Table of Content

27 January 2025, Volume 0 Issue 01

Previous Issue   

Stance Detection with LoRA-based Fine-tuning General Language Model

HAN Xiaolong, ZENG Xi, LIU Kun, SHANG Yu

2025, 0(01):  1-6.  doi:10.3969/j.issn.1006-2475.2025.01.001

Asbtract ( 232 )   PDF (2429KB) ( 184 )  

References | Related Articles | Metrics

 Stance detection is a key task in natural language processing， which determines the stance of an author based on text analysis. Text stance detection methods transition from early machine learning methods to BERT models， and then evolve to the latest large language models such as ChatGPT. Distinguishing from the closed-source feature of ChatGPT， this paper proposes a text stance detection model， ChatGLM3-LoRA-Stance， by using the domestic open-source ChatGLM3 model. In order to apply large models in professional vertical fields， this paper uses LoRA efficient fine-tuning method. Compared with P-Tuning V2 efficient fine-tuning method， LoRA is more suitable for zero-shot and few-shot text stance detection tasks in text. The paper uses the publicly available VAST dataset to fine-tune the ChatGLM3 model， evaluating the performance of existing models in zero-shot and few-shot scenarios. Experimental results indicate that ChatGLM3-LoRA-Stance model has significantly higher F1 scores than other models on zero-shot and few-shot detection tasks. Therefore， the results verify the potential of large language models on text stance detection tasks， and suggest that that the use of LoRA efficient fine-tuning technology can significantly improve the performance of ChatGLM3 large language model in text stance detection tasks.

Seismic Velocity Building Based on Interactive Attention DeMulti Unite 

LI Canwei, WU Chunlei, LU Jing, WANG Chunlin, ZHU Mingfei

2025, 0(01):  7-14.  doi:10.3969/j.issn.1006-2475.2025.01.002

Asbtract ( 79 )   PDF (2750KB) ( 96 )  

References | Related Articles | Metrics

 The velocity parameter of seismic wave runs through the whole seismic exploration process and is very important for seismic imaging. Due to the low computational efficiency of traditional seismic velocity building methods， this paper proposes a seismic velocity building based on interaction attention DeMulti unite， namely IADMU model. Two parts of the network are proposed: DeMulti Unite Convolution （DMUC） and Interactive Attention Block（IAB）. The original seismic shot record passes through the DMUC to extract its advanced feature information from the global and local dimensions. The deconvolution and interactive attention modules are then fed to directly predict the corresponding velocity models with the help of cross-channel interactions. This paper conducts a lot of experiments on simulated data and SEG salt data sets， ablation experiment proves the effectiveness of the proposed DMUC and IAB； Comparative experiments demonstrated that， compared to U-Net， Res-UNet， and DeepLabV3 networks， this network exhibited superior performance on both datasets， validating the superiority of the proposed network in this paper.

Real Time IoT Data Processing System Based on StarRocks

DONG Yizhou1, PAN Weihua2, ZHANG Nan1, MENG Zhuang1

2025, 0(01):  15-19.  doi:10.3969/j.issn.1006-2475.2025.01.003

Asbtract ( 81 )   PDF (900KB) ( 70 )  

References | Related Articles | Metrics

With the popularization and application of IoT technology， a large amount of real-time data needs to be processed and analyzed. Due to the massive and real-time characteristics of IoT data， traditional databases cannot meet the requirements of data storage scale and data processing efficiency， this article proposes a distributed real-time IoT data processing system based on StarRocks. The system utilizes the distributed architecture of StarRocks to build underlying data storage. By introducing message queues and data merging batch submission technology， it ensures fast data writing. At the same time， through storage strategies optimization， index optimization， and materialized view technology， rapid processing and querying of large-scale real-time data have been achieved. The powerful data compression capability of the system also effectively saves data storage space. This framework supports horizontal scaling in data storage scale， improving availability and robustness. Through experimental analysis， the system has significant advantages over traditional distributed databases in terms of data writing， data querying， and data compression.

