Computer and Modernization

Entity-integrated Summarization Model Based on Improved Graph2Seq

TAO Yuan, QIAN Huimin

2025, 0(06): 1-8. doi:10.3969/j.issn.1006-2475.2025.06.001

Asbtract ( 125 )

PDF (762KB) ( 132 )

References | Related Articles | Metrics

Abstract: In order to address the issues caused by high computational resource consumption and limited attention on key named entities， a novel summarization model named Entity-Sparse-Attention Graph-to-Sequence （ESG2S）， based on the Graph2Seq model， is proposed in this paper. Firstly， a graph data is created from a syntactic dependency graph enhanced by the extracted entity nodes from the original text. Secondly， this graph data is then input into an encoder to learn the textual structure. Finally， the encoded graph data is fed into an LSTM decoder integrated with Symmetric Divergence-Enhanced Sparse Attention to generate multiple summaries. Experiments on the CNN/DM dataset show that this model outperforms several recent mainstream methods and is effective in preserving entity information， resulting in summaries with better readability and comprehensiveness.

Student Classroom Behavior Detection Based on Improved YOLOv7-tiny Algorithm

BAI Jia, GU Lin

2025, 0(06): 9-15. doi:10.3969/j.issn.1006-2475.2025.06.002

Asbtract ( 91 )

PDF (934KB) ( 147 )

References | Related Articles | Metrics

Abstract: In response to the challenges of small and densely packed targets， as well as recognition performance affected by various occlusions and lighting interferences in student classroom behavior detection， this paper proposes an improved algorithm for student classroom behavior detection based on YOLOv7-tiny， termed YOLOv7-tiny-SCBD. Firstly， a convolutional block attention module is introduced to extract more feature information in both channel and spatial dimensions， improving the model's recognition performance. Secondly， by utilizing content-aware feature recombination and upsampling operators， the network's receptive field is enlarged， and upsampling is dynamically performed according to the input feature map information to obtain clearer and more realistic upsampling results. Finally， the Inner-MPDIoU loss function is employed to comprehensively and accurately evaluate the similarity between predicted boxes and ground truth boxes， thereby enhancing the model's detection accuracy. Experimental results demonstrate that the proposed algorithm achieves a mAP@0.5 of 81.6% on the SCB3-U dataset， which is a 3.1 percentage points improvement over the original YOLOv7-tiny algorithm. With 6.19 M parameters， 13.5 calculations， and a detection frame rate of 85.39 frames per second， can effectively detect student classroom behavior.

Privacy-preservation Models for Identifying Financial Fraud in Enterprises

XUE Wenxue, XIE Yingde

2025, 0(06): 16-20. doi:10.3969/j.issn.1006-2475.2025.06.003

Asbtract ( 97 )

PDF (755KB) ( 126 )

References | Related Articles | Metrics

Abstract: To avoid legal and economic barriers to sensitive data and effectively identify financial fraud in enterprises， a machine learning based financial fraud identification model was constructed using the concept of privacy protection. Based on the financial and non-financial information of 19，334 samples from 2012 to 2021， the Internet information was introduced to build a privacy-preservation oriented identification model of enterprise financial fraud （Hetero-SBoost and Hetero-NN）. The results show that the proposed model after the introduction of Internet information， the performance of the optimized models were 7%~10% higher than that of the traditional models， indicating that the introduction of Internet information helps to improve the recognition effect and further optimize the model on the basis of compliance. To further verify the accuracy of that the proposed model in practical applications， a comparison was made between 3，452 samples from 12 companies in Shandong and the results of advanced models （DeepProtect， Starlite）. The results indicate that Hetero-SBoost ensures the overall performance of the model and has better robustness. This paper completes the financial fraud recognition modeling without disclosing data. The introduction of Internet information and privacy protection verifies the effectiveness of the identification model.

Hierarchical Classification Algorithm for Marine Organisms

XIE Peidong, CHENG Yuanzhi, XU Haotian

2025, 0(06): 21-27.

