Table of Content

10 October 2026, Volume 46 Issue 10

Previous Issue   

Artificial intelligence

Survey on differentiable neural architecture search methods

Han WANG, Ruiqi YU, Shuang REN

2026, 46(10):  3065-3075.  DOI: 10.11772/j.issn.1001-9081.2025091152

Asbtract ( )   HTML ( )   PDF (809KB) ( )  

Figures and Tables | References | Related Articles | Metrics

By automatically designing neural network structures， Neural Architecture Search （NAS） overcomes the trial-and-error limitations of traditional manual design， thereby substantially reducing the reliance of deep learning model development on professional experience， while enhancing the model performance significantly. Among various methods， Differentiable Neural Architecture Search （DNAS）， which leverages the one-shot framework and gradient descent optimization， has drastically shortened the search time and has emerged as a main research direction in current research. However， a systematic review of differentiable methods and their developmental trajectories is still absent. To bridge this gap and provide a reference for related studies， recent DNAS methods were synthesized and analyzed. First， the fundamental theories and benchmark datasets of DNAS algorithms were introduced， followed by a survey of the existing challenges of DNAS. Then， a hierarchical analysis of current DNAS algorithms was conducted. Finally， the future research directions of DNAS were discussed.

Graph convolutional recommender system integrating contrastive learning and global graph data augmentation

Long CHEN, Xinmin HOU

2026, 46(10):  3076-3082.  DOI: 10.11772/j.issn.1001-9081.2025091126

Asbtract ( )   HTML ( )   PDF (662KB) ( )  

Figures and Tables | References | Related Articles | Metrics

With the rapid growth of information， recommender systems have played a crucial role in improving user experience and solving information overload. However， the traditional recommendation methods still face limitations in addressing data sparsity， cold-start problems， and model interpretability. To tackle these challenges， a graph contrastive learning convolutional neural network recommender system based on global graph data augmentation， called GACL-Rec （Global Augmented Contrastive Learning for Recommendation）， was proposed. First， global augmentation was performed at the node， edge， and feature levels to enhance the model’s robustness under sparse data conditions. Second， a multi-granularity graph contrastive learning framework was designed， in which the consistency and discriminability of the embeddings were strengthened through contrastive learning tasks at the node， subgraph， and graph levels. Experimental results on several public datasets， including MovieLens-1M and Amazon-Books， demonstrate that the proposed recommender system outperforms the traditional models such as Matrix Factorization （MF） and LightGCN in terms of recommendation accuracy， robustness， and performance in cold-start scenarios， particularly achieving superior results on sparse datasets. Furthermore， parameter sensitivity analyses and ablation study results verify the effectiveness of the global augmentation strategy and the multi-granularity contrastive learning mechanism. It can be seen that the proposed model has significant advantages in addressing key challenges in recommender systems， such as data sparsity and cold-start problems.

Heterogeneous graph neural network based on multi-head relation transformation

Huajuan HUANG, Yongkang FANG, Xiuxi WEI

2026, 46(10):  3083-3092.  DOI: 10.11772/j.issn.1001-9081.2025081065

Asbtract ( )   HTML ( )   PDF (1404KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Concerning the problem that the existing Heterogeneous Graph Neural Networks （HGNNs） rely on manually constructed meta-paths and fail to fully capture complex relation semantics， a Multi-Head Relation-Transformed Heterogeneous Graph neural Network （MHRTHGN） method was proposed. In this method， independent relation transformation matrices were constructed for each type of relation， a multi-head attention mechanism was introduced to characterize multi-relation features precisely， and a relation-specific multi-head attention mechanism was designed to fuse multi-relation information at both node-level and relation-level. At the same time， a Feed-Forward Network （FFN） was combined to further enhance semantic expressiveness capability of the model. Experimental results show that on the ACM， IMDB， and DBLP datasets， the proposed method improves the accuracies of node classification and clustering tasks by 3-5 percentage points on average compared with representative baseline models such as HAN （Heterogeneous graph Attention Network） and Graph Convolutional Network （GCN）， validating its effectiveness and performance superiority in multi-relation heterogeneous graph learning. It can be seen that MHRTHGN method can model multi-relation information without relying on meta-path prior， providing a more generalizable solution with higher application potential for heterogeneous graph representation learning.

Multi-model collaborative continual learning framework based on filtering mechanism

Yuzhen WU, Chen PAN, Qiannan LI

2026, 46(10):  3093-3100.  DOI: 10.11772/j.issn.1001-9081.2025091087

Asbtract ( )   HTML ( )   PDF (1131KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the difficulty of balancing stability and plasticity in the existing continual learning methods， a multi-model collaborative continual learning framework based on filtering mechanism was proposed. The framework is composed of a main model， an auxiliary model， and a student model. The main model was designed to focus on predicting and outputting high-confidence results， the auxiliary model was used to memorize hard samples that the main model failed to learn， and the student model was employed to inherit and integrate the recognition ability of a cascaded teacher model constructed from the main and auxiliary models. The three models were optimized collaboratively during each task process. Experimental results show that in task-incremental scenarios， the proposed framework achieves higher AAC （Average ACcuracy）， BWT （BackWard Transfer）， and FWT （ForWard Transfer） than typical methods such as EWC （Elastic Weight Consolidation）， GEM （Gradient Episodic Memory）， and iCaRL （incremental Classifier and Representation Learning） on both the Fashion-MNIST and CIFAR-10 datasets， demonstrating stronger stable memory and adaptive capabilities； in class-incremental scenarios， the framework has good applicability on both the 10-class Fashion-MNIST and the 100-class CIFAR-100 datasets， with its AAC higher than that of the comparative methods significantly.

Policy optimization method integrating overestimated and underestimated value functions

Ziheng ZHANG, Jin QIN

2026, 46(10):  3101-3106.  DOI: 10.11772/j.issn.1001-9081.2025091175

Asbtract ( )   HTML ( )   PDF (1931KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In Reinforcement Learning （RL）， the estimation bias of value functions is a critical challenge that limits the improvement of algorithm performance. It is caused by factors such as uncertainty in environmental dynamics， function approximation errors， and data distribution shifts， and makes the policy optimization direction deviate from the true optimal solution and even leads to divergence in the training process. Therefore， a Policy Optimization method Integrating overestimated and underestimated Value Functions （POIVF） was proposed. Meanwhile， a dynamic weight adjustment method was designed， through which the weight of the overestimated value function was adjusted dynamically by real-time approximation of the underestimation degree of the current underestimated value function. Additionally， a method for selecting underestimated value functions dynamically was presented. To verify the effectiveness of POIVF， the method was combined with the TD3 （Twin Delayed Deep Deterministic policy gradient）， TADD （Triplet Average Deep Deterministic policy gradient algorithm）， and QMD3 （Quasi-Median Delayed Deep Deterministic policy gradient） algorithms， respectively， and was tested and evaluated on six control tasks in OpenAI Gym. Experimental results show that POIVF achieves performance improvements in most tasks.