A Food Safety Risk Warning Method Based on BP Neural Network

XU Shengchao, CHEN Fuqiang

2025, 0(01):  20-24.  doi:10.3969/j.issn.1006-2475.2025.01.004

Asbtract ( 82 )   PDF (885KB) ( 63 )  

References | Related Articles | Metrics

 This paper proposes a food safety risk warning method based on BP neural network. Firstly， a comprehensive food safety risk assessment index system is constructed through screening and integration， and the weights of each indicator are calculated by using the three scale analytic hierarchy process. Then， 7 key indicators are selected as input indicators for early warning research to fully reflect the food safety risk situation. Finally， the BP neural network is used for food safety risk warning， and the connection weights and thresholds of the BP neural network are optimized through genetic algorithm to make the output results closer to the expected output. The experimental results show that the risk warning time data range of this method is 7.23 s~10.23 s， and the accuracy data range of risk warning is 94.05% ~98.44%， indicating that the BP neural network has good practical application effect in food safety risk warning and has practicability.

Optimization and Deployment of Object Detection Algorithm Based on Domestic AI Chips

CHEN Siyun1, MA Huaibo2, ZHANG Huajun2, LAN Zining2, CHEN Wenxin2 , HU Jie1, CHANG Sheng1

2025, 0(01):  25-29.  doi:10.3969/j.issn.1006-2475.2025.01.005

Asbtract ( 111 )   PDF (2281KB) ( 76 )  

References | Related Articles | Metrics

 At present， various types of neural networks have gradually been widely applied in all aspects of society. The performance of neural network models largely depends on the quality of their training strategies， and their deployment cannot be separated from the support of corresponding hardware platforms. In order to ensure the information security and development of the electronic information industry in China under the current situation， it is urgent to replace relevant domestic AI chips. Taking the replacement of domestic AI chips as the starting point， this article explores the deployment process of neural network algorithms on domestic platforms based on the Quanai QA-200RC development kit. The improvement of YOLOv6 neural network training and host program optimization are carried out according to specific task requirements. With real-time detection through cameras， target detection of rocket debris is achieved， the frame rate is 30 FPS， the mAP_0.5 is 90.1%， and the power consumption is 8.1 W， which meets the requirements for completing object detection tasks on edge platforms and is helpful for promoting the application of domestic chips in related fields.

Hyperspectral Image Denoising Using Low Rank Tensor Decomposition and Weighted Group Sparse Regularization

WANG Yefang1, JIA Xiaoning1, 2, CHENG Libo1, 2, LI Zhe1, 2

2025, 0(01):  30-36.  doi:10.3969/j.issn.1006-2475.2025.01.006

Asbtract ( 112 )   PDF (2260KB) ( 68 )  

References | Related Articles | Metrics

 Hyperspectral images have significant reference value in fields such as environmental monitoring， remote sensing science， and medical imaging. However， the imaging process is susceptible to contamination by mixed noise due to limitations in the imaging acquisition equipment and adverse weather conditions， leading to a significant decline in image quality. To tackle this problem， we propose a denoising model for hyperspectral images based on low rank tensor decomposition and weighted group sparsity-regularized. Specifically， to effectively retain the edge information of the hyperspectral image and extract sparse structural features， we propose a group sparse regularization method based on the [l2,1] norm， which aims to weight and constrain the differential images in the spatial and spectral directions. Then， a combined approach is proposed， which utilizes the [l1] norm and Frobenius norm， to effectively eliminate complex mixed noise in the images， thereby enhancing the overall image quality. Furthermore， we use ADMM algorithm to solve the model proposed in this paper. Experimental evaluations of the model are conducted using both simulated and real data， and the results demonstrate the superiority of the proposed model over the baseline model in terms of various evaluation metrics， particularly the proposed model has obvious advantages in hyperspectral image recovery.

