Computer and Modernization

Intent-based Lightweight Self-Attention Network for Sequential Recommendation

HE Sida, CHEN Pinghua

2024, 0(12): 1-9. doi:10.3969/j.issn.1006-2475.2024.12.001

Asbtract ( 62 )

PDF (578KB) ( 60 )

References | Related Articles | Metrics

The parameters of the self-attention calculation mechanism in the existing sequence recommendation models are too large， and there is insufficient preference information in the user's shopping intention. This paper proposes an intent-based lightweight self-attention network for sequential recommendation. On the basis of the traditional product sequence embedding， the model introduces intention sequence embedding to further explore the conversion patterns between sequences. At the same time， in order to reduce the computational complexity of self-attention between pairwise products/intentions in the sequence， a convolutional segmentation sampling module is designed to divide the user behavior sequence and intention sequence into multiple segments， mapping user interests to multiple sequence segments. Comparative experiments are conducted on three public datasets， MovieLens-1M， Yelp， and Amazon-Books. Compared with baseline models， the self-attention mechanism is applied to capture the dependency between segments， effectively reducing the number of parameters. The results show that the hit rate， normalized discounted cumulative gain and mean reciprocal ranking are increased by 5.32%， 4.40% and 5.51% on the MovieLens-1M dataset， 30.93%， 22.73% and 28.84% on the Yelp dataset， and 7.78%， 11.55% and 13.98% on the Amazon-Books dataset， which verify the effectiveness of the model proposed in this paper.

Entity Linking Method Based on Topics and Description Information

ZHENG Jiuchao, ZHAO Xinyuan

2024, 0(12): 10-14. doi:10.3969/j.issn.1006-2475.2024.12.002

Asbtract ( 36 )

PDF (657KB) ( 47 )

References | Related Articles | Metrics

Entity linking is widely applied in fields such as information mining and question-answering systems，playing a pivotal role in constructing knowledge graphs. However， it is noted that the majority of current entity linking methods inadequately leverage candidate entity information and merely implicitly consider the relationships between entities within the global model. In response， we propose an entity linking approach named TopDEL， which integrates topics and descriptive information. TopDEL leverages descriptive entity information to aid in the selection of words with significant relevance to the entities from their surrounding context. Concurrently， the BERTopic topic model is incorporated into the local model to extract topics from documents. The word distribution under each topic is then utilized to represent the relationships among various entities for entity linking purposes. Experimental results conducted on four publicly available datasets underscore the efficacy of the TopDEL method.

Fashion Clothing Pattern Generation Based on Improved Stable Diffusion

ZHAO Chenyang, XUE Tao, LIU Junhua

2024, 0(12): 15-23. doi:10.3969/j.issm.1006-2475.2024.12.003

Asbtract ( 38 )

PDF (1223KB) ( 32 )

References | Related Articles | Metrics

Dress pattern is a window for people to show their personality and fashion. In recent years， with the continuous development of multimodal technology， text-based dress pattern generation has been well studied. However， the existing methods have not been well applied due to the problems of combining poor semanticity and low resolution. After the large-scale language-image pre-training model CLIP was proposed， various pre-training diffusion models combined with CLIP for text-image generation tasks have become the mainstream methods in this field. However， the original pre-training models have poor generalization ability to the downstream task， relying solely on the pre-training model does not allow flexible and accurate control of the color and structure of the dress pattern， and its large number of parameters is difficult to re-train from scratch. To solve the above problems， this study designs a Stable Diffusion-improved network FT-SDM-L （Fine Tuning-Stable Diffusion Model-Lion）， which uses the dress image text dataset to update the weights of the cross-attention module in the original model. The experimental results show that the ClipScore and HPS v2 scores of the fine-tuned model are improved by 0.08 and 1.22 on average， which validates the important ability of this module in combining textual information. Subsequently， to further enhance the model’s feature extraction and data mapping capabilities in the apparel domain， a lightweight adapter， Stable-Adapter， was designed to be added to the module’s output location to maximize the sensing of changes in the input cues. By adding only 0.75% extra parameters to the adapter， the ClipScore and HPS v2 scores of the model can be further improved by 0.05， 0.38. Good results are achieved in terms of fidelity and semantic consistency of clothing pattern generation

