Computer and Modernization

An Active Distribution Network Dynamic Fault Recovery Method Based on Rolling Optimization

YU Zhiwen, ZHAO Ruifeng, LAN Tian, LI Qian, LI Haobin

2025, 0(08): 1-9. doi:10.3969/j.issn.1006-2475.2025.08.001

Asbtract ( 88 )

PDF (1653KB) ( 117 )

Related Articles | Metrics

Abstract: To fully harness the emergency power supply potential of distributed photovoltaics in the fault recovery stage of the distribution network， mitigate the impact of uncertainty in source-load output on the reliability of control schemes， this paper proposes an active distribution network dynamic fault recovery method based on a rolling optimization framework. Initially， employing the RTH-CNN-LSTM algorithm for short-term forecasting of distributed photovoltaic output to obtain predicted output during the fault recovery phase. Subsequently， considering the time required for fault line inspection， material constraints， and operational safety constraints of nodes， voltages， and power flow in the distribution network during the fault recovery phase， the fault recovery model for the distribution network is constructed from the perspectives of load recovery rate， strategic economics， and system reliability， integrating distributed photovoltaic forecast output， switch states， and network topology. Lastly， utilizing the rolling optimization framework， the Red-Tailed Hawk （RTH） optimization algorithm is employed to dynamically solve the fault recovery model， obtaining the optimal action strategy for line maintenance sequence， switch status， and load shedding amount. Using the IEEE 123-node test distribution network as a case study， simulation results demonstrate that the proposed rolling optimization method effectively reduces the impact of distributed photovoltaic output uncertainty on the effectiveness of recovery schemes， significantly enhancing the reliability and efficiency of fault recovery schemes.

Chinese Entity Relation Joint Extraction Method Based on Deep Learning

WEI Huimin1, 2, ZHOU Jiake1, 2, WEN Yongjun1, 2, TANG Lijun1, 2

2025, 0(08): 10-15. doi:10.3969/j.issn.1006-2475.2025.08.002

Asbtract ( 99 )

PDF (1440KB) ( 101 )

References | Related Articles | Metrics

Abstract: Entity-relationship extraction is an important part of artificial intelligence technologies such as building knowledge graphs and improving search engine efficiency. Due to the complexity， ambiguity， and implicit nature of Chinese text composition， the process of Chinese entity relationship extraction is prone to entity overlapping， entity nesting， and information redundancy. Therefore， this paper proposes a deep learning-based joint extraction model of Chinese entity relations（SRGP). The model firstly encodes the input text， obtains the set of specific relations through the specific relation prediction network， fuses the set of specific relations with the input text into the entity recognition module through the attention mechanism， and reduces the redundant computation in the extraction of Chinese entity relations. For the problems of insufficient extraction of overlapping entities and inaccurate recognition of nested entities， the global pointer annotation strategy based on specific relations is proposed by utilizing the idea of global normalization under the constraints of a specific set of relations. Two general Chinese datasets， DUIE1.0 and CMeIE， are selected respectively， and this paper’s model， SRGP， is compared with the typical models of entity-relationship joint extraction， such as CopyRE， PRGC， and CasRel， for the comparison experiments， and the experimental results show that this paper’s model achieves F1 values of 61.3% and 80.1% on the two datasets， which are respectively 1.5 and 2.2 percentage points higher than those of the best-performing baseline models CasRel and PRGC.

Goal Driven Recommendation-oriented Dialog Generation Method

JING Qingwu, CHEN Hongjun, GAO Di, ZHOU Meimei

2025, 0(08): 16-23. doi:10.3969/j.issn.1006-2475.2025.08.003

Asbtract ( 87 )

PDF (2250KB) ( 90 )

