Table of Content

20 November 2025, Volume 0 Issue 11

Previous Issue   

Real-time Secure Compression Method for Protection and Control Messages in 5G Power Distribution Networks

MEI Qin1, LIU Zhiren2, ZHAO Chenyang3, CHEN Yihan1, SU Mingrui3, WANG Gang3, WANG Junbo3

2025, 0(11):  1-9.  doi:10.3969/j.issn.1006-2475.2025.11.001

Asbtract ( 25 )   PDF (3933KB) ( 62 )  

References | Related Articles | Metrics


Abstract: With the steady advancement of the new power system construction， using 5G communication networks to carry power distribution network protection and control service data has become a research and application hotspot in recent years. However， the traffic and security issues in 5G network operation have hindered the large-scale promotion of this technology. This paper proposes a joint design scheme for real-time protection and control message compression and encryption in 5G power distribution networks. Aiming at the data characteristics of GOOSE messages， an inter-frame compression method combining direct transmission and compression is designed. Considering the short packet characteristics of data to be encrypted and the real-time requirements， an improved design of dual session key pools and dual-threaded periodic key update mechanism is carried out based on the ZUC stream cipher algorithm. To address potential abnormal situations such as frame loss and out-of-order in 5G transmission， a key stream synchronization mechanism and a key pool restart mechanism are designed. The system is implemented relying on the domestic edge computing platform. The experimental results show that the proposed method fully meets the real-time requirements of protection and control services， significantly reduces the protection and control data traffic in 5G power distribution networks， and achieves autonomous and controllable system security. This paper provides a new idea for the promotion and application of 5G-based protection and control scheme in power distribution networks.



 Generative Artificial Intelligence Oriented Privacy Data Protection Method

XU Shengchao, LYU Junmin, JIANG Darui

2025, 0(11):  10-16.  doi:10.3969/j.issn.1006-2475.2025.11.002

Asbtract ( 23 )   PDF (1151KB) ( 34 )  

References | Related Articles | Metrics

Abstract: Regarding the serious privacy breach risk caused by generative artificial intelligence，we focus on data privacy protection in generative artificial intelligence. This includes four topics: data privacy desensitization in generative AI， data privacy anonymization processing of generative AI， data privacy enhancement of generative AI， and intelligent monitoring and warning of abnormal data behavior using generative AI. Data desensitization is completed based on generative adversarial networks， the generator is capable of generating highly realistic data samples，however， the discriminator is difficult to distinguish authenticity， thereby enhancing the security of private data. Sensitive information of private data is anonymized based on k-anonymity algorithm to further improve the security of private data； the differential privacy technology of generative artificial intelligence model  is utilized to enhance data privacy and achieve the purpose of protecting private information； Formulate monitoring and early warning procedure for abnormal behavior during access， transmission and processing of private data is formulated to prevent the occurrence of security risk events of private data. The experimental results show that the maximum similarity between the generated data and the original privacy data reaches 98%， the minimum retention of sensitive information of privacy data reaches 3.21%， the maximum noise ratio of privacy data reaches 33%， and the area under ROC curve is larger.



Multi-strategy Dung Beetle Optimization Algorithm for Optimizing UAV Trajectory Planning Problem

YANG Huimin, YANG Jin, SUN Yujie

2025, 0(11):  16-31.  doi:10.3969/j.issn.1006-2475.2025.11.003

Asbtract ( 20 )   PDF (13265KB) ( 33 )  

References | Related Articles | Metrics

Abstract: UAV trajectory planning is an important challenge for UAVs to perform complex missions， which involves finding optimal paths in dynamic and uncertain environments， and the global search and local development problems in UAV trajectory planning have always been the focus of research， based on which this paper designs a multi-strategy hybrid optimization algorithm combined with the dung beetle optimization algorithm to improve the algorithm. In order to improve the global search and local exploitation ability of the algorithm， this paper adopts Latin hypercubic sampling to generate the initial population， adaptive variable spiral strategy to adjust the search direction， and optimal domain perturbation strategy to improve the overall convergence performance. In order to avoid the algorithm falling into the local optimum， Brownian motion and Levy flight strategy are added to dynamically adjust the algorithm development. Experimental results show that the MSIDBO proposed in this paper significantly improves the efficiency and accuracy of trajectory planning based on the guarantee of path length， smoothness and obstacle avoidance performance. The algorithm demonstrates superior global search capability and local optimization capability in uncertainty and dynamic environments， and is applicable to a variety of complex mission scenarios.