A Method of Using Compound Event Probability Operation to Solve Problem of Negative Information Blocking Maximization

WANG Xiaohang1, LI Yongjie1, YU Lei1, FAN Xiao2

2024, 0(12): 24-33. doi:10.3969/j.issn.1006-2475.2024.12.004

Asbtract ( 21 )

PDF (1197KB) ( 23 )

References | Related Articles | Metrics

While online social networks provide people with convenient information interaction， they also widely spread negative information thus cause panic in the society. Therefore， it is urgent to take reasonable and effective strategies to block the spread of negative information in the network to the greatest extent. In COICM model， this paper studies the problem of negative information blocking maximization and designs a method to compute the positive（negative） activation probabilities of nodes based on maximum influence in-arborescence， and then proposes a heuristic algorithm to solve this problem. The core idea is that， firstly， distinguishing the state of nodes in the local impact in-tree， that is， the node is positive（negative） activated at the current time， has been positive or negative activated before the current time and remains inactive until the current time， and the five states constitute the sample space of the events occurring at the node up to the current time. Then use the compound event probability operation method to work out the probability expression of positive （negative） activation of the node at the current time as well as calculate the negative activation probability of the root node through recursive calculation. Finally， take the sum of the negative activation probabilities of all nodes in the network as the influence of the negative seed set. The algorithm uses the greedy framework to iteratively select the node with the largest negative information blocking as the node to propagate positive information. Compared with existing algorithms on four real social network datasets of different sizes， the results show that the proposed algorithm has better negative information blocking effect， and can be applied to large-scale networks.

Gesture Recognition Method Based on WiFi and Prototypical Network

HUANG Tingpei1, MA Lubiao1, LI Shibao2, LIU Jianhang1

2024, 0(12): 34-39. doi:10.3969/j.issn.1006-2475.2024.12.005

Asbtract ( 41 )

PDF (1535KB) ( 32 )

References | Related Articles | Metrics

WiFi-based gesture recognition plays an important role in touchless human-computer interaction. However， existing WiFi-based gesture recognition systems faced the challenges of small data amount and poor cross-domain performance. In order to solve the above problems， the captured raw WiFi channel state information （CSI） is denoised by CSI Ratio， the extracted phase and converted into CSI images， which is transformed into an image classification problem. Then the transformed images are fed into the prototypical network （PN） for small sample cross-domain gesture recognition， and an enhanced Convolutional Block Attention Module （CSI-CBAM） is added to the PN feature extraction network to improve the gesture representation learning. Extensive experiments were conducted on the Widar3.0 dataset. The experimental results showed that when each class in support set reaches four labeled samples， the system average recognition accuracies are 93.54%， 91.28%， 91.99%， and 89.16% for cross-environment， cross-user， cross-location， and cross-orientation， respectively. Average cross-domain accuracy is higher than 90%， the proposed method only required a small number of labeled samples to achieve high accuracy cross-domain recognition.

Anomaly Detection of Network Traffic Based on Autoencoder

LYU Meijing1, NIAN Mei1, ZHANG Jun1, 2, FU Lusen1

2024, 0(12): 40-44. doi:10.3969/j.issn.1006-2475.2024.12.006

Asbtract ( 32 )

PDF (629KB) ( 28 )

References | Related Articles | Metrics

In the face of increasingly complex network traffic and data structures with increasing dimensions， the existing traffic anomaly detection schemes have problems such as high false positive rate， low efficiency and poor practicability. To solve these problems， an autoencoder based network traffic anomaly detection model is proposed. Firstly， the model extracts the features of network traffic based on random forest algorithm and selects the optimal collection， and divides the feature vector set into several subsets by hierarchical clustering algorithm to reduce the feature dimension. Then the feature subset is processed in parallel by the autoencoder and the RMSE value is calculated. The maximum average RMSE value of multiple experiments is defined as the normal flow threshold. The average RMSE value and threshold of the test data are used to determine the abnormal traffic. The experimental results show that the recall rate of this model is 4.3 percentage points higher than that of the traditional anomaly detection method， and the running time is reduced by about 37%.

