled sites and achieve the highest inference accuracy using the subsequent joint inference algorithm at low sampling rates.
Research on Recommendation Algorithm Based on Embedding User Behavior Sequence Feature Enhancement
CAO Tianruo, LI Jingyue
Computer Science. 2025, 52 (11A): 240400141-5.  doi:10.11896/jsjkx.240400141
Abstract PDF(2133KB) ( 10 )   
References | Related Articles | Metrics
With the rapid development of the internet,various functional apps have emerged,allowing people to carry out various activities online.All kinds of goods,news,advertisements,and other information continue to be generated and spread on the internet.At the same time,engineers in the field of recommendation algorithms are constantly collecting useful features to iteratively optimize the algorithm’s effectiveness.From early collection of portrait features,evolving to user behavior logs,historical beha-vior statistics,and the current research on user behavior sequence features,the field of recommendation algorithms has now esta-blished a complete feature engineering paradigm.In recent years,it has been discovered that the historical behavior sequence of users is a very important feature.However,the semantic embeddings that can be obtained solely by item IDs are very limited and cannot automatically intersect with other related information,resulting in limited benefits in algorithm effectiveness.Since last year,the introduction of language models has achieved significant results in both the academic and industrial spheres,and engineers in the field of recommendation algorithms have also made some attempts.We propose a recommendation algorithm enhancement based on language models to enhance user behavior sequence features.By leveraging the semantic analysis and logical reasoning ability of language models,pre-training representation learning of user behavior sequence features is used to achieve feature enhancement,ultimately improving the effectiveness of recommendation algorithm models.
Guided Diffusion Sequence Recommendation Methods
LI Bo, MO Xian
Computer Science. 2025, 52 (11A): 241200062-6.  doi:10.11896/jsjkx.241200062
Abstract PDF(2536KB) ( 12 )   
References | Related Articles | Metrics
With the dynamic change of user behavior preference,traditional sequential recommendation methods face the challenge of difficult to capture the change of user intention.In order to solve this problem,this study proposes guided diffusion sequence recommendation method(GDRec),which aims to achieve accurate capture of the user’s current intention by embedding the target item representation into the diffusion model.Specifically,the GDRec model includes the following key components:a sequence encoder,a cross-attention conditional denoising decoder,and a cross-divergence objective.The sequence encoder gradually generates the user preference representation to capture the dynamic relationship between the historical sequence and the current target.The cross-attention conditional denoising decoder removes the noise in the embedded representation and improves the prediction accuracy of the next target item.The cross-divergence objective,on the other hand,empowers the model with ranking capabilities,ensuring high quality representations,and embedding target item representations to guide the diffusion process.Finally,a large number of experiments on Amazon Office and Tools datasets prove that GDRec is superior to the existing advanced methods in multiple evaluation indicators,showing its superior performance in sequence recommendation tasks.In addition,ablation experiments and hyperparameter analysis further verify the effectiveness and stability of the model.
Label Data Compression Algorithms Based on BWT,MTF and ANS
LIAO Rui, TANG Jie, LIANG Tongjia, ZHENG Xinlei, WANG Binyi, QI Zhiqiang
Computer Science. 2025, 52 (11A): 241000081-6.  doi:10.11896/jsjkx.241000081
Abstract PDF(2387KB) ( 7 )   
References | Related Articles | Metrics
Using some rule sets it is possible to convert some information into specific content stored in a labeling code of a certain length.When there is a lot of information,it is more difficult to use the labeling code.Compression of tag code data can reduce the overhead of storing information and facilitate identification.In order to realize the compression of such data,this paper improves ANS based on BWT and MTF to form a lossless compression algorithm applicable to labeled data,which provides lossless data compression of labeled code to a certain extent,which is conducive to the storage and identification of labeled code information and promotes the use of labeled code.
Remaining Useful Life Prediction of Lithium-ion Batteries Based on Frequency-channelAttention Mechanism and MSCNet
LU Shiyu, WANG Hairui, ZHU Guifu, LI Yalong
Computer Science. 2025, 52 (11A): 241200041-8.  doi:10.11896/jsjkx.241200041
Abstract PDF(4848KB) ( 9 )   
References | Related Articles | Metrics
To address issues such as inaccurate feature extraction,significant data noise,and low precision in tracking capacity degradation trends in lithium-ion battery capacity estimation,a novel model combining the Frequency Channel Attention Mechanism(FCA) and Multi-Scale Inter-Series Correlations Net(MSCNet) is proposed.The model is designed in three stages.Firsty,raw sensor data are preprocessed to remove noise.Secondly,prominent periodic patterns are extracted through frequency domain analysis using the frequency-enhanced attention mechanism.Finally,the multi-scale outputs are aggregated by MSCNet,reducing model parameters while improving effectiveness of feature extraction.Experiments based on publicly available CALCE and NASA datasets demonstrate that the proposed model reduces relative error(RE) in battery life prediction by 10%~20% compared to existing algorithms,enabling more accurate tracking of battery degradation trends.
Deep Neural Network-based Resource Allocation for Large-scale Operation Simulation
YE Shuai, LI Hao, SHI Peiteng, HUANG Yulin
Computer Science. 2025, 52 (11A): 241000036-5.  doi:10.11896/jsjkx.241000036
Abstract PDF(2672KB) ( 10 )   
References | Related Articles | Metrics
With the development of artificial intelligence,operation experiments tend to be intelligent.Large-scale operation simulation is an important support for conducting intelligent operation experiments and an effective means to solve problems such as multiple variables and complex combinations in operation experiments.It has the characteristics of large sample size and high speed requirements.The high-speed operation of massive simulation samples depends on the efficient scheduling of high-perfor-mance hardware clusters,which faces the problems of large differences in computing resource requirements and difficult manual allocation.How to accurately predict and dynamically allocate the resources required for each sample is the key to improving the efficiency of large-scale simulation.This paper proposes a deep neural network(DNN)-based resource prediction model for large-scale operation simulation.The method firstly constructs a deep neural network in-loop simulation resource management architecture.Secondly,it constructs a deep neural network prediction model by extracting features and learning from combat simulation sample files.During the operation of large-scale simulation,it achieves accurate prediction and dynamic allocation of massive ope-ration simulation job resources by online predicting the computing resources required for each sample.Test results show that in a typical operation experiment simulation scenario with thousands of samples,theproposed prediction model reduces the completion time by 20.8% on 10 high-performance server nodes compared to traditional configuration methods.
Study of Temporal Uncertainty in Infix/Postfix Trace Alignment Conformance
GAO Lingting, YE Jianhong, JIANG Wenhui, HUANG Yifan
Computer Science. 2025, 52 (11A): 241200200-8.  doi:10.11896/jsjkx.241200200
Abstract PDF(2741KB) ( 10 )   
References | Related Articles | Metrics
Alignment is a type of conformance checking technique that involves checking the modeled process behavior against the process behavior recorded in the event data.Due to hardware failures,software errors,and other factors,temporal data recordings show diversity,including different accuracies and errors,leading to temporal uncertainty in the recorded data.This paperconsi-ders the infix/postfix traces containing temporal uncertainty,and proposes a trace fragment alignment method based on temporal uncertainty,which addresses the traditional trace fragment alignment method that cannot effectively deal with uncertainty,and solves the problems of insufficient alignment accuracy and low computational efficiency of the traditional alignment due to temporal uncertainty.Specifically,firstly,uncertain traces are processed and behavior net are generated.Secondly,markings of the process model are computed and auxiliary nets are constructed.Finally,synchronous product nets are constructed to compute the trace fragment alignment with time uncertainty.The proposed method broadens the application scope of the alignment technique,enabling the alignment to adapt to and handle data containing temporal deviations,and enhancing the stability and robustness of the alignment algorithm in the face of imperfect data.Experimental results show that the proposed method improves the alignment accuracy and effectively reduces the computational complexity compared to the traditional method when dealing with uncertainty.
Comparative Study of Missing Value Imputation Methods from Perspective of Interpretability
LI Yi, WANG Tongxin, PANG Bozhong
Computer Science. 2025, 52 (11A): 241100156-8.  doi:10.11896/jsjkx.241100156
Abstract PDF(3124KB) ( 9 )   
References | Related Articles | Metrics
With the widespread application of deep learning,high-quality tabular data is crucial for model performance.However,missing values can severely disrupt the underlying data structure and distribution.Although numerous imputation methods exist,current research predominantly focuses on imputation accuracy,lacking a systematic evaluation of how imputation outcomes affect the interpretability of downstream models.This paper proposes a framework for evaluating missing value imputation methods from the perspective of model interpretability.Firstly,it explores the advantages of deep generative models in learning complex data distributions to generate high-quality imputed values.Next,it constructs various missing data scenarios and employs Shapley values as a core metric to quantitatively compare the impact of different imputation methods on model feature importance explanations.Experimental results demonstrate that:1)Deep generative models can effectively learn the sample distribution and excel at preserving data structure and informational integrity.2)There is no direct correlation between imputation accuracy and the stability of model explanations;the choice of imputation method significantly alters the final Shapley values.3)As the proportion of missing data increases,the differential impact of various imputation methods on model interpretability becomes more pronounced.This study reveals the latent impact of missing value imputation on model interpretability and provides empirical evidence and a new evaluation perspective for selecting appropriate imputation strategies in interpretability-critical scenarios.
Research on Demand Forecasting for Aviation Spare Parts Based on Machine Learning
WANG Rui, WANG Zhikai, ZHONG Yiming, SUN Hui, YANG Kaixin
Computer Science. 2025, 52 (11A): 241100116-9.  doi:10.11896/jsjkx.241100116
Abstract PDF(5586KB) ( 15 )   
References | Related Articles | Metrics
To scientifically and accurately forecast the demand for aviation spare parts in airline inventories and develop rational material plans,this paper proposes a machine learning-based method.The approach considers factors such as spare part prices,importance,maintenance interval times,and installation quantities.Firstly,principal component analysis(PCA) and K-means clustering are applied to reduce dimensionality,enabling the visualization and classification of spare parts with different demand patterns.Then,it establishes hybrid kernel extreme learning machine(HKELM) and random forest(RF) models for multivariate regression predictions on the categorized data.To address the challenge of parameter selection in the prediction process,the sparrow search algorithm(SSA) is employed to iteratively optimize the optimal parameters of the two models.Finally,the method is validated with a case study using real operational data from a specific airline fleet.Its performance is compared with backpropagation(BP) neural networks,support vector machines(SVM),and least squares support vector machines(LSSVM).The results indicate that the proposed forecasting method achieves favorable outcomes and provides valuable guidance for airline material planning efforts.