Few-Shot Patent Classification Method Integrating Multi-dimensional Prompts and Multi-level Label Expansion

YOU Xindong, ZHAO Yuxian, LYU Xueqiang, LIU Boshan

2025, 0(11):  32-40.  doi:10.3969/j.issn.1006-2475.2025.11.004

Asbtract ( 16 )   PDF (1302KB) ( 32 )  

References | Related Articles | Metrics

Abstract: To promote the integration of industry， academia， and research and drive the development of emerging and future industries， it is necessary to classify university patents according to industrial needs. However， currently， there is a lack of patent classification resources for emerging industries， and the cost of data annotation is high. Therefore， this paper proposes a few-shot patent classification method that integrates multi-dimensional prompts and multi-level label word expansion for emerging industry patent classification. This method uses BERTopic for topic clustering to obtain the topic keywords in patent texts and uses GLM-4 to extract professional terms from patent texts to help the model understand patents from multiple dimensions at the macro and micro levels. It uses Masked Language Modeling （MLM） and ChatGPT to expand the label word space from multiple levels and provide more abundant and semantically deep label words for the prompt learning model. Experiments are verified on the constructed few-shot patent classification data set and achieve better classification results than baseline models. Moreover， the effect is better than that of the GLM-4 large language model， verifying the effectiveness of the proposed method in few-shot patent classification.



Improved V-SLAM Method for Transmission Line Inspection UAV Based on LightGlue Network

ZHU Wenji1, BAN Weihua2, ZOU Lin3, LIU Xu3

2025, 0(11):  41-48.  doi:10.3969/j.issn.1006-2475.2025.11.005

Asbtract ( 16 )   PDF (4338KB) ( 27 )  

References | Related Articles | Metrics

Abstract: To address the problem that the localization accuracy of inspection UAVs is affected by the change of illumination and view angle when they perform visual simultaneous localization and mapping （V-SLAM） in the transmission line environment，an improved Stereo V-SLAM method based on LightGlue network is proposed. Firstly， a SuperPoint feature extraction network is used to extract feature points that are more robust to changes in illumination and viewing angle. Then， the feature matching module is improved by LightGlue network combined with optimized parallel image pyramid model to improve the precision of image feature matching and the real-time performance of algorithm operation. Finally， the point cloud map is converted to an octree map to reduce the memory overhead. The experimental results show that the proposed algorithm is more adaptable to changes in illumination and viewing angle； in the EuRoc dataset test， the localization accuracy is improved by about 28.34% compared with OpenVSLAM， and the real-time performance is improved by 33.30% compared with SL-ORB-SLAM2. In the field experiment， the localization accuracy is improved significantly，and the octree map reduces the memory footprint by about 47.28% compared to the point cloud map. In summary， the algorithm proposed can adapt to the transmission line environment， complete accurate localization in real-time and construct octree maps， which has good engineering application prospect.



Recognition of Composite Power Disturbances Based on Feature Selection and SCNGO-HKELM Algorithm

GUO Wangyong, HUANG Kun, LIANG Jiaben

2025, 0(11):  49-57.  doi:10.3969/j.issn.1006-2475.2025.11.006

Asbtract ( 19 )   PDF (3167KB) ( 27 )  