Joint intent-slot recognition method based on prompt-guided generation and tri-mutual attention

Shangzhi TENG, Jin HUANG, Xueqiang LYU, Xindong YOU

2026, 46(10):  3107-3116.  DOI: 10.11772/j.issn.1001-9081.2025091092

Asbtract ( )   HTML ( )   PDF (1065KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the semantic ambiguity and difficulty in candidate selection caused by the diverse results generated by Large Language Models （LLMs） in intent detection and slot filling tasks， a joint intent-slot recognition method based on prompt-guided generation and tri-mutual attention mechanism was proposed， named PGT-Joint （Prompt-guided Generation and Tri-mutual attention for Joint intent-slot recognition）. First， the prompts were constructed based on domain knowledge to guide the LLM in generating multiple candidate intent-slot combinations. Second， a tri-mutual attention mechanism was designed to model the high-order semantic relationships among user questions， candidate intents， and slots explicitly. Finally， a contrastive learning ranking model was used to filter and optimize the candidate results. Experimental results on a Mini/Micro LED domain question-answering dataset show that compared with current mainstream models except Global-Locally Graph Interaction Network （GL-GIN）， such as BERT-Intent and BERT-Slot， PGT-Joint improves the intent detection accuracy by 0.4 to 3.8 percentage points and the slot filling F1 score by 4.1 to 6.6 percentage points. Notably， the most significant gain is in overall accuracy， with an improvement of 5.2 to 10.2 percentage points. In addition， ablation experimental results verify the key roles of prompt generation and tri-mutual attention mechanism in performance improvement， indicating that PGT-Joint can enhance semantic understanding capabilities in domain-specific dialogue systems effectively.

Zero-shot knowledge graph relation prediction method via dual-channel embedding fusion

Bing GAO, Jinzhuo LIN, Qijie ZOU, Yongqiang YAO

2026, 46(10):  3117-3123.  DOI: 10.11772/j.issn.1001-9081.2025080945

Asbtract ( )   HTML ( )   PDF (743KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Zero-shot knowledge graph relation prediction aims to infer triplets with unseen relations using the existing fact triplets. The existing methods often underutilize entity descriptions， only considering text information and graph structure information of relations. Text information for relations only relies on dataset-provided text descriptions， so it is difficult to capture the detailed semantics of relations and the overall semantic information of knowledge graph triplets. To address this issue， a zero-shot knowledge graph relation prediction method via dual-channel embedding fusion was proposed. First， a structural encoder and a semantic encoder were used to generate structural and semantic features of triplet entities. Then， the obtained features were fused via a dual-channel entity multimodal information fusion module， and the text generation ability of Large Language Models （LLMs） was used to generate description information after enhancing relation text semantics. Third， through training， semantically enhanced relation features and multimodal entity features were aligned semantically in vector space. Finally， for zero-shot relations， zero-shot knowledge graph relation prediction was realized according to semantically enhanced text description information. Experimental results on two real datasets show that compared to the optimal baseline method， on the NELL_ZS dataset， the proposed method improves Mean Reciprocal Rank （MRR）， Hit@1， Hit@5， and Hit@10 by 8.7， 7.2， 9.5， and 11.2 percentage points， respectively； on the WN18RR_ZSL dataset， the proposed method improves MRR， Hit@1， Hit@5， and Hit@10 by 6.6， 9.2， 3.9， and 0.8 percentage points， respectively， demonstrating the effectiveness of the proposed method.

Speech emotion recognition with global-aware fusing binary multi-scale features

Zheng WANG, Jiahui NI, Fanlong ZHANG

2026, 46(10):  3124-3132.  DOI: 10.11772/j.issn.1001-9081.2025080935

Asbtract ( )   HTML ( )   PDF (1557KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In response to the problems that traditional Convolutional Neural Networks （CNNs） have difficulty in capturing global emotional information in Speech Emotion Recognition （SER） due to limited receptive fields， and that global self-attention mechanisms have excessive computational overhead on resource-constrained devices such as mobile terminals， a lightweight and high-precision SER model suitable for edge deployment — a Global-Aware Fusion network with Binary Multi-scale Feature Representation （GAF-BMFR） was proposed. In this network， a Binary Neural Network （BNN） was used as the backbone to reduce model complexity. The core parts include the following ones： a Binary Multi-Scale Block （BMSB） was designed， in which parallel 3×3 and 5×5 binary convolutions were used to extract fine-grained and large-receptive-field features synchronously in the time-frequency domain； a lightweight Global-Aware Block （GAB） was introduced， which utilized fully connected layers and 1×1 convolutions to replace the high-cost self-attention mechanism， thereby achieving global correlation modeling and adaptive selection of cross-scale features. Furthermore， a union loss function of Class-Balanced Focal， Log-Margin， and Multi-Scale binary Regularization （CB-FLM-MSR） was designed to alleviate problems such as class imbalance and ambiguous emotional boundaries. Experimental results on public datasets show that GAF-BMFR has superior performance with low resource consumption. Taking the script part of the IEMOCAP dataset as an example， compared to the typical Attention Pooling Convolutional Neural Network （APCNN） model， GAF-BMFR has the Weighted Accuracy （WA） improved by 25.23 percentage points， and the Unweighted Accuracy （UA） improved by 24.18 percentage points. Concurrently， GAF-BMFR has the number of model parameters reduced by approximately 24.04% significantly， compared to its full-precision version. It can be seen that GAF-BMFR model enhances the accuracy and robustness of SER effectively while reducing the number of parameters and computational overhead significantly， providing a feasible lightweight solution for deploying efficient affective computing applications in resource-constrained scenarios.

Criminal psychological attribution classification model based on TF-Chi probability features and hybrid prompt learning

Congcong LI, Guandong GAO, Ke XIAO, Bowei SHEN

2026, 46(10):  3133-3142.  DOI: 10.11772/j.issn.1001-9081.2025081057

Asbtract ( )   HTML ( )   PDF (891KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Analyzing the causes of criminal psychology is crucial for assessing criminal risks and develop psychological correction plans. Traditional models have limitations in comprehending criminal psychology texts， revealing correlations between prompt vectors and downstream tasks， and identifying the importance of key terms dynamically. Therefore， a CP-BERT （Criminal Psychology-Bidirectional Encoder Representations from Transformers）-based attribution classification model for criminal psychology was proposed， enabling four-class classification of psychological causes of crime. First， an 8.03-million-character criminal psychology corpus was constructed for domain-adaptive pre-training， and semantic modeling capability of the model was enhanced. Second， a hybrid prompt learning method was employed to divide prompt templates into prefix and suffix segments： highly relevant prior information was injected into the prefix segment during the initialization phase， while random initialization was performed on the suffix segment by using normal distribution to supplement information uncovered by the prefix segment， and optimal proportions were determined experimentally. Third， an improved TF-Chi （Term Frequency-Chi-Square） association probability feature method was proposed， in which numerical features were mapped to information relevance probability values dynamically， and feature fusion was achieved by constructing a weight matrix. Finally， an annotated dataset consisting of 3 484 personal and criminal histories was constructed for model training and evaluation. Experimental results demonstrate that this model has an accuracy of 91.25% and an F1 score of 91.56%， significantly outperforming multiple baseline models such as ChatGLM and DeepSeek in all metrics， verifying its effectiveness in criminal psychological attribution tasks.

Agricultural user identification method integrating image-text and large language model

Yimeng SHANG, Zenglin CHI, Hongming ZHANG, Bin LIU, Guoqiang HU, Dangdang NIU

2026, 46(10):  3143-3152.  DOI: 10.11772/j.issn.1001-9081.2025091082

Asbtract ( )   HTML ( )   PDF (1106KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The difficulty of relying on historical interaction data for effective modelling in the cold-start stage of recommender systems constrains the quality and precision of personalized services severely. To address this issue， user profile pictures and nicknames obtained from login recommender systems were utilized， and a multimodal identification model combining user profile pictures and nicknames， ResBERT-MMNet （ResNet-BERT MultiModal Network）， was proposed. In the model， a multi-layer stacked bidirectional residual cross-modal attention mechanism was designed to achieve image-text semantic interaction， thereby enhancing the understanding of user identities. Furthermore， to address the limitation of incorrect predictions produced by multimodal models under low-confidence conditions， a Large Language Model （LLM）?based prediction of agricultural user identity types was introduced， and the DeepSeek-R1-Distill-Qwen-7B model was used for Low-Rank Adaptation （LoRA） fine-tuning， so as to infer user identity types from user nicknames， and the calculation results of LLM were employed as a supplementary decision-making module to improve overall system robustness. Experimental results on a self-built dataset demonstrate that ResBERT-MMNet outperforms other models in multi-type agricultural user identification tasks with the identification precision and F1-score reaching 67.47% and 68.05%， respectively. After integrating with the fine-tuned LLM， the model achieves an identification precision of 69.73% and an F1-score of 70.06%， indicating strong practical applicability. It can be seen that the proposed model provides the prior data in the cold-start stage of agricultural recommender systems and offers a technical solution for user profiling and precision services.