Research on Hybrid Methods for Predicting Students’ Online Learning Performance Based on Generative Model
DUAN Chao, WANG Yiqing, WANG Jie, ZHANG Mingyan
Computer Science. 2025, 52 (11A): 250200029-9.  doi:10.11896/jsjkx.250200029
Abstract PDF(2969KB) ( 6 )   
References | Related Articles | Metrics
Learning performance prediction can help teachers to intervene in time by using the learning behavior data of students on the online learning platform to identify at-risk students,but it faces the problem of data imbalance,which makes it particularly difficult to accurately identify at-risk students.Addressing the issues that the mainstream deep generative model VAE,cannot guarantee the rationality of generated samples in current solution strategies,and GAN tends to introduce new errors when processing time-series data,with either over-training or under-training of the generator or discriminator leading to a decline in the quality of generated data,this paper proposes a new prediction method for student learning performance based on generative mo-dels.Firstly,the VAE based on bidirectional long short-term memory(BiLSTM) is utilized to initialize the GAN,enabling it to start training from a more stable point while also better understanding the correlation and periodic characteristics between subsequent data points in the student behavior sequence.Secondly,a multi-head attention mechanism is introduced in the discriminator part to enhance its ability to distinguish between real data and generated data,and then continue to game with the generator.Finally,the deep generative model and the classical resampling strategy SMOTE are integrated based on the idea of Blending ensemble learning,which effectively combines the advantages of data and algorithm to improve the overall generation ability of the model.A large number of experimental results on real student data sets show that the model can generate high-quality data to improve the recognition ability of the prediction model for at-risk students,and is superior to the baseline method in multiple evaluation indicators.
Personalized Multi-attribute Airline Itinerary Recommendation System by Graph ConvolutionalNeural Network
PENG Mingtian, WANG Weishuai, TIAN Feng, LI Jiangtao, LU Yan, MA Shuyan, ZHU Honglin, LIU Chi
Computer Science. 2025, 52 (11A): 250200088-6.  doi:10.11896/jsjkx.250200088
Abstract PDF(3499KB) ( 8 )   
References | Related Articles | Metrics
The rapid expansion of the aviation market has made flight choices increasingly complex,making it difficult for passengers to choose the best option from a massive amount of information.Existing flight recommendation systems mostly use static methods of sorting by price,time or punctuality,which make it difficult to take into account the personalized needs of users and the complexity of multi-connection flight combinations.In response to this situation,this paper proposes an air itinerary recommendation system based on graph convolutional neural networks,which uses graph structure data processing to improve recommendation accuracy and personalized effects.The system builds a graph structure model of flight data,refines the key attributes of flights,and converts users’ historical ticket purchasing behavior into interactive information between graph nodes.Through layer-by-layer feature aggregation through graph convolutional neural network,the high-order relationship between users and flight attributes is captured.Experiments show that the proposed model effectively combines user preferences and flight static attributes,significantly improves the performance and accuracy of the recommendation system,and provides users with better itinerary suggestions
Simplest Complete Cooperative Combination and Concept Reduction
MA Wensheng, HOU Xilin
Computer Science. 2025, 52 (11A): 250100053-5.  doi:10.11896/jsjkx.250100053
Abstract PDF(1708KB) ( 10 )   
References | Related Articles | Metrics
Based on the utilization relation between users and granulars of big data for a given task,a definition for the cooperative unit is established,and its aggregation is termed the cooperative combination.Criteria are outlined to define a complete cooperative combination based on its inclusion of all elements within the utilization relation.Additionally,a cooperative combination is designated as the simplest complete cooperative combination if it is complete and none of its proper subsets share this property.Finally,the concept reduction algorithm within formal concept analysis is applied to identify the simplest complete cooperative combination for a given task.
Fairness-enhancing Decision Tree Algorithm
JIANG Wenhui, YE Jianhong, GAO Lingting, HUANG Yifan
Computer Science. 2025, 52 (11A): 241200119-9.  doi:10.11896/jsjkx.241200119
Abstract PDF(2724KB) ( 13 )   
References | Related Articles | Metrics
In the field of machine learning,the problem of intrinsic biases in models has received increasing attention,and these biases often originate from imbalances in the training data or flaws in the algorithm design,which lead to unfair treatment of certain groups in the prediction results.To address this problem,this paper proposes a fairness-enhanced decision tree algorithm,which effectively reduces the imbalance in the data by introducing a fairness preprocessing method,and changes the traditional decision tree splitting criterion by integrating classification accuracy and fairness in the splitting criterion of the decision tree.The proposed method aims to achieve the fair distribution of prediction results among different groups,reduce the bias in model decision-making,and ensure that all individuals are treated fairly.Experimental results show that the proposed method demonstrates good performance under multiple fairness metrics,significantly reduces the prediction bias among different groups,and exhibits stronger fairness bias-correction performance than the existing traditional algorithms.
Analysis of Opinion Dynamics Based on Sensitivity to Opinion Disparity and Trust in Opinion Leaders
ZHANG Weijing, GAO Yanping
Computer Science. 2025, 52 (11A): 250100007-9.  doi:10.11896/jsjkx.250100007
Abstract PDF(5013KB) ( 10 )   
References | Related Articles | Metrics
In social networks,individual attributes exert a pivotal influence on the evolution of group opinions.To delve into this phenomenon,this paper extends the traditional Hegselmann-Krause(HK) model by incorporating two parameters:individual sensitivity to opinion divergence and the degree of trust in opinion leaders,proposes a new model for opinion dynamics.Individual sensitivity to opinion disparity refers to the degree to which individuals are sensitive to the opinion differences of others when updating their own views.This sensitivity is quantified by a sensitivity coefficient,with a higher coefficient indicating a greater propensity for individuals to communicate and interact with others whose views are close to their own.Such a mechanism may precipitate opinion polarization,as individuals are more inclined to interact with those sharing similar perspectives,thereby reinforcing their existing opinions.The trust individuals place in opinion leaders describe the degree to which individuals rely on opinion leaders when forming their opinions.In the model,each individual may accept the influence of opinion leaders’ opinions with different levels of trust.The paper first conducts a brief theoretical analysis of the model and then explores the impact of these two attributes on opinion evolution through simulation experiments in scale-free networks.The results show that the higher the sensitivity of individuals to opinion differences,the greater the divergence of opinion values and the longer the convergence time.The higher the trust individuals place in opinion leaders,the faster the group opinions will converge towards the opinions of the opi-nion leaders.Subsequently,the paper increases the number of opinion leaders and constructs an improved HK model with two opinion leaders.Through simulation experiments,the paper analyzes the impact of the proportion of individuals receiving opinions from opinion leaders and the trust individuals place in opinion leaders on opinion evolution.The experimental results indicate that the higher the trust individuals place in opinion leaders,the more easily the group opinions will align with the opinions of the opinion leaders,and the faster the convergence speed of group opinions.Meanwhile,the higher the proportion of individuals recei-ving opinions from opinion leaders,the more easily the evolution process of group opinions will be dominated by the opinions of the opinion leaders,and the final stable state of group opinions will be closer to the opinions of the opinion leaders.
Multiple Attention Mechanism News Recommendation Approach with Hypergraph Learning
MENG Xiangfu, WANG Wanchun, ZHANG Yumeng, FAN Wenyi
Computer Science. 2025, 52 (11A): 250200067-7.  doi:10.11896/jsjkx.250200067
Abstract PDF(2668KB) ( 10 )   
References | Related Articles | Metrics
In personalized news recommendation,graph structures are often utilized to establish interaction relationships between users and news,however conventional graph structures mostly overlook the high-order association information among clicked news items.Furthermore,existing methods typically rely on a single vector to learn user interest representations and candidate news representations,leading to inadequate modeling.To address these issues,a multiple attention mechanism news recommendation model approach with hypergraph learning is proposed.Firstly,a candidate news hypergraph is constructed,leveraging a hypergraph attention network to capture high-order correlations between candidate news and their semantically similar news,thereby enriching the semantics of candidate news.Secondly,a news-topic hypergraph is built to model user interests,employing a neural network architecture with multiple attention mechanisms to explore deep,fine-grained user interest features.Lastly,an activation unit is introduced to further extract user interests from candidate news,enhancing recommendation accuracy.The experiments on the MIND-small and MIND-large datasets confirm the effectiveness of the proposed approach.
Granular Perception Machine Based on GRAM Matrix
WU Shaohua, CHEN Yuming
Computer Science. 2025, 52 (11A): 241200110-7.  doi:10.11896/jsjkx.241200110
Abstract PDF(2185KB) ( 7 )   
References | Related Articles | Metrics
The perceptron is a simple linear classifier and is also the cornerstone of SVM and deep neural networks.However,most complex problems are often nonlinear,and the perceptron performs poorly in handling such issues.Therefore,this paper introduces granular computing theory,whereby training samples are granulated into feature granules and feature granular vectors based on reference samples.The paper defines a granular GRAM matrix and proposes a granular perceptron model based on the GRAM matrix.This model optimizes the dual form of the perceptron to construct a new granular perceptron model.To better handle nonlinear classification problems,a kernel function is introduced to construct a kernel GRAM matrix based on granular vectors,and the loss function and learning method of the GRAM granular perceptron are provided.Finally,experiments analyze the model’s convergence,nonlinear processing capability,the number of reference samples,and the classification performance of models.The result shows the effectiveness and correctness of the model of GRAM granular perceptron.