Multi-Target Adversarial Cross-domain Recommendation Based on Attributed Heterogeneous Graph

YUAN Jie, ZHU Yan

2025, 0(01):  37-43.  doi:10.3969/j.issn.1006-2475.2025.01.007

Asbtract ( 69 )   PDF (1564KB) ( 66 )  

References | Related Articles | Metrics

Cross-domain recommendation technique plays a crucial role in addressing the challenges of cold start and data sparsity in recommendation systems. However， existing methods often make the simplifying assumption that users’ preferences remain consistent across different domains. This assumption overlooks the inherent heterogeneity of user preferences， resulting in suboptimal recommendation performance. Furthermore， most approaches focus solely on recommendation tasks between domains， lacking a natural extension to multi-domain scenarios. In this spaper， we propose a novel approach called MAAH（Multi-Target Adversarial Cross-domain Recommendation based on Attributed Heterogeneous Graph）. The method leverages the structural information encoded in an attributed heterogeneous graph to represent users and items. By capturing both the homogeneity and heterogeneity of user behavior across domains， MAAH provides a more comprehensive understanding of user preferences. Importantly， we introduce adversarial learning to further integrate and discriminate user preferences， thereby enhancing recommendation effectiveness in each domain. Notably， MAAH achieves multi-objective cross-domain recommendation， addressing the limitations of existing methods. Experimental results on publicly available datasets demonstrate that MAAH effectively mitigates data sparsity and offers a promising solution to the cold start problem in cross-domain recommendation scenarios. 

Point Cloud Data Classification Method of PointNet++ with Position Adaptive Convolution

YAN Xiaoqi, PENG Yiqing, REN Xiaoling

2025, 0(01):  44-49.  doi:10.3969/j.issn.1006-2475.2025.01.008

Asbtract ( 76 )   PDF (4866KB) ( 66 )  

References | Related Articles | Metrics

Aiming at the problem of low classification accuracy of point cloud data in complex scenes， a PointNet++ deep neural network model based on position adaptive convolution is proposed. Since adaptive position convolution has strong ability to capture fine-grained local features and can fully obtain the spatial variation and geometric structure feature information of three-dimensional point clouds， on the basis of PointNet++ network， the proposed method in this paper first obtains the key points through the farthest point sampling， and then uses the K nearest neighbor method to realize grouping according to the key points， and then using the adaptive position convolution replaces the MLP in the original method to extract the local features of each group， and finally completes the point cloud classification. The proposed method was compared on two public point cloud datasets S3DIS and Semantic3D. Experimental results show that the overall accuracy and mIoU of the proposed method on the indoor dataset S3DIS are about 2.7 percentage points and 3.2 percentage points higher than PointNet++ network， respectively， and the overall accuracy and mIoU of the outdoor dataset Semantic3D are about 2.5 percentage points and 2.1 percentage points higher than PointNet++.

Academic Recommendation System Based on Knowledge Graph and Semantic Information

ZHANG Yue, LI Huayu, ZHANG Zhikang, SHEN Xinyi

2025, 0(01):  50-58.  doi:10.3969/j.issn.1006-2475.2025.01.009

Asbtract ( 124 )   PDF (1930KB) ( 111 )  

References | Related Articles | Metrics

In the diverse domains of the Internet， facing the ever-increasing volume of data， there is a growing need for recommendation systems to provide users with personalized information. Utilizing knowledge graphs can enhance the accuracy， diversity， and interpretability of these systems. Addressing the current methods’ limitations in accurately capturing genuine user preferences during propagation， and their lack of attention to the utility of semantic information， this paper proposes an advanced paper recommendation algorithm based on semantic features and knowledge graphs. This method employs the BERT model to extract semantic features from paper abstracts， and uses knowledge graphs for collaborative propagation to obtain entity representations of users and items. During propagation， user preferences are accurately transmitted through multi-head attention， and an attention aggregation network is differentiated between entity representation sets at each layer， the importance of initial information is emphasized. Performance evaluations on three public datasets demonstrate that the model proposed in this paper， compared to the selected optimal baseline models， achieves an increase of 0.010、0.018  and 0.007 in AUC， and 0.007 、0.008 and 0.008 in F1 score， respectively， thereby showing the effectiveness and the superiority of the algorithm proposed in this paper. 