References | Related Articles | Metrics

Abstract: The task of recommendation-oriented dialog generation aims to achieve accurate recommendations by obtaining user preferences through human-computer dialog interactions. In response to the problem of limited dialog recommendation types and low quality of generated replies in existing research， this paper proposes a Goal Driven Recommendation-oriented Dialog Generation model （GDRDG） based on the Unified Language Model pre-training （UniLM）. The model comprises a text representation module， a multi-head encoding module， a decoding module， and a specialized attention masking mechanism. The text representation module uses UniLM to vectorize the input text， ensuring that the model captures deep semantic features of the text. The multi-head encoding module employs a multi-head self-attention mechanism to capture global contextual information， enhancing the coherence and relevance of the generated responses. The decoding module generates the target of the current dialogue round and the response based on this target， ensuring that the reply is consistent with the context and guides the conversation towards the intended goal. The special attention masking mechanism is used to control the information flow during the decoding process， ensuring that the model focuses only on information relevant to the current round， thereby improving the quality of the response. Experimental results demonstrate that the proposed GDRDG model outperforms existing methods in metrics such as BLEU， Distinct， F1， and Hit@1， thereby validating the model’s effectiveness and advancement.

Cross-domain Book Recommendation Method Based on Heterogeneous Information Network

SHI Fengyuan1, 2, MAO Yi3, JIAO Lei4

2025, 0(08): 24-30. doi:10.3969/j.issn.1006-2475.2025.08.004

Asbtract ( 90 )

PDF (2568KB) ( 97 )

References | Related Articles | Metrics

Abstract: Aiming at the problem that the accuracy of the recommendation algorithm is reduced due to data sparsity and cold start in the learning data domain in the current student book recommendation model， a cross-domain recommendation method HeterCDR （Heter Cross-Domain Recommendation） based on heterogeneous information networks is proposed. The modeling of source domain information is realized by introducing the translation distance model to construct a heterogeneous information network， and the DANN model is used to realize the migration of source domain information to the target domain. The heterogeneous information network and cross-domain recommendation are combined to improve the accuracy of target domain recommendation. The experimental data set uses the relevant data of students from grade 20 to grade 21 of a higher vocational college. The experimental results show that compared with other recommendation models， the hit rate of the HeterCDR model is improved by about 3.35% on average， the NDCG index is improved by about 2.8%， and the RMSE index is reduced by about 2.65%.

Discovery of Fine-grained Subjective Perception of Urban Blocks in User Comment Data

2025, 0(08): 31-38. doi:10.3969/j.issn.1006-2475.2025.08.005

Asbtract ( 72 )

PDF (2313KB) ( 68 )

References | Related Articles | Metrics

（1. School of Computer Science and Engineering， Shenyang Jianzhu University， Shenyang 110168， China；
2. Liaoning Provincial Key Laboratory of Big Data Management and Analysis for Urban Construction， Shenyang 110168， China；
3. Shenyang Branch of National Special Computer Engineering Technology Research Center， Shenyang 110168， China）
Abstract： As one of the important dimensions for evaluating urban construction and planning， the subjective perception of urban blocks will help to create a more humane and livable urban space. Based on the representation learning technology and combined with the content of user comments， this paper discovers the structural characteristics between the neighborhood and the subjective perception， and solves the problems of coarse granularity and lack of data in the existing neighborhood perception. Firstly， a fine-grained perception category system is proposed， which analyzes the perception dictionary and POI （Points of Interest） user comment data， analyzes the subject words by LDA and combined with hierarchical clustering to construct a fine-grained street perception category. Secondly， in order to solve the problem of incomplete data， a similarity-weighted k-nearest neighbor filling method is designed， which effectively supplementes the missing POI evaluation content through the information of large category， small category， and geographical location. Finally， the autoencoder is used to transform the neighborhood perception into potential feature vectors. The real dataset of Beijing is used to evaluate the grading and ranking of housing prices in the neighborhood， and the effectiveness of the proposed method is verified.

Anomalous Event Prediction Approach Using Graph Neural Network

LI Yan, MAO Jiaming, WANG Ziying, GU Zhimin, JIANG Haitao

2025, 0(08): 39-47. doi:10.3969/j.issn.1006-2475.2025.08.006

Asbtract ( 96 )

PDF (2546KB) ( 92 )