Network & Communication
Service Migration Path Selection Method Based on Interest and Mobility Perception in EdgeComputing Environment
DAI Mengxuan, XIA Yunni, MA Yong, MA Yuyin, DONG Yumin, LIU Hui, CHEN Peng, SUN Xiaoning, LONG Tingyan
Computer Science. 2025, 52 (11A): 250200002-8.  doi:10.11896/jsjkx.250200002
Abstract PDF(4107KB) ( 10 )   
References | Related Articles | Metrics
Mobile Edge Computing(MEC),as an innovative technology,deploys computing resources at the network edge to provide users with low-latency computing and storage services.In this research field,user mobility has consistently been a focal point,with existing work primarily focusing on analyzing and utilizing the movement trajectories between users and edge servers.However,such approaches often overlook users’ points of interest(POI) data and lack effective handling of migration failures,resulting in low service hit rates and high migration costs.Recent research has discovered that beyond mobility information,users’ points of interest data can also be effectively integrated and utilized.Addressing this finding,this paper proposes an Interest and Mobility-aware Service Path Migration(IMSPM) method.This approach fuses trajectory prediction models with user interest prediction models to achieve optimized target server selection and reliable,low-cost service migration path planning.Experimental results demonstrate that compared to traditional methods that rely solely on mobility information,IMSPM exhibits significant advantages in multiple performance metrics,including service hit rate and service migration frequency
Station Deployment Optimization Algorithm for High-precision AOA Positioning
DING Lei, REN Lu, HOU Xuan, ZHANG Dongpo, ZHU Li’na
Computer Science. 2025, 52 (11A): 241100046-9.  doi:10.11896/jsjkx.241100046
Abstract PDF(3192KB) ( 10 )   
References | Related Articles | Metrics
Aiming at the problems of low iteration efficiency and easy local convergence of the optimized deployment scheme for location of direction-finding station,this paper summarizes the impact of site layout on system positioning performance.Subsequently,based on optimization theory,the paper describes and establishes an AOA positioning optimization layout model,consi-dering constraint factors such as communication network connectivity and system effectiveness in real-world positioning scenarios.By utilizing geometric precision factor values and penalty functions as objective functions for the optimization problem,an improved particle swarm optimization algorithm is employed to solve it.Finally,through theoretical analysis and simulations,this algorithm’s effectiveness is demonstrated.
Traffic Scheduling Mechanism for Time-sensitive Networks Based on Satisfiability Modulo Theories
XU Jing, LIU Chunlong, HUO Jiahao, HUANGFU Wei
Computer Science. 2025, 52 (11A): 250300028-6.  doi:10.11896/jsjkx.250300028
Abstract PDF(3266KB) ( 10 )   
References | Related Articles | Metrics
Globally,the fusion of industrialization,informatization,and intelligence is significantly impacting various industries,especially in fields requiring stringent latency,such as in-vehicle systems and avionics.TSN has emerged as a key technology for achieving deterministic low-latency communications.However,TSN’s current network-level traffic scheduling mechanisms struggle to fully meet the complex priority demands of upper-layer services.To address this issue,this paper proposes an SMT-based TSN scheduling mechanism called SMT-TAS.By incorporating an SMT solver into the existing TAS model,and designing a traffic scheduling algorithm based on priority satisfaction rate,SMT-TAS enables real-time generation of optimal scheduling schemes based on dynamic business scenarios.Experimental results demonstrate that compared to traditional TAS methods,SMT-TAS improves the average priority satisfaction rate by approximately 20%,significantly enhances system schedulability,and reduces end-to-end latency by around 10%,demonstrates outstanding performance in terms of solving efficiency.Furthermore,it exhibits higher stability and reliability in large-scale tasks,effectively meeting various TSN scheduling constraints,providing strong support for the further development and application of TSN.
Multi-dimensional Performance Evaluation Approach Based on Tor Over QUIC
QI Jianshe, YANG Xiaohan, ZHOU Dacheng
Computer Science. 2025, 52 (11A): 241200080-6.  doi:10.11896/jsjkx.241200080
Abstract PDF(2880KB) ( 8 )   
References | Related Articles | Metrics
Tor network,as one of the most popular anonymity networks,uses TCP protocol as the transport layer protocol,and this choice leads to problems such as head-of-line blocking,unfair bandwidth allocation,and inefficient congestion control,which seriously affects the performance and scalability of the Tor network,and there are researches on Tor Over QUIC mode using QUIC protocol to solve these problems.However,the single performance evaluation index in Tor Over QUIC mode focuses only on delay and security evaluation,which is difficult to reflect the comprehensive impact of protocol upgrade on the core characteristics of anonymity network,resulting in unclear direction of protocol optimization and lack of data support for deployment decisions.This lack of evaluation dimension not only restricts the full play of the advantages of QUIC protocol,but also affects the user’s willingness to adopt the protocol due to the performance shortcomings,which easily affects the promotion and use of Tor Over QUIC.Therefore,this paper proposes a multi-dimensional performance evaluation method based on Tor Over QUIC,which comprehensively evaluates the performance of Tor Over QUIC mode from multiple dimensions such as latency,anonymity,security,robustness,and usability,in order to guide the deployment and use of Tor Over QUIC.Comparative experiments on Tor network and Tor Over QUIC network show that the evaluation method is effective and practical.
Adaptive Gradient Sparsification Approach to Training Deep Neural Networks
HUANG Xinli, GAO Guoju
Computer Science. 2025, 52 (11A): 250100106-6.  doi:10.11896/jsjkx.250100106
Abstract PDF(2911KB) ( 11 )   
References | Related Articles | Metrics
Top-k Sparsification Method with error compensation is one of the state-of-the-art technologies in the training of distributed deep neural networks(DNNs).This technique aims to reduce the amount of communication by dynamically transmitting only parts of the gradients in each iteration,with the amount of transmitted gradients depending on the value of k.Although a smaller k can speed up training time,it may degrade the test accuracy,even with error compensation,known as the speed-accuracy dilemma.Based on the observation that the increase speed of the training accuracy and test accuracy have a dynamic correlation over time,this paper presents AdaTopK－an adaptive Top-k compressor with convergence guarantees.AdaTopK can dynamically adjust the value of k to accelerate the training speed while keeping or enhancing the test accuracy.Extensive experiments in the static and dynamic network scenarios show that AdaTopK can reduce 29% training time over the baseline without compression,while reducing 15% training time over DC2.
Bluetooth-PDR Multi-sensor Fusion Indoor Positioning Method Based on UKF
MAO Dongfang, JIANG Guoping
Computer Science. 2025, 52 (11A): 250100083-4.  doi:10.11896/jsjkx.250100083
Abstract PDF(1895KB) ( 9 )   
References | Related Articles | Metrics
To improve the accuracy and stability of indoor positioning,this paper proposes an indoor positioning method based on UKF and by integrating bluetooth fingerprint database and PDR.Firstly,by collecting signal strength data from different locations,a bluetooth fingerprint database is constructed.Secondly,the use of multiple sensors such as built-in accelerometers and gyroscopes in mobile phones for pedestrian dead reckoning is employed.Then,the UKF approach is used for fusion to overcome the disadvantage of cumulative errors in pedestrian dead reckoning,thus achieving high-precision indoor positioning.It has the characteristics of low cost,high sensitivity,good stability,and simple positioning methods.Finally,the experimental simulation results demonstrate the effectiveness of this method.
Redundancy Compression Strategy in Cooperative Perception Services Based on Value ofInformation
WANG Ruijia, SHEN Zhen, LI Junjie, DING Lei
Computer Science. 2025, 52 (11A): 241100009-6.  doi:10.11896/jsjkx.241100009
Abstract PDF(3081KB) ( 9 )   
References | Related Articles | Metrics
Connected andAutonomous Vehicles(CAVs) leverage Vehicle-to-Everything(V2X) communication and 6G sensor data to enable Cooperative Perception Services(CPS).In road environments,multiple CAVs may simultaneously perceive and share information about the same object.This results in the exchange of significant amounts of irrelevant and redundant information within the V2X network,leading to additional communication overhead.To address this issue,a redundancy compression strategy based on the Value of Information(VoI) is proposed.Firstly,the value of perception information is quantified through mathematical methods.Then,when a CAV sends an upload request to the base station,the VoI is aggregated at the base station.Subsequently,CPS satisfaction is formulated as a maximization problem under the control of the base station,which is solved using a Simulated Annealing(SA) algorithm.This strategy enables the base station to optimally control the information uploaded by CAVs,maximizing the utility of cooperative perception and minimizing redundancy in the V2X network.Simulation results show that compared to existing strategies,the proposed approach effectively reduces target redundancy,achieving an average reduction in transmission delay by 22.3% and improving CPS quality by 21.6%.
MEC Network Task Offloading and Migration Strategy Based on Optimization Model
YU Ping, YAN Hui, BAO Jie, GENG Xiaozhong
Computer Science. 2025, 52 (11A): 241200215-6.  doi:10.11896/jsjkx.241200215
Abstract PDF(2609KB) ( 10 )   
References | Related Articles | Metrics
The research on task offloading and migration strategies in optimization model-driven Mobile Edge Computing(MEC) networks is conducted against the backdrop of the surge in Internet of Things(IoT) devices and the widespread adoption of 5G technology.MEC significantly reduces data transmission latency and cloud-side pressure by migrating computing resources to the network edge.This study proposes a series of task offloading and migration strategies and validates their effectiveness through performance evaluations.Experimental results demonstrate that the proposed strategies optimize key performance indicators in typical application scenarios: latency is reduced by approximately 25%,energy consumption is decreased by 30% and task throughput is increased by 20%.Specific optimizations include dynamic resource scheduling for load balancing and improved offloading efficiency;a Quality of Service(QoS)-guaranteed migration mechanism to ensure service stability; and cross-layer optimization design to enhance multi-task collaboration capabilities.Additionally,machine learning-based prediction techniques are employed to dynamically adapt to network fluctuations,thereby improving system flexibility.The research conclusions indicate that the optimization model offers significant advantages in ensuring efficient resource allocation and task real-time performance,thereby enhancing the service quality and user experience of MEC networks.The strategies can be widely applied in heterogeneous networks and dynamic environments,demonstrating potential for further expansion.
