Computer Science

Performance Skeleton Analysis Method Towards Component-based Parallel Applications

FU Tian-hao, TIAN Hong-yun, JIN Yu-yang, YANG Zhang, ZHAI Ji-dong, WU Lin-ping, XU Xiao-wen

Computer Science. 2021, 48 (6): 1-9. doi:10.11896/jsjkx.201200115

Abstract

PDF(2185KB) ( 1352 )

References | Related Articles | Metrics

Performance skeleton analysis technology (PSTAT) provides input parameters for performance modeling of parallel applications by describing the program structure of parallel applications.PSTAT is the basis of performance analysis and performance optimization for large-scale parallel applications.Aiming at a kind of component-based parallel applications in the field of numerical simulation,based on the dynamic and static application structure analysis technology oriented to general program binary file,this paper proposes and implements an automatic performance skeleton generation method based on “component-loop-call” tree.On this foundation,a performance skeleton analysis toolkit CLCT-STAT(Component-Loop-Call-Tree SkeleTon Analysis Toolkit) is developed.This method can automatically identify the function symbols of component class members in component-based applications,and generate the performance skeleton of parallel application with component as the smallest unit.Compared with the method of manual generation of performance skeleton by analytical modeling,the proposed method can provide more program structure information and save the cost of manual analysis.

Adaptive Tiling Size Algorithm for 3D Stencil Computation on SW26010 Many-core Processor

ZHU Yu, PANG Jian-min, XU Jin-long, TAO Xiao-han, WANG Jun

Computer Science. 2021, 48 (6): 10-18. doi:10.11896/jsjkx.200700059

Abstract

PDF(2242KB) ( 1302 )

References | Related Articles | Metrics

Stencil computation is an important part of scientific computing and large-scale applications.Tiling is a widely-used technique to explore the data locality of Stencil computation.In the existing methods of 3D Stencil optimization on SW26010,time tiling is rarely used and manual tuning is needed for tiling size.To solve this problem,this paper introduces time tiling method and proposes an adaptive tiling size algorithm for 3D Stencil computation on SW26010 many-core processor.By establishing a performance analysis model,we systematically analyze the influence of tiling size to the performance of 3D Stencil computation,identify the performance bottleneck and guide the optimization direction under the hardware resource constraints.Based on the performance analysis model,the adaptive tiling size algorithm provides the predicted optimal tiling size,which can be helpful to deploy 3D Stencil rapidly on SW26010 processor.3D-7P Stencil and 3D-27P Stencil are selected for experiment.Compared with the result lacking of time tiling,the speedup rates of the above two examples with optimal tiling size given by our algorithm can reach 1.47 and 1.29,and the optimal tiling size in experiment is consistent with that given by our model,which verify the proposed performance analysis model and tiling size adaptive algorithm.

List-based Software and Hardware Partitioning Algorithm for Dynamic Partial Reconfigurable System-on-Chip

GUO Biao, TANG Qi, WEN Zhi-min, FU Juan, WANG Ling, WEI Ji-bo

Computer Science. 2021, 48 (6): 19-25. doi:10.11896/jsjkx.200700198

Abstract

PDF(2039KB) ( 970 )

References | Related Articles | Metrics

Parallel computing is an important means to improve the utilization rate of system resources.More and more systems on multi-processor chip meet the requirements of different computing tasks by integrating processors with different functional characteristics.A heterogeneous multiprocessor system-on-chip (DPR-HMPSoC) with dynamic partial reconfigurability is widely used because of its good parallelism and high computing efficiency,while the software/hardware partitioning algorithm with low complexity and high solving performance is an important guarantee for giving full play to its computational performance advantages.The existing related software/hardware partitioning algorithms have high time complexity and insufficient support for the DPR-HMPSoC platform.In response to the above problems,this paper proposes a list heuristic software/hardware partitioning and scheduling algorithm.By constructing a scheduling list based on task priority,a series of operations such as task scheduling,mapping and FPGA dynamic partial reconfigurable area partitioning are completed.It introduces software application mode-ling,computing platform modeling,and the detailed design scheme of the proposed algorithm.The simulation experiment results show that the proposed algorithm can effectively reduce the solution time compared with the MILP and ACO algorithms,and the time advantage is proportional to the task scale.In terms of scheduling length,the average performance of the proposed algorithm is improved by about 10%.

Implementation of Transcendental Functions on Vectors Based on SIMD Extensions

LIU Dan, GUO Shao-zhong, HAO Jiang-wei, XU Jin-chen

Computer Science. 2021, 48 (6): 26-33. doi:10.11896/jsjkx.200400007

Abstract

PDF(1971KB) ( 1599 )

References | Related Articles | Metrics

The basic mathematical function library is a critical soft module in the computer system.However,the long vector transcendental function on the domestic Shenwei platform can only be implemented indirectly by cyclic utilizing the system scalar function currently,thus limiting the computing capability of the SIMD extensions of Shenwei platform.In order to solve this problem effectively,this paper implements the long vector transcendental function based on lower-level optimization of SIMD extensions of Shenwei platform and proposes the floating-point computing fusion algorithm for solving the problem that the two-branch structure algorithm is difficult to vectorize.It also proposes the implementation method of higher degree polynomials based on the dynamic grouping of Estrin algorithm,which improves the pipelining performance of polynomial assembly evaluation.This is the first time to implement the long vector transcendental function library on the Shenwei platform.The providedfunction interfaces include trigonometric functions,inverse trigonometric functions,logarithmic functions,exponential functions,etc.The experimental result shows that the maximum error of double precision version is controlled below 3.5ULP (unit in the last place),and the maximum error of single precision version is controlled below 0.5ULP.Compared with the scalar function of Shenwei platform,the performance is significantly improved,and the average speedup ratio is 3.71.

Implementation and Optimization of Floyd Parallel Algorithm Based on Sunway Platform

HE Ya-ru, PANG Jian-min, XU Jin-long, ZHU Yu, TAO Xiao-han

Computer Science. 2021, 48 (6): 34-40. doi:10.11896/jsjkx.201100051

Abstract

PDF(2223KB) ( 1284 )

References | Related Articles | Metrics

The Floyd algorithm for finding shortest paths in a weighted graph is a key building block which is used frequently in a variety of practical applications.However,the Floyd algorithm cannot scale to large-scale graphs due to its time complexity.Its parallel implementations for different architectures are thus proposed and have been proved effective.To address the mismatching between existing ineffective parallel implementation of the Floyd algorithm and domestically designed processors,this paper implements and optimizes the Floyd algorithm targeting the Sunway platform.More specifically,this paper implements the algorithm using the programming model designed for the heterogeneous architecture of the Sunway TaihuLight,and captures the performance bottleneck when executed on the target.This paper next improves the performance of the Floyd algorithm by means of algorithmic optimization,array partitioning and double buffering.The experimental results show that the implementation of the Floyd algorithm on the Sunway platform can achieve the highest speedup of 106X over the sequential version executed on the managing processing element of the SW26010 processor.

Automatic Porting of Basic Mathematics Library for 64-bit RISC-V

CAO Hao, GUO Shao-zhong, LIU Dan, XU Jin-chen

Computer Science. 2021, 48 (6): 41-47. doi:10.11896/jsjkx.201200058

Abstract

PDF(1983KB) ( 882 )

References | Related Articles | Metrics

Subject to the core technology and intellectual property rights and other objective conditions,the research and development of domestic independent chip is highly restricted.RISC-V has the advantages of simplicity and modularity as an open source instruction set architecture(ISA),and it will become a new choice of domestic processor.As one of the most basic core software libraries of computer system,basic mathematics library is particularly important to the software ecological construction and healthy development of domestic processors.However,RISC-V has no relevant basic mathematics library at present.Therefore,this paper aims at porting basic mathematics library based on domestic Shenwei processor to the 64-bit RISC-V platform.In order to solve the problem of efficient transportation of the library,an automatic porting framework is designed at first,which can achieve high scalability through loose coupling between functional modules.Secondly,based on the characteristics of 64-bit RISC-V ISA,a global active register allocation method and a hierarchical instruction selection strategy are proposed.Finally,the framework is applied to bring about the transportation of some typical functions in the Shenwei basic mathematics library.Test results show that the ported functions are working correctly and the performance is improved compared with GLIBC.