References | Related Articles | Metrics

Abstract: Log-oriented anomalous event prediction provides important support for security diagnosis， intelligent operation and maintenance of complex systems. Most of the existing mainstream anomalous event prediction methods based on deep learning technology capture sequence features from the local perspective of event sequence segments and the feature types are relatively single， resulting in low prediction accuracy. To solve this problem， an anomalous event prediction method based on graph neural network is proposed. The log sequence is represented as a graph structure with log events as nodes and the relationship between events as edges， so that it can simultaneously depict the log sequence from the perspective of semantics， statistics and event relationship to capture its spatio-temporal dynamic characteristics to improve the prediction performance. On this basis， the anomaly prediction task is transformed into a graph classification problem， and an anomaly prediction model based on graph isomorphic network is established by training graph neural network， which can more accurately capture the difference between the log sequence before the failure and the log sequence under normal conditions， and further improve the performance of anomaly prediction. The experimental results on three benchmark datasets show that the average F1 of the proposed method is 0.958， which is better than the comparison methods， and it can accurately predict anomalous events for early warning.

A Intrusion Detection Method Based on Imbalanced Power Communication Traffic

XIE Shanyi1, 2, WANG Zhongao1, 2, ZHAN Congcong1, 2, LI Xingwang1, 2, XIA Haoran3

2025, 0(08): 48-56. doi:10.3969/j.issn.1006-2475.2025.08.007

Asbtract ( 68 )

PDF (2880KB) ( 70 )

References | Related Articles | Metrics

Abstract: With the proliferation of the Internet， cybersecurity issues have become increasingly prominent. Ensuring network security is crucial within electric power communication networks. However， one challenge faced by these networks is the disparity in the volume between normal and abnormal traffic， as well as the uneven distribution among different types of abnormal traffic. To address this issue， this paper proposes an intrusion detection method for imbalanced electric power communication traffic， named GSMOTE-EAVA. GSMOTE-EAVA firstly utilizes Recursive Feature Elimination for data preprocessing and feature selection by calculating the importance of features to identify the most critical ones. Secondly， to tackle the challenge of data imbalance， a Gaussian noise-based SMOTE algorithm is employed to augment the communication traffic data， thus enhancing the neural network model’s ability to learn and adapt to various situations. Finally， an ensemble adaptive voting algorithm based on classifiers like decision trees， random forests， KNN， and DNN is designed to implement intrusion detection in electric power communication network traffic. Through experiments on the IEC 60870-5-104 intrusion detection dataset and CICIDS2017 dataset， the proposed model significantly improves the detection rate of small sample categories in the dataset under four classifications， and can effectively identify and deal with abnormal traffic in the power communication network.

Malicious Code Homology Analysis Method Based on ATT&CK Framework and Bert Model

ZHENG Xiaoyu, LIN Jiuchuan, CHEN Wenxuan, YAO Xinyu

2025, 0(08): 57-62. doi:10.3969/j.issn.1006-2475.2025.08.008

Asbtract ( 78 )

PDF (2355KB) ( 139 )

References | Related Articles | Metrics

Abstract: At present， malware attacks are one of the main threats to cyberspace security. By analyzing malicious programs from known organizations and determining the homology of unknown malicious programs based on similar characteristics， it is helpful to identify unknown malicious programs and attribution attack organizations. However， the existing homology analysis models have some problems， such as high complexity of manual feature extraction， inadaptability to large-scale analysis scenarios， low efficiency， and lack of in-depth consideration of the transmission relationship between attack behaviors. This paper proposes a homology recognition model based on the ATT&CK framework and the Bert（bidirectional encoder representation from transformers） model， which solves the problem of low homology recognition accuracy caused by code confusion and polymorphism in the face of static features through high-dimensional attack techniques and tactics in the ATT&CK framework. The Bert model is used to effectively integrate the multi-dimensional features of malicious code， and solve the problem of insufficient sequence modeling by recurrent neural network-based analysis methods. Experimental results show that the proposed scheme can effectively identify the homology between malicious codes.