UAV Logistics Network Planning Method Considering Demand and Range
WEN Haolin, LIANG Xin, CHEN Tong, LI Yuqi
Computer Science. 2025, 52 (11A): 250200042-5.  doi:10.11896/jsjkx.250200042
Abstract PDF(2804KB) ( 9 )   
References | Related Articles | Metrics
To address the challenge of limited drone range in executing long-distance transportation tasks,a segmented relay-based drone logistics network planning method is proposed,utilizing a genetic algorithm.This method comprehensively considers logistics demand and drone range limitations,aiming to minimize construction and operational costs.By designing network connectivity judgment loops,evaluation index judgment loops,and iterative objective judgment loops,it achieves intelligent optimization of the drone logistics network structure.The algorithm incorporates four operators that effectively control the number and location of relay nodes,enabling the planning solution to progressively approach the global optimum.Through the construction of an agent-based simulation model,it is demonstrated that the drone logistics network generated by this method can economically and efficiently meet logistics demand,particularly suitable for long-distance urban transportation or logistics network planning in remote mountainous areas.This method provides an innovative solution for the green and efficient operation of drone logistics networks.
Computer Software & Architecture
Review of Development and Application of Software Defect Prediction Techniques in IndustrialInternet of Things Environment
DENG Tao, DENG Ye
Computer Science. 2025, 52 (11A): 250200052-11.  doi:10.11896/jsjkx.250200052
Abstract PDF(2254KB) ( 13 )   
References | Related Articles | Metrics
In the context of IIoT,the generation of vast amounts of software code data necessitates effective analysis through advanced SDP techniques.These techniques not only enable the rapid identification of anomalies but also facilitate comprehensive investigations into potential issues,as even minor deviations can lead to significant code failures.This paper systematically reviews over 61 relevant articles published between 2018 and 2025,highlighting the primary challenges and recent advancements in SDP within IIoT.Various perspectives on SDP technologies are explored,including statistical methods,machine learning approaches,and model-oriented techniques.Future research should prioritize the dynamics of defect patterns in complex heterogeneous environments,address the challenges of data scarcity and high labeling costs,and balance the trade-off between real-time processing and resource constraints.Additionally,the interpretability of models and user cognitive understanding must be enhanced to improve system comprehensibility and operational robustness.A comprehensive analysis of existing datasets related to IIoT is also presented,laying a solid foundation for further research in this critical area.
Semantic Variations Based Defect Generation and Prediction Model Testing
GUO Liwei, WU Yonghao, LIU Yong
Computer Science. 2025, 52 (11A): 241200059-7.  doi:10.11896/jsjkx.241200059
Abstract PDF(2384KB) ( 8 )   
References | Related Articles | Metrics
In recent years,machine learning techniques have made significant advancements in defect prediction within software development,enabling the automatic detection of errors in large-scale codebases.These advancements are expected to enhance the reliability,security,and overall quality of software.Defect prediction models can autonomously identify whether code contains errors.However,existing models,while having certain advantages,also exhibit limitations.They often fail to accurately identify vulnerabilities or incorrectly label defective code segments as problem-free.Currently,there is a lack of systematic empirical studies on the quality of defect detection models.The existing method,DPTester,assesses the effectiveness of defect models by generating defective code through modifications to if conditions in the code.However,the defect code produced by this method is overly simplistic,and the evaluation scenarios do not cover a wide range of models,including the latest large language models.To address this gap,this paper proposes an improved method called DefectGen.This new approach introduces multiple strategies to generate defect code that more closely reflects real-world issues.Furthermore,the evaluation of defect models includes large language mo-dels.Experimental results indicate that DefectGen significantly enhances the ability to generate complex defect code compared to previous methods,producing 1.2 times more defective code from a single correct code instance.When testing the CodeT5+,CodeBERT,and GPT-4o models,the proportions of incorrect defect predictions were found to be 62%,78%,and 30%.Additionally,DefectGen demonstrates higher efficiency in both test input generation and defect detection phases,with generation and detection times of 0.003 seconds and 0.02 seconds per test input.These results suggest that DefectGen not only effectively exposes the limitations of existing models but also provides new opportunities for improving defect prediction models and enhancing software quality assurance processes.
Advances in Automatic Software Defect Location Techniques
FANG Jinqiu, YUN Guorong, ZHAO Haiyong, XIE Haomeng
Computer Science. 2025, 52 (11A): 250200024-14.  doi:10.11896/jsjkx.250200024
Abstract PDF(3598KB) ( 9 )   
References | Related Articles | Metrics
Software defect localization has become an important research topic in the field of software debugging,and automated software defect localization aims to improve the degree of automation of defect location,help developers locate possible defects in large-scale software more efficiently,so as to achieve the purpose of optimal allocation of test resources.With the defect location technology of automatic software as the core,the relevant research results are systematically sorted out.Firstly,according to whether it is necessary to run test cases,the positioning technology is divided into static positioning technology and dynamic positioning technology,the representative algorithms of each type of technology are summarized,and the influence of deep learning on related algorithms is discussed.Secondly,according to the difference in the number of defects contained in the software system,the related algorithms are compared from the perspectives of single defect and multiple defects.Finally,the commonly used evaluation datasets and evaluation indicators of related algorithms are given,some challenges in this field are pointed out,and some directions worthy of further research are prospected.
Multi-agent Collaborative Code Generation Technology Driven by Large Language Models
XIA Peng, ZHANG Yijun, QI Ji
Computer Science. 2025, 52 (11A): 241200033-9.  doi:10.11896/jsjkx.241200033
Abstract PDF(2567KB) ( 13 )   
References | Related Articles | Metrics
In code generation tasks,pretrained large language models and agents have become key technologies for improving the quality and efficiency of code generation.However,when facing complex programming problems,intelligent agents based on large language models still struggle to provide effective solutions.This paper proposes a framework of multi-agent collaborative code generation to solve complex programming problems through the collaboration among agents,which includes four stages:problem analysis,task planning,code generation,and code debugging.The different base model strategies for agents based on open-source LLMs are proposed and the impact on system performance is tested.Additionally,an iterative programming paradigm incorporating reflection and debugging loops is introduced to optimize code generation based on feedback from each stage.Experimental results demonstrate that the multi-agent collaborative approach achieves significant performance improvements compared to traditional direct code generation methods across multiple datasets.Particularly,the hybrid model strategy achieves optimal perfor-mance on all tested datasets.Performance on test datasets is further improved with the adoption of reflection and debugging loops.
Speculative Control Flow Vectorization Method for SIMD
HAN Lin, WU Ruofeng, LIU Haohao, NIE Kai, LI Haoran, CHEN Mengyao
Computer Science. 2025, 52 (11A): 241100012-7.  doi:10.11896/jsjkx.241100012
Abstract PDF(1943KB) ( 13 )   
References | Related Articles | Metrics
SIMD automatic vectorization is an important means to give full play to the computing power of processors and improve the performance of applications,but the existence of control flow brings great challenges to automatic vectorization.The traditional control flow quantization method relies on IF transformation technology,but this technology also brings the problem of low efficiency of code execution.Therefore,in order to alleviate this problem,a speculative vectorization method of control flow for SIMD is proposed.The method detects the predicate-related region in vector code,uses the cost model to guide the implementation of the speculative transformation for branch consistency in the region,and eliminates the useless predicate execution at runtime,thus eliminating the problem of low code efficiency caused by redundant computation.The work of this method is based on the current mainstream GCC10.3 compiler.The experiment selected the industry-recognized SPEC CPU 2006 test set topic and the TSVC test set of testing vectorization ability.The results showed that the performance of SPEC2006 test set 481 topic was improved by 10% after using this method.The acceleration ratio of typical TSVC_2 test cases can reach more than 20%.Experimental results on standard test sets show that this method can effectively improve the execution efficiency of GCC compiler’scontrol flow quantization code
Information Security
Survey on Fuzz Testing Techniques for Network Protocols
HAN Luchao, ZHANG Wei
Computer Science. 2025, 52 (11A): 241100173-9.  doi:10.11896/jsjkx.241100173
Abstract PDF(2181KB) ( 10 )   
References | Related Articles | Metrics
Fuzz testing,as one of the automatic bug detection techniques,has found a large number of bugs in recent years by continuously inputting random or semi-random variant data to the target under test,leading to anomalies or crashes of the target under test.This paper focuses on fuzz testing for network protocol software implementation,and systematically analyses and summarises the research results related to network protocol fuzz testing in recent years.First,the basic process of network protocol fuzz testing is taken as the traction,and the working principle of fuzz testing technology in the testing phases of protocol message pre-processing,test case generation,fuzz test execution,and test anomaly monitoring and other testing phases are elaborated,as well as the progress of its representative research work.Then,through the application of mainstream protocol fuzz testing tools in the large-scale integrated testing of multiple network protocols,the evaluation and validation of mainstream network protocol fuzz testing tools are realised.Finally,the research results in recent years are analysed and summarised.Then,the evaluation and validation of the current mainstream network protocol fuzz testing tools are achieved through the application of mainstream protocol fuzz testing tools in the large-scale integrated testing of multiple network protocols;finally,the future development direction and challenges of network protocol fuzz testing technology are summarised and outlooked.
Research on Malware Classification Algorithm Based on Instruction Flow Graph
XING Yuyang, WANG Baohui
Computer Science. 2025, 52 (11A): 240800062-6.  doi:10.11896/jsjkx.240800062
Abstract PDF(2559KB) ( 7 )   
References | Related Articles | Metrics
In recent years,malicious codes have become increasingly rampant,with both the quantity and types showing a rapid growth trend.Therefore,machine learning methods have been widely introduced to improve the efficiency of malicious code identification and classification.This paper focuses on the multi-classification task of malicious codes,adopts static analysis methods,and combines technologies such as disassembly,graph construction,as well as graph theories to extract features from the original files of malicious code samples.Based on the traditional CFG features and bytecode features,the IFG feature is proposed.The IFG feature,CFG feature,and bytecode feature are respectively used to train machine learning models for a horizontal comparison experiment.From the training effect:Compared with the CFG feature,using the IFG feature,the model’saccuracy rate increases by about 5%;compared with the bytecode feature,using the IFG feature,the model’s accuracy rate increases by 0.3%,and the mo-del’s training time is shortened by more than 60%.