Large language model information extraction for aeronautical data link

Kui TANG, Shuanliang GAO, Yang LI, Geng LI, Jiaxing ZHANG, Yuanfang ZHU, Yongming FU, Ruiqi MAO, Hongru LIANG

2026, 46(10):  3153-3160.  DOI: 10.11772/j.issn.1001-9081.2025091104

Asbtract ( )   HTML ( )   PDF (802KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In critical task scenarios such as aeronautical data link command parsing in the military domain， information extraction systems must achieve both high accuracy and structural consistency to ensure the safety and reliability of mission execution. However， the traditional extraction methods based on rules or few-shot example prompts often suffer from severe extraction results’ hallucinations when dealing with complex semantics， flexible syntax， and highly constrained structures. To address this issue， a Retrieval-Augmented Generation Information Extraction （RAG-IE） algorithm based on Large Language Model （LLM） was proposed by combining retrieval-augmented technology with information extraction technology. First， a meta-language library oriented to the target information types was constructed to provide standardized extraction templates. Then， efficient template retrieval was achieved based on an inverted index by combining keyword matching with semantic expansion strategies. Finally， the prompts were designed to guide the LLM to convert natural language into structured information， thereby enabling highly controllable and accurate information extraction of natural language commands in aviation data link scenarios. Experimental results show that this algorithm is superior to the existing information extraction algorithms， reducing the hallucinations in the information extracted by LLMs greatly. Meanwhile， this algorithm has strong domain independence and can be extended to other scenarios requiring highly reliable extraction by reconstructing the meta-language library， providing a generalizable technical pathway for building intelligent and verifiable information processing systems.

Flight ground service milestone event prediction based on CTrans-ATGCN

Xia FENG, Qi WANG, Mingze ZHANG, Haichao ZUO, Guoyu WANG

2026, 46(10):  3161-3170.  DOI: 10.11772/j.issn.1001-9081.2025080942

Asbtract ( )   HTML ( )   PDF (1410KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the challenges of severe data missingness， high uncertainty in influencing factors， and limited prediction accuracy in flight ground service milestone event prediction， a flight ground service milestone event prediction framework named CTrans-ATGCN was proposed， including two core components： missing data imputation and milestone event prediction. First， based on the Transformer architecture， a convolutional layer was introduced to extract the temporal and topological features in flight ground service milestone event data， and a self-attention mechanism was employed to model potential global dependencies among different milestone events， thereby constructing CTrans （Convolutional Transformer）， a method for flight ground service data imputation that integrates convolutional self-attention mechanism. On this basis， a milestone event prediction method， Attention-based Temporal Graph Convolutional Network （ATGCN）， was constructed. In this method， a Temporal Graph Convolutional Network （T-GCN） was utilized to capture spatio-temporal dependencies in historical milestone event data， and by combining an attention mechanism with a scoring function， the weights among different milestone event nodes were adjusted dynamically， so as to achieve high-precision prediction for ground service milestone events. Experimental results on a 2023 real-world dataset from an airport in North China show that in a non-random missing scenario with a 70% missing rate， the proposed imputation method reduces the Mean Absolute Error （MAE） by approximately 14.91% compared to the suboptimal model GraphSAGE （Graph Sample and AggreGatE）； on low-missing-rate datasets， the proposed “imputation + prediction” method achieves prediction performance similar to that of Sliding Bi-Directional Graph Convolutional Neural network Model （SBDGCNM）； on datasets with 70% missing rate， CTrans-ATGCN achieves the MAE and Root Mean Square Error （RMSE） of 14.76 and 18.42， respectively， outperforming all baseline methods. It can be seen that this framework provides an effective method for flight ground service prediction under missing data， and offers reliable support for efficient scheduling of ground personnel and equipment.

Data science and technology

Traffic flow prediction model with spatial-temporal awareness and self-supervised Transformer

Jinxiang WEI, Leichun WANG, Xiangfeng LU, Ruirui SONG, Yunping HE, Xiaomeng LIU

2026, 46(10):  3171-3179.  DOI: 10.11772/j.issn.1001-9081.2025091098

Asbtract ( )   HTML ( )   PDF (916KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the limitations of the existing traffic flow prediction methods in capturing dynamic patterns of road network topology and local fine-grained features， a traffic flow prediction model with Spatial-Temporal Awareness and self-supervised Transformer （STA-Autoformer） was proposed. In the model， by integrating self-attention mechanisms with self-supervised learning strategies， the accuracy of spatial-temporal traffic flow prediction was improved. First， spatial and temporal self-attention mechanisms were integrated to overcome the limitation of the existing methods in capturing long-range spatial-temporal dependencies. Second， a dynamic topology awareness module was designed to address the local impact of emergencies. Finally， a multi-scale temporal awareness module was constructed， in which temporal enhancement and diffusion mechanisms were used to extract temporal features from periodic and local features， respectively. Experimental results demonstrate that compared to the best-performing baseline model ST-SSL （Spatio-Temporal Self-Supervised Learning）， STA-Autoformer reduces the Mean Absolute Error （MAE） in inflow prediction by 3.0% and 1.8%， respectively， while lowering the Mean Absolute Percentage Error （MAPE） by 5.8% and 1.4%， respectively， on the NYCBike1 and NYCBike2 datasets； on the NYCTaxi dataset， STA-Autoformer reduces the MAE in inflow and outflow predictions by 1.4% and 1.3%， respectively. It can be seen that STA-Autoformer achieves significant performance improvements across multiple real-world traffic datasets.

Mesoscale modeling-based multi-agent simulation method for dynamic traffic

Rui ZHANG, Zhicheng ZHENG, Lihu PAN, Huanyuan GUO, Linliang ZHANG, Huimin YAN

2026, 46(10):  3180-3190.  DOI: 10.11772/j.issn.1001-9081.2025091132

Asbtract ( )   HTML ( )   PDF (1280KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the difficulty in multi-scale collaborative modeling and the insufficient interaction continuity in traffic simulation systems， a multi-scale micro-meso-macro integrated traffic simulation framework was proposed. In the framework， with a core logic of data-driven processing， scale classification， policy execution， trajectory control， simulation feedback， a modular modeling architecture was constructed for multi-source information fusion and behavior policy decoupling； by establishing a mesoscale bridging layer， the bidirectional information propagation between micro behaviors and macro trends was achieved， and a Social Value Orientation （SVO） swarm Monte Carlo Tree Search （MCTS） cooperative decision-making mechanism was incorporated. At the same time， at the macroscopic level， a dynamic OD （Origin-Destination） matrix generation module was constructed to empower the simulation system with adaptive control capabilities to the temporal evolution of traffic flow. Experimental results demonstrate that the proposed method achieves substantial improvements in both behavioral rationality and scenario adaptability of the system while maintaining simulation efficiency. Compared with the traditional single-scale models and single-vehicle decision models without collaborative strategies， the method reduces the conflict rate by 19.8% on average， and increases the Throughput Ratio （TR） by 34.6% comprehensively， providing an effective technical path for intelligent coordination and resource optimization in complex traffic scenarios.