References | Related Articles | Metrics

Abstract: In this paper， a method for identifying the compound disturbances of power quality based on the selection of time-frequency domain features and the improved Northern Goshawk Optimization （SCNGO）-Hybrid Kernel Extreme Learning Machine （HKELM） is proposed to address issues such as the singularity of feature indicators， complexity of classifier network structure， and difficulty in hyperparameter tuning during the process of disturbance recognition in the electric power system. Firstly， focusing on 9 typical compound disturbances of power quality， their mathematical signal characteristics models are constructed， analyzing the time-frequency domain characteristics of various types of disturbances. Based on this analysis， 19 time-frequency domain indicators for disturbance feature extraction are proposed. Subsequently， considering the impact of feature indicator redundancy on the accuracy of disturbance recognition， Kernel Principal Component Analysis （KPCA） algorithm is utilized for feature indicator selection to establish an optimal indicator set. Finally， a disturbance classifier based on SCNGO-HKELM is introduced. Through the SCNGO algorithm， adaptive adjustment of the kernel function hyperparameters and weight coefficients of HKELM is achieved， enhancing the classifier’s generalization ability while ensuring its learning capability and improving the accuracy and efficiency of disturbance recognition. Experimental results demonstrate that the proposed method achieves an identification accuracy of 97.64% for the 9 classes of typical compound power quality disturbances， with stable classification accuracy in different noise environments， validating the effectiveness and accuracy of the proposed method.



Crowd Counting Estimation Algorithm of Railway Stations Based on Improved P2PNet

WAN Chengkai1, AN Gaoyun2, CUI Lan3

2025, 0(11):  58-64.  doi:10.3969/j.issn.1006-2475.2025.11.007

Asbtract ( 26 )   PDF (4063KB) ( 40 )  

References | Related Articles | Metrics

Abstract: An improved P2PNet based algorithm for estimating the number of people in railway stations is proposed. The algorithm has made significant modifications and optimizations to the traditional P2PNet algorithm. Firstly， the algorithm model adopts BiFPN path aggregation to enhance the network’s ability to fuse feature maps of different scales， solving the problem of large differences in personnel size and scale in images. Secondly， the network introduces the A-SPP structure into the low-level feature maps to increase the receptive field range and enhance its ability to extract multi-scale features. Thirdly， the CSAM attention mechanism is adopted before the OutLayer in the network to dynamically adjust the importance of each channel and spatial position in the feature map， which more effectively regresses the position and classification of people in the image. Finally， Focal Loss replaces traditional cross entropy in loss function to solve the problems of imbalanced positive and negative samples and imbalanced difficult and easy samples. The results of comparative experiments on publicly datasets and proprietary datasets show that this algorithm outperforms current advanced algorithms in terms of mean absolute error. In actual station video scenarios， this algorithm can accurately estimate the number of people.



Substation Equipment Defect Detection Based on Lightweight YOLOv8

SUN Erjie1, ZHANG Qifeng2, WANG Deqing3

2025, 0(11):  65-70.  doi:10.3969/j.issn.1006-2475.2025.11.008

Asbtract ( 19 )   PDF (2990KB) ( 36 )  

References | Related Articles | Metrics

Abstract: With the development of smart grid， the stable operation of substation equipment becomes particularly important. However， due to the large size of existing detection models and the high requirements for the deployment of edge devices， the traditional defect detection methods face challenges in application. In order to solve this problem， this paper proposes a lightweight substation equipment defect detection model based on improved YOLOv8. Firstly， the Backbone and Neck parts of the model are optimized by introducing C2f_Faster block and Slim-Neck structure to reduce redundant calculation and memory access and solve the problem of low detection accuracy caused by feature redundancy. Secondly， the design of the Detect_G module further improves the speed and accuracy of the model detection. Finally， a multi-scale attention mechanism based on cross-space is introduced to enhance the detection ability of equipment defects in small target substations. Experimental results show that the proposed algorithm achieves 92.56% mAP， 5.9 M model parameter count and 345.6 fps detection speed on the substation defect dataset， and its performance is superior to other mainstream algorithms such as SSD， Faster R-CNN， YOLOv4， YOLOv7 and YOLOv8.