Research on Conflict-type Group Decision-making Method Based on Dynamic Triangular FuzzyNumbers and Improved TOPSIS Method
WANG Keke, AI Wei, YIN Yanyan, QIAN Qian
Computer Science. 2025, 52 (11A): 241000119-8.  doi:10.11896/jsjkx.241000119
Abstract PDF(2815KB) ( 9 )   
References | Related Articles | Metrics
Currently,methods utilizing triangular fuzzy numbers and TOPSIS for group decision-making often solely consider the expert evaluation information collectively provided by the expert group,neglecting the fact that different experts may have va-rying preferences for the same matter,as well as differing personal weights among experts.Therefore,based on dynamic triangular fuzzy numbers and the refined TOPSIS method,this study extends the collective evaluation information provided by the expert group to individual evaluations from different experts.It proposes methods for judging and resolving conflicts between individual expert evaluations and group decision information.Furthermore,it employs practical cases to validate the scientific validity and effectiveness of the proposed methods.In this study,several experts are invited to evaluate various candidate schemes.The study calculates the Euclidean distance,grey correlation degree,and relative closeness of each expert to the positive and negative ideal schemes for different candidate schemes.Subsequently,by incorporating the individual weights of each expert,it computes the group relative closeness of each scheme.Additionally,it determines the threshold and conflict measure values for conflict detection.If a decision conflict arises,the respective experts revise their evaluation information and implement disciplinary measures by reducing the individual weights of the concerned experts.The study then recalculates the relative closeness of each expert to different candidate schemes and the final group relative closeness.Decisions are made based on the final group relative closeness,leading to the selection of the optimal scheme.
Lightweight Memory Safety Runtime Detection Method Combined with Static Analysis
MAO Ruiqi, CHEN Zhe
Computer Science. 2025, 52 (11A): 241100060-8.  doi:10.11896/jsjkx.241100060
Abstract PDF(2247KB) ( 13 )   
References | Related Articles | Metrics
Memory safety issues,such as buffer overflow,have long troubled C language developers.Runtime detection is a reliable solution to C language memory safety problems,but it introduces significant runtime overhead.Existing methods to reduce runtime overhead for memory safety detection may be incompatible with existing code,depend on manual annotations,introduce false negatives and positives,or fail to ensure timing consistency between illegal memory access and error reporting.This paper proposes a lightweight runtime detection method for stack memory regions,which combines static analysis to replace certain runtime metadata lookups with compile-time metadata checks,and replaces most high-overhead detection function calls with inline Boolean condition checks.The method also uses on-demand interprocedural alias analysis to extend detection to interprocedural and whole-program analysis.A prototype tool,LISA(Lightweight Inline Safety Assertion),was implemented with static analysis and detection code instrumentation based on the C language abstract syntax tree.Experiments show that LISA reduces runtime detection overhead by an average of 36%,with only about 0.5% additional space overhead.Furthermore,LISA addresses compatibility with existing code,enhances runtime detection effectiveness,and ensures real-time memory safety,overcoming limitations of previous methods.
Signcryption Scheme Based on SM9 Domestic Cryptographic Algorithm
XIE Zhenjie, LUO Youqiang, ZHAO Fangfang, REN Shuai
Computer Science. 2025, 52 (11A): 241200049-8.  doi:10.11896/jsjkx.241200049
Abstract PDF(1802KB) ( 9 )   
References | Related Articles | Metrics
Signcryption is a cryptographic technique that combines digital signature and encryption,reducing computational and communication overhead compared to executing them separately.The SM9 domestic cryptographic algorithm,developed indepen-dently in China as an identity-based cryptographic algorithm,is not only secure and efficient,but also effectively lowers the costs associated with public key infrastructure construction and certificate management.Addressing the inefficiencies in computational performance and signcryptext length in existing schemes,a new signcryption scheme based on the SM9 algorithm is proposed.By innovatively designing the signcryption secret key tuple,and combining the key and signature information into a single element,the scheme significantly reduced computational complexity and compressed the signcryptext length.Under the random oracle model,the scheme is proven to have IND-CCA and EUF-CMIA security based on the Gap-q-BDHI and q-SDH hard problems,respectively.Theoretical analysis and experimental tests confirme that the proposed scheme improved signcryption and decryption verification efficiency by 67% and 62%,respectively,compared to the existing similar scheme,while reducing the signcryptext length by 25%.
Design and Application of Decision Tree Algorithms for Privacy-preserving
LI Jincheng, LI Yingna, FU Guoqing
Computer Science. 2025, 52 (11A): 241200115-9.  doi:10.11896/jsjkx.241200115
Abstract PDF(2548KB) ( 11 )   
References | Related Articles | Metrics
In the digital era,data has emerged as a critical asset.Data sharing not only fuels advancements in the artificial intelligence sector,but also poses the threat of privacy violations.Fully Homomorphic Encryption(FHE) technology offers a secure solution for executing various machine learning algorithms on encrypted data,bypassing the risks associated with data exposure.Nonetheless,operations on encrypted data demand a significant computational overhead,prompting the need for algorithms to be redesigned with FHE optimization in mind.This paper introduces a novel privacy-preserving decision tree scheme based on the CKKS fully homomorphic encryption algorithm.It utilizes a low-degree approximate step function and a lightweight interaction protocol to supplant complex nonlinear operations,enabling the training and inference of decision trees directly on encrypted data.Extensive experiments on four benchmark UCI datasets reveal that the proposed scheme achieves an average AUC of 0.92 and an average F1-Score of 0.77,outperforming both the PrivaTree and SecDT schemes while also exhibiting greater stability.
P-DAG:An Efficient and Secure Blockchain System Based on Parallel Chain
JIANG Lingyun, LIU Guanhao, YANG Jinglin, XU Jia
Computer Science. 2025, 52 (11A): 241000174-6.  doi:10.11896/jsjkx.241000174
Abstract PDF(2550KB) ( 10 )   
References | Related Articles | Metrics
Although blockchain systems based on tree structures leverage the concurrency of tree graphs to significantly improve throughput,they still face numerous security challenges that need to be addressed.In response to the issue where such tree-based blockchain systems are vulnerable to liveness attacks,leading to a failure in ledger state convergence,a scalable and highly secure blockchain system called P-DAG(Parallel-Directed Acyclic Graph) is proposed,featuring high throughput and low confirmation latency.This system adopts a ledger structure with multiple parallel chains and decouples block creation from the process of adding blocks to the chain,thus distributing the computational power of malicious nodes to enhance the overall security of the system.By utilizing the randomness and uniform distribution properties of hash values,a hash-based random weight assignment mechanism is designed to reduce the convergence time of each chain and the block confirmation latency.Theoretical analysis and simulation experiments show that P-DAG achieves throughput similar to Conflux,but reduces ledger convergence time by approximately 50% and block confirmation latency by about 30% compared to Conflux.
Adversarial Attack on Vertical Graph Federated Learning
BAI Yang, CHEN Jinyin, ZHENG Haibin, ZHENG Yayu
Computer Science. 2025, 52 (11A): 241200220-10.  doi:10.11896/jsjkx.241200220
Abstract PDF(3491KB) ( 10 )   
References | Related Articles | Metrics
Graph vertical federated learning(GVFL) is a distributed machine learning approach that integrates graph data with vertical federated learning,widely applied in fields such as financial services,healthcare,and social networks.This method not only preserves privacy but also leverages data diversity to significantly enhance model performance.However,studies indicate that GVFL is vulnerable to adversarial attacks.Existing adversarial attack methods targeting graph neural networks(GNN),such as Gradient Maximization Attack and Simplified Gradient Attack,still face challenges when applied in the GVFL framework.These challenges include low attack success rates,poor stealth,and inapplicability under defense conditions.To address these issues,this paper proposes a novel adversarial attack method for GVFL,termed Node and Feature Adversarial Attack(NFAttack).NFAttack designs node and feature attack strategies to conduct efficient attacks from multiple dimensions.The node attack strategy evaluates node importance using degree centrality metrics and disrupts high-centrality nodes by connecting a certain number of fake nodes to form adversarial edges.Meanwhile,the feature attack strategy introduces hybrid noise-composed of random noise and gradient noise-into node features,thereby affecting classification results.Experiments conducted on six datasets and three GNN models demonstrate that NFAttack achieves an average attack success rate of 80%,approximately 30% higher than other me-thods.Furthermore,NFAttack maintains strong attack performance even under various federated learning defense mechanisms.
Attacking Image Manipulation Localization Model by Eliminating Semantic Features
JIANG Weihao, LIU Bo
Computer Science. 2025, 52 (11A): 241100104-7.  doi:10.11896/jsjkx.241100104
Abstract PDF(3231KB) ( 9 )   
References | Related Articles | Metrics
At present,the public is increasingly concerned about the image tampering technology because it will cause ethical and security issues.Deep neural networks can be used to locateimage tampering areas.However,with the development of deep neural networks,adversarial attacks against them have also developed,and these attack methods have also promoted the research on the robustness of the model.Existing adversarial attack methods mainly focus on tampering trace features,but different Image Manipulation Localization models focus on different tampering trace features,resulting in insufficient migration ability of adversarial attacks.Since convolutional neural networks or Transformer networks can also extract semantic features,and Image Manipulation Localization models often use these models as baseline models,which would inevitably extract some semantic features when extracting tampering features.In order to improve the generalization ability of adversarial samples,a attack method is proposed,focusing on eliminating the semantic features of tampered images,training a semantic segmentation network as the attack target,and proposing a loss function for attacking intermediate semantic features,making it difficult for the model to identify the semantic information of the tampered part of the image.This attack method has better transfer ability,can hide perturbations and genera-te more aggressive adversarial samples.It has been proven in multiple experiments that it can attack most existing models and outperform other adversarial attack methods,and provides novel insights for the image manipulation localization.