Analysis of Topics on Database Systems in Stack Overflow

LIU Yun-han, SHA Chao-feng, NIU Jun-yu

Computer Science. 2021, 48 (6): 48-56. doi:10.11896/jsjkx.200800217

Abstract

PDF(1824KB) ( 887 )

References | Related Articles | Metrics

Database management system has been a more mature software system,but software developers still encounter a variety of problems when using database systems to manage or analyze data.They would access Stack Overflow or other CQA forums to seek solutions.In this paper,94473 database related questions are obtained on Stack Overflow.Applying the LDA topic model on the dataset and grouping these questions into 25 topics,the results show that the developers’ questions can be classified as “table”“SQL” and “SELECT” etc.By studying the prevalence and difficulty of different database-related topics,it is found that a topic such as “SQL” is more popular.In addition,three different databases MySQL,Oracle and MongoDB are also studied,and the topic distribution of questions related to different database systems is analyzed in this paper.The findings of this paper will help to understand the challenges faced by database developers and thus provide suggestions for updating database system versions,design of database courses and even research questions in the field of database.

Study on Why-not Problem in Skyline Query of Road Network Based on Query Object

ZHU Run-ze, QIN Xiao-lin, LIU Jia-chen

Computer Science. 2021, 48 (6): 57-62. doi:10.11896/jsjkx.200700016

Abstract

PDF(1960KB) ( 736 )

References | Related Articles | Metrics

With the rapid development of information technology,data becomes an important strategic resource.How to use big data to query is the research content of many scholars.At the same time,when the queried objects are not selected,how to use big data to meet the query requirements of users has also become an important research direction.Based on the analysis of the shortcomings of the existing algorithms,according to the characteristics of the real life queries,this paper studies the why-not problem in the Skyline query of the road network based on the query object,and puts forward the attribute optimization algorithm for this problem.The algorithm includes modifying the spatial and non spatial attributes of why-not point,as well as modifying the location of query center.Considering the actual situation,the time attribute is considered separately rather than simply as one dimension of non spatial attribute.The algorithm adopts pruning strategy to improve the efficiency.Finally,the real road network data and the generated interest point data set are used for comparative experiments.The results show that the method of modifying the spatial and non spatial attributes at the same time in a specific period of time can effectively solve this problem.

Data Analysis of OpenReview

ZHANG Ming-yang, WANG Gang, PENG Qi, ZHANG Yan-feng

Computer Science. 2021, 48 (6): 63-70. doi:10.11896/jsjkx.200500138

Abstract

PDF(2426KB) ( 1316 )

References | Related Articles | Metrics

The current academic evaluation system in China has been criticized for several years.It is crucial to build a fair,impartial and open academic evaluation system for creating a good academia environment.In recent years, the emergence of OpenReview,an open review website for academic papers,has brought a new idea to the evaluation of academic papers.It employs double-blind review process and makes all the submissions and reviews publicly accessible,which strengthens the supervision of the review process and improves paper review’s fairness and openness,and makes OpenReview widely used in top AI conferences.This paper collects 5527 submissions and their 16853 reviews from the OpenReview platform and performs several big data analysis tasks.It mainly focuses on the submissions from Chinese scholars and the reviews written by Chinese scholars,and obtains seve-ral interesting results.These results are helpful for understanding the characteristics of Chinese scholars and can provide insightful suggestions to improve our academic evaluation system.

Improved KNN Time Series Analysis Method

HUANG Ming, SUN Lin-fu, REN Chun-hua , WU Qi-shi

Computer Science. 2021, 48 (6): 71-78. doi:10.11896/jsjkx.200500044

Abstract

PDF(2767KB) ( 1796 )

References | Related Articles | Metrics

Recently,with the rise of data mining and machine learning,the research about time series analysis has become more and more abundant.As a classic method of machine learning,KNN(K-Nearest Neighbor)is widely used in various fields of time series analysis due to its simplicity and high prediction accuracy.However,the original KNN algorithm has some limitations in predicting time series.The prediction effect of directly using Euclidean distance as a measure of similarity is not ideal,and it cannot adapt to the prediction of time series with overall trends.This paper proposes an improved KNN algorithm named TSTF-KNN(Time Series Trend Fitting KNN).It improves the effect of KNN similarity measurement by normalizing the feature sequence at each moment,so that it can search for similar feature sequences more effectively.In addition,this paper adds error terms to the prediction result to adjust the prediction result so that it can predict the result more effectively.In order to verify the effectiveness of the method,this paper selects 4 public data sets from the kaggle public data sets,and preprocesses the 4 data sets to obtain 5 time series for the experiment.Then,this paper uses TSTF-KNN,KNN,single-layer LSTM(Long Short-Term Memory) neural network and ANN(Artificial Neural Network) to perform prediction experiments on 5 processed time series,analyze the prediction results,and compare the mean square error(MSE),which verifies the effectiveness of this method.Experimental results show that this method can effectively improve the accuracy and the stability of the KNN regression method for time series prediction,so that it can better adapt to the prediction scenarios of time series with overall trend changes.

Kernel-preserving Embedding Based Subspace Learning

HE Wen-qi, LIU Bao-long, SUN Zhao-chuan, WANG Lei, LI Dan-ping

Computer Science. 2021, 48 (6): 79-85. doi:10.11896/jsjkx.200900014

Abstract

PDF(2305KB) ( 786 )

References | Related Articles | Metrics

Subspace learning is an important research subject in the field of feature extraction.It maps the original data into a low-dimensional subspace through a linear or nonlinear transformation,and preserves the geometric structure and useful information of the original data as much as possible in this subspace.The performance of subspace learning mainly depends on the design of similarity measure and the graph construction for feature embedding.Aiming at the two issues,a novel kernel-preserving embedding based subspace learning(KESL) method is proposed,which can adaptively learn the similarity information from data and construct the kernel-preserving graph.First,to tackle the problem that the traditional dimension reduction methods cannot preserve the inner structure of high-dimensional nonlinear data,our algorithm introduces the kernel function and minimizes the reconstruction error of samples,which is beneficial for mining the data structural relationship for classification.Then,aiming at the limitation that existing graph-based subspace learning methods mainly concern the similarity information of the samples within a class,our algorithm uses the learned similarity matrices to construct intra-class and inter-class graphs,respectively.Thus,in the projected subspace,the kernel-preserving relationship of the samples in the same class can be strengthened,while the kernel-preserving relationship of the samples from different classes can be largely inhibited.Finally,through the joint optimization of kernel preserving matrix and graph embedding,the desired projection under the optimal representation can be dynamically solved.Expe-rimental results on several datasets show that the proposed algorithm is competitive to the state-of-the-art subspace learning algorithms in various classification tasks.

Kernel Subspace Clustering Based on Second-order Neighbors

WANG Zhong-yuan, LIU Jing-lei

Computer Science. 2021, 48 (6): 86-95. doi:10.11896/jsjkx.200800180

Abstract

PDF(2834KB) ( 850 )

References | Related Articles | Metrics

The processing of high-dimensional data sets is the focus of computer vision.Subspace clustering is one of the most widely used methods to achieve high-dimensional data clustering.The traditional subspace clustering assumes that the data comes from different linear subspaces,and different subspace regions do not overlap.However,real data often do not meet these two constraints,which affects the effect of subspace clustering.In order to deal with these two problems,this paper introduces a kernelized subspace to solve the nonlinear problem of subspace data,and introduces the second-order neighbors of the subspace coefficient matrix to deal with the overlapping subspace problem.Then a three-step clustering algorithm based on second-order neighbors of the kernelized subspace is designed.Firstly,the self-similarity coefficients of the kernelized subspace data are obtained.Secondly,the overlapping regions of the subspaces are eliminated.Finally,the coefficient matrix is spectrally clustered.In this paper,the designed subspace clustering algorithm is first tested on three artificial data sets,and then the experiment is performed on 12 real data sets,including face,scene characters and biomedical data sets.Experimental results show that the proposed algorithm has certain advantages over the latest algorithms.