Asbtract ( 65 )

PDF (732KB) ( 84 )

References | Related Articles | Metrics

Abstract: The vast number of marine organisms，coupled with the high degree of morphological similarity among different species，poses challenges for their identification and classification.Current research methods primarily utilize convolutional neural networks and self-attention mechanisms to extract features and directly perform classification. However，this approach overlooks the potential hierarchical structure that may exist among categories.To address this issue， a novel hierarchical classification algorithm is proposed， which integrates convolution and self-attention mechanisms.This method fully exploits the advantages of convolution in capturing local features at shallow layers and self-attention in capturing global features at deeper layers，naturally combining the two. Additionally，based on prior biological knowledge，we construct a hierarchical structure among marine organism categories and create branches at both deep and shallow levels to utilize hierarchical relationships for predictions ranging from coarse to fine categories.To further enhance the interaction between deep and shallow layer information，we propose a dynamic connectivity pattern， enabling the network to obtain information of different granularity across different hierarchical levels. Finally，we introduce a category relation enhancement module at the end of the network to assist the network in learning hierarchical semantic relationships，thereby achieving more accurate classification. Experimental results demonstrate that the proposed algorithm outperforms existing classification methods.

Multi-Channel Speech Separation Method with Self-guided Transformer

TAN Yingwei, DING Xuefeng

2025, 0(06): 28-33. doi:10.3969/j.issn.1006-2475.2025.06.005

Asbtract ( 90 )

PDF (979KB) ( 88 )

References | Related Articles | Metrics

Abstract: In the field of speech processing， multi-channel speech separation technology aims to effectively separate the speech signals of different speakers from multi-channel mixed speech. However， existing methods have shortcomings in handling long-distance dependency relationships between multi-channel feature points. In response to this issue， this study proposes a novel multi-channel speech separation method based on self-guided Transformer （SG-former）， which aims to construct an adaptive fine-grained global attention mechanism. The core mechanism of SG-former is to reassign tokens through saliency maps. Under this framework， significant regions can extract key information at a fine-grained level， while secondary regions adopt a coarse-grained extraction approach to reduce computational costs. The generation of saliency maps relies on the hybrid-scale self-attention mechanism， which can accurately capture the long-distance dependency relationships between multi-channel feature points. To verify the effectiveness of the proposed method， experiments were conducted on a spatialized WSJ0-2MIX database. The experimental results show that the SG-former method has a significant advantage in Signal to Distortion Ratio Improvement （SDRi） compared to the baseline Beam Guided TasNet method， achieving a 20.34 dB improvement. This result fully demonstrates the superiority of SG-former in dealing with multi-channel speech separation problems， especially in establishing long-distance dependency relationships. The experimental results show that this method outperforms existing technologies in terms of performance， providing new ideas and methods for research in the field of multi-channel speech separation.

Optimized Scheduling of Microgrid Based on Improved Harris Eagle Algorithm

DUAN Hongjin, ZHANG Yi

2025, 0(06): 34-41. doi:10.3969/j.issn.1006-2475.2025.06.006

Asbtract ( 87 )

PDF (1361KB) ( 79 )

References | Related Articles | Metrics

Abstract： In order to reduce the operating cost of microgrid and solve the problem that the traditional Harris Eagle algorithm is easy to fall into local optimum， a new Harris Eagle algorithm based on refractive imaging inverse learning strategy and hybrid butterfly algorithm was proposed， and the improved algorithm was simulated through 11 test functions. Then， the improved Harris Eagle algorithm is used to solve the problem of island optimal scheduling of microgrids， and the simulation results show that the improved Harris Eagle algorithm can reduce the operating cost of microgrids.