Image Tampering Detection and Self-recovery Algorithm Based on Quantum-secure and FragileWatermarking
CHEN Hongxiang, CHEN Guo, ZHANG Hui, WU Meiqi, DING Qiqi, LUO He, WANG Hao, GU Linming, LUO Huiheng, WANG Jinghan
Computer Science. 2025, 52 (11A): 250900081-8.  doi:10.11896/jsjkx.250900081
Abstract PDF(2820KB) ( 8 )   
References | Related Articles | Metrics
In the context of remote monitoring and unattended inspection in power systems,images have become a crucial medium for recording and transmitting defect information.However,when transmitted over public channels,these images are highly susceptible to malicious tampering and forgery,posing serious threats to system security and fault response efficiency.To address this issue,this paper proposes an image tamper detection and self-recovery method tailored for power communication scenarios.The proposed approach constructs a perfect hashing-based authentication model using a quantum random number generator,ensuring high randomness and unpredictability to enhance resistance against tampering.It further integrates a block-level image matching strategy and a SPIHT encoding algorithm to generate both authentication and recovery data.These data are embedded into the original communication image as fragile watermarks,enabling precise localization and recovery of tampered regions.The embedding key is securely shared via a quantum key distribution protocol,effectively preventing key leakage or manipulation during transmission.Experimental results on standard image datasets demonstrate that the proposed method outperforms existing schemes in terms of tamper detection accuracy,recovery performance,security,embedding capacity,and visual quality,making it well-suited for integrity protection and trusted authentication of power image transmission.
Information Level Inference Method for Data Aggregation Based on Granular Association
LI Jinhui, CAO Lifeng, WANG Xiaoqin, BAI Jinlong, CHEN Yang
Computer Science. 2025, 52 (11A): 241200047-8.  doi:10.11896/jsjkx.241200047
Abstract PDF(3030KB) ( 11 )   
References | Related Articles | Metrics
To address the issue of sensitive information leakage through the existence of big data aggregation,this study analyzes the correlation between data deeply and proposes an information level inference method for data aggregation based on granular association.The method mines highly associated data objects based on the dependencies of data attributes,and then deduces the possibility of inferring highly sensitive information from data aggregation when users access the multi-information system based on the fuzzy set possibility measurement of the sensitivity level of the associated attributes of the data objects.This approach aids in establishing access policies for users,controlling the control the analysis of associated data,and reducing the risk of information leakage.
Research on Generating Adversarial Network Traffic Based on Generative Adversarial Network
YANG Lin, LIN Honggang
Computer Science. 2025, 52 (11A): 241200189-9.  doi:10.11896/jsjkx.241200189
Abstract PDF(3699KB) ( 12 )   
References | Related Articles | Metrics
Adversarial network traffic plays a crucial role in fields such as device privacy protection and network security.How-ever,current adversarial network traffic generation methods lack constraints on quality,resulting in generated traffic that deviates from the original traffic characteristics,thereby losing its adversarial capability in practical applications.Therefore,this paper proposes a GAN-based adversarial network traffic generation method,which improves the generator design.The convolutional neural network is employed to extract abstract representations of original traffic features,and perturbations are generated through basic iterative algorithms to ensure that the perturbations maintain the characteristics of the original traffic.The generator loss function is optimized to achieve minimal differences between the generated traffic and the original traffic.Additionally,a perturber module is introduced,utilizing a grid search algorithm to assign weights to perturbations and optimize parameter combinations,ensuring the diversity of the generated traffic.To comprehensively consider the impact of feature space distance differences and relative change rates on generation quality,a relative difference disturbance metric is proposed to more accurately evaluate the differences between adversarial network traffic and the original traffic.Experimental results show that,within an effective perturbation range,compared to other methods,the adversarial network traffic generated by this method maintains a high deception rate for target classification models while producing smaller L∞ disturbance and relative difference disturbance values,and exhibiting higher similarity to the original traffic,effectively improving the generation quality of adversarial network traffic.
Classification of Encrypted Application Traffic Enhanced by Multi-level GraphRepresentation
WANG Zhihong, LIU Shengran, CHI Zegui, YANG Ying
Computer Science. 2025, 52 (11A): 241200126-7.  doi:10.11896/jsjkx.241200126
Abstract PDF(2824KB) ( 7 )   
References | Related Articles | Metrics
With the increasing demand for privacy protection and data security,more and more applications and services use traffic encryption technology.While protecting users’ privacy,it also provides convenience for illegal users,seriously threatening network security defense and supervision.Due to the limitation of single and multiple records in representation,this paper proposes a model of encrypted application traffic enhanced by the multi-level graph representation.The proposed method constructs packet graphs based on multi-type interactive information in a single record,such as payload length,direction,sequence,and cluster information.Furtherly,multi-record graphs are constructed based on flow sequence association to break through the limitation of a single record.Finally,the graph neural network is introduced to realize the representation of traffic based on packet graphs and record graphs.Experiments are carried out on the ISCX VPN-nonVPN 2016 dataset,which is a widely used open-source dataset in the encrypted traffic classification area.Experimental results show the overall classification accuracy of the proposed method on VPN and non-VPN reach 98.1% and 89.2% respectively,and the F1 score is significantly improved compared with Text-based-CNN,k-GNN etc.
Traffic Generation Methods Based on Temporal Generative Adversarial Networks
LIN Honggang, LI Yuhan
Computer Science. 2025, 52 (11A): 250200021-8.  doi:10.11896/jsjkx.250200021
Abstract PDF(3634KB) ( 13 )   
References | Related Articles | Metrics
With the advancement of network traffic analysis techniques,the demand for high-quality synthetic traffic data conti-nues to grow.However,most existing traffic generation methods primarily focus on packet-level features while neglecting temporal dependencies,leading to suboptimal performance.To address this limitation,an enhanced traffic generation approach based on TimeGAN is proposed.This method employs a Gated Recurrent Unit(GRU) to jointly capture both packet-level and temporal features.Additionally,a multi-head local-global attention mechanism is integrated to improve feature fusion and achieve balanced modeling of traffic characteristics.A periodicity-aware discriminator and dynamic decoding strategy are further introduced to ge-nerate variable-length traffic sequences while preserving periodic patterns.The generated traffic is evaluated in terms of usability and similarity,with experimental results demonstrating superior performance across multiple metrics compared to existing me-thods.This approach effectively enhances the quality of synthetic traffic and provides a more realistic emulation of actual network behavior.
MOBSF_rule Based Android Malware Detection Method
CHEN Weiguo, ZHANG Gaofeng, JIA Sheng, XU Benzhu, ZHENG Liping
Computer Science. 2025, 52 (11A): 250200120-11.  doi:10.11896/jsjkx.250200120
Abstract PDF(3609KB) ( 9 )   
References | Related Articles | Metrics
In the field of Android application security research,a highly effective method within static analysis involves utilizing reverse engineering tools to decompile the application and subsequently extracting the Function Call Graph(FCG) from the decompiled code files,which serves as a primary feature for malware identification.Notably,FCG subgraphs based on sensitive APIs have been widely validated.However,the majority of existing research efforts in this area rely on older sets of sensitive APIs and have not continued to update with the iteration of system APIs.Through experimentation,it has been discovered that when using traditional sensitive API sets to extract feature nodes from the application’s FCG,many cases fail to obtain the desired feature nodes.For instance,with the iterative updates of the Android system,there are significant API adjustments and replacements,or dynamic implicit invocation of system APIs can be achieved using reflection mechanism(Reflect)-related technologies.In response to this,based on the latest comprehensive research framework for Android applications,this paper proposes an Android malware detection method that extracts FCG subgraphs using MOBSF_rule.The method first generates the FCG from the decompiled code files of the application.Then,it utilizes the MOBSF_rule set to extract feature nodes,generating five-node,six-node,and seven-node graphs containing these feature nodes,and statistically analyzing the occurrence frequency of different configuration subgraphs.Finally,the frequency matrix is input into the machine learning method for training and inference.Compared to existing sensitive API sets,the proposed method has the following advantages.1)The MOBSF_rule filtering rule set demonstrates outstanding performance in extracting feature nodes,effectively extracting key API features including reflection mechanisms,component interactions,signature verification,network communication,and client/server(C/S) architecture communication.Compared to traditional sensitive API sets,the effective rate of feature extraction in the latest malware datasets has increased by 69.765%.2)The MOBSF_rule set shows excellent capability in extracting feature nodes across different time tags,exhibiting strong stability.It can not only adapt to the continuous updates of the Android system but also maintain a highly consistent feature extraction capability across different versions.Between 2012 and 2022,compared to traditional sensitive API sets,the overall variance in feature extraction effectiveness over multiple years decreases by 98.747%.3)The method employs the Stacking ensemble learning approach,and compared to the random forest ensemble learning method and the multilayer perceptron method,the accuracy rate has increased by 4.32%.
Lightweight Aeronautical Broadband Communications System Security Authentication Protocol
CHEN Hongwei, YUE Meng
Computer Science. 2025, 52 (11A): 241200183-7.  doi:10.11896/jsjkx.241200183
Abstract PDF(4719KB) ( 12 )   
References | Related Articles | Metrics
Currently,aeronautical communications system is facing the challenge of spectrum saturation and lack relevant security standards and defense measures,which makes aeronautical communications data vulnerable to illegal interception and tampering.To enhance the security of aeronautical communications system and optimize the efficiency of ground-to-air communications,a lightweight aeronautical broadband communications system security authentication protocol based on symmetric encryption has been proposed.This protocol aims to ensure that aircraft and ground gateway can mutually authenticate identities and securely transmit data.By relying on pre-shared root keys and message authentication code verification operations,the protocol has an average message overhead of 59 bytes,effectively circumventing the certificate management issues faced by existing aeronautical communications system with public key authentication schemes while maintaining its lightweight nature,and providing a robust network security barrier for the identity authentication process.Through formal modeling verification with the Scyther tool and system packet analysis testing,the protocol’s high efficiency and reliability in bandwidth-limited aeronautical communications environments have been confirmed.