PCB Defect Detection Method Based on Improved YOLOv7

ZHANG Simin, LIU Xinmei, YIN Junling, LI Baoling

2024, 0(12): 45-52. doi:10.3969/j.issn.1006-2475.2024.12.007

Asbtract ( 40 )

PDF (2764KB) ( 33 )

References | Related Articles | Metrics

A PCB defect detection method based on an improved version of YOLOv7 has been proposed to address the issues of inaccurate detection， slow detection speed， and low recognition accuracy in traditional network models. Firstly， this method replaces CatConv with partial convolution PConv from FasterNet in the original YOLOv7 model to reduce memory access and parameter quantity， thereby improving detection speed. Secondly， a bidirectional feature pyramid network （BiFPN） is introduced into the head network of the YOLOv7 model to achieve multi-scale feature fusion for PCB defect detection， enhancing the model’s detection accuracy. The FasterNet module is then fused with BiFPN to form the YOLOv7+FasterNet+BiFPN model for PCB defect detection， enhancing the model's capability to express defect features. Finally， the original CIoU loss function is improved to XIoU loss function， which not only improve the convergence speed of the model and its resistance to perturbations on small bounding boxes， but it also accurately measures the accuracy and localization precision of the bounding box predictions. The experimental results show that the improved YOLOv7 model achieves an mAP@0.5 of 95.7% and a recall rate of 98.0% on the test set. Compared to the original YOLOv7 model， the mAP@0.5 value and recall rate have increased by 7 and 2 percentage points， respectively. The detection time is only 21.7 ms. Additionally， the computational complexity of FLOPs has also decreased by 6.5 G compared to the original model. The proposed method outperforms traditional network models in terms of detection speed， recall rate， and accuracy， providing an effective solution for PCB defect detection.

Oil and Gas Well Production Prediction Model Based on Empirical Wavelet Transform

ZHANG Xiaodong1, BAI Guangzhi1, LI Min1, LI Haoyang2

2024, 0(12): 53-58. doi:10.3969/j.issm.1006-2475.2024.12.008

Asbtract ( 22 )

PDF (1321KB) ( 22 )

References | Related Articles | Metrics

Oil and gas well production prediction is of great significance for efficient development of oil and gas resources. A two-channel production prediction model incorporating empirical wavelet transform （EWT） and convolutional bi-directional long and short-term memory network is proposed to address the problem of strong nonlinearity and difficulty in prediction of production data due to inter-opening production and other artificial operational factors. One part of the model uses EWT to decompose gas production data， and the decomposed subseries are extracted in the time and frequency domains using a bi-directional long and short-term memory network （BiLSTM）； the other part of the model uses a one-dimensional convolutional neural network （1D-CNN） to extract local time-series features from the multidimensional historical production data， and then uses BiLSTM combined with a self-attentive mechanism to extract the output features from the 1D-CNN module output features to mine the global features of gas well production data. Finally， the features of the two parts of the model are fused to generate the final prediction results. Experimental modeling analysis is carried out using the late production history data of a gas well， and it is found that the prediction results of this method are more accurate for complex production sequences with frequent manual measures， which verifies the feasibility of applying this method to actual production prediction in oil fields.