Transient Fault Detection for Low-orbit Internet Communication System Based on CatBoost

XIE Zetao, ZHUANG Yi

2025, 0(01):  59-66.  doi:10.3969/j.issn.1006-2475.2025.01.010

Asbtract ( 48 )   PDF (2039KB) ( 38 )  

References | Related Articles | Metrics

 Low-orbit Internet communication systems are of great significance in the field of modern communications， but due to their highly complex characteristics， the detection of transient system faults has always been a challenging problem. This paper analyzes the fault sources of possible transient faults in the low-orbit Internet communication system， establishes a transient fault model of the low-orbit Internet communication system， and proposes a CatBoost-based transient fault detection method for the low-orbit Internet communication system with the features of efficiency and accuracy. Firstly， fault is injected into the compiled intermediate code， and relevant instruction features are extracted through control flow graph and propagation path analysis; Secondly， the CatBoost machine learning algorithm is used to train the fault prediction model； Finally， the instructions are reinforced with partial redundancy according to the prediction results of the model to realize autonomous fault detection. Comparative experimental results show that the instantaneous fault prediction model based on CatBoost proposed in this article has a higher detection rate and lower space-time overhead.

BD Based Lattice Reduction Assisted Continuous Interference Cancellation Detection Algorithm

LIU Haitao, FENG Fan

2025, 0(01):  67-73.  doi:10.3969/j.issn.1006-2475.2025.01.011

Asbtract ( 58 )   PDF (1359KB) ( 34 )  

References | Related Articles | Metrics

In multi-user MIMO-OFDM communication systems， the increasing number of access users and transmitting/receiving antennas leads to severe multi-user interference and multi-stream interference. Therefore， a multi-user block diagonalization（BD）approach is employed to perform signal detection for multiple users on the same frequency band. An algorithm for continuous interference elimination based on BD and lattice reduction is proposed. Initially， the total channel for multiple users is decomposed into several mutually independent single-user subchannels using two singular value decomposition techniques. In theory， this step can completely eliminate multi-user interference. Subsequently， lattice reduction and QR sorting techniques are applied to optimize the decomposed channels， enhancing the orthogonality of the channel matrix while reducing computational complexity. Finally， a continuous interference elimination detection based on the minimum mean square error （MMSE） criterion is performed on the received signals at the receiver. This process systematically restores the original signals， eliminating multi-stream interference for each user. Simulation experiments demonstrate that the proposed algorithm significantly improves the detection performance of the communication system without increasing computational complexity. At a bit error rate of 10-3， the performance enhancement is 5.5 dB compared to traditional multi-user BD detection algorithms and around 4 dB compared to combined BD-based multi-user MMSE and zero-forcing （ZF） detection algorithms.

Optical Fiber Timing Synchronization in Power Systems Based on ARIMA-ELM Algorithm

LI Qiusheng1, LI Wei1, ZHOU Dongxu1, GUO Chuang1, JIN Wenxin 2

2025, 0(01):  74-79.  doi:10.3969/j.issn.1006-2475.2025.01.012

Asbtract ( 60 )   PDF (1904KB) ( 35 )  

References | Related Articles | Metrics

The real-time and synchronization of power fiber optic communication networks directly affect the efficiency of the entire power grid communication network. Unreasonable handling of the timing characteristics of power fiber optic communication may lead to collaboration issues between power grid equipment. This paper aims to achieve high-precision determination of time signals based on the wavelength of fiber optic transmission. Firstly， the differential autoregressive integrated moving average mode （ARIMA） is used to sequence the input fiber optic data and calculate the wavelength and transmission speed to accurately predict the nodes for fixed time data transmission. Secondly， extreme learning machines algorithm （ELM） is applied to extract and analyze spatiotemporal features of data， ensuring that data from different sources can maintain synchronization when input into optical fibers， achieving dual synchronization in both time and space. Finally， the data transmission node is determined based on the transmission wavelength of the optical fiber. The experimental analysis results indicate that fiber optic data transmission can achieve high-quality full-time synchronization between power grid control equipment and the dispatch and command center.