Power usage effectiveness prediction of small air-cooled data centers based on LSTM-Transformer

Yuxuan TENG, Congfeng JIANG, Junming LIU, Chuanqi XU

2026, 46(10):  3191-3198.  DOI: 10.11772/j.issn.1001-9081.2025080944

Asbtract ( )   HTML ( )   PDF (961KB) ( )  

Figures and Tables | References | Related Articles | Metrics

With the continuous expansion of data center scale， Power Usage Effectiveness （PUE） has become an important indicator for measuring energy efficiency of data centers. Most existing research focuses on hyperscale data centers， while small air-cooled data centers have significant differences in scalability， load fluctuations， network infrastructure， and management approaches， leading to relatively insufficient related prediction studies. Therefore， a PUE prediction model of small air-cooled data centers based on LSTM-Transformer architecture was proposed. The model was designed to combine the capability of Long Short-Term Memory （LSTM） network in extracting time-series features with the advantage of Transformer in modelling global dependencies， so that multi-source heterogeneous energy data were integrated effectively， thereby enhancing the characterization and prediction capabilities of PUE variation patterns. Experimental results on a real data center dataset show that the Mean Squared Error （MSE）， Mean Absolute Error （MAE）， and Coefficient of Determination （R2） of the proposed model are 0.000 7， 0.021 4， and 0.958 3， respectively； compared with the LSTM model， the proposed model has the MSE and MAE reduced by 53.33% and 28.43%， respectively， and R2 increased by 0.051 6； compared with the Transformer model， the proposed model has the MSE and MAE reduced by 78.79% and 45.41%， respectively， and R2 increased by 0.136 3， verifying the effectiveness of the proposed hybrid architecture in improving PUE prediction accuracy. It can be seen that the proposed model can provide reliable support for the energy efficiency optimization and operation management of small air-cooled data centers.

Mamba-ADCL： time series anomaly detection method based on contrastive learning and Mamba

Zhihong ZHAO, Xushuo LIU

2026, 46(10):  3199-3205.  DOI: 10.11772/j.issn.1001-9081.2025080958

Asbtract ( )   HTML ( )   PDF (750KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of the existing time series anomaly detection methods， such as the heavy reliance on labeled data and the insufficient modeling capability in complex scenarios， a Mamba-ADCL （Anomaly Detection in time series via Contrastive Learning with Mamba） was proposed. In the method， a dual-branch Mamba structure was designed， in which a shared-weight framework and a contrastive learning strategy were employed to enhance the feature discrepancies between normal and abnormal samples effectively， thereby improving model robustness and generalization. Meanwhile， a representation method with collaborative time-domain and frequency-domain modeling was introduced， in which time series data was transformed into frequency representations and combined with time-domain data to construct a dual-view contrastive learning mechanism jointly， thereby enhancing the recognition ability for abnormal patterns. Experimental results on four public datasets， MSL， SMAP， SWaT， and PSM show that the F1-score of Mamba-ADCL is improved by 2.22， 0.22， 3.86， and 0.35 percentage points compared with that of AnomalyTrans （Anomaly Transformer）， and by 2.43， 0.52， 0.70， and 0.28 percentage points compared with that of TFMAE （Temporal-Frequency Masked AutoEncoder）， which verifying the effectiveness of Mamba-ADCL in time series anomaly detection.

Multivariate time series data analysis method based on transposed self-attention mechanism and convolutional autoencoder

Sheng LIU, Yinghua TONG

2026, 46(10):  3206-3214.  DOI: 10.11772/j.issn.1001-9081.2025080971

Asbtract ( )   HTML ( )   PDF (841KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the limitations of the existing Transformer models in trend information extraction and the distortion in modeling variable correlations caused by sensor heterogeneity， a multivariate time series data analysis method based on Transposed Self-Attention （TSA） mechanism and Convolutional AutoEncoder （CAE） was proposed. In the method， a transposed one-dimensional CAE structure was constructed to reconstruct time series data features in the time dimension， which captured long-term trend information in multivariate time series data effectively. Meanwhile， the TSA mechanism was introduced into the multivariate correlation analysis process， which improved the modeling accuracy of inter-variable relationships significantly. Comparative experimental results on prediction performance of the proposed method and recent mainstream methods showed that the proposed method achieved an average Mean Squared Error （MSE） of 0.545， which was 11.4% lower than that of iTransformer and 7.4% lower than that of DLinear. Furthermore， anomaly detection experimental results on five public datasets showed that the proposed method achieved an average F1 score of 84.92%， outperforming all the baseline methods. Prediction experimental results on a real-world water quality monitoring dataset showed that the proposed method reduced the average Mean Absolute Error （MAE） and MSE by 46.2% and 42.9%， respectively， compared to Autoformer， demonstrating its effectiveness in practical applications.

Cyber security

Survey of remote access Trojan detection methods based on traffic analysis

Bin ZHANG, Wenyong WANG, Qixun ZHU, Jiaxu QIU, Jiahe ZHANG, Conghui LIAO, Ziyu LIAO

2026, 46(10):  3215-3226.  DOI: 10.11772/j.issn.1001-9081.2025091099

Asbtract ( )   HTML ( )   PDF (918KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In recent years， Remote Access Trojans （RATs） have been widely used by hackers for malicious activities such as data theft， system compromise， and ransomware attacks， and their stealth and persistence make them a major threat in the field of cybersecurity. Since the traditional RAT detection methods have low detection rates and high false negative rates when facing encrypted traffic and protocol obfuscation， RAT detection based on traffic analysis has emerged as a cybersecurity research hotspot due to its deep insight capability into communication behaviors. Through extensive literature research， a detailed introduction to the latest research progress using traffic analysis to identify RATs was given. First， a brief overview of the communication principles and detection methods of RAT was provided. Second， the current RAT detection methods using traffic analysis were summarized as the focus. Third， the specific application scenarios of RAT detection were summarized from a practical application perspective. Finally， the problems in the current research process were discussed， and the future development directions were proposed.

Information security risk assessment based on multi-objective programming and evidential reasoning theory

Jingping WANG, Yanting LIU, Haiming LONG, Yingjun ZHANG

2026, 46(10):  3227-3232.  DOI: 10.11772/j.issn.1001-9081.2025091123

Asbtract ( )   HTML ( )   PDF (525KB) ( )  

Figures and Tables | References | Related Articles | Metrics

With the rapid development of mobile Internet and the emergence of massive big data， information security risks are increasingly prominent， so the construction and improvement of a scientific and objective security risk assessment system have great theoretical and practical significance. However， the existing assessment methods face challenges such as the representation of uncertain information， the objective quantification of expert and attribute weights， and the fusion of uncertain information. To address these challenges， an information security risk assessment model based on multi-objective programming and evidential reasoning theory was proposed. First， interval-valued intuitionistic fuzzy numbers were used to represent uncertain information， and three basic operational laws were defined. Second， expert and attribute weights were quantified objectively based on multi-objective programming. Third， evidential reasoning fusion rules in the interval-valued intuitionistic fuzzy environment were introduced to solve the fusion problem of uncertain information. Finally， the ideal solution approaching method TOPSIS （Technique for Order Preference by Similarity to Ideal Solution） was referenced for comprehensive information security assessment. Experimental results verify the effectiveness and feasibility of the proposed model in information representation， weight quantification， and risk assessment.