Short-term Load Forecasting in Industrial Parks with Multi-scale Time Coding

WANG Haiyang, GONG Tongxin, YANG Jintao, CHEN Zailong

2024, 0(12): 59-65. doi:10.3969/j.issn.1006-2475.2024.12.009

Asbtract ( 28 )

PDF (756KB) ( 28 )

References | Related Articles | Metrics

To enhance the accuracy of short-term load prediction in industrial parks， a model based on complete ensemble empirical mode decomposition with adaptive noise with auto-encoder and convolutional neural network-Transformer is proposed. The model addresses the issues of coupling， nonlinearity， and stochasticity of short-term loads. Given that sudden events and emergencies in real scenarios can cause abnormal fluctuations in load data， the sliding time window method is used to firstly detect and correct any anomaly data. Secondly， the frequency domain decomposition algorithm is utilized to resolve the coupling of the load data by dividing the historical load data into multi-scale frequency domain components. Thirdly exogenous features with high correlation to be selected load fluctuations are generated using auto-encoder and feature engineering methods and used as inputs along with the components. Then a convolutional neural network is used to analyze latent features and fuse them with the inputs. The results are fed into the Transformer network， which combines its coding capability and multi-attention mechanism to capture the characteristics of the time series. The final prediction result is obtained by super-imposing the final output of each sub-module. Using the real load dataset as an example， the results demonstrate that the proposed model significantly enhances short-term load forecasting accuracy.

Safety Helmet Wearing Detection Algorithm for Complex Construction Scenes

LIU Yunhai1, Feng Guang1, WU Xiaoting2, YANG Qun2

2024, 0(12): 66-71. doi:10.3969/j.issn.1006-2475.2024.12.010

Asbtract ( 35 )

PDF (1052KB) ( 28 )

References | Related Articles | Metrics

In view of complex background interference and foreign object occlusion in the construction scenes， which reduces the accuracy of helmet wearing detection， we propose a safety helmet wearing detection algorithm for complex construction scenes. This paper improves the YOLOv5 algorithm， adding the Coordinate Attention （CA） mechanism， replacing the first two layers in the backbone network using the Stem Block， applying a Decoupled detection Head （DH） structure with the addition of the Coordinate Attention mechanism. At the same time， an additional large-scale feature extraction layer is added. Results on the helmet dataset show that the improved CADH-YOLOv5 algorithm with a mean detection precision of 91.2% can significantly improve the performance of safety helmet wearing detection for complex construction scenes， which is superior to similar algorithms， and has limited real-time performance.

DSMSC Based on Attention Mechanism for Remote Sensing Image Scene Classification

LIU Baobao, YANG Jingjing, TAO Lu, WANG Heying

2024, 0(12): 72-77. doi:10.3969/j.issn.1006-2475.2024.12.011

Asbtract ( 35 )

PDF (1060KB) ( 21 )

References | Related Articles | Metrics

To address the issue of limited classification accuracy in remote sensing image scene classification， arising from the complex background and varying scales of scene objects， this paper introduces a remote sensing image scene classification model based on a depthwise separable multiscale dilated feature fusion network with an attention mechanism. Firstly， this model employs a feature extraction module built on depthwise separable convolutions， allowing the extraction of deep-level image features while minimizing the parameter count. Subsequently， a multiscale dilated convolution module is used to expand the network’s receptive field， enabling the extraction of both global and contextual features from remote sensing images. Finally， the attention mechanism is used to make the network focus on important features， and the extracted features are input into a Softmax classifier for the purpose of classification. We validate the proposed model on two datasets， AID and WHU-RS19， for remote sensing scene classification. Experimental results demonstrate that， in comparison to baseline models such as AlexNet， VGG-16， and ResNet18， the proposed model achieves an accuracy improvement to 93.32% on AID and 91.15% on WHU-RS19， while maintaining a relatively lower parameter count. The proposed model holds significant theoretical implications for remote sensing image scene classification.