Real-Time Traffic Classification Method Based on High-dimensional Feature#br# Dimensionality Reduction and Clustering

XIAO Junbi, FU Tianqi

2025, 0(01):  80-85.  doi:10.3969/j.issn.1006-2475.2025.01.013

Asbtract ( 64 )   PDF (2059KB) ( 45 )  

References | Related Articles | Metrics

This paper proposes a real-time traffic classification model based on high-dimensional feature reduction clustering to address the problem of traditional network traffic classification methods being unable to avoid the impact of unknown traffic on classification and difficult to achieve real-time traffic classification. First， a CNN network model is built to extract high-dimensional features from traffic data， and save feature vectors. Then， UMAP is used to reduce the dimensionality of feature vectors， and the DBSCAN clustering algorithm is used to classify traffic， which effectively reduces the impact of unknown traffic on the model while achieving application-level classification. This paper proposes a time-delay control mechanism based on flow consistency， which borrows the idea of TCP congestion control mechanism and greatly reduces the time for traffic classification， making the model proposed in this paper able to meet the requirements of real-time traffic classification. At the same time， this paper collects a set of application-level traffic data sets in a real network. The experimental results on public data sets and the data set buiding by this paper show that the accuracy of this paper’s method is approximately 98% in the known data set， and when the unknown traffic is close to 50%， the accuracy remains at around 80%， and it can meet the requirements of real-time classification.

Efficient Board Games Algorithm with Integrated Strategy Value Network 

ZHOU Yi1, TIAN Yongshen1, QIU Yufeng2, GAO Hua1

2025, 0(01):  86-93.  doi:10.3969/j.issn.1006-2475.2025.01.014

Asbtract ( 61 )   PDF (4184KB) ( 54 )  

References | Related Articles | Metrics

 Board games always have been a focus of deep reinforcement learning research due to their complex board configurations and rules， which require a lot of time to find optimal solutions. Current algorithms for chess games use action probability distribution-based methods for action selection during self-play， which leads to inefficient exploration and exploitation. They also require separate neural network computations for strategy and value， resulting in low sample usage and long training times. This paper proposes an efficient chess game algorithm that combines strategy-value networks， replacing the original action selection method with the Geng-Bellman maximum value method. It balances exploration and exploitation in action search using ε-greedy and simulated annealing algorithms. Experimental results show that compared to various classical chess game algorithms， the proposed algorithm achieves a win rate of over 90% against traditional algorithms. Moreover， using Gumbel-max method during training leads to significantly higher Elo ratings compared to traditional action selection methods with low Monte Carlo simulation counts. With training reaching 3000 Elo ratings， the proposed algorithm can save 50% of the time. 

Ground Penetrating Radar Pipeline Object Detection Method Based on Improved YOLOv8

LI Xi, PAN Yu

2025, 0(01):  94-99.  doi:10.3969/j.issn.1006-2475.2025.01.015

Asbtract ( 99 )   PDF (1351KB) ( 78 )  

References | Related Articles | Metrics

Addressing the issues of traditional ground penetrating radar （GPR） pipeline object detection methods， such as the inability to precisely locate pipelines， time-consuming and inefficient interpretation processes， and interference from complex background noise， this paper designs a GPR image pipeline object detection method based on YOLOv8， with improvements made to the original YOLOv8 network. First， the PConv operator is introduced into the backbone network to make the network structure more lightweight， speeding up the model’s processing speed， and reducing redundant computations and memory access. Second， the Triplet Attention module is introduced to enhance the model’s feature extraction ability across different dimensions， improving object detection accuracy in complex backgrounds. Lastly， the bounding box loss function is replaced with Wise-IoU to improve the regression performance and robustness of the bounding boxes. This paper conducts experiments using a GPR pipeline dataset， and the results show that the improved model proposed in this paper achieves better performance in terms of detection accuracy and computational cost.

Metal Gear Surface Defect Detection Algorithm Based on Improved YOLOv8s

TU Fuquan, QI Yanqi, LIU Jian, WANG Shufeng

2025, 0(01):  100-106.  doi:10.3969/j.issn.1006-2475.2025.01.016

Asbtract ( 91 )   PDF (1316KB) ( 64 )  

References | Related Articles | Metrics

 Aiming at the existing problems such as low identification accuracy， slow detection speed and difficult deployment in real-time detection of metal gear surface defects， a metal gear defect detection algorithm YOLO-GEAR is proposed to improve the efficiency and accuracy of detection and ensure product quality. Firstly， the lightweight module C2f-Faster is designed in the feature extraction layer， which greatly reduces the number of parameters and calculation amount of the model， so as to improve the model detection speed. Secondly， EMA attention module is added to improve the efficiency and accuracy of feature extraction. Finally， the bidirectional feature fusion structure BiFPN is introduced to enhance the feature fusion capability. The experimental results show that the average accuracy of the proposed algorithm on the test set is increased by 3.2% compared with the improvement before， the detection speed reaches 153.8 FPS， and the memory of the network model is only 6.2 MB. It is verified that the algorithm has the advantages of high recognition accuracy， fast detection speed， and small model memory ratio， which is helpful for the realization of industrial deployment.