Improved SOD Algorithm with Cross Modal Interaction and Multi Scale Aggregation

WANG Jingpeng, CUI Yuyong, CAI Changlin, HE Ming’ao, LI Yinghao, TANG Zhonghe

2025, 0(11):  71-79.  doi:10.3969/j.issn.1006-2475.2025.11.009

Asbtract ( 17 )   PDF (4367KB) ( 28 )  

References | Related Articles | Metrics

Abstract: Salient object detection （SOD） is an important research direction in the field of computer vision， which aims to identify and segment the most noteworthy objects in a scene. It is difficult for the single-modal SOD algorithm to achieve effective detection results after the image information is disturbed by illumination and out-of-focus， while the multi modal detection algorithm has the problems of large difference in feature information， low effectiveness of cross-modal feature fusion， and low feature utilization rate between different levels. In order to solve the above problems， this paper proposes an improved SOD algorithm based on cross modal interaction and multi scale aggregation. The algorithm adopts a dual-loop cross-modal interaction mechanism to fuse RGB image features and thermal infrared image features in a cooperative incentive learning manner， and the information receptive field amplification mechanism is used to fuse spatial and channel information of different dimensions between the two modal information at the same level. The multi-scale aggregation mechanism mines the features of different depths of the network model， transmits and connects， aggregates the shallow fine grained information and the deep coarse grained abstract information， and finally obtains the object detection results. ResNet， VGGNet and DenseNet are used for feature extraction， and the detection effects are compared through experiments. Experiments on a variety of targets in outdoor scenes are carried out to verify the algorithm and qualitative and quantitative analysis， and the results show that our algorithm achieves good detection accuracy and detection effect， and the overall performance is better than that of the existing SOD model.



LSGI-YOLOv8: Ceramic Tile Surface Defect Detection Algorithm Based on Lightweight YOLOv8

YANG Anbo1, ZHONG Guoyun1, LIU Meifeng2, XI Chao2, ZHANG Wei1, DING Peng1

2025, 0(11):  80-88.  doi:10.3969/j.issn.1006-2475.2025.11.010

Asbtract ( 21 )   PDF (3795KB) ( 37 )  

References | Related Articles | Metrics

Abstract: Aiming at the problem of low detection accuracy and slow detection speed of ceramic tile defect detection caused by small defects and complex patterns on the surface of ceramic tiles， a ceramic tile surface defect detection algorithm based on improved YOLOv8 is proposed. The algorithm designs a spatial pooling layer module LSPPF incorporating the LSKA （Large Separable Kernel Attention） mechanism to improve the multi-scale feature extraction capability of the model； a lightweight feature fusion network SimBiFPN is designed to reduce the number of jump connections in the weighted bidirectional feature pyramid network， enhance the model’s representation of small target defect features， and reduce the number of model parameters; a D-Slim-Neck module based on lightweight convolution GSConv （Grouped Shuffle Convolution） and depth-separable convolution （Depthwise Separable Convolution， DSConv） is proposed for replacing the traditional Slim-Neck module to significantly reduce the model computation and complexity; to address the problem of different shapes of tile surface defects， a loss function Inner_CIoU with an auxiliary box is proposed to control the size of the auxiliary frame by adjusting the scale factor to further improve the detection accuracy of the algorithm. The experimental results show that the mean average precision of the improved algorithm reaches 90.3%， which is 2.1 percentage points higher than YOLOv8， and the number of parameters and calculations are reduced by 6.7% and 13.6% respectively， achieving tile surface defect detection performance comparable to other advanced algorithms.