DDoS attack detection method based on DeepInsight and hybrid quantum convolutional neural network

Huaijin LI, Shibin ZHANG, Min YANG

2026, 46(10):  3233-3241.  DOI: 10.11772/j.issn.1001-9081.2025091091

Asbtract ( )   HTML ( )   PDF (1080KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Distributed Denial of Service （DDoS） attacks remain one of the primary threats faced by modern networks due to their high stealthiness and polymorphism. With the development of the Internet of Things （IoT） and cloud computing， DDoS traffic exhibits characteristics such as suddenness， heterogeneity， and rapid evolution， posing challenges for traditional detection methods in complex attack identification and sample imbalance. To overcome these limitations， a DDoS detection method combining DeepInsight and Classical-to-Quantum Convolutional Neural Network （DI-C2Q-CNN） was proposed. First， DeepInsight technology was utilized to map multi-dimensional time-series network traffic data into structured grayscale images intelligently， and by reconstructing spatial relationships between features explicitly， the identifiability of attack patterns was enhanced. Second， a Classical-to-Quantum Convolutional Neural Network （C2Q-CNN） hybrid architecture was constructed， integrating classical Convolutional Neural Network （CNN） with quantum convolution modules. By leveraging the representational capability of quantum superposition states and high-dimensional feature mapping， the higher-order representation of image features and the ability to learn complex patterns were enhanced， thereby improving the capability to identify complex network attack patterns. Additionally， to address training bottlenecks caused by small samples and class imbalance， multiple transfer learning solutions were introduced to enhance model adaptability and robustness. Experimental results demonstrate that the proposed method outperforms mainstream models such as Support Vector Machine （SVM） and CNN-LSTM on the UNSW-NB15 dataset， even under extreme sample compression and class imbalance conditions， showing good performance and application potential in DDoS attack detection tasks.

Fault-tolerant method for industrial IoT edge computing based on sparse graph attention

Yuqing NIU, Zhiyong ZHANG, Zhongya ZHANG, Hang ZHANG

2026, 46(10):  3242-3249.  DOI: 10.11772/j.issn.1001-9081.2025091151

Asbtract ( )   HTML ( )   PDF (822KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of node overload and failure caused by resource constraints in Industrial Internet of Things （IIoT） edge computing environments， this paper proposed a fault-tolerant method based on edge intelligence — Sparse Graph ATtention-based Generative Adversarial Network （SGAT-GAN）. This method integrated a fault prediction mechanism and employed a Fault Embedding Encoder （FEE） to achieve real-time fault detection， classification， and prediction， thereby enhancing the system’s fault-tolerance capability. On this basis， this method combined Generative Adversarial Network （GAN） with a preemptive migration strategy to select the optimal strategy for intelligent fault recovery， thereby providing an efficient and intelligent solution for IIoT edge computing systems. Experimental results in a Raspberry Pi-based edge environment demonstrate that this method outperforms four state-of-the-art methods — Proactive Coordinated Fault Tolerance （PCFT）， Clustering-based Multiple Objective Dynamic Load Balancing technique （CMODLB）， self-supervised deep proxy-based edge computing fault tolerance method DeepFT， and Preemptive migration prediction Generative Adversarial Network （PreGAN） — in terms of average detection precision and average classification accuracy of faults. Specifically， compared with the best-performing state-of-the-art method PreGAN， the proposed method achieves improvements of 1.96 and 1.62 percentage points in these two metrics， respectively， validating its effectiveness and superiority in enhancing the reliability of IIoT edge computing systems.

Advanced computing

Zynco： distributed service framework for runtime state coevolution

Zhongyi LIU, Mengxiao LI, Cheng LIN, Feng TIAN, Dan FU, Wei SHAN, Zheng WANG, Jian YANG

2026, 46(10):  3250-3259.  DOI: 10.11772/j.issn.1001-9081.2025080953

Asbtract ( )   HTML ( )   PDF (1170KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The distributed computing systems place greater demands on incremental state recognition， node consistency synchronization， and seamless service switching during continuous task running. However， the existing mechanisms still have limitations in awareness of cluster state structure， node consistency control， and controllable service version switching， which hinder stable service running under frequent state evolution. Therefore， Zynco was proposed， a lightweight and distributed framework for runtime state coevolution. Zynco was designed with three core mechanisms： a fingerprint-based and structure-aware incremental recognition mechanism， by which only effective state changes were transmitted and a sliding version chain was maintained； a feedback protocol combining User Datagram Protocol （UDP） multicast with Transmission Control Protocol （TCP） control channels， to achieve efficient distribution and acknowledgment of state among multiple nodes； a three-layer service structure with listener-based routing， thereby enabling smooth service switching and consistent rollback. Experimental results show that， in a 20-node cluster， Zynco requires only 27 s to complete state synchronization for incremental updates involving 54 GB of file data， without task interruption or state tearing during service switching； under a 1 000 QPS （Queries Per Second） load， Zynco has the response jitter time below 8 ms. It can be seen that Zynco is efficient， stable， controllable， and can provide good operational support for state?driven distributed high?performance services in massive?data collaborative update scenarios.

Improved hippopotamus optimization algorithm based on coupling interference and covariance evolution

Miaomiao LIU, Jiahao XI, Qiang ZHANG, Jingfeng GUO, Wenxiang SHI

2026, 46(10):  3260-3271.  DOI: 10.11772/j.issn.1001-9081.2025081050

Asbtract ( )   HTML ( )   PDF (3222KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of uneven population distribution in the initialization phase， susceptibility to the local optimum in complex search space， and insufficient optimization capability in the later exploitation phase of Hippopotamus Optimization （HO） algorithm， an improved HO algorithm based on Latin hypercube sampling， Neural population dynamic optimization with coupled interference and Covariance matrix adaptation （LNC_HO） was proposed. First， in the initialization phase， Latin hypercube sampling was employed to optimize the initial population distribution， thereby enhancing population diversity and global coverage. Second， in the exploration phase， a coupled interference strategy was combined to strengthen the algorithm’s global search capability in high-dimensional space， thereby reducing the risk of falling into the local optimum. Finally， in the exploitation phase， a Covariance Matrix Adaptation Evolution Strategy （CMA-ES） was introduced to adjust the covariance matrix and step-size parameters adaptively， thereby enhancing the algorithm’s dynamic adaptability to search space structures and improving both convergence speed and accuracy in the late-stage local search. Experimental results on twelve multi-type benchmark functions in CEC2017， compared with seven mainstream swarm intelligence algorithms， show that LNC_HO ranks first in terms of optimal fitness value， and achieves an improvement of 42.22% in convergence accuracy and reduces standard deviation by 27.37% compared to the HO algorithm， indicating its superior convergence performance and stability， thereby validating the effectiveness of the improvement strategies. When LNC_HO is applied to the pressure vessel design engineering optimization problem， it achieves a cost reduction of 8.28% compared to the HO algorithm in 30 independent runs， validating its practical value in real-world engineering applications.

APINN-based method for solving integro-differential-algebraic equations

Xiaoqing ZENG, Juan TANG, Huafeng XIAO, Manqi XIE, Kai LUO

2026, 46(10):  3272-3279.  DOI: 10.11772/j.issn.1001-9081.2025091145

Asbtract ( )   HTML ( )   PDF (794KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues of high computational and storage costs， poor stability of stiff systems， and difficulties in handling algebraic constraints in the traditional numerical methods for Integro-Differential-Algebraic Equations （IDAEs）， an Auxiliary Physics-Informed Neural Network （APINN）-based method for solving IDAEs， named APINN-IDAE， was proposed. The core idea of this method is to use the multi-task learning mechanism of APINN， introducing auxiliary network outputs to specifically approximate the integral history terms， so that the intractable integral constraints in the original IDAE system were transformed into differential relations that are easier to optimize， thereby avoiding the computationally expensive numerical integration in the traditional methods. The differential variables were handled by the main network output， while the integral terms and algebraic constraint residuals were processed by the auxiliary outputs， respectively. In addition， a unified weighted loss function was constructed， incorporating the differential equations， algebraic equations， auxiliary variable differential relations， and initial conditions. Experimental results on several typical IDAE examples （including nonlinear， variable-coefficient， and strongly coupled systems） show that the predicted solutions of the proposed method highly coincide with the exact solutions within the solution domain. And the relative L₂ errors for both differential and algebraic variables reach the order of 10?3 to 10??， without observable error accumulation over time. It can be seen that APINN-IDAE retains the mesh-free and automatic differentiation advantages while using auxiliary variables to avoid explicit differentiation of integral terms， thereby reducing the computational complexity of integral operators and improving gradient stability， offering potential robustness for stiff or high-index IDAE systems， and providing a new perspective for solving multi-physics coupling problems.