Optimization of Blockchain Dynamic Sharding and Cross-shard Transaction Protocol Based on Greedy Strategy
AI Yuan, LI Jiahao, ZHAO Yitao, HU Kai
Computer Science. 2025, 52 (11A): 250100133-8.  doi:10.11896/jsjkx.250100133
Abstract PDF(3425KB) ( 12 )   
References | Related Articles | Metrics
An optimized dynamic sharding algorithm,along with a cross-sharding transaction protocol,is proposed to tackle the challenges associated with the sharding mechanism in blockchain technology.These challenges include load imbalance,the complexity of verifying cross-sharding transactions,and ensuring the atomicity of such transactions.To address these issues,this paper develops a dynamic slicing algorithm utilizing a greedy strategy,which adjusts the slicing dynamically through weight calculations to achieve load balancing based on blockchain transaction data.Additionally,to resolve the atomicity and latency issues of cross-slicing transactions,it introduces an innovative cross-slicing transaction protocol and a slice migration strategy.This approach ensures the atomicity of cross-slicing transactions by incorporating a transaction locking and rollback mechanism.Experimental results indicate that this method significantly reduces transaction latency.
Quantum Circuit Optimization for Simplified AES Cryptographic Algorithm
DING Lang, LUO Qingbin, LYU Yi, ZHENG Yuanmeng, LIAO Haoyu
Computer Science. 2025, 52 (11A): 250100075-7.  doi:10.11896/jsjkx.250100075
Abstract PDF(2596KB) ( 12 )   
References | Related Articles | Metrics
AES is currently the most widely used internationally standardized block cipher algorithm.The National Institute of Standards and Technology(NIST) uses the quantum security strength of AES as a reference for evaluating the security of post-quantum cryptography.Therefore,implementing the quantum circuit of the AES algorithm and analyzing its quantum security has become a research hotspot in cryptography.However,since implementing the AES algorithm’s quantum circuit requires hundreds of qubits and tens of thousands of quantum gates,making the implementation and optimization of Simplified AES quantum circuits has become an important research direction.This study successfully implements the S-box quantum circuit using the DORCIS tool based on the S-box lookup table,decomposes the CCCNOT gate into four Toffoli gates by borrowing one qubit,and avoids swap gates in shift operations by permuting variables.Additionally,an 8-qubit S-box quantum circuit for key expansion is designed and implemented using the Boolean expression of the S-box.The optimized S-AES quantum circuit is verified in the Qiskit Aer simulator,requiring only 32 qubits,51 NOT gates,220 CNOT gates,and 120 Toffoli gates.Compared to existing me-thods,this approach significantly reduces quantum resource consumption,enhancing the efficiency of implementing the Simplified AES quantum circuit.
Demand Response Scheme for Low Voltage Users Based on Light Weight Blockchains
CHANG Ningyuan, HUANG Ting, ZHANG Huang
Computer Science. 2025, 52 (11A): 250200125-8.  doi:10.11896/jsjkx.250200125
Abstract PDF(2065KB) ( 9 )   
References | Related Articles | Metrics
The rapid development of smart grids has enhanced the communication between power companies and electricity users,making demand response for low-voltage users a highly promising smart grid service.In recent years,efforts to strengthen smart grid functionalities using blockchain technology have increasingly drawn attention.However,issues such as energy consumption and user privacy concerns introduced by blockchain have become unavoidable.This paper proposes a low-voltage user demand response solution based on a lightweight blockchain,which features low energy consumption and ensures user data privacy.To address the fairness issues in blockchain and data security concerns of the main chain arising from the low energy consumption and the lightweight blockchain structure,the paper further introduces illegal manipulation monitoring,distributed hash tables,and regional authentication algorithms to support the solution’s normal operation.The application of cloud computing in smart grids has also developed rapidly,as it can maximize resource integration and address the distributed computing challenges posed by the massive data in smart grids.
Research on Security Performance Evaluation and Verification of Video Surveillance NetworkBased on Variable Fuzzy Theory
WANG Keke, BIAN Yue, YIN Yanyan
Computer Science. 2025, 52 (11A): 241100095-6.  doi:10.11896/jsjkx.241100095
Abstract PDF(1979KB) ( 11 )   
References | Related Articles | Metrics
At present,research on the security of video surveillance network systems still remains at the level of security indicator systems and indicator weights.In order to further carry out research on the security of video surveillance network systems,a variable fuzzy evaluation model was established to evaluate the security performance of video surveillance network systems.Regret theory was introduced to verify the evaluation results,and actual cases were used to verify the effectiveness of the evaluation model.This study uses variable fuzzy theory to calculate the comprehensive relative membership degrees of fuzzy evaluation mo-del,neuron excitation function model,TOPSIS ideal point model,and fuzzy optimization evaluation model.Then,the evaluation results corresponding to these four models are obtained,and their arithmetic mean is calculated to obtain the final evaluation level.The verification method based on regret theory used in this study can select the optimal evaluation scheme,and thus verify the evaluation results using the variable fuzzy evaluation model.The evaluation model proposed in this study helps to take targeted measures for weak links in video surveillance networks,improve the overall security of video surveillance network systems,and provide guarantees for the security construction and use of video surveillance systems.
Fraud User Detection Based on Heterogeneous Information Network with Knowledge Graph Eembedding
LYU Shuqi, ZHANG Yunfeng
Computer Science. 2025, 52 (11A): 250400085-7.  doi:10.11896/jsjkx.250400085
Abstract PDF(2374KB) ( 12 )   
References | Related Articles | Metrics
In the scenario of credit payment service,the detection of fraudulent users has always been a research hotspot.In the deep learning method,heterogeneous information networks are usually used to model different types of node objects and their interaction relations.For example,nodes are used to represent users and merchants in the payment service scenario,and edges are used to represent the interaction relations between nodes,so as to make full use of the structural information of the graph.However,when capturing node feature information,many models that have been proposed often only focus on the end nodes of the meta path and ignore the information of the middle nodes of the meta path,which will lead to the problem of information loss.Therefore,this paper proposes a heterogeneous graph fraud user detection model based on knowledge graph embedding.Firstly,it introduces the knowledge graph embedding method as the meta path internal aggregation encoder.Different from the method of only focusing on the upper nodes of the meta path,the meta path internal aggregation coder will pay attention to the intermediate nodes of the meta path when obtaining the node information,so as to gather the node information on the whole meta path,which can effectively solve the problem of information loss.Moreover,it designs a multi-layer fusion attention mechanism to simulate users’ preferences for attributes and meta paths from the node and path levels,and analyzes the importance of features from the perspective of fusion at the global level.The experimental results on different types of data sets show that the proposed model achieves relatively good results compared with many existing fraud detection methods.
Cooperative Defense Method for Network Space Object of Power Monitoring System
LI Xiaogeng, HAN Xiao, XIAO Haiyi
Computer Science. 2025, 52 (11A): 241200158-7.  doi:10.11896/jsjkx.241200158
Abstract PDF(2303KB) ( 9 )   
References | Related Articles | Metrics
The power monitoring system is the core facility for ensuring stable power supply.Currently,most of the network security defense measures for power monitoring systems are based on fixed strategies,which often lack specificity for the current system environment and security events.Moreover,implementing such defense strategies can also have a significant impact on the normal operation of system business.To solve the above problem,a cooperative defense method for network space object is proposed.Firstly,in order to block network threats,IP tracing technology is used to redraw the attack path,taking into account the number of hops between nodes and the attacked object in the attack path,as well as the network traffic at nodes.A fitness function is constructed,and the optimal blocking position is determined based on the idea of improved genetic algorithm.Secondly,based on the types of objects,it formulates defense strategies for classifying objects,introduces a defense action correlation calculation model,and determines specific defense actions.Simulation experiments show that the proposed network space objectco-operative defense method has significant advantages in selecting and executing defense actions,as well as defense effectiveness,which can minimize the impact of defense actions on normal system operations.
Privacy-preserving Cross-certificate System Authentication and Access Control Model for Material Supply Chain
YANG Ke, GUO Qinglei, SHEN Yiming, BAI Neng, SONG Wenting, WANG Weiyu
Computer Science. 2025, 52 (11A): 250100131-10.  doi:10.11896/jsjkx.250100131
Abstract PDF(3022KB) ( 11 )   
References | Related Articles | Metrics
Electronic data management has emerged as a pivotal tool for government and corporate procurement in the process of modernizing material supply chains,which plays a decisive role in boosting market competitiveness and ensuring fairness in transactions.However,electronic procurement faces several challenges,especially during the bidding process.These include the varying credibility of diverse electronic data sources,the difficulty of protecting bidders’ privacy,and potential risks of collusion between bidders and procurers.Blockchain technology,with its decentralized structure,distributed ledger,and high transparency,aligns well with the distributed nature of participants in electronic procurement.To address these challenges,this paper proposes a privacy-preserving multi-entity cross-certificate authentication and access control model for material supply chains.The system utilizes smart contracts to maintain a trust list and employs efficient Merkle tree signatures for certificate issuance with minimal storage overhead.This ensures the trustworthiness of both certificate authorities and bidders,while optimizing the cross-domain certificate verification process,further enhancing the transparency and consistency of authentication.Additionally,the system employs attribute-based encryption to encrypt and store sensitive data from bidders.Fine-grained access control is implemented to allow only authorized procurers to access the necessary information,effectively preventing collusion risks before the bid opening and ensures fairness and transparency in the bidding process.Rigorous security analysis and simulation tests demonstrate that the proposed solution not only supports multi-entity cross-certificate system authentication,but also safeguards bidder privacy,providing flexible and robust access control.