Cauchy Non-negative Matrix Factorization for Data Representation

DUAN Fei, WANG Hui-min, ZHANG Chao

Computer Science. 2021, 48 (6): 96-102. doi:10.11896/jsjkx.200700195

Abstract

PDF(2105KB) ( 720 )

References | Related Articles | Metrics

As an important matrix factorization model,non-negative matrix factorization(NMF) is widely used in the fields of data mining and machine learning.It is often used to extract low-dimensional,sparse,and meaningful features from a collection of non-negative data vectors.Although effective in some scenarios,standard NMF utilizes squared Frobenius norm to quantify the reconstruction residual,which make it very sensitive to non-gaussian noises and outliers.As many real data inevitably contain various kind of noises,it is desirable to use robust version of NMF which is insensitive to non-gaussian noise and outliers.This paper proposes to use the Cauchy function to measure the quality of approximation instead of squared Euclidean distance for each sample and take the dependencies between different feature dimensions into account.Based on the theory of half-quadratic programming,this paper derives multiplicative updating rules to solve the proposed model effectively.To verify the effectiveness of the proposed approach,extensive unsupervised clustering experiments are conducted on several benchmark face image datasets.The experimental results show that the proposed model is robust to head poses variations,lighting,and emotion changes.Further,our model achieves consistently good performance when the parameter c varies in a large range on all three benchmark datasets.

Learning Global Guided Progressive Feature Aggregation Lightweight Network for Salient Object Detection

PAN Ming-yuan, SONG Hui-hui, ZHANG Kai-hua, LIU Qing-shan

Computer Science. 2021, 48 (6): 103-109. doi:10.11896/jsjkx.200600068

Abstract

PDF(3655KB) ( 1062 )

References | Related Articles | Metrics

To solve the problems of insufficient feature fusion and redundant models in salient object detection algorithms,this paper proposes a novel globally guided progressive feature aggregation network for lightweight salient object detection.Firstly,the lightweight feature extraction network MobileNetV3 is used to extract different levels of features of the image.Then,the lightweight multi-scale receptive field enhancement module is applied to further enhance the global representation of the highestlevel feature extracted by MobileNetV3.Finally,the progressive feature aggregation module is utilized to progressively fuse high-level and low-level features from top to bottom and the common cross entropy loss function is used to optimize these fused features in multiple stages,so as to obtain the saliency maps from coarse to fine.The whole network is an absolute end-to-end framework without any pre-processing and post-processing.Extensive experiments on six benchmark datasets demonstrate the superiority of the proposed method against other 10 methods in terms of metrics such as PR Curve,F-measure,S-measure and MAE.At the same time,the model is only about 10MB and can run at a speed of 46FPS on a GTX2080Ti GPU when processing a 400×300 image.

Temporal Consistency Preserving Multi-Mask Sparse Deep Representation for Visual Tracking

GUO Wen, YIN Tong-ling, ZHANG Tian-zhu, XU Chang-sheng

Computer Science. 2021, 48 (6): 110-117. doi:10.11896/jsjkx.200800212

Abstract

PDF(4834KB) ( 1009 )

References | Related Articles | Metrics

Building a model that can not only fully consider the discriminability of the object appearance,but also keep the temporal consistency of the features in tracking process is the key to solve the tracking problem.In order to improve the discrimination of feature representation and alleviate the degradation of feature in tracking process,a novel temporal consistency preserving multi-mask sparse deep representation method for visual tracking is proposed in the paper.Firstly,multi-task sparse deep expression learning method is constructed by using different feature attributes of deep convolution features on different layers to fully explore the correlation of multi-source information.Secondly,the temporal consistency constrained regularization term is constructed by the residual of relevant frames,which can compensate for the degradation of tracking process features and improve the temporal consistency of tracking features.Numerous experimental results on Benchmark show that this algorithm has better tracking effectiveness and stability than the current state-of-the-art methods in complex background,fast motion and other situations.

Crowd Counting Method Based on Cross-column Features Fusion

LI Jia-qian, YAN Hua

Computer Science. 2021, 48 (6): 118-124. doi:10.11896/jsjkx.200700107

Abstract

PDF(2418KB) ( 821 )

References | Related Articles | Metrics

Crowd counting is a challenging subject in computer vision and machine learning.Due to the phenomenon of crowd scale change and scene occlusion,the counting accuracy is low.A crowd counting method based on cross-column features fusion,called cross-column features fusion network(CCFNet),is proposed in this paper.CCFNet fuses features from multiple columns and different receptive fields,and combines with the dilate convolution employing coprime expansion rate.Therefore,CCFNet can not only increase the receptive field but also ensure the continuity of information,so as to adapt to the huge changes in the crowd size better.At the same time,the attention model is introduced to guide the network to focus on the head position in the images.According to the attention score graph,different weights are assigned to different positions to highlight the crowd and weaken the background.Finally,a high-quality density map is obtained.In comparative experiments on the current mainstream population counting datasets,the mean absolute error(MAE) reaches 63.2 and 8.9 on the A and B subsets of the ShanghaiTech dataset,222.1 on the UCF_CC_50 dataset,and 7.1 on the WorldExpo’10 dataset.The results show that the proposed method has better counting accuracy and can adapt to different scenes.Especially for scenes with large scale variation,its effect is better than most of the pre-vious algorithms.

Generation of Realistic Image from Text Based on Feature Fusion

XU Ze, SHUAI Ren-jun, LIU Kai-kai, MA Li, WU Meng-lin

Computer Science. 2021, 48 (6): 125-130. doi:10.11896/jsjkx.200400107

Abstract

PDF(2804KB) ( 1153 )

References | Related Articles | Metrics

Recent challenging task of synthesizing images from text descriptions based on the generative adversarial network(GAN) has shown encouraging results.These methods can produce images with general shapes and colors,but often produce global images with unnatural local details and distortions.This is due to the inefficiency of the convolutional neural network in capturing high-level semantic information for pixel-level image synthesis and the fact that the generator-discriminator in a rough state generates flawed results for lack of detail,which then serves as input to the final result.We propose a generative adversarial network based on feature fusion,which introduces multi-scale feature fusion by embedding residual block feature pyramid structure,generates the final fine image directly by adaptive fusion of these features,and produces a 256px×256px realistic image with only one discriminator.The proposed method is verified on the flower data set Oxford-102 and Caltech bird database CUB,and the quality of generated images is evaluated by using Inception Score and FID.The results show that the quality of the generated images produced by the proposed method is better than images produced by some classical methods.

Image Recognition with Deep Dynamic Joint Adaptation Networks

LIU Yu-tong, LI Peng, SUN Yun-yun, HU Su-jun

Computer Science. 2021, 48 (6): 131-137. doi:10.11896/jsjkx.210100008

Abstract

PDF(2300KB) ( 1051 )

References | Related Articles | Metrics

Compared with the traditional image recognition methods,the depth network can extract the features with better representational ability,so as to obtain better recognition effect.In reality,most of the data provided by tasks are unlabeled or partially labeled,which makes it difficult for deep network to learn.The knowledge learned from the source domain is used for the learning of the target domain by means of transfer learning,which can alleviate this problem.In order to overcome the image-data diffe-rence between the source domain and the target domain in the transfer process,an image recognition method based on deep dyna-mic joint adaptation networks is proposed.During the training of the transfer networks,the dynamic joint adaptation method is used to realize the data distribution adaptation in the multi-layer network structure.Then the entropy minimization principle is used for the target classifier to pass through the low-density area of the target domain.At last,the image classification and recognition are realized.The experimental results show that,with this method,the average accuracy of the three transfer tasks based on 24 kinds of plant disease provided by the 2018 AI challenge competition are 97.27%,94.25% and 93.66%,which are better than other algorithms.A large number of empirical results show that the transfer learning framework based on the deep networks,meanwhile,using dynamic joint adaptation and entropy minimization principle can recognize images accurately.