SAR Ship Detection Algorithm Based on Improved YOLOv8

GU Yue, DENG Songfeng, SHEN Ji, MU Wentao, ZHAO Enqi

2024, 0(12): 78-83. doi:10.3969/j.issn.1006-2475.2024.12.012

Asbtract ( 58 )

PDF (1264KB) ( 36 )

References | Related Articles | Metrics

To enhance the accuracy of ship target detection in SAR images， especially when facing challenges such as uneven target sizes， dense distributions， and complex backgrounds， an improved YOLO-3M ship target detection algorithm based on YOLOv8 is proposed. Firstly， the algorithm introduces a Multi-Scale Dilated Convolution Block （MSDB） into the backbone network， which uses convolutions with different dilation rates to extract multi-scale features， thereby enlarging the receptive field without increasing computational costs. Secondly， a Multidimensional Collaborative Attention （MCA） mechanism is incorporated into the neck network to capture key features across the channel， height， and width dimensions， facilitating interaction between different dimensional information and helping the network to effectively focus on key parts within complex backgrounds. Finally， an MPDIoU loss function is introduced in the detection head to address issues with existing loss functions that struggle to effectively detect when the predicted bounding box and the actual bounding box have the same aspect ratio but completely different widths and heights. Experimental results on the SSDD dataset show that the YOLO-3M algorithm achieves higher accuracy and average precision while effectively reducing the number of parameters and computational requirements， making the model more lightweight and suitable for resource-constrained environments. Additionally， there is a significant improvement in reducing false positives and false negatives in complex ship detection scenarios.

Detection and Recognition Algorithms for Chinese and English Scene Text Images

WANG Yanyuan, MAO Zhengchong

2024, 0(12): 84-90. doi:10.3969/j.issn.1006-2475.2024.12.013

Asbtract ( 33 )

PDF (1514KB) ( 29 )

References | Related Articles | Metrics

The complex background of scene text images makes it challenging for detection algorithms to locate text regions accurately， leading to difficulties in recognition. To simultaneously detect and recognize scene text content in both Chinese and English languages， and improve the accuracy of detection and recognition， an improved algorithmic model TD-ABCNetv2 based on ABCNetv2 network is proposed. Addressing the issue of variations in text features such as shape， arrangement， and font， this model adopts SKNet as the backbone network and introduces the Selective Kernel module to help the network learn features of different scales， accommodating texts of various scales， shapes， and orientations. Considering the different character sizes and intervals of Chinese and English scene texts， the ECA attention module is added to the FPN structure to integrate the channel information more effectively， enhance the network’s sensitivity to different features， and make the feature fusion more targeted. Additionally， the CIoU loss function is introduced to more accurately measure the degree of overlap between bounding boxes， adapt to changes in the shape of the text， and enhance the generalization ability of the model. The experimental results show the proposed model is validated through experiments on several public datasets.

Design of Semi-physical Simulation Test System for Unmanned Sailboat

WAN Bing1, 2, 3, ZHAO Wentao4, PAN Duotao1, ZHAO Zhengtao2, 3, SUN Zhaoyang2, 3, YU Jiancheng2, 3

2024, 0(12): 91-99. doi:10.3969/j.issn.1006-2475.2024.12.014

Asbtract ( 26 )

PDF (1610KB) ( 29 )

References | Related Articles | Metrics

Focusing on some potential vulnerabilities of the electronic control system of unmanned sailboats， a low-cost semi-physical simulation test system for unmanned sailboats was designed. Using real test data， hypothetical data or the sensor data carried by the platform itself， the operation logic and the operating status of the hardware actuator of the unmanned sailboat under specific working conditions are verified. It provides a reference scheme for solving potential vulnerabilities before practical engineering application.The system is composed of the existing “Seagull” unmanned sailboat electronic control system and Python to build a host mechanism， and the communication between the upper and lower computers is carried out by the agreed communication command protocol. Support fixed format configuration files and test data files， and design test conditions， processes and contents by writing file content. Based on the sea trial data of the “Seagull” unmanned sailboat， the orientation， fixed point and trajectory tracking operation logic of the unmanned sailboat under different working conditions have been verified. The system has low cost， flexible test methods， low test environment requirements and no need to purchase additional equipment， which has considerable application scenarios in actual engineering.