Optimization Configuration Method of Urban Power Grid Reactive Power Compensation Devices Based on Improved AOA Algorithm

QIAO Rongfe, DONG Xin, KAN Changtao, LI Guang, TAO Qizuo

2025, 0(06): 42-50. doi:10.3969/j.issn.1006-2475.2025.06.007

Asbtract ( 82 )

PDF (1219KB) ( 73 )

References | Related Articles | Metrics

Abstract: With the mass access of rail transit and new energy in urban power grid， the active and reactive power transmission of urban power grid is serious， resulting in frequent voltage exceedance of urban power grid， and bringing major hidden dangers to the safe and stable operation of power grid. This paper presents an optimal allocation method of reactive power compensation equipment in urban power grid based on improved arithmetic optimization algorithm. Firstly， a comprehensive index of voltage stability is constructed by using reactive voltage sensitivity index and local voltage stability index to determine the position of reactive power compensation points. Secondly， the capacity optimization strategy of multi-objective reactive power compensation equipment is proposed， which takes into account the network active power loss， voltage deviation and equipment investment cost. Based on the traditional Arithmetic Optimization Algorithm （AOA）， improvements such as Kent chaos mapping， compound cycloidal method， Sparrow elite variation and Cauchy variation were introduced to improve the convergence speed and accuracy of the algorithm. Finally， an improved arithmetic optimization algorithm is used to solve the optimal configuration of reactive power compensation equipment. In this paper， IEEE30 nodes are used as simulation computing power for analysis， and the experimental results verify the effectiveness and rationality of the proposed method， which can improve the rationality of reactive power compensation equipment allocation in urban power grids.

Multi-UAV Path Planning Based on Improved Fireworks Algorithm

YANG Jin, CHEN Buqian

2025, 0(06): 51-55. doi:10.3969/j.issn.1006-2475.2025.06.008

Asbtract ( 74 )

PDF (789KB) ( 78 )

References | Related Articles | Metrics

Abstract: In this paper， the path planning of multiple UAVs in complex environment is studied， and the objective function satisfying the constraints is designed. At the same time， aiming at the shortcomings of fireworks algorithm in global convergence and local convergence， an improved fireworks algorithm is proposed. In the process of generating variation spark， Levy variation is used instead of Gaussian variation to generate variation spark， and the problem that the algorithm is easy to fall into local optimal at the origin is avoided. In addition， in order to improve the locality of the fireworks algorithm and the information exchange between individuals， a deep information exchange strategy is introduced to select the next generation of sparks. By comparing the improved fireworks algorithm with other intelligent optimization algorithms， the simulation results show that the improved fireworks algorithm has significant advantages in convergence speed and stability， and can better carry out multi-UAV path planning.

Cyberspace Situational Awareness Method Based on Traffic Volume

KUANG Ye1, ZHOU Mo2, LIU Ceyue1

2025, 0(06): 56-60. doi:10.3969/j.issn.1006-2475.2025.06.009

Asbtract ( 60 )

PDF (1361KB) ( 128 )

References | Related Articles | Metrics

Abstract: Network anomalies can seriously influence the normal operation of infrastructure and cause huge financial losses. Existing studies perceived the changes of network state by observing the changes of network traffic. However， due to the lack of quantitative indicators， these cyberspace situational awareness methods have the problems of high false positive rate and low accuracy. This paper proposes a new cyberspace situational awareness method to address these challenges. Specifically， we use the public data set to obtain the traffic information of the links in the detection area. Firstly， a traffic critical value determination algorithm based on Wilson score is proposed to calculate the corresponding traffic threshold of each link. Secondly， a network state awareness scheme is proposed， which perceives the change of network state by monitoring whether the traffic volume of any link in the detection area is lower than the critical value and lasts for 11 minutes. Finally， this paper perform experiments with data from public measurement infrastructures （RIPE Atlas） to evaluate the performance of our approach， and the results show that our approach can effectively perceive the changes of network state.

Simulation of Incremental Encryption of Privacy Information in Mobile Edge Databases

MA Lin

2025, 0(06): 61-64. doi:10.3969/j.issn.1006-2475.2025.06.010

Asbtract ( 67 )

PDF (403KB) ( 160 )

References | Related Articles | Metrics

Abstract: The mobile edge database has redundant amount of information， and the additional key sequence is less flexible and adaptive， which leads to lower privacy information hiding degree and poor incremental encryption effect. For this reason， the incremental encryption method for mobile edge database privacy information is proposed. Based on the mobile edge database， clustering and encoding of private information， combined with the information coupling spatio-temporal evolution trend， in a chaotic mode， semantic transformation processing of private data， the introduction of priority key and mobile edge database proxy service， to determine the similarity of the keyword coordinate value， the generation of data keys. Combined with the priority encryption strategy to audit the encryption process， the incremental encryption of privacy information is completed. After Simulation testing， the information hiding degree is always controlled above 96%， and the change of ASCII code value is more stable， basically stable at about 9. The results show that the method can improve the information hiding degree and the encryption effect is better.