Edge Detection in Images Corrupted with Noise Based on Improved Nonlinear Structure Tensor

SONG Yu, SUN Wen-yun

Computer Science. 2021, 48 (6): 138-144. doi:10.11896/jsjkx.200600017

Abstract

PDF(4377KB) ( 839 )

References | Related Articles | Metrics

The performance of existing edge detection methods in images corrupted by noise is not satisfying.Aiming at the edge detection problem in images corrupted by moderate noise,an image edge detection method based on improved nonlinear structure tensor using steering kernel is proposed.First the tensor products of the noisy image are computed.Then the tensor products are diffused according to the derivatives of the image which depends on the tensor product itself.The diffusivity matrix in the diffusion equation is composed of the tensor products which are spatially adaptive averaged using a steering kernel instead of isotropic filtered using a gaussian kernel.Finally,the eigenvalues and eigenvectors of the diffused tensor products are computed in order to detect the image edge.Experimental results show that,compared with image edge detection methods based on linear structure tensor,nonlinear structure tensor diffused according to the derivatives of the tensor,nonlinear structure tensor diffused according to the derivatives of the image,the proposed method can get clearer edges with smaller amount of noise.

Image Shadow Removal Algorithm Based on Generative Adversarial Network

SHI Heng, ZHANG Ling

Computer Science. 2021, 48 (6): 145-152. doi:10.11896/jsjkx.200900109

Abstract

PDF(4425KB) ( 1034 )

References | Related Articles | Metrics

Although the existing learning-based shadow removal methods have made some progress,these methods mainly focus on the image itself and do not well explore other information related to shadow.These methods often result in problems such as image texture blurring or illumination inconsistency.To solve these problems,this paper proposes a new network based on genera-tive adversarial network (GAN) to remove shadows in the image.First,it uses an encoder-decoder to obtain a coarse shadow removal result.Then,it optimizes the coarse result by utilizing the shadow-related residual information to produce more realistic and natural shadow removal image.Our generator contains three encoder-decoder structures.The first encoder-decoder is used to restore the illumination of the image and generate a coarse shadow removal result.To solve the problem of color and illumination inconsistency,the residual information is input into the following encoder-decoder to correct the coarse results.Furthermore,the third encoder-decoder is used to refine the details in the image,which can avoid texture inconsistency between shadow regions and nonshadow regions.The discriminator is used to identify the authenticity of the image shadow removal result.Experiments show that the proposed method achieves the best RMSE value in both shaded region and non-shaded region,and effectively solves the problem of texture blur.In addition,the shadow removal image generated by the proposed method is much closer to the real image,which proves the effectiveness and feasibility of the proposed method.

Method of CNN Flag Movement Recognition Based on 9-axis Attitude Sensor

ZHONG Yue, FANG Hu-sheng, ZHANG Guo-yu, WANG Zhao, ZHU Jing-wei

Computer Science. 2021, 48 (6): 153-158. doi:10.11896/jsjkx.200500005

Abstract

PDF(3339KB) ( 967 )

References | Related Articles | Metrics

Different from the traditional method of flag movement recognition of optical fiber sensor,image recognition and kinect depth image,this paper proposes a method of flag movement recognition based on 9-axis attitude sensor.The data of 3-axis acce-leration,3-axis angular velocity and 3-axis magnetic decrement are collected by wearing a 9-axis attitude sensor at the wrist;based on the CNN classification model,the algorithm of data preprocessing and classification recognition is improved;in the data preprocessing stage,the wavelet decomposition and reconstruction functions are used to carry out the high-frequency denoising and low-frequency information extraction of the collected 9-axis data,and the dimension and the length of each action sample are unified through time series windowing and segmentation;in the feature extraction stage,the constructed network models of double convolution layer,single pooling layer and single full connection layer are used to extract the features of the reconstructed data;in the stage of classification and recognition,a CrossEntropy-Logistic joint loss function is proposed to carry out iterative training for 5 actions.The experimental results show that the use of wavelet decomposition and reconstruction detcoef function coefficient of low frequency detail of signals is extracted by using one-dimensional CNN data feature extraction.The training and testing accuracy obtained by the fusion analysis of the predicted loss value and the predicted probability through CL joint loss function is the highest in comparison with various methods.The average training recognition rate can reach more than 99% and the testing accuracy can reach 94%.

Frontiers in Neural Question Generation:A Literature Review

QIU Jia-zuo, XIONG De-yi

Computer Science. 2021, 48 (6): 159-167. doi:10.11896/jsjkx.201100013

Abstract

PDF(2747KB) ( 1610 )

References | Related Articles | Metrics

Question generation means that the machine actively asks a natural language question by given a passage.Neural question generation is trained in a completely end-to-end training mode,using neural networks to convert documents and answers to questions,which is an emerging and important research direction in natural language processing.This paper first gives a brief introduction to neural question generation,including basic concepts,mainstream frameworks,and evaluation methods.Then,it introduces the key issues of question generation,including input modeling,long document processing,multi-task learning,and the application of machine learning,other issues and improvements.Finally,it introduces the relationship between question generation and question answering,as well as future research of question generation.

Meta-reinforcement Learning Algorithm Based on Automating Policy Entropy

LU Jia-you, LING Xing-hong, LIU Quan, ZHU Fei

Computer Science. 2021, 48 (6): 168-174. doi:10.11896/jsjkx.200600133

Abstract

PDF(2263KB) ( 1133 )

References | Related Articles | Metrics

Traditional deep reinforcement learning methods rely on a large number of samples and are difficult to adapt to new tasks.By extracting prior knowledge from previous training tasks,meta reinforcement learning provides a fast and effective me-thod for agents to adapt to new tasks.Meta deep reinforcement learning based on maximum entropy reinforcement learning framework optimizes strategies by maximizing expected reward and strategy entropy.However,the current meta reinforcement learning algorithms based on the maximum entropy reinforcement learning framework generally adopt fixed temperature parameters,which is unreasonable in the multi-task scenario of meta reinforcement learning.To solve this problem,an adaptive adjustment strategy entropy algorithm is proposed.Firstly,by limiting the entropy of the strategy,the original objective function optimization problem is transformed into a constrained optimization problem.Then,the dual variable in the constrained optimization problem is taken as the temperature parameters,and the updated formula is obtained by solving the dual variable by Lagrangedualmethod.According to the updated formula,the temperature parameters will be adjusted adaptively after each round of meta trai-ning.Experimental data show that the average score of the proposed algorithm on Ant -Fwd-back and Walker-2D increases by 200,the meta training efficiency improves by 82%,the strategy convergence on Human-Direc-2D requires 230 000 training steps,and the convergence speed increases by 127%.Experimental results show that the proposed algorithm has higher meta training efficiency and better stability.