Optimization and Evaluation of 3D Gaussian Splatting with Dynamic Opacity Filtering

WAN Siyao1, FANG Zhiwei1, ZHOU Ting1, LYU Hanyun2, LUO Zheng3, CHEN Siyuan1

2025, 0(11):  89-96.  doi:10.3969/j.issn.1006-2475.2025.11.011

Asbtract ( 24 )   PDF (15047KB) ( 30 )  

References | Related Articles | Metrics


Abstract: 3D Gaussian splatting is an efficient modeling and rendering technique that demonstrates significant advantages in handling complex scenes and real-time rendering. However， when facing insufficient viewpoint coverage or complex lighting conditions， issues such as noise interference and inadequate 3D structure representation arise due to limited data support and algorithmic constraints， negatively impacting model accuracy and visual quality. To address these challenges， this paper presents an innovative 3D Gaussian iterative optimization method that introduces a filtering mechanism balancing density and opacity thresholds， significantly suppressing noise and enhancing the model’s geometric expressiveness in complex scenes. Furthermore， this paper presents a 3D structure evaluation system to assess the geometric accuracy and structural performance of Gaussian ellipsoids in 3D reconstruction. Experiments conducted on the DTU， Mip-NeRF 360， and Tanks & Temples datasets demonstrate that the proposed method not only maintains high visual quality in novel view generation but also surpasses existing baseline models in terms of 3D geometric accuracy. Additionally， it outperforms two state-of-the-art 3D Gaussian splatting-based mesh extraction methods: Surface-Guided Gaussian Reconstruction （SuGaR） and Gaussian Mesh Synthesis （GaMeS）. Through verification， the proposed filtering mechanism exhibits excellent adaptability across different scenes， effectively handling various complex lighting conditions and scene structures.



Fault Diagnosis Method for Charging Pile Based on Vision Transformer

QIU Xinyu1, CHEN Xiao1, CHEN Mingming2, GAO Hui3, MENG Ziyue3

2025, 0(11):  97-105.  doi:10.3969/j.issn.1006-2475.2025.11.012

Asbtract ( 25 )   PDF (3274KB) ( 32 )  

References | Related Articles | Metrics

Abstract: With the increasing penetration rate of electric vehicles and the increasing popularity of charging facilities， the difficulty of operation and maintaining electric vehicle charging facilities is also increasing. The fault diagnosis of Direct Current （DC） charging piles is a crucial part of the operation and maintenance of electric vehicle charging facilities. Detecting early faults of charging piles in time is of great significance in eliminating the risk of charging pile failures and ensuring the stable operation of charging piles. The existing fault diagnosis methods not only require the assistance of high-cost specialized equipment but also have high requirements for data sampling and feature extraction； Although the fault diagnosis method based on neural networks and its variants has high performance， its training process requires high-quality labelled data， and it is not easy to obtain enough labelled data for training. Therefore， a fault diagnosis method based on Vision Transformer （ViT） is proposed in this paper only by using the voltage and current signals collected by the DC charging pile itself. In this method， the low-frequency sampling signals of voltage and current of the charging pile are converted into time series images， and the ViT model is used for feature learning. In this process， the pre-training technology is used to transfer the cross-domain feature representation knowledge into the fault diagnosis model， so that the ViT model can be fine-tuned on a relatively small labelled data set， thus achieving better performance on limited data and alleviating the need for labelled data. The experimental results show that the average accuracy of the fault diagnosis model is 92.2%， which meets the practical requirements. The method proposed in this paper supports online diagnosis and does not depend on special equipment， so it has a good popularization prospects.



Transformer Fault Diagnosis Method Based on Multi-source Data Fusion and Optimized Deep Belief Network

XU Bo1, WEI Yijun1, ZENG Han2, KANG Futian3, HUANG Ying1, DENG Fangming2

2025, 0(11):  106-111.  doi:10.3969/j.issn.1006-2475.2025.11.013

Asbtract ( 17 )   PDF (1285KB) ( 35 )  

References | Related Articles | Metrics

Abstract: To address the issues of single data sources， low diagnostic accuracy， and slow convergence speeds in existing transformer fault diagnosis methods， this paper proposes a transformer fault diagnosis model based on multi-source data fusion technology and a deep belief network optimized by the black widow algorithm. Firstly， the oil chromatography data are preprocessed； secondly， the deep neural network （DNN） and convolutional neural network （CNN） are used to extract features from the oil chromatography data and the image data， respectively； then the features extracted by DNN and CNN are fused， and the fused features are obtained； next， the fused features are input into the deep belief network （DBN）， and the parameters of the DBN network are optimized using the black widow optimization method， and then the classification results of fault diagnosis are finally output. The final experimental results demonstrate that the proposed method achieves high precision， recall， and accuracy in fault diagnosis， exhibits rapid convergence， and demonstrates strong adaptability to complex diagnostic environments. The average precision， recall， and accuracy rates are 93.67%， 94.32%， and 94.73% respectively， indicating its suitability for transformer fault diagnosis applications.