Unauthorized Construction Recognition Algorithm of UAV Aerial Photography Based on Deep Learning

SUN Yibo, MOU Li

2025, 0(06): 65-70. doi:10.3969/j.issn.1006-2475.2025.06.011

Asbtract ( 94 )

PDF (1089KB) ( 103 )

References | Related Articles | Metrics

Abstract: Considering the issues of missed detection， slow detection efficiency， and high detection difficulty in the current traditional manual method for detecting illegal buildings in high-rise structures， this paper proposes a YOLOv5s-based algorithm for illegal building detection. Firstly， the coordinate attention mechanism is incorporated into the original framework's backbone section to enhance detection accuracy. Secondly， the bidirectional feature pyramid network （BiFPN） structure is introduced to improve both feature extraction ability and fusion capability across different layers of the model. The loss function is replaced with SIoU to address the problem of matching angles between predicted and actual boxes. Experimental results demonstrate that compared to other commonly used models on our self-built dataset， precision P and mAP value reach 91.36% and 83.45%， respectively. The improved algorithm enhances performance while maintaining high computational speed， meeting both accuracy and timeliness requirements for UAV aerial detection of illegal buildings.

Pancreatic Image Segmentation Approach Based on Improved SegFormer

LIANG Panru, XIN Guojiang, DING Changsong

2025, 0(06): 71-78. doi:10.3969/j.issn.1006-2475.2025.06.012

Asbtract ( 92 )

PDF (1049KB) ( 112 )

References | Related Articles | Metrics

Abstract: To address the issue of low segmentation accuracy in CT images due to the small volume of the pancreas and significant individual differences in its position and shape， we propose an improved SegFormer-based method for pancreatic image segmentation. Prior to model training， we construct candidate regions based on the distribution of the pancreas and perform cropping to effectively reduce background interference and lower the input image resolution. Next， we employ the SegFormer network and introduce an encoding resolution enhancement strategy by adjusting the downsampling ratio to increase the size of the encoder's output feature maps， which retains more detail information to better handle morphological variations of the pancreas. We then incorporate residual polarized self-attention modules to compute channel and spatial attention on the encoded features， highlighting key characteristics of the pancreatic region while suppressing the activation of irrelevant features， thus improving the model's segmentation accuracy. The proposed method achieved an average DSC of 85.5% on the NIH dataset， with a parameter count of 3.91 M and a computational load of 6.89 G FLOPs， indicating its effectiveness in the pancreatic segmentation task and its potential for clinical applications.

Complex Fire Detection Based on Deformable Convolutions and Attention Mechanisms

HAO Rongrong1, MA Qiaomei1, TAN Yajun1, SHI Huanyin2

2025, 0(06): 79-85. doi:10.3969/j.issn.1006-2475.2025.06.013

Asbtract ( 92 )

PDF (1155KB) ( 114 )

References | Related Articles | Metrics

Abstract: To address the issues of flame and smoke occlusion， difficulty distinguishing between flame/smoke and background， and low detection accuracy in complex fire scenarios， a detection model called CFD-YOLO（Complex Fire Detection YOLO）has been proposed. This model is based on the YOLOv8 framework， with several enhancements. Firstly， in the backbone network， the model combines Deformable Convolution DCNv4 with the C2f module， leveraging DCNv4's unconstrained dynamic weighting mechanism to significantly improve the capture of complex deformations and non-rigid features in fire images. Secondly， in the neck section， a cross-attention-based deep semantic fusion module， PSFM， is introduced to achieve adaptive feature enhancement by deeply fusing different feature layers of fire images. Finally， in the head section， the occlusion-aware attention SEAM is used to improve the detection head， allowing it to effectively handle occlusions of flames and smoke in complex environments. The loss function employed is SlideLoss， which dynamically adjusts the positive and negative sample coefficients to reduce false detection rates. The experimental results showed that the mAP index reached 80.33% and 88.28% respectively in the self built dataset and public dataset， which were 3.85 and 3.91 percentage points higher than the original YOLOv8 network. Compared with the current mainstream models， it also has good detection performance.