Multi-attribute decision-making method based on utility three-way decision in Pythagorean fuzzy environments

Fuyue TAO, Lei WANG, Min LI

2026, 46(10):  3280-3287.  DOI: 10.11772/j.issn.1001-9081.2025091111

Asbtract ( )   HTML ( )   PDF (711KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The traditional Multi-Attribute Decision-Making （MADM） methods are based on two-way decision， which involve only acceptance or rejection. Three-Way Decision （3WD） reduces decision risks by introducing a deferred decision. In the current complex and dynamic decision-making environments， applying 3WD to provide solutions to MADM problems has emerged as a hot research topic. Therefore， a utility three-way multi-attribute decision-making model based on Pythagorean Fuzzy Numbers （PFNs） similarity was proposed. First， the relative utility function was extended to the Pythagorean fuzzy environment， and a similarity measure of PFNs was integrated， to establish a similarity-based relative utility function. Then， the ideal superior and inferior reference sequences were obtained using the score functions， and grey relational coefficients were computed on the basis of distance measures between PFNs， thus proposing a method for calculating conditional probabilities based on Grey Relational Analysis （GRA）. Based on this， a ranking strategy was introduced via expected utility， thereby obtaining a three-way multi-attribute decision-making model based on PFN similarity. The feasibility of the proposed method was validated through a site selection assessment case for new municipal solid waste treatment plants. Experimental results demonstrate that the proposed method can categorize alternatives precisely and reduce decision risks. Compared to the traditional MADM methods， it incorporates objective weights with 3WD classification to overcome the limitation of only ranking； compared to the three-way MADM methods， its GRA optimized conditional probabilities are more consistent with the classification logic of 3WD.

Computer software technology

Multi-strategy compiler fuzzing system based on large language models

Donghui LI, Yingying LI, Chunyan ZHANG, Peng LIU

2026, 46(10):  3288-3297.  DOI: 10.11772/j.issn.1001-9081.2025091078

Asbtract ( )   HTML ( )   PDF (972KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the common issues in the existing compiler fuzzing， such as insufficient test case diversity， high testing costs， and heavy reliance on manual intervention， a multi-strategy compiler fuzzing system based on Large Language Models （LLMs）， DHFUZZ （Deep & Hybrid FUZZer）， was proposed. In the system， multi-strategy test case generation was integrated with LLMs， thereby adjusting prompts to guide the LLM in generating diverse test cases， and the custom code generators were combined to construct test code with specific functions rapidly. Additionally， mutation strategies were incorporated to further enhance test case diversity and improve code coverage， thereby exploring potential compiler execution paths more effectively， so as to trigger deep vulnerabilities. At the same time， an end-to-end automated testing process was designed to fully automate the process from test case generation and compiler execution to bug detection. Experimental results demonstrate that DHFUZZ can generate test cases covering diverse syntactic structures and semantic features， and improves code coverage significantly in tests on mainstream compilers such as GCC （GNU Compiler Collection） and Clang， demonstrating its vulnerability detection potential， and providing an efficient and low-cost automated testing solution for compiler reliability assurance.

Unit test generation method based on path constraint sharding driven large language model

Xiaolong XU, Junfeng WANG, Peng WU, Xiansheng CAO

2026, 46(10):  3298-3306.  DOI: 10.11772/j.issn.1001-9081.2025091085

Asbtract ( )   HTML ( )   PDF (2297KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Automated unit test generation is the key to modern software development to improve development efficiency and ensure software quality. Large Language Models （LLMs） are applied to automated test case generation because of their good code understanding ability. However， when dealing with complex functions， it is difficult for LLMs to cover deep branch paths. Therefore， PyULLM method was proposed， which combined path constraint sharding with LLM generation ability to solve the above problems. Specifically， all the path constraints were collected systematically by preorder traversal of the abstract syntax tree. On this basis， a fine-grained relationship between code lines and path constraints was established. According to the obtained relationship， intelligent partitioning of path constraint set was performed. LLM was able to focus on specific path constraints by this partitioning mechanism， which improved the coverage of generated unit test cases in complex scenarios significantly. Experimental results show that compared with the SOTA（State Of The Arts） tool Pynguin， PyULLM method improves the Line Coverage ratio （LC） by 24.16 percentage points and the Branch Coverage ratio （BC） by 26.61 percentage points； compared with the current SOTA method CodaMosa， PyULLM method has the LC improved by 19.06 percentage points， and the BC increased by 21.72 percentage points. It can be seen that PyULLM method can generate unit test cases for complex functions effectively.

Multimedia computing and computer simulation

Lightweight network based on segmentation guidance and frequency domain enhancement for wheat ear counting

Yuxi LI, Yaying WANG, Shiqin TIAN, Bochuan ZHENG

2026, 46(10):  3307-3316.  DOI: 10.11772/j.issn.1001-9081.2025091115

Asbtract ( )   HTML ( )   PDF (3198KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Convolutional Neural Networks （CNNs） and Transformers based wheat ear counting methods suffer from difficulty in modeling long-range dependencies， low computational efficiency， and frequent false dense positives or false negatives. To address these issues， a wheat ear counting method named PCLMamba was proposed. This network is a lightweight Mamba network based on segmentation guidance and frequency domain enhancement. First， Grouping Directional Mamba Model （GDMM） was introduced to process image sequences in groups and send them to selective scanning in different orders in a parallel way， thereby capturing long-range dependencies efficiently in a lightweight way. Second， the local features and globally scanned features were fused via a Channel Attention Fusion Module （CAFM）. Finally， to solve the problems of dense false positives and negatives， a Segmentation module （Segmenter） was designed to supply prior knowledge， to suppress the background and enhance the foreground. Besides， a Frequency Spatial Enhancement （FSE） module was designed to enhance intermediate features in both spatial and frequency domains， thereby further boosting counting accuracy. Experimental results show that the Mean Absolute Errors （MAEs） of PCLMamba on the MTC， WED， and SHC datasets are 4.4， 4.2， and 15.4， respectively， and the Root Mean Square Errors （RMSEs） are 7.7， 5.3， and 18.3， respectively. On the WED dataset， compared with CountMamba， the MAE and RMSE are reduced by 20.8% and 18.5%， respectively； on the MTC dataset， compared with TasselNet， which has a similar number of parameters， the MAE and RMSE are reduced by 33.3% and 22.2%， respectively. PCLMamba improves the accuracy of wheat ear counting with a relatively small number of parameters， providing a feasible lightweight path for applying state-space models to dense crop counting and offering technical support for automatic field wheat ear counting in smart agriculture.

Gait recognition method based on multi-feature fusion using frequency-modulated continuous wave radar

Jiawei JIANG, Jian GUO, Mingwei YUAN, Zhichun LI, Lijuan SUN

2026, 46(10):  3317-3323.  DOI: 10.11772/j.issn.1001-9081.2025091139

Asbtract ( )   HTML ( )   PDF (927KB) ( )  

Figures and Tables | References | Related Articles | Metrics

A gait recognition method based on multi-feature fusion using Frequency-Modulated Continuous Wave （FMCW） radar was developed to overcome the low recognition accuracy of the existing millimeter-wave point cloud gait recognition methods caused by relying on static spatial features excessively and lacking modeling of dynamic motion information. First， the three-dimensional coordinates of the point cloud were aligned spatially to obtain a unified spatial representation. On this basis， the velocity and signal-to-noise ratio features of target points were fused， and the spatial structure and dynamic information were modeled jointly. Second， cross-spatio-temporal gait features were further extracted by multi-scale convolution and pooling modules. Experimental results show that the proposed method achieves high recognition accuracy on multiple public datasets and a self-built dataset， with an average accuracy of 86.3% and a maximum accuracy of 92.2%. Across different scenarios， it has the recognition accuracy improved by 0.90% to 32.08% compared with comparative methods such as mmGaitNet and PointMLP. It can be seen that through multi-dimensional feature fusion and multi-scale modeling， the proposed method enhances the representation capability of gait dynamic features effectively， and improves the accuracy and robustness of millimeter-wave radar gait recognition.