A Data Privacy Protection Scheme Based on Federated Learning

CHENG Yuwen, JING Yijun, SHI Zicheng, JING Changqiang, GUO Feng, WU Chuankun

2025, 0(08): 63-69. doi:10.3969/j.issn.1006-2475.2025.08.009

Asbtract ( 85 )

PDF (3019KB) ( 88 )

References | Related Articles | Metrics

Abstract: The current healthcare data domain faces the issue of data silos， which restricts the flow and sharing of data among different institutions and hinders cross-institutional treatment for patients. To address this problem， this paper proposes a privacy protection scheme based on federated learning （Federated Learning with Schnorr Zero-knowledge Based Identity Authentication and Differential Privacy Protection， FL-SZIDP）. Firstly， a data-sharing framework based on federated learning is established. Secondly， to defend against adversaries attempting to steal original data through reverse attacks， differential privacy noise is added to the model parameters uploaded by each participant. To prevent malicious participants from joining the federated learning process， identity authentication based on Schnorr zero-knowledge proof is performed， ensuring the credibility of the participants’ identities. Finally， the effectiveness of the proposed algorithm is verified using the MNIST data set. The experimental results show that the scheme FL-SZIDP ensures accuracy while protecting privacy.

Machine-vision based Geometric and Color-difference Inspection System for Solid Wood Floor

LIANG Zhantao1, HUANG Guohao1, HUANG Chengzi1, YIN Jianxin3, GAO Xingyu2, HUANG Yang1

2025, 0(08): 70-75. doi:10.3969/j.issn.1006-2475.2025.08.010

Asbtract ( 72 )

PDF (2959KB) ( 67 )

References | Related Articles | Metrics

Abstract: With the naturally grown colors and textures， the solid wood floor is an excellent material for interior decoration. In order to meet the artistic effects required for specific interior decoration， the geometry and of solid wood floor need to be precisely inspected and colors need to be well-coordinated. So， it is necessary to measure the geometrical dimensions of solid wood floor and classify the colors during the manufacturing process to meet the personalized needs of customers. However， the traditional manual method is constrained by factors such as labor intensity， work efficiency， and detection objectivity. It is difficult to meet the growing demand for industrial automation and intelligent development. In this paper， the geometric size and color-difference is detected by a developed machine-vision based system for solid wood floor. The algorithm for measuring the size and the color-difference is optimized. In the geometric dimension measurement， the influence of the random attitude of the wooden floor on the measurement accuracy during the production process is considered. The 3D attitude of the wooden floor is sensed and corrected in real time by using the ArUco （Augmented Reality University of Cordoba） code， which improves the accuracy of the dimension measurement； the gradient autocorrelation operator is used in color-difference identification and integrated to remove the wood texture. The color-difference is defined in Lab color space. The proposed method is verified and validated experimentally. It enables the detection of under-machined and over-machined contours using a benchmark plate. The relative accuracy of dimensional detection is about 0.8%， and the color-difference of the wooden floor is well characterized and identified.

Improved Classroom Behavior Detection Algorithm for YOLOv8

SU Yansen, MOU Li

2025, 0(08): 76-81. doi:10.3969/j.issn.1006-2475.2025.08.011

Asbtract ( 92 )

PDF (1863KB) ( 96 )

References | Related Articles | Metrics

Abstract: Aiming at the problems of low detection accuracy of student classroom behavior under monitoring and difficulty in deploying models， an improved YOLOv8 algorithm is proposed for detecting student behavior. Firstly， the YOLOv8 backbone network is improved by introducing the Swin Transformer network as the backbone feature extraction network to reduce information loss and improve the effectiveness of feature extraction. Secondly， to enhance the model’s attention to the features of distant targets， a flexible dual channel attention mechanism EMA is introduced， which makes the model focus more on targets with fewer pixels at long distances and improves detection accuracy. Finally， in the Neck section， the Slim Neck design paradigm containing GSConv is used to make lightweight improvements to the model. The experimental results on the SCB-Dataset3 dataset show that the improved model has a parameter count of 3.3 M and a computational load of 11.1 GFLOPs， respectively， with a detection accuracy of 88.75%. Compared with the original model， the parameter count is reduced by 40.7%， the computational load is reduced by 15.9%， and the detection accuracy is improved by 7.7 percentage points. This achieves good detection accuracy while achieving model lightweighting.