Image Quality Assessment with Multi-scale Interaction and Cross-channel Attention

QIN Hongzheng1, 2, WANG Tonghan1, 2, JIA Huizhen1, 2

2025, 0(06): 86-91. doi:10.3969/j.issn.1006-2475.2025.06.014

Asbtract ( 77 )

PDF (681KB) ( 195 )

References | Related Articles | Metrics

Abstract: Authentic distortion-oriented no-reference image quality assessment is still challenging due to the variety of distortion types， complexity， and localized noise. As various regions within an image contribute differently to the overall perception of quality， it is problematic to characterize image distortion solely through global feature representation. A no-reference image quality assessment method with multi-scale interaction and cross-channel attention is proposed. Firstly， multi-scale interaction blocks are constructed based on the Transformer that utilizes a convolutional attention mechanism， to enhance the exchange of local and global information among features of varying scales. Meanwhile， a multi-scale dynamic connectivity approach is designed to avoid underutilization or overutilization of multi-scale features. Secondly， a cross-channel attention module is employed to further facilitate the information complementarity and integration of cross-scale features at the channel level. Finally， quality scores are obtained through regression. The proposed method is evaluated against widely recognized no-reference image quality assessment methods across five publicly available image quality evaluation datasets， and the test results demonstrate a potential improvement of 2% in performance compared to the best-performing method. Furthermore， the number of parameter is decreased by 3% and 74%， while the computational complexity by 11% and 74%， respectively， in comparison to the methods based on self-attention and transpositional attention mechanisms. The result demonstrates that the proposed approach achieves advanced quality assessment performance and strong generalization capabilities while keeping the complexity low.

Multi-Result Voting Fusion Method for License Plate Recognition Based on Character Similarity

CHEN Ze1, LI Xiying1, 2, JIANG Qianyin3, LIN Qunxiong4, SUN Quanzhong4

2025, 0(06): 92-100. doi:10.3969/j.issn.1006-2475.2025.06.015

Asbtract ( 90 )

PDF (1051KB) ( 119 )

References | Related Articles | Metrics

Abstract: License plate recognition （LPR） is an important tool for vehicle information recognition. However， it faces some problems such as similar character confusion and unmeasurable credibility of recognition results. In order to improve the usability of LPR algorithms， this paper proposes a multi-result voting fusion method based on character similarity. The method calculates the voting weight of the LPR results according to the character similarity and Bayes theorem， and selects the largest voting weight as the voting result. In the process of calculating the voting weight， the plate weighted voting method and character weighted voting method are proposed according to the characteristics of license plate. Between them， the plate weighted voting method takes the whole license plate as the voting unit， and the character weighted voting method takes a single license plate character as the voting unit. At the same time， voting result confidence is calculated to measure the credibility of the voting results， and it is the theoretical minimum voting accuracy. The experimental results show that the proposed method can effectively reduce the confusion of character recognition. Compared with the traditional weighted voting method， the voting accuracy and error-prone character voting accuracy of the proposed method are increased by 0.78 percentage points and 2.18 percentage points respectively. The experimental results also show that the proposed voting result confidence calculation method is effective and stable. Compared with the traditional weighted voting method， the proposed method can better reflect the credibility of the voting results.

Annotation of Analog Circuit Structures via Graph Attention Networks

LI Xinpeng, TONG Minglei

2025, 0(06): 101-105. doi:10.3969/j.issn.1006-2475.2025.06.016

Asbtract ( 77 )

PDF (429KB) ( 79 )

References | Related Articles | Metrics

Abstract：Automated annotation of circuit structures can generate hierarchical representations of analog circuit networks， thereby advancing the development of automated analog circuit design tasks. This paper introduces a graph attention network-based model that transforms circuit netlists into graph structures， proposes a feature extraction strategy to learn and predict the circuit structures composed of nodes in the netlists， and presents a method for quickly generating a large number of SPICE circuit netlists to provide ample data for training the graph model. Experiments compared the recognition effects of graph convolutional networks， graph isomorphism networks， and GraphSAGE on the same dataset. The results show that the model outperforms the other models in accuracy， precision， and mean average precision， achieving 90.9%， 91.6%， and 91.9%， respectively. These results demonstrate the superiority of the model in capturing circuit connections， especially in terms of its effectiveness in processing complex circuit diagrams.