Dynamic Fleet Scheduling of Arctic Shipping Routes Under Beidou Third-generation Service

LYU Hongjun, SHEN Lixin, LIU Lu, KANG Yipei

2025, 0(11):  112-118.  doi:10.3969/j.issn.1006-2475.2025.11.014

Asbtract ( 22 )   PDF (1595KB) ( 31 )  

References | Related Articles | Metrics

Abstract: Employing the Arctic shipping route as a regular cargo route can not only fully leverage its unique geographical location and potential shipping capacity but also effectively mitigate the global supply chain crisis amid the complex international situation. Nevertheless， compared to other sea areas， the Arctic shipping route is characterized by high-frequency and unpredictable environmental disturbance factors， such as sea ice variations， temperature changes， and visibility alterations. Cargo ships need to maintain real-time communication to cope with the aforementioned crises; however， there is currently a lack of an immediate， safe， and reliable information interaction mechanism. With the official commissioning of the Beidou-3 satellite system covering the Arctic shipping route， the short message communication of Beidou-3 has emerged as an effective means for constructing the information interaction mechanism. On this basis， this paper presents a solution to the dynamic dispatch optimization problem （ADSOP） of the Arctic shipping route fleet under the Beidou-3 service: A multi-base station type fleet information service framework dedicated to the dispatch of the Arctic shipping route fleet is designed based on the Beidou-3 satellite navigation system. Under the service of this framework， a generalized spatio-temporal network model for the dynamic dispatch of the Arctic shipping route fleet is devised to analyze different scenarios of the fleet carrying out cargo tasks. An improved differential evolution algorithm （DE） is adopted for model solving. Finally， AIS vessel data of an Arctic route from a certain shipping company is utilized for case analysis. The research results indicate that under the particular navigation environment conditions of the Arctic shipping route， the proposed scheme can serve as a significant reference for carriers when designing the cargo dispatch plan for the Arctic shipping route fleet.



Causal Inference Algorithm Driven by Denoising Diffusion Probabilistic Model

ZHANG Hongtao, ZHU Yongzhong, ZHANG Yuxuan, XIA Shiyuan

2025, 0(11):  119-126.  doi:10.3969/j.issn.1006-2475.2025.11.015

Asbtract ( 14 )   PDF (976KB) ( 31 )  

References | Related Articles | Metrics

Abstract：A new causal inference algorithm is proposed to address the limitations of traditional methods， which fail to account for the impact of interventions on structural causal models and suffer from large inference bias and low stability when handling high-dimensional， complex causal data. This algorithm integrates a denoising diffusion probabilistic model. First， it analyzes the effects of different types of interventions on both the structural causal model and the diffusion model to improve algorithm interpretability. Next， the likelihood lower bound of the model under the causal Markov assumption is theoretically derived， and a diffusion sampling process that includes causal parents is constructed. Finally， by combining the variational structure of the diffusion-based causal model， the algorithm can sample from both original and intervention data and perform counterfactual estimation， thus simplifying model training complexity and enhancing robustness. In comparison with other algorithms， simulation results show that under different structural causal model assumptions， the proposed algorithm reduces the maximum mean discrepancy by 10%~41% and the mean squared error of counterfactual estimation by 9%~63%. Empirical results indicate a reduction of 11% in the maximum mean discrepancy and 5% in the mean squared error of counterfactual estimation. Experiments demonstrate that this algorithm effectively handles complex data structures and noise distributions， significantly improving the accuracy and stability of both sampling and counterfactual estimation.