Dual-backbone Network Insulator Defect Detection Algorithm Based on YOLOv5

GAO Lisha1, PENG Guozheng2, WANG Chong1, HAN Shuo1, XIANG Nan1, ZHANG Qizhe2, ZHU Yuefei3, CHENG Xu3

2025, 0(08): 82-88. doi:10.3969/j.issn.1006-2475.2025.08.012

Asbtract ( 86 )

PDF (3290KB) ( 96 )

References | Related Articles | Metrics

Abstract： The current insulator defect detection algorithm has problems such as insufficient feature extraction， poor small target detection effect， and unbalanced overall algorithm performance. To address the above problems， this paper proposes a dual-trunk network insulator defect detection algorithm based on YOLOv5. Firstly， this paper designs a new dual-trunk feature extraction network GELAN-Ghost suitable for YOLOv5. This module can more fully extract feature information while maintaining lightweight. Secondly， using the idea of inverted residual structure， a plug-and-play efficient multi-scale attention module iEMA is designed in the neck network part of the algorithm. Finally， a new dynamic detection head DynamicHeadv3 is designed to replace the original detection head to extract richer features and enhance the perception ability of the model. The experimental results show that the improved model has an accuracy improvement of 1.4 percentage points， a parameter amount and a computational amount reduced by 46% and 33% respectively， and the detection speed has also been improved to a certain extent. The performance has achieved a good balance， which is more in line with the needs of drone and edge insulator defect detection.

Classification and Diagnosis of Breast Cancer Histopathological Images Based on Transfer Learning

LI Jia

2025, 0(08): 89-96. doi:10.3969/j.issn.1006-2475.2025.08.013

Asbtract ( 85 )

PDF (1980KB) ( 86 )

References | Related Articles | Metrics

Abstract: Breast cancer is one of the diseases that significantly affects women’s health， and there is currently no substantial cure for it. In recent years， with the rapid development of artificial intelligence technology， especially the potential advantages of deep learning in requiring less manual intervention during image feature extraction， this technology used in breast cancer detection to facilitate early diagnosis and personalized treatment plans， thereby improving patient survival chances. This paper selects the large publicly available breast cancer histopathological image dataset BreakHis to study breast cancer detection methods based on deep convolutional neural networks， including the application of three pre-trained models: GoogLeNet， ResNet50， and EfficientNet， constructing transfer learning techniques for both binary and multi-class classification. Experimental results show that for binary classification， the validation accuracies achieved by the above three networks are 97.01%， 97.35%， and 97.13% respectively; for multi-class classification， the validation accuracies are 87.78%， 92.33%， and 93.59% respectively， with ResNet50 showing relatively better classification performance. Finally， the challenges and future research directions of breast cancer detection based on deep learning models are analyzed.

Optimizing Wavelet Neural Network for Pure Milk Recognition using Improved Genetic Algorithm

HU Shaowen, HUANG Langxin, YU Lihui, SHI Weili, JIANG Nan, SUN Wu, LUO Shuhuan

2025, 0(08): 97-103. doi:10.3969/j.issn.1006-2475.2025.08.014

Asbtract ( 59 )

PDF (2859KB) ( 81 )

References | Related Articles | Metrics

Abstract： To overcome the difficulty of visually distinguishing pure milk types and non-destructive testing of pure milk， an improved genetic algorithm-optimized wavelet neural network pure milk recognition algorithm is proposed， which can effectively ameliorate the accuracy and recognition efficiency of traditional wavelet neural network recognition algorithms. Firstly， a genetic algorithm is employed to the traditional wavelet neural network recognition algorithm， which is utilized to optimize the weight， threshold， and wavelet basis function translation and contraction factor parameters in the wavelet neural network to improve the accuracy of the recognition algorithm. In addition， a cyclic perturbation strategy has been adopted to the algorithm， greatly reducing the number of iterations required for optimal results， thereby improving the recognition efficiency of the algorithm. In the algorithm experiment section of this article， 200 sets of pure milk samples of the same brand but different types are selected as experimental samples， and near-infrared spectroscopy technology is used to obtain absorbance data of all milk samples in the wavelength range of 4000～10000 cm−1. Subsequently， to improve the training efficiency of milk data， principal component analysis algorithm is utilized to extract feature data with high cumulative contribution rate. The extracted principal component feature data are preliminarily trained and tested utilizing the proposed algorithm. The experimental results show that adding genetic algorithm could improve the accuracy from 97.5% to 100%. After adding the cyclic perturbation strategy， the number of training iterations could be greatly reduced， and the convergence speed of the algorithm could be greatly improved. Therefore， the pure milk recognition algorithm proposed in this article can effectively and non destructively distinguish pure milk.