Residential Microgrid Scheduling Algorithm Based on Deep Reinforcement Learning

GUO Xinyi1, JIANG Kai1, YANG Pengwei2, CHEN Geng2

2025, 0(06): 106-113. doi:10.3969/j.issn.1006-2475.2025.06.017

Asbtract ( 56 )

PDF (683KB) ( 74 )

References | Related Articles | Metrics

Abstract: Residential microgrid scheduling is a complex task aimed at optimizing grid operation under highly variable power demands. Due to the system's complexity， uncertainty， and dynamics， efficient and accurate decision-making is essential. This paper proposes a DQN-based model， DQN for Load Scheduling （DQN-LS）， which accurately schedules residential loads across different times and nodes while considering constraints such as line capacity， generation limits， and node demand. It also factors in seasonal， temporal， and weather variations to ensure fast and precise decisions. To evaluate its performance， DQN-LS is compared with leading methods， including CNN-LSTM-based agents， Monte Carlo Tree Search （MCTS）， and Multi-Agent Soft Actor-Critic （MASAC）. Experimental results， using real datasets and simulations， demonstrate that DQN-LS outperforms baseline models in average reward， reward variance， scheduling efficiency， and constraint violations， confirming its effectiveness and superiority in residential microgrid scheduling.

LDA-IDF Process Log Anomaly Detection Method for Smart Grids

LIU Shaojun, SHA Yitian, JIN Qianqian, CHEN Peng, WU Yue

2025, 0(06): 114-119. doi:10.3969/j.issn.1006-2475.2025.06.018

Asbtract ( 77 )

PDF (538KB) ( 94 )

References | Related Articles | Metrics

Abstract: With the expansion of smart grid systems， the process log data has grown exponentially， making traditional log anomaly detection methods inadequate for handling such massive and complex data volumes. To promptly detect abnormal processes in smart grid systems and ensure their safe and stable operation， this paper proposes LogLDAIDF， a process log anomaly detection method based on LDA-IDF. The method first extracts LDA topic features from process log sequences， then introduces the IDF （Inverse Document Frequency） method to weight the topic features， and selects the top-k weighted topics as the final process log topic features， finally combining with a Bi-LSTM deep learning model for process log anomaly detection. Experimental results on two real-world datasets， HDFS and OpenStack， demonstrate that our proposed method significantly outperforms existing methods， achieving F1 scores of 0.987 and 0.969， respectively， results show its effectiveness and practicality in process log anomaly detection.

Virtualization-based Resource Isolation Scheduling Mechanism for Data Middle Platform

LIU Zihan, SHEN Li, XI Mengting, ZHU Jiajia, LU Jiaxin, CHA Junjie

2025, 0(06): 120-126. doi:10.3969/j.issn.1006-2475.2025.06.019

Asbtract ( 84 )

PDF (865KB) ( 71 )

References | Related Articles | Metrics

Abstract：In a data middle platform environment， departmental systems often have significant differences in architecture and tech stacks. The collected data comes from various databases， resulting in high data heterogeneity and different resource needs. This necessitates resource isolation to ensure no interference between departments' data and computing resources. To handle changing resource demands due to business changes， resource scheduling is required. This paper proposes a virtualization-based isolation solution using the Kubernetes container management platform. Each department has its own virtual environment for resource isolation and management. An improved Kubernetes load balancing strategy is introduced， considering CPU， memory， bandwidth， and disk I/O resources. This strategy enhances resource allocation and system performance. Experiments show a 40% improvement over existing Kubernetes load balancing strategies， demonstrating its effectiveness in maintaining balance in the virtual container system.

Table of Content