Blind image super-resolution reconstruction method by integrating degradation estimation and dynamic residual-in-residual dense blocks

Xin ZHANG, Huawei YI, Mengyuan ZHAO, Yanfei WANG, Linchen ZHANG

2026, 46(10):  3324-3331.  DOI: 10.11772/j.issn.1001-9081.2025081031

Asbtract ( )   HTML ( )   PDF (2689KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Blind image super-resolution reconstruction focuses on recovering Low-Resolution （LR） images into clear High-Resolution （HR） images in real-world scenarios. Most existing degradation estimation networks cannot accurately estimate the degradation information of LR images， and the traditional Residual-in-Residual Dense Blocks （RRDBs） in super-resolution networks have poor generalization ability， which fail to adapt to complex degradation types. To address the above issues， a blind image Super-Resolution reconstruction method by Integrating Degradation estimation and Dynamic residual-in-residual dense blocks （IDD-SR） was proposed. First， an integrating degradation estimation network was proposed and used for accurate estimation of image degradation information. Then， a Dynamic Residual-in-Residual Dense Block （DRRDB） was proposed to improve the universality of the super-resolution reconstruction network. Finally， accurate prediction of complex degradation information and high-quality super-resolution image reconstruction were achieved through end-to-end joint optimization. Experimental results on the Set14， BSD100， and DRealSR datasets demonstrate that the Peak Signal-to-Noise Ratio （PSNR） values of 4× super-resolution of IDD-SR are improved by 0.90 dB， 0.71 dB， and 0.97 dB， respectively， compared with those of PDM-SRGAN. It can be seen that the proposed method has superior detail recovery capability and artifact suppression effect.

Small-target-oriented intelligent diagnosis model for cotton leaf diseases

Abudukelimu HALIDANMU, Gengrong ZHANG, Yuhan LIANG, Anran WEI, Abulizi ABUDUKELIMU

2026, 46(10):  3332-3341.  DOI: 10.11772/j.issn.1001-9081.2025091148

Asbtract ( )   HTML ( )   PDF (1138KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Recently， intelligent management of crop diseases has made considerable progress. However， cotton leaf disease still poses significant challenges due to small target sizes and blurred leaf boundaries， and there is a lack of models that integrate “recognition-diagnosis-decision-making.” Therefore， an intelligent diagnosis model for cotton leaf diseases was proposed. First， a dataset containing 1 320 entries of cotton cultivation question and answer was constructed to provide data basis for subsequent fine-tuning of the Tongyi Qianwen large language model. Second， the Multi-Scale Simple Attention Module （MSSimAM） was introduced at the end of the YOLOv12n backbone to highlight multi-scale disease-related features and suppress redundant information. Third， in the detection head， the Adaptive Threshold Focal Loss （ATFL） function was employed to address the imbalance between easy and hard samples， thereby forming the MN-YOLO model. Finally， MN-YOLO model was integrated with the fine-tuned Tongyi Qianwen large language model to generate disease prevention and treatment measure recommendations based on the detection results. Experimental results indicates that MN-YOLO model achieves a mean Average Precision at Intersection over Union （IoU） threshold of 0.5 （mAP₅₀） of 96.9% while maintaining a low parameter count， representing increases of 2.0 percentage points in mAP₅₀， 0.5 percentage points in precision， and 4.5 percentage points in recall compared with YOLOv12n. In generalization experiments， the proposed model also demonstrates strong adaptability and stability across different datasets. The results indicate that this diagnosis model for cotton leaf diseases combines disease detection and treatment measure recommendation， improving detection precision while providing effective support for smart agriculture.

Wheat stripe rust recognition method based on multi-scale feature fusion and attention enhancement

Huimin HUANG, Pengfei WANG, Ying ZHAO, Aihong YUAN, Jin HU, Jianfeng WANG

2026, 46(10):  3342-3353.  DOI: 10.11772/j.issn.1001-9081.2025091184

Asbtract ( )   HTML ( )   PDF (5155KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems of complex backgrounds， significant variations in lesion scale， and subtle differences between adjacent categories in multi-level recognition of wheat stripe rust， a Multi-Scale feature fusion and Attention-enhanced Residual Network （MSA-ResNet） was proposed. First， residual connections were employed to alleviate the gradient vanishing during deep network training and improve the feature representation capability of the model. Second， a top-down multi-scale feature fusion strategy was adopted to enhance the consistency of features at different scales through semantic alignment， thereby improving the model’s adaptability to lesion size variations and morphological complexity. Finally， a Convolutional Block Attention Module （CBAM）， integrating channel and spatial attention mechanisms， was introduced to enhance feature responses in key lesion regions and suppress interference from background elements such as weeds and soil. Experimental results on the Yellow-Rust-19 public dataset and a self-built dataset show that MSA-ResNet achieves the accuracies of 95.67% and 97.17%， respectively， outperforming EfficientNet-CBAM by 1.67 and 0.84 percentage points， respectively. These results indicate that the proposed MSA-ResNet better meets practical application requirements and provides a superior solution for multi-level recognition of wheat stripe rust under field conditions.

Frontier and comprehensive applications

Evolution and prospect of deep learning in meteorological data assimilation

Haiyan LYU, Lijun ZHOU, Jie ZHANG, Yuan ZHAO, Lina WANG

2026, 46(10):  3354-3364.  DOI: 10.11772/j.issn.1001-9081.2025091135

Asbtract ( )   HTML ( )   PDF (834KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Data Assimilation （DA） is the core of modern Numerical Weather Prediction （NWP） systems. However， traditional methods are constrained by bottlenecks such as non-linear and non-Gaussian assumptions as well as high computational cost. Deep Learning （DL）， with its powerful non-linear fitting and end-to-end optimization capabilities， offers a revolutionary paradigm to address these challenges. Therefore， from a computer science perspective， this paper reviewed systematically and analyzed a three-stage evolutionary trajectory for the application of DL in meteorological DA： component replacement， paradigm shift， and system reconfiguration. First， it elaborated on the component replacement stage， in which DL serves as surrogate models to replace observation operators and correct model errors， thereby addressing traditional bottlenecks. Second， it provided an in-depth analysis of the paradigm shift stage， where generative models， represented by Variational Auto Encoders （VAEs） and diffusion models， reconstruct the assimilation problem as a probabilistic generation task. Third， it discussed the system reconfiguration stage， in which Artificial Intelligence （AI） meteorological foundation models such as FengWu and FuXi were adopted， and the assimilation module was embedded deeply， so as to construct a closed-loop system of integrated assimilation-forecast. Finally， the paper explored the core challenges in this field， including physical consistency， data dependency， and operational deployment， and prospected future research directions， such as neuro-symbolic fusion and open benchmark construction.