Two Level Double Association-Based Evolutionary Algorithm for Multi-Objective Optimization

WANG Rongchen, LIU Junhua

2025, 0(08): 104-114. doi:10.3969/j.issn.1006-2475.2025.08.015

Asbtract ( 74 )

PDF (1548KB) ( 67 )

References | Related Articles | Metrics

Abstract: In this paper， Two level Double Association-based Evolutionary Algorithm for multi-objective optimization（T-DAEA） is proposed to solve the curse of dimensionality in multi-objective optimization problem as the number of objectives increases. The proposed angle-based bi-association strategy considers the empty subspace and associates it with the solution with the highest fitness， increasing the probability of exploring unknown regions. In addition， a new quality assessment scheme is designed to quantify the quality of each solution in the subspace by first measuring the convergence and diversity of each solution， then designing dynamic penalty coefficients to balance the convergence and diversity by penalizing the global diversity distribution of the solutions， and performing an adaptive hierarchical ordering of the solutions that have completed the assessment to ensure the selection of the best solution. The performance of the proposed algorithm， T-DAEA， is validated and compared with four state-of-the-art multi-objective evolutionary algorithms on a number of well known benchmark problems （up to 20 objectives）. Experimental results show that the algorithm is highly competitive in terms of both convergence enhancement and diversity maintenance.

5G Terminal Signaling Outlier Detection Algorithm Based on Joint Distance and Density

WEI Xiaogang1, MEI Wenming2, HU Youjun1, TU Zhengwei1, WANG Rui3

2025, 0(08): 115-118. doi:10.3969/j.issn.1006-2475.2025.08.016

Asbtract ( 61 )

PDF (1305KB) ( 75 )

References | Related Articles | Metrics

Abstract: In order to quickly and accurately detect the abnormal behavior of 5G terminal signaling， this paper proposes a 5G terminal signaling outlier detection algorithm based on joint distance and density. Initially， preprocessing is applied to the 5G terminal signaling data collected by software to reduce the dimensionality of the data samples. When the number of data samples is less than 1000 after dimensionality reduction， the KNN outlier detection algorithm based on average distance is used to detect the outlier points of the data samples； when the number of data samples is greater than or equal to 1000， the COF outlier detection algorithm based on density is used to detect the outlier points. Simulation results show that the algorithm proposed in this paper outperforms the traditional benchmark algorithm in terms of accuracy， precision， and recall， and has a faster response time.

Research on Multi-objective Coordinated Control Method of User-side Energy Storage Based on Harmony System

LI Gang, SUN Shaobin, KONG Wei, ZHAO Lei

2025, 0(08): 119-126. doi:10.3969/j.issn.1006-2475.2025.08.017

Asbtract ( 72 )

PDF (2731KB) ( 63 )

References | Related Articles | Metrics

Abstract: In the context of a high proportion of new energy sources， optimizing the cooperative operation of energy storage systems can enhance the overall economic and environmental benefits and contribute to sustainable development. This paper proposes a multi-objective coordinated control strategy for the problem of a single optimization scenario of user-side energy storage systems. The distributed coordinated control of user-side energy storage systems is realized by improving the adaptive step-size method of the alternating direction multiplier method （ADMM）. Firstly， a distributed framework of user-side energy storage based on Harmony systems is constructed， and a coordinated control strategy for these systems is designed. Subsequently， a multi-objective optimization model encompassing the costs associated with electrical energy usage， environmental impact， and energy storage operation is developed. To address the issue of disparate convergence rates observed during multi-objective optimization iterations， an adaptive step-size distributed iterative optimization method based on ADMM is employed. This approach effectively enhances operational efficiency while safeguarding the privacy of the user-side energy storage system. The experimental results demonstrate that the proposed distributed framework significantly enhances the operational efficiency， while the multi-objective coordinated control strategy markedly enhances the system gain.

Table of Content