Intrusion Detection Method for Power Monitoring System Based on Multi-source Network Data
JIANG Yakun, LIN Xu
Computer Science. 2025, 52 (11A): 241200157-7.  doi:10.11896/jsjkx.241200157
Abstract PDF(3121KB) ( 11 )   
References | Related Articles | Metrics
With the continuous advancement of informatization,networking,and intelligence in the power system,the power monitoring system is facing increasingly severe network security threats.It is particularly important to conduct a comprehensive and in-depth analysis of the multi-source data covered by the power monitoring system network,taking into account various factors such as network asset security risks,user behavior,and business characteristics.Based on this,a multi-source data cleaning me-thod and intrusion detection method for power monitoring systems are proposed.The improved maximum correlation and minimum redundancy algorithm is used to select the multi-source security data features of the power monitoring system network,retain appropriate security data features,and use a network intrusion detection model to detect and classify multi-source network security data,effectively solving the problem of complex feature attributes of multi-source data in power monitoring systems leading to decreased accuracy of model classification in the later stage.Simulation experiments show that the proposed multi-source data feature selection method and intrusion detection algorithm have significantly improved the detection rate and classification accuracy of attacks on power monitoring systems.
Interdiscipline & Application
Design of Centralized Management and Control Platform for Multi-system HeterogeneousAerospace TT&C and Data Transmission Resources
MIAO Lin, SHEN Hongjing, WANG Li, CAO Yiwen, LU Chongyu
Computer Science. 2025, 52 (11A): 250200110-6.  doi:10.11896/jsjkx.250200110
Abstract PDF(3973KB) ( 19 )   
References | Related Articles | Metrics
Aerospace TT&C(Trascking,Telemetry and Command)and data transmission resources are indispensable for satellite missions.Currently,decentralized and independent monitoring is the primary method for managing ground station resources.However,the traditional model of deploying one monitoring client per terminal is unable to meet the demands of managing large-scale heterogeneous resources.To address issues such as the proliferation of monitoring terminals,inconsistent alert mechanisms,and insufficient scalability,a centralized management and control platform(CMC-Platform) for multi-system heterogeneous aerospace TT&C and data transmission resources has been designed.The integrated feature extraction method,based on MBSE(Mo-del-Based Systems Engineering) technology,is used to establish parameter extraction rules to shield the differences between resources,and the failure warning monitoring based on the state machine model for multiple policies is implemented based on the integrated feature parameters,which solve the problem of analyzing the parameters of heterogeneous resources concisely and efficiently.The CMC-Platform can collect,extract and centrally display the condition of resources in real time,as well as carry out failure warning and emergency control during mission execution.The operational trial demonstrates that the platform achieves centralized control of over 100 heterogeneous resources from a single terminal,cutting emergency response time by 63.1%.
Research on Tourist Trip Design Problem of Theme Parks Based on Adaptive Large Neighborhood Search
ZHANG Shuzhu, LI Yongmei
Computer Science. 2025, 52 (11A): 250300080-7.  doi:10.11896/jsjkx.250300080
Abstract PDF(2301KB) ( 10 )   
References | Related Articles | Metrics
In recent years,theme parks have emerged as a popular form of tourism and have gradually become a favored choice for leisure and vacation among tourists.However,how to reasonably plan itineraries to maximize the visitor experience within limited time has become an important issue in the operation and management of theme parks.In response to this phenomenon,this paper proposes a new method for designing tourist itineraries,aiming to maximize visitors’ play experience.First,based on the operational characteristics of theme parks,interest points within the park are classified,and corresponding profit functions are constructed for each category to quantify visitors’ play experiences.At the same time,considering the negative impact of waiting time and travel time on visitor experience during actual tours,these are incorporated as constraints into the model.On this basis,a mixed-integer linear programming model is constructed,with precise modeling of the time window constraints for interest points to better reflect the actual operational scenario.To effectively solve this complex optimization problem,this paper proposes an improved adaptive large neighborhood search algorithm,which significantly enhances the solution quality by dynamically adjusting the search strategy.Through extensive numerical experiments,the effectiveness and feasibility of the proposed model and algorithm are systematically verified.Finally,taking Shanghai Disneyland Resort as a case study,the proposed model and algorithm are applied to real-world scenarios,and the results show that the method can significantly improve the visitor experience,providing scientific decision-making support for theme park operations and management.
Mechanical Stability Evaluation Method of Power Transformer Considering Short-circuit Reclosing
HAO Yuefeng, LIU Jun, XU Kui, XU Shurong, SHI Lintao, LU Yuchu
Computer Science. 2025, 52 (11A): 240800082-11.  doi:10.11896/jsjkx.240800082
Abstract PDF(7804KB) ( 13 )   
References | Related Articles | Metrics
With the expansion of power grid capacity and coverage,short-circuit faults at transformer outlets are becoming more and more frequent.In order to reduce the influence of not considering the secondary impact of reclosing on the anti-short-circuit check of power transformer,based on the fourth strength theory and the axial and radial instability criteria,a mechanical stability evaluation method of power transformer considering the influence of short-circuit reclosing is proposed.Firstly,the field-circuit coupling method is used to simulate the short-circuit reclosing condition,and the current size,magnetic flux leakage distribution and electromagnetic force density of the transformer winding during the short-circuit process are calculated.On this basis,the two-dimensional axisymmetric simplification of the finite element model is carried out,and the high-order polynomial fitting is used to construct a surrogate model for the rapid prediction of the magnetic flux leakage density of the transformer winding,so as to improve the calculation efficiency of the performance check of the power transformer under short-circuit conditions,and provide a new way for the rapid evaluation of the short-circuit resistance of the transformer.Then,the displacement and residual deformation of the winding under the action of von Mises stress and short-circuit electrodynamic force are analyzed by electromagnetic-structural field sequential coupling.Finally,the mechanical stability of transformer under short-circuit electromagnetic load is checked by combining the minimum yield failure coefficient and multi-dimensional structural instability criterion,and the construction scheme of dynamic stability evaluation and early warning system of in-service transformer based on cloud-edge coordination is expounded.The implementation of this evaluation method effectively improves the accuracy of transformer winding stability check,and provides scientific support for transformer maintenance and fault prevention.
Prediction of Short-and-Medium Term Photovoltaic Power Generation Based on Improved ModernTCN
ZHANG Yuechao, AN Guocheng, SUN Chenkai
Computer Science. 2025, 52 (11A): 241000164-7.  doi:10.11896/jsjkx.241000164
Abstract PDF(3113KB) ( 11 )   
References | Related Articles | Metrics
The medium- and short-term forecasting of photovoltaic(PV) power generation enables real-time monitoring of power fluctuations and reduces the impact of power volatility on PV systems.However,extreme weather changes and equipment faults or aging often lead to missing data in PV power generation records.This paper proposes an improved ModernTCN time series forecasting model to address this issue.The model firstly utilizes the multi-scale instance module and bias module of BiTGraph to handle missing data in the raw data,enhancing the spatiotemporal receptive field of the input.Then,ModernTCN’s dilated convolutions improve the ERF,allowing the model to better capture short- and medium-term dependencies in time series data as well as cross-variable dependencies in multivariate time series.The improved model supports medium- and short-term time series forecasting,even with minor data gaps in PV generation records,and is validated on three power-related datasets.Experimental results demonstrate that,compared to BiTGraph and ModernTCN models,the BG-ModernTCN model achieves an average reduction in mean squared error by 11.9% and mean absolute error by 12.8%.
Research on Emergency Rescue Quadcopter UAV Safety Control Based on Feedforward PID
MENG Dongyue, HUANG Yuchuan, HAN Guoxiang, LI Hongchen, WANG Pengfei
Computer Science. 2025, 52 (11A): 241200203-9.  doi:10.11896/jsjkx.241200203
Abstract PDF(4636KB) ( 12 )   
References | Related Articles | Metrics
To ensure the safe and reliable operation of quadcopter unmanned aerial vehicles at emergency rescue sites,and to address the non-standard interaction of quadcopter unmanned aerial vehicle information systems at emergency rescue sites,this paper proposes an emergency rescue quadcopter unmanned aerial vehicle system rescue site spatial information body,and verifies the integrity and confidentiality of the emergency rescue command center control module and control signals based on cryptographic methods.At the same time,in order to ensure stable flight of quadcopter drones according to command signals,a quadcopter drone dynamic model is established,and a Feedforward PID based quadcopter drone pose controller is adopted.Finally,a joint experiment is conducted using Python’s Cryptography and Matlab’s Simulink.The results indicate that the password module used in the study can complete authenticity and confidentiality verification,and the pose controller can also effectively achieve patrol operations at accident rescue sites.This study provides effective theoretical support and technical assurance for the safety and reliability of quadcopter unmanned aerial vehicles in practical emergency rescue applications.
Research on Timeliness of Information Flow of Main and Auxiliary Business of Plant and StationUnder Comprehensive Monitoring
ZHOU Zheng, CHEN Chen, WANG Song, SONG Xiaofan
Computer Science. 2025, 52 (11A): 241100111-6.  doi:10.11896/jsjkx.241100111
Abstract PDF(2153KB) ( 9 )   
References | Related Articles | Metrics
In order to solve the problem of timeliness of the information flow of main and auxiliary services of plant and station under comprehensive monitoring,the topological model of the information flow of the main and auxiliary services of the plant and station is constructed,the source node,the network node and the host node are abstracted into the transmission structure of the information flow,and the physical connection matrix and the virtual connection matrix are used to represent the physical and logical connection states between the nodes.In four typical business scenarios:circuit breaker failure,network transmission,one-key sequential control,and active and auxiliary linkage,the information flow transmission path and delay characteristics are quantitatively analyzed,and the path bottleneck and optimization requirements in the case of multiple events are identified.The results show that the delay of different service flows increases significantly due to the competition of switch ports,especially in the scenario of high interval number,the path delay shows an increasing trend.Therefore,it is concluded that reasonable optimization of network topology and configuration of priority scheduling strategies can significantly improve the information flow transmission efficiency of primary and secondary services.
渝ICP备16013121号-4
Add:18# Honghu West Rd., Yubei ,Chongqing,China    Post Code: 401121
Tel: 023-63500828/023-67039612/023-67039623
E-mail: jsjkx12@163.com
Powered by Beijing Magtech Co., Ltd.