Randomized Spanning Tree Based Redundant Fault Tolerance Algorithm for SDN Control Plane

GUO Xiaofeng, ZHUANG Yi

2025, 0(01):  107-112.  doi:10.3969/j.issn.1006-2475.2025.01.0.017

Asbtract ( 56 )   PDF (1661KB) ( 38 )  

References | Related Articles | Metrics

 In this study， a random spanning tree-based controller layout algorithm DRT2CA （Minimum Two Covering Algorithm Based on Dynamic Random Spanning Tree） is proposed for the control plane fault-tolerant layout problem in software-defined networks. The algorithm aims to minimize the number of controllers and reduce the deployment cost of the control plane under the premise of guaranteeing redundancy fault tolerance. By continuously generating random spanning trees and employing greedy strategies for controller layout search on the tree， the DRT2CA algorithm achieves the minimum redundancy fault-tolerant coverage with fewer controllers and effectively improves the system resource utilization. The experimental results show that under different network scales and controller capacities， the DRT2CA algorithm is able to realize fault-tolerant control plane layout with fewer controllers deployed compared to existing redundant controller deployment algorithms， has higher redundancy layout efficiency， and is able to provide an innovative solution for constructing an efficient and reliable SDN control plane.

Improved Underwater Target Detection Algorithm Based on YOLOv8

LIU Fei, YANG Degang, ZHANG Xin, QIN Jing

2025, 0(01):  113-119.  doi:10.3969/j.issn.1006-2475.2025.01.018

Asbtract ( 134 )   PDF (3309KB) ( 87 )  

References | Related Articles | Metrics

Aiming at the problem of low detection accuracy caused by false detection and missing detection in underwater image target detection， this paper proposes a lightweight underwater image target detection algorithm based on improved YOLOv8n， aiming to improve the detection accuracy of underwater target images. Firstly， the backbone network in YOLOv8 is replaced by the residual network ResNet10 to enhance the feature extraction capability of the backbone network. Secondly， the large convolution kernel attention mechanism is used to improve the fast feature pyramid module to improve the model’s ability to fuse multi-scale features. Then， the C2f module of the original model is replaced by the generalized efficient layer aggregation network in the latest YOLOv9 algorithm， so that the model can maintain high accuracy while reducing computing costs. Finally， the new loss function Inner-SIoU is used to improve the generalization ability of the model and accelerate the convergence speed of the model. Through experiments， on the URPC2020 underwater image target detection dataset， the improved algorithm mAP50 has reached 86.2%， 2.6 percentage points higher accuracy than the original model. Compared with the advanced YOLOv8s and YOLOv7 tiny detectors， as well as the research work in the same field， the method proposed in this paper has achieved higher detection accuracy.

Identity Authentication Scheme Based on National Cryptographic Algorithm in No Connection

WANG Hong, ZHAO Yuxin

2025, 0(01):  120-126.  doi:10.3969/j.issn.1006-2475.2025.01.019

Asbtract ( 58 )   PDF (1567KB) ( 43 )  

References | Related Articles | Metrics

 In traditional identity authentication schemes， methods such as passwords， tokens and biometric features often require users to authenticate themselves while having an electromagnetic connection with the server. These pose security risks of authentication information being intercepted and attacked. Therefore， the article designs an identity authentication scheme based on national cryptographic algorithm in no connection. Through the coordination of QR codes and national cryptographic algorithm， the scheme forms a complete closed-loop process to verify the user’s identity in the state of no connection between the authenticator and the user. Compared with traditional identity authentication schemes， the connectionless authentication scheme proposed in this paper can effectively avoid the danger of potential electromagnetic attacks. The scheme has the characteristics of simple principle， higher security and less investment， which is practiceable in non-contact access control， information system login verification and other scenarios.