Helmet Wearing Detection in Electric Power Field Based on#br# Millimeter-wave Radar and Visual Fusion

CHEN Liang, LI Cheng, YI Wei, XIONG Wei, WANG Xiaofan, TANG Haidong

2024, 0(12): 100-107. doi:10.3969/j.issn.1006-2475.2024.12.015

Asbtract ( 35 )

PDF (1785KB) ( 23 )

References | Related Articles | Metrics

The failure of power site operators to wear safety helmets is one of the important causes of safety accidents. In order to prevent the recurrence of similar accidents， using millimeter-wave radar and vision fusion technology， a set of intelligent recognition algorithms that can detect whether workers are wearing safety helmets in real time based on edge computing equipment has been developed. Firstly， the conversion relationship between the coordinate systems is calculated and the joint calibration is carried out to achieve spatial fusion， Secondly， synchronization of radar and visual data is realized by timestamp matching. Then， the millimeter-wave radar data is preprocessed and the image region of interest （ROI） is calculated; finally， based on the YOLO v5 model， the improved lightweight network ShuffleNetv2 is used as the backbone network and the loss function is replaced to improve the network operation speed， and the personnel wearing safety helmets are detected in the ROI. The experimental platform was built in the electric power field and the algorithm was compared with the existing pure vision scheme. The results showed that the proposed method was slightly improved compared with the existing advanced methods in terms of detection accuracy， and greatly improved compared with the pure vision scheme in terms of real-time performance， which could realize real-time detection at the operation site.

Research Progress in Ultra Short-term Power Load Forecasting Technology

WU Xiuling1, ZHOU Sheng1, WANG Chunjuan1, YU Cuizhuo2, LIU Hao3

2024, 0(12): 108-115. doi:10.3969/j.issn.1006-2475.2024.12.016

Asbtract ( 37 )

PDF (573KB) ( 30 )

References | Related Articles | Metrics

Ultra-short term forecasting has wide applications in various fields. Accurate prediction of power load in the ultra-short term is of great significance for real-time scheduling and resource allocation in the power sector. Effective power scheduling enhances the consumer experience while avoiding resource wastage. With the increasing diversification of power consumption structure in our country， accurately predicting power load has become a challenging task. This article introduces the application scenarios of ultra-short term power load forecasting， as well as the current difficulties and challenges faced. From a technical perspective， the mainstream methods for ultra-short term power load forecasting are categorized into traditional forecasting methods， intelligent forecasting methods， and combination forecasting methods. Each category is further divided and classified based on implementation approaches. Furthermore， the principles of some representative models within each category are explained. Finally， the advantages and limitations of these three types of methods are compared and summarized. A table is provided to intuitively demonstrate the characteristics of some mentioned models. Reasonable suggestions for future research directions in ultra-short term power load forecasting are also proposed.

SDN Failure Detection and Recovery Scheme Based on State-aware Data Plane

XIAO Junbi, QIU Yi

2024, 0(12): 116-123. doi:10.3969/j.issn.1006-2475.2024.12.017

Asbtract ( 19 )

PDF (649KB) ( 26 )

References | Related Articles | Metrics

The processing strategy when a node or link fails in an SDN network is usually based on static flow table rules. Due to the limitation of the separation of the SDN data plane and control plane architecture， it is not allowed to store any network flow state information in the data plane， which results in the lack of autonomous fault decision-making capability of the data plane， and to address this problem， a state-aware data plane-based SDN fault detection and recovery scheme， SAFDR， is proposed. SAFDR defines the relationship between data flow processing primitives and state transitions in the data plane， and forms a state-aware data plane fault detection and recovery method to improve the flexibility of data plane processing and adaptive capability in case of faults.The link quality analysis model proposed in SAFDR dynamically updates the set of standby links according to the network state， which can efficiently solve the link congestion problem caused by the dynamic network changes. In this paper， we use simulation experiments to demonstrate that SAFDR meets the requirements of carrier-grade fault recovery time and has higher link utilization than other fault recovery solutions.

Table of Content