Intelligent Prediction Model of Tool Wear Based on Deep Signal Processing and Stacked-ResGRU

HU De-feng, ZHANG Chen-xi, WANG Shi-tao, ZHAO Qin-pei, LI Jiang-feng

Computer Science. 2021, 48 (6): 175-183. doi:10.11896/jsjkx.210100101

Abstract

PDF(2491KB) ( 1371 )

References | Related Articles | Metrics

Intelligent monitoring of tool wear is an important factor affecting the intelligent development of modern machinery industry.Most machining tools collect signals through sensors and establishes their relationship to obtain the tool wear prediction without interrupting the processing process.Automatic tool change or alarm will be carried out according to whether the wear threshold is reached to realize online intelligent monitoring.It is an urgent problem to extract effective feature information from sensor signals and build a model which can predict the tool wear quickly and accurately.Therefore,based on the above problems,this paper proposes a tool wear prediction method based on deep signal processing and Stacked-ResGRU.In signal processing,the BiGRU-Self Attention(BGSA)network combing bidirectional GRU and self-attention is designed to obtain dynamic temporal features and reflect the influence degree of different features,which can improve modeling efficiency.At the same time,the Stacked-ResGRU model is proposed to predict the real-time tool wear value,and the network structure is optimized by residual structure to accelerate the convergence speed of the model.Experimental studies for tool wear prediction in a dry milling operation are conducted to demonstrate the viability of this method.Through the experimental results and comparative analysis,the proposed method can effectively characterize the tool wear degree,greatly reduce the prediction errors,and achieve a precise prediction effect.

Fault Prediction Method Based on Improved RNN and VAR for Ship Equipment

ZENG You-yu, XIE Qiang

Computer Science. 2021, 48 (6): 184-189. doi:10.11896/jsjkx.200700117

Abstract

PDF(3103KB) ( 942 )

References | Related Articles | Metrics

Aiming at the problem that the existing multivariable time series prediction methods cannot be applied to the multi-sensor fault prediction of ships,an improved recurrent neural network and vector autoregressive fault prediction method for ships equipment is proposed.This method can not only learn the interdependence of multiple variables and the long-term dependence of time series,but also help to reduce the insensitivity of traditional neural network to the input scale of time series prediction.Firstly,the data of normal state and fault state are extracted from the ship history database and converted into the input of the supervised learning problem.Then,the complex correlation between ship variables is captured by the attention mechanism.The nonlinear and linear relationship of ship time signals are captured by inputting the output of attention mechanism into recurrent neural network and vector autoregression.Finally,the outputs of recurrent neural network components and the outputs of vector autoregressive components are processed as the final prediction results.The experimental results show that the proposed method is more stable in the training process of ship equipment fault prediction,and the root-mean-square error of the test results below 1.2.It can more accurately predict the trend of ship equipment properties and fault occurrence.

Reverse Diagnostic Method Based on Vehicle EMC Standard Test and Machine Learning

LEI Jian-mei, ZENG Ling-qiu, MU Jie, CHEN Li-dong, WANG Cong, CHAI Yong

Computer Science. 2021, 48 (6): 190-195. doi:10.11896/jsjkx.200700204

Abstract

PDF(2394KB) ( 756 )

References | Related Articles | Metrics

The rapid development of intelligent vehicles not only improves electromagnetic compatibility(EMC) testing technology,but also brings new challenges to vehicle EMC design,which is benefited from test data-oriented troubleshooting.With the increase in electronic complexity,vehicle on-board system designers should confront with more and more EMC failure possibilities and they are in need of effective EMC failure diagnosis approach.However,EMC fault diagnosis is difficult due to the distinguishing features of EMC test dataset,such as small sample,nonlinear,high dimensions,etc.In view of this situation,this paper puts forward a feature extraction algorithm for electromagnetic compatibility test data based on years of rectification experience of EMC test engineers,and uses the valuable feature data extracted from the test data to set up a support vector machine(SVM) two classification model.Corresponding application effect is displayed.In order to verify the effectiveness of the proposed method,this paper adopts the naive Bayesian classification model for comparison.The experimental results show that the proposed method can match the demand of EMC fault diagnosis for intelligent vehicles.

Text Matching Fusion Model Combining Multi-granularity Information

LYU Le-bin, LIU Qun, PENG Lu, DENG Wei-bin , WANG Chong-yu

Computer Science. 2021, 48 (6): 196-201. doi:10.11896/jsjkx.200700100

Abstract

PDF(1830KB) ( 1421 )

References | Related Articles | Metrics

Conventional text matching methods are basically divided into representational text matching models and interaction-based text matching models.Since the representation-based text matching model is easy to lose semantic focus and the interaction-based text matching model ignores global information,a text matching fusion model combining multi-granularity information is proposed in this paper.This model fuses two text matching models through interactive attention and expressing attention,and then uses convolutional neural networks to extract multiple different levels of granularity information presented in the text.Then the local important information and global semantic information can be captured.The experimental results on three different text matching tasks show that the proposed model outperform other optimal models by 5.3%,0.4%,1.5% on the NDCG@5 evaluation index respectively.By extracting multiple granularity information of the text and combining interactive attention and expressed attention,the proposed model can effectively pay attention to the text information of different levels,and solve the problem of losing semantics and ignoring global information during the text matching process in the traditional models.

Novelty Detection Method Based on Global and Local Discriminative Adversarial Autoencoder

XING Hong-jie, HAO ZhongHebei

Computer Science. 2021, 48 (6): 202-209. doi:10.11896/jsjkx.200400083

Abstract

PDF(2516KB) ( 995 )

References | Related Articles | Metrics

Generative Adversarial Nets(GAN) and Adversarial Autoencoder(AAE) have been successfully applied to image ge-neration.Moreover,adversarial network can learn the data features contained within the given samples in an unsupervised manner.However,when the conventional adversarial networks are applied to novelty detection,they may obtain poor classification results.The reasons lie in two aspects.One is that GAN belongs to generative models,while novelty detection models are usually classified as discriminative models.The other is that the existing AAEs use the intermediate vectors of autoencoder as the discriminant inputs,which makes the reconstruction outcomes for the given data are not satisfying.Therefore,an AAE based on double discriminators is proposed to make it fit for tackling the novelty detection problems.The double discriminators of the proposed model have different discriminative capabilities,i.e.,local discriminative capability and global discriminative capability.Experimental results on datasets of MNIST,Fashion-MNIST and CIFAR-10 show that the proposed method can effectively avoid the mode collapse problem that may occur during training.In addition,in comparison with its related approaches,the proposed method achieves a better detection performance.

Relative Risk Degree Based Risk Factor Analysis Algorithm for Congenital Heart Disease in Children

XU Hui-hui, YAN Hua

Computer Science. 2021, 48 (6): 210-214. doi:10.11896/jsjkx.200500082

Abstract

PDF(1656KB) ( 661 )

References | Related Articles | Metrics

The analysis of disease-related risk factors is an important part of application of data mining theory in the medical field,which is helpful for doctors to analyze causes of disease and carry out effective work of disease prevention and control.But disease data in the medical field have their own characteristics,such as high imbalance,which means that most valuable information is contained in the attribute items with a small support.It is easy to lose important information when applying the classical association rule algorithm based on the support directly.Therefore,based on the knowledge of medical field and the common statistical standard of medical field——Relative Risk,this paper proposes a mining algorithm for high relative risk itemsets(MARRI) and two corresponding pruning methods,which are interaction pruning and sample number pruning,and verifies the algorithm on the dataset of children’s congenital heart disease.Experimental results show that the algorithm is effective to mine the information in low support items and disease-related factors mined out are more valuable.