Research progress and future prospects of multimodal depression detection based on deep learning

Junxiao REN, Xiaoyan LAN, Xuyang SHI, Qiang LI

2026, 46(10):  3365-3374.  DOI: 10.11772/j.issn.1001-9081.2025091147

Asbtract ( )   HTML ( )   PDF (922KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Effective detection of depression is crucial for early identification of potential depression risks. Traditional methods rely on clinical interviews， which are highly subjective and prone to misdiagnosis and delayed treatment. With the development of artificial intelligence， Deep Learning （DL） technology， with its superior feature representation and modeling capabilities， is becoming a core method for depression detection gradually. However， most existing reviews still focus on single-modality or traditional methods， lacking a systematic summary of multimodal DL methods in the field of depression detection， while multimodal data has significant advantages in providing complementary clues for depression. Therefore， typical DL models and multimodal fusion strategies were reviewed systematically， and the core techniques and latest advancements in this field were summarized. First， this paper reviewed the evolution of depression detection methods from single-modality to multimodality， and formed a development trajectory based on datasets and methods. Second， it focused on discussing popular DL models and multimodal fusion strategies in depression detection applications， compared the performance differences of different models and fusion strategies， and reviewed the advantages and limitations of different multimodal fusion techniques. Finally， it summarized the challenges and potential directions for future innovation in this field.

SiamGAN-CNNBiLSTM： respiratory state classification system with WiFi signal-based data augmentation

Rong JI, Tianyu HUANG, Yanlei GE, Xingbin LI, Bo SUN, Yingqi LI

2026, 46(10):  3375-3383.  DOI: 10.11772/j.issn.1001-9081.2025081062

Asbtract ( )   HTML ( )   PDF (1519KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In the non-contact respiratory state monitoring based on WiFi Channel State Information （CSI）， respiratory state data are easily affected by environmental interference during acquisition， resulting in degraded signal quality. Meanwhile， due to constrained acquisition conditions and privacy protection， the number of samples is often insufficient， leading to a few-shot problem， and the generated samples usually lack consistency with real samples in key discriminative features， thereby reducing the effectiveness of data augmentation. In addition， the small differences among different respiratory states further increase the difficulty of classification and limit the depth of research on multi-state respiratory recognition. To address these challenges， an SiamGAN-CNNBiLSTM-based respiratory state recognition system was proposed. First， a multi-state respiratory dataset consisting of normal breathing， apnea， and sneezing was constructed， and a systematic CSI signal preprocessing pipeline was designed to improve signal quality. Second， an SiamGAN module integrating Generative Adversarial Network （GAN） and Siamese network was developed， so that through the adversarial mechanism of generator and discriminator， samples with higher realism were generated. At the same time， a Siamese encoder was introduced to extract 128-dimensional intrinsic features of the signals， and a Mean Squared Error （MSE） constraint was imposed in the feature space to enhance the consistency between generated and real samples in key discriminative features. By this strategy， the discriminability and generalization of generated samples were strengthened， and the physical properties of the original signals were preserved， so as to avoid the introduction of classification-irrelevant high-frequency noise. Finally， a deep classification network combining Convolutional Neural Network （CNN） and Bidirectional Long Short-Term Memory （BiLSTM） network was constructed to fully mine the spatio-temporal features of respiratory signals， and achieve high-precision recognition of multiple respiratory states. Experimental results show that in the office scenario， SiamGAN-CNNBiLSTM achieves excellent performance with an accuracy improvement of 2.2 percentage points， compared with CNN-BiLSTM method， and an accuracy improvement of 8.3 percentage points， compared with the Transformer method； in corridor scenario， SiamGAN-CNNBiLSTM achieves an accuracy of 96.7%， representing a decrease of only 1.6 percentage points compared with that in office scenario， indicating the robustness of the proposed method across different environments.

Real-time dynamic multi-source multi-sink path planning method for indoor fire scenarios

Tao HAN, Litao HAN, Ran JI

2026, 46(10):  3384-3390.  DOI: 10.11772/j.issn.1001-9081.2025091071

Asbtract ( )   HTML ( )   PDF (1644KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Real-time dynamic optimal path planning is critical for facilitating rapid evacuation of personnel in indoor fire scenarios. However， due to the rapid development and high dynamics of indoor fires， the existing evacuation path planning algorithms cannot achieve efficient dynamic path planning in multi-source multi-sink environments. To address this issue， an indoor fire emergency evacuation path planning method based on an improved D^*Lite algorithm was proposed， namely Rectangular Constrained Shared Map D^*Lite （RCSM-D^*Lite） algorithm. First， a search rectangle constraint was introduced to reduce the number of traversed nodes. Second， to account for the rapid spread of fire， a dual-layer search mechanism was adopted to ensure that evacuees maintain a safe distance from hazardous fire regions. Finally， a virtual node was introduced to unify exits， and a shared environment map was incorporated to share information such as map layout， fire hazard levels and other fire dynamics， thereby improving algorithm efficiency. Path planning experiments with 20 persons were carried out on a 40×160 map. The results show that the improved D^*Lite algorithm reduces the time computation by 72.1% compared with the original D^*Lite algorithm. Simulation results under dynamic fire scenarios verify that this algorithm allows evacuees to keep a safe distance from hazardous fire regions and reach exits safely， providing support for indoor emergency evacuation simulation and planning.

Fault warning of feedwater pumps in power plant based on multi-head attention and new gated unit

Jingyang WANG, Hao WANG, Bin ZHANG, Wei SHI, Min HUANG, Yang ZHANG

2026, 46(10):  3391-3399.  DOI: 10.11772/j.issn.1001-9081.2025080937

Asbtract ( )   HTML ( )   PDF (1452KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To achieve early fault warning for feedwater pumps in power plants， a feedwater pump fault warning model， MHA-NGU， based on Multi-Head Attention （MHA） mechanism and New Gated Unit （NGU） was proposed. MHA mechanism was employed to capture long-range dependencies in the time series data， thereby enhancing key dimensions through multi-perspective feature interaction. NGU is an improved Recurrent Neural Network （RNN） proposed in this paper， in which the output gate and reset gate were designed with two levels， thereby avoiding the degradation in global dependency capture capability caused by single?level gates’ over?reliance on local information. Additionally， a T-transformation module was embedded within NGU to further alleviate the gradient vanishing problem of the model. Experimental validation was conducted using data from a 1 030 MW thermal power unit in Zhejiang Province as an example. The results demonstrate that the proposed model increases the Coefficient of Determination （R2） by 4.38% and 1.90% compared with GRU and the second-best model， MHA-GRU， respectively. Furthermore， it triggers a warning 5.39 hours earlier than the GRU model， realizing an early fault warning for feedwater pumps effectively.

Fault diagnosis method integrating dual-threshold preliminary screening and edge-level dynamic graph convolution

Wanxiong MIN, Yicheng YAN, Jiuyin WANG, Yukuan LIANG, Wenyang LIANG, Zuozhou PAN, Yong LIAO

2026, 46(10):  3400-3406.  DOI: 10.11772/j.issn.1001-9081.2025091114

Asbtract ( )   HTML ( )   PDF (977KB) ( )  

Figures and Tables | References | Related Articles | Metrics

During operation， power transformers emit vibration signals continuously， which contain many pulses and fluctuations caused by mechanical faults. These signals serve as the primary data source for assessing the operating condition of power transformers. To address the issues of low diagnostic accuracy， poor real-time performance， and inability to handle multi-fault scenarios effectively in the traditional power transformer fault diagnosis techniques， the voiceprint characteristics of internal vibration acoustic signals in power transformers were analyzed， and a fault diagnosis method integrating dual-threshold preliminary screening and edge-level dynamic graph convolution was constructed on the basis of acoustic features. First， a normal/fault signal classification module was proposed to assess the real-time operating condition of power transformers using acoustic short-time energy and zero-crossing rate. Second， an Edge-Level Dynamic Graph Convolutional Network （ELDGCN） was designed to learn spatial correlations among voiceprint signals adaptively by optimizing edge weights. Finally， experiments were conducted under real-world power transformer scenarios. The results show that the proposed method improves the fault identification efficiency significantly， with the identification rate for datasets containing mixed faults reaching 98.17%.