Pyramid Evolution Strategy Based on Dynamic Neighbor Lasso

ZHANG Qiang, HUANG Zhang-can, TAN Qing, LI Hua-feng, ZHAN Hang

Computer Science. 2021, 48 (6): 215-221. doi:10.11896/jsjkx.200400115

Abstract

PDF(2747KB) ( 744 )

References | Related Articles | Metrics

The optimization problem is one of the common problems in the engineering field,the essence of most engineering problems is the function optimization problem.Pyramid evolution strategy(PES) algorithm can effectively set a balance between “exploitation” and “exploration” as well as “competition” and “cooperation” when solving function optimization problems,but there are still some shortcomings,such as slow convergence speed,low accuracy,and easy to fall into a local optimal.In order to solve these shortcomings,this paper proposes a pyramid evolution strategy based on dynamic nearest neighbor lasso(DNLPES).The DNLPES algorithm adaptively controls the selection range parameters of the target individual group based on the evolution.At the same time,the Euclidean distance is used to measure the difference between individuals in the target individual group.The difference information between individuals is used to guide the cooperation between individuals,the population evolution is completed by continuously generating new individuals and eliminating the individuals with poor fitness value.The DNLPES algorithm improves the accuracy of the algorithm by making full use of the difference information between individuals in the population and enhancing the cooperation between individuals.Comparing the DNLPES algorithm and the 7 algorithms on 9 test functions,experimental result shows that the DNLPES algorithm has a certain competitiveness in solving accuracy.Compared with the stan-dard PES algorithm,the DNLPES algorithm has obvious advantages in solving accuracy and convergence speed.

Distributed Representation Learning and Improvement of Chinese Words Based on Co-occurrence

CAO Xue-fei, NIU Qian, WANG Rui-bo, WANG Yu, LI Ji-hong

Computer Science. 2021, 48 (6): 222-226. doi:10.11896/jsjkx.200900140

Abstract

PDF(1298KB) ( 1028 )

References | Related Articles | Metrics

The co-occurrence matrix of words and their contexts is the key to learning the distributed of words.Different methods can be used to measure the association between words and their contexts when constructing a co-occurrence matrix.In this paper,we firstly introduce three association measures of words and their contexts,construct corresponding co-occurrence matrices and learn the distributed representations of words under a unified optimization framework.The results on semantic similarity and word analogy show that GloVe method is the best.Then,we further introduce a hyperparameter to calibrate the co-occurrences of the words and their contexts based on the Zip’f distribution,and present a method for solving the estimated value of the hyperparameter.The obtained distributed representations of words based on the improved method indicate that the accuracy of the word analogy task has increased by 0.67%,and it is significant under the McNemar test.The correlation coefficient on the word simila-rity task has increased by 5.6%.In addition,the distributed representations of the words learned by the improved method is also applied to the semantic role identification task as the initial vector of word feature,and the F1 value obtained is also increased by 0.15%.

Parallel Pruning from Two Aspects for VGG16 Optimization

LI Shan, XU Xin-zheng

Computer Science. 2021, 48 (6): 227-233. doi:10.11896/jsjkx.200800016

Abstract

PDF(1754KB) ( 927 )

References | Related Articles | Metrics

In recent years,much of pruning for convolutional neural network is based on the norm value of the filter.The smaller the norm value is,the smaller the impact on the network after clipping.This idea can make full use of the numerical characteristics of the filter.However,it ignores the structural characteristics of the filter.Based on the above viewpoint,this paper applies AHCF(Agglomerative Hierarchical Clustering method for Filter) to vgg16.Then,a parallel pruning method from two aspects is proposed to prune the filter from both numerical and structural perspectives.This method reduces the redundant filters and the parameters in the VGG16 network.Besides,it improves the classification accuracy,meanwhile keeping the learning curve of the original network.On CIFAR10 dataset,the accuracy of the proposed method is 0.71% higher than that of the original VGG16 network.On MNIST,the accuracy of the proposed method is as good as the original network.

Heuristic Construction of Triadic Concept and Its Application in Social Recommendation

LIU Zhong-hui, ZHAO Qi, ZOU Lu, MIN Fan

Computer Science. 2021, 48 (6): 234-240. doi:10.11896/jsjkx.200500136

Abstract

PDF(1719KB) ( 758 )

References | Related Articles | Metrics

Formal concept analysis is a knowledge discovery method that has great achievements in theory and application.Recently,with the emergence of three-dimensional data,triadic formal concept analysis has been developed.However,there are few researches and applications in this field,especially it has not been applied to recommendation systems.This paper proposes an efficient triadic concept set construction method and applies it to social recommendation.Firstly,the heuristic information is designed to generate a set of triadic concepts covering all users.Heuristic information aims to construct strong concepts with a certain scale of extension and intension.Then,appropriate social relations are screened through the attributes of the proposed items,and the recommendation prediction is realized by combining the popularity of the items in the concept.Three experiments are carried out in real data set and sampled data set respectively.In the first experiment,the number of triadic concepts and running time constructed by the heuristic method and ∨_oc operation are compared respectively.The concepts constructed by the ∨_oc operation do not significantly improve the recommendation effect.The second experiment compares the accuracy,recall rate and F1 of the recommendation effect.It reveals that increasing the number of conditions can effectively improve the recommendation effect.The results of the last experiment show that the recommendation effect of the new algorithm is better than that of KNN and GRHC.

Joint Question Answering Model Based on Knowledge Representation

LIU Xiao-long, HAN Fang, WANG Zhi-jie

Computer Science. 2021, 48 (6): 241-245. doi:10.11896/jsjkx.200600011

Abstract

PDF(1516KB) ( 857 )

References | Related Articles | Metrics

Question answering system based on knowledge base aims to extract answers directly from the knowledge base by parsing users’ natural language question sentences.Currently,most knowledge based question answering models follow the two steps of entity detection and relationship recognition,but such methods ignore the structural information contained in the know-ledge base and the connection between the two tasks.In this paper,a joint question answering model based on knowledge representation is proposed.First,the knowledge representation model is used to map the entities and relationships in the knowledge base to a low-dimensional vector space,then the question sentences are embedded into the same vector space through neural network,and the entities in the question sentences are detected at the same time.The semantic similarity between knowledge base triples and question sentences is measured in the vector space,so that knowledge base embedding and multi-task learning are introduced into the task of knowledge based question answering.The experimental results show that the proposed model can greatly improve the training speed,and the accuracies of entity detection and relationship recognition task reach the mainstream level.It is proved that knowledge embedding and multi-task learning can improve the performance of knowledge based question answe-ring task.

Detecting Blocking Failure in High Performance Interconnection Networks Based on Random Forest

XU Jia-qing, HU Xiao-yue, TANG Fu-qiao, WANG Qiang, HE Jie

Computer Science. 2021, 48 (6): 246-252. doi:10.11896/jsjkx.201200142

Abstract

PDF(2404KB) ( 803 )

References | Related Articles | Metrics

High performance interconnection network is the key to high speed collaborative parallel of all nodes in high perfor-mance computer system.During the operation and maintenance of high performance interconnection networks,it is found that the network port blocking fault caused by the deterioration of link quality is difficult to locate and greatly affects the availability of the whole system.With the advent of the era of artificial intelligence,intelligent operation and maintenance has played an important role in network operation and maintenance.However,research on intelligent operation and maintenance based on high performance interconnection network is relatively few.This paper is based on a large amount of data and rich experience accumulated by operations staff in the self-development of high speed interconnection network operation and maintenance.It proposes a supervised random forest method for network blocking detection.The experimental results show that the proposed method has an ave-rage accuracy of 93.7% while maintaining an average recall rate of 95%,and can effectively solve the problem of network blocking detection.

Fault-tolerant Routing Algorithm in BCube Under 2-restricted Connectivity

YI Yi, FAN Jian-xi, WANG Yan, LIU Zhao, DONG Hui

Computer Science. 2021, 48 (6): 253-260. doi:10.11896/jsjkx.200900203

Abstract

PDF(1895KB) ( 772 )

References | Related Articles | Metrics

BCubeis a data center network with good performance.Compared to traditional tree-shaped data center network,BCube has not only higher fault-tolerance performance but also superior scalability.Currently,the research on BCube can be reduced to logical structure (a special generalized hypercube),in which the switches are regarded as transparent devices.In actual application,with the continuous increase of network scale,vertex failure has become inevitable.Therefore,it is meaningful to study the fault-tolerant routing of the network.At present,there is some research on fault-tolerant routing of BC_n,k.However,the fault-tolerant routing of BC_n,k under the 2-restricted connectivity has not been studied yet.Before proposing the algorithm,the 2-restricted connectivity of BC_n,k is firstly proved which is 3(k+1)(n－1)－2n for k≥3 and n≥3.Then a fault-tolerant routing algorithm with a time complexity of O(κ(BC_n,k)³) is proposed where κ(BC_n,k)=(k+1)(n－1) denotes the connectivity of BC_n,k.The proposed algorithm can find a fault-free path between any two distinct fault-free vertices when the number of faulty vertices is less than 3(k+1)(n－1)－2n and every fault-free vertex has at least two fault-free neighbors.

Load Balancing Mechanism for Bandwidth and Time-delay Constrained Streaming Media Server Cluster

ZHENG Zeng-qian, WANG Kun, ZHAO Tao, JIANG Wei, MENG Li-min

Computer Science. 2021, 48 (6): 261-267. doi:10.11896/jsjkx.200400131

Abstract

PDF(2135KB) ( 707 )

References | Related Articles | Metrics

Overall load capacity of streaming media server cluster is largely affected by its service delay and bandwidth load balancing.Therefore,how to improve the real-time capability of service and balance the bandwidth load are the keys to improve the streaming media server cluster service capabilities.This paper proposes a load balancing mechanism for bandwidth and time-delay constrained streaming media server cluster.Through discretizing bandwidth of server and task,the mechanism builds the server and task state sets.And it uses genetic algorithms to calculate and store the optimal allocation scheme in each state offline to speed up the online task assignment scheme calculation while effectively allocating tasks with different bandwidth requirements to each server to optimize the cluster load.Results of simulation show that the mechanism can effectively balance the bandwidth load and reduce the number of failed tasks on the basis of having a calculation delay similar to the round-robin algorithm and least connections algorithm,thereby improving the overall service quality and ability.

Cluster-based Topology Adaptive OLSR Protocol for UAV Swarm Network

SUN Yi-fan, MI Zhi-chao, WANG Hai, ZHAO Ning

Computer Science. 2021, 48 (6): 268-275. doi:10.11896/jsjkx.200500130

Abstract

PDF(2626KB) ( 1069 )

References | Related Articles | Metrics

To address the routing problem in the UAV swarm network,the OLSR protocol announces Hello and Topology Update(TC) messages in the form of fixed period to maintain network topology information.However,in the case of frequent changes in network structure,such periodic announcements cannot respond to changes in the network topology in time,resulting in significant network performance degradation.This paper proposes a new cluster-based optimized link state routing protocol(CB-OLSR),which is composed of intra-cluster and inter-cluster routing.CB-OLSR dynamically adjusts the broadcast cycle of Hello and other control messages according to the network topology changes,so as to update the network status in time and obtain better network performance.At the same time,the short and direct connection path within the cluster is optimized.If the source node and target node are neighbors within two hops,the hierarchy structure is ignored and bypass shortcut is used to forward message directly,so as to reduce the load on the cluster head and extend its life.Through the simulation experiment on EXata simulation platform,the results show that CB-OLSR is significantly better than OLSR in terms of data packet loss rate and throughput,so it is more suitable for UAV swarm network.

Dynamic Loading Algorithm for Docker Container

LIU Bang-bang, YI Guo-hong, HUANG Zu-yuan

Computer Science. 2021, 48 (6): 276-281. doi:10.11896/jsjkx.200500152

Abstract

PDF(1749KB) ( 1153 )

References | Related Articles | Metrics

To improve anti concurrency and the average response time of docker container server cluster,a DLOAD(Dynamic Loading Algorithm) for dynamically loading container server is designed.Based on the WRR load algorithm,this algorithm refe-rences the concept of real-time weight,makes up for the shortcomings of WRR algorithm in weight setting,and optimizes the load algorithm of docker container server.DLOAD algorithm will record the resource information of the server,take the ratio of container connections,CPU utilization,memory utilization,network IO,disk IO and average response time as parameters,calculate the real-time weight of container,and record the real-time weight in the weight table of the load server.After the load server queries the weight table,it calls the WRR algorithm and recommends the best Docker container server ID to load.Through the simulation experiment,the average response time and throughput of docker container server before and after the improvement are analyzed and compared.It is concluded that the improved DLOADalgorithm can improve the average response time and anti concurrency of the server more efficiently than other algorithm,and improve the performance of the container server.

Non-linear Load Capacity Model of Complex Networks

WANG Xue-guang, ZHANG Ai-xin, DOU Bing-lin

Computer Science. 2021, 48 (6): 282-287. doi:10.11896/jsjkx.200700040

Abstract

PDF(2307KB) ( 1205 )

References | Related Articles | Metrics

The study of network formation mechanism,geometric property,evolution rules,network structure analysis,behavior prediction and control gives rise to the discipline of complex network,and cascade failure process of complex network has always been concerned.This paper presents a non-linear load capacity model with two variable parameters,which is more suitable for real network,to solve the cascading failures problem of complex networks.Simulations on four different networks verify the effectiveness of the proposed model.The results show that the model can better defend against cascading failures,and has a lower investment cost and a better performance in the case of higher robustness.

Matrix Theory Aided Convergence Analysis of Consensus Behavior in FANET with Beacon Loss

HUANG Xin-quan, LIU Ai-jun, LIANG Xiao-hu, WANG Heng

Computer Science. 2021, 48 (6): 288-295. doi:10.11896/jsjkx.201000137

Abstract

PDF(2195KB) ( 883 )

References | Related Articles | Metrics

Flying Ad-Hoc Network(FANET) which is a field of wireless Ad-Hoc network formed by small unmanned aerial vehicles(UAVs),is critical in achieving UAV swarm system.Beacon mechanism in FANET plays the fundamental role in performing consensus behavior of UAV swarm.However,the wireless link failures of practical FANET will introduce beacon loss which will affect the convergence rate or convergence time,which describe how fast all UAV states reach the common value.To achieve optimal consensus behavior,it is important to know how beacon mechanism affects the consensus behavior.To solve above-mentioned problem,this paper has investigated the analytical relation between the convergence rate/time of consensus behavior and the beacon loss probability.In the analytical work,information flow topology at each period is modeled by random directed graph,and one indicator matrix weighted is designed to model the Laplacian matrix of the graph.Based on the knowledge of matrix theory and spectrum radius of a matrix,the analytical work in this paper firstly gives an analytical expression of expected consensus va-lue of the consensus process.Utilizing the expected consensus value,a novel quantification of convergence rate based on the expected final consensus value is provided.Different from existing analytical work,the convergence rate is quantified based on the expected consensus value,rather than to the average value of all states at each period.Finally,utilizing the knowledge of matrix theory and spectrum radius of a matrix,the proposed analytical work analyzes the relation between the convergence rate/time and the beacon loss probability.Simulation results show that,the proposed analytical model can accurately capture the convergence rate along with time in practical FANETs.Moreover,the proposed model can accurately capture the effect of average link failure probability on each link,initial state distribution and the number of UAVs on the convergence time.Moreover,compared with exis-ting analytical model,the proposed analytical model can capture the convergence performance in practical FANET more precisely.

Study of Cellular Traffic Prediction Based on Multi-channel Sparse LSTM

ZHANG Zheng-wan, WU Di, ZHANG Chun-jiong

Computer Science. 2021, 48 (6): 296-300. doi:10.11896/jsjkx.210400134

Abstract

PDF(2530KB) ( 886 )

References | Related Articles | Metrics

Next-generation cellular networks play an important role in network management and service provision,so that predictive analysis of mobile network traffic is becoming more and more important to our daily life.To predict the urban cellular traffic,this study designs a traffic prediction model based on multi-channel sparse LSTM.Compared with multilayer perceptron networks or other neural network structures,LSTM is very suitable for processing time series data.Therefore,the designed multi-channel method can effectively capture multi-source network traffic information,and its sparse method can adaptively assign different weights to different traffic time nodes,so that the ability of the deep neural network model to capture important features is improved.This study evaluates the performance of the proposed method with respect to the single-step and multi-step prediction using the cellular traffic data set in Milan,Italy.The experiment results show that the proposed method is more accurate than the benchmark methods.In addition,this study reports the impact of different sampling settings of cellular traffic on the storable length and accuracy of the LSTM network model.

Geographic Local Differential Privacy in Crowdsensing:Current States and Future Opportunities

WANG Le-ye

Computer Science. 2021, 48 (6): 301-305. doi:10.11896/jsjkx.201200223

Abstract

PDF(1838KB) ( 1257 )

References | Related Articles | Metrics

Geographic privacy protection is one of the key design issues in crowdsensing.Traditional protection mechanisms need to make assumptions on adversaries’ prior knowledge to ensure protection effect.Recently,a breakthrough in the privacy research community,namely local differential privacy (LDP),is introduced into crowdsensing for location protection,which can provide theoretically guaranteed protection effect regardless of adversaries’ prior knowledge,without requiring trustful third parties.This paper conducts a concise review of the works applying this new privacy-preserving technique in crowdsensing.For diverse existing Geo-LDP (geographic LDP) mechanisms that serve different crowdsensing tasks,this paper analyzes their characteristics and extracts common design considerations in practice.It also points out potential research opportunities in the future study.

Cognitive Mechanisms of Fake News

WU Guang-zhi, GUO Bin, DING Ya-san, CHENG Jia-hui, YU Zhi-wen

Computer Science. 2021, 48 (6): 306-314. doi:10.11896/jsjkx.201200194

Abstract

PDF(1426KB) ( 1571 )

References | Related Articles | Metrics

The advent of the social media era not only accelerates the flow of information,but also provides a breeding ground for the rapid spread of fake news.Fake news may seriously interfere with the perception of the masses` cognition,causing the masses to make wrong decisions,disrupt social order,interfere with political elections,and cause many negative social effects.Existing research lacks a summary of the cognitive mechanism of fake news.In order to explore the psychological and neural basis of reading fake news,this paper has a deeper understanding of the source,spread and social impact of fake news,so as to provide guidance for correcting fake news.This paper defines the cognitive mechanism of fake news,and summarizes two methods for studying the cognitive mechanism of fake news:cognitive experiment method and data analysis method.Cognitive experiment methods are summarized into four parts:internal mental state,external social environment,correcting fake news,and cross-domain cognitive mechanism.The data analysis method is summarized into two parts:the cognitive mechanism of political fake news and the cognitive mechanism of non-political fake news.Finally,three points of thinking are put forward for the future research direction,namely,rumor-defying strategies,deep fake news cognitive mechanism mining,and epidemic-type fake news cognitive mechanism research.

Identity-based Encryption Scheme Based on R-SIS/R-LWE

QIAN Xin-yuan, WU Wen-yuan

Computer Science. 2021, 48 (6): 315-323. doi:10.11896/jsjkx.200700215

Abstract

PDF(1569KB) ( 1096 )

References | Related Articles | Metrics

The identity-based encryption(IBE) by lattice can effectively resist quantum attacks,and this mechanism takes users’ identity information as public keys,which can ease the management of public key infrastructure(PKI) with an extremely large number of users.The lattice-based IBE system is an improvement of the traditional PKI to solve some problems in the Internet of Things(IoT) environment.However,previous IBE schemes based on lattices are cumbersome,and there are few implementations of these schemes.Aiming at this problem,this paper proposes an IBE scheme based on R-SIS and R-LWE with advantages of low expansion rate,which is secure against IND-sID-CPA.Firstly,a block reusing technology is proposed to reuse a ciphertext block for auxiliary decryption which occupies a significant amount in storage so that the expansion rate of ciphertext decreases and the encryption efficiency improves in a large extent.Then,by using a compression algorithm and introducing a plaintext expansion parameter,the two indicators of the scheme have been further optimized.Next,the scheme’s security,correctness,and computing complexity are analyzed through rigorous theoretical derivation,and numerical experiments with Maple give the optimal parameter values of this scheme under three scenarios.Finally,the new scheme is implemented with C++,and the performance of the scheme and the BFRS scheme in three scenarios are compared.Experiments and comparisons show that,while ensuring the correctness and security,this scheme improves the encryption and decryption efficiency of the original scheme and reduces the ciphertext expansion rate effectively.

Improved Negative Selection Algorithm and Its Application in Intrusion Detection

JIA Lin, YANG Chao, SONG Ling-ling, CHENG Zhenand LI Bei-jun

Computer Science. 2021, 48 (6): 324-331. doi:10.11896/jsjkx.200400033

Abstract

PDF(2958KB) ( 917 )

References | Related Articles | Metrics

As a typical algorithm of artificial immune system,negative selection algorithm(NSA) is widely used in intrusion detection.Aiming at the problems of low accuracy,high false alarm rate and high redundancy of detector set in the traditional negative selection algorithm,an improved negative selection algorithm is proposed and applied to the intrusion detection.The main idea is as follows:first,non-self-antigens is clustered by density peak clustering algorithm to generate a known detector,which can detect the known invasion behavior.Then the abnormal point is defined and it is taken as the center of candidate detector preferentially to calculate and generate unknown detector,which can detect unknown intrusion behavior,so as to reduce the randomness of detector generation.In the experimental stage,AC(accuracy) and FA(false alarm) are selected as evaluation indexes.The algorithm has been simulated on the KDDCUP99 and CSE-CIC-IDS2018 data sets,and the experimental results show that the algorithm has lower false alarm rate and higher accuracy rate on the two data sets,which verifies the proposed improved method has a better detection effect.

Malicious User Detection Method for Social Network Based on Active Learning

ZHANG Ren-zhi, ZHU Yan

Computer Science. 2021, 48 (6): 332-337. doi:10.11896/jsjkx.200700151

Abstract

PDF(1969KB) ( 905 )

References | Related Articles | Metrics

As a classification task,malicious user detection needs to label training samples.However,the scale of social networks is usually large,and it costs a lot to label all samples.In order to find out the more worthy samples in the case of limited labeled budget,and make full use of unlabeled samples to improve the detection performance of malicious users,a detection method based on graph neural network and active learning is proposed.The method is divided into two parts:detection module and active lear-ning module.Inspired by Transformer,the detection module improves the graph neural network GraphSAGE,flattens the aggregation process of each order neighbors of its nodes,so that higher-order neighbors can directly aggregate to the central node and reduce the information loss of high-order neighbors.Then,through ensemble learning,the extracted representations are used from different perspectives to complete the detection task.The active learning module measures the value of unlabeled samplesaccor-ding to the results of ensemble classification,and alternately uses detection module and active learning module in the sample labeling stage to guide the process of labeling sample,which is more conducive to the model classification.In the experimental stage,AUROC and AUPR are used as evaluation indexes to verify the effectiveness of the improved detection module on a real large-scale social network data set,and the reasons for the improvement are analyzed.Then,compared with the existing two similar active learning methods,the experimental results show that the proposed method has better classification performance in the case of labeling the same number of training samples.

Network Security Situation Assessment Based on Genetic Optimized PNN Neural Network

WANG Jin-heng, SHAN Zhi-long, TAN Han-song, WANG Yu-lin

Computer Science. 2021, 48 (6): 338-342. doi:10.11896/jsjkx.201200239

Abstract

PDF(1851KB) ( 826 )

References | Related Articles | Metrics

In order to improve the performance of network security situation assessment,this paper presents a network security situation assessment method based on genetic optimization probabilistic neural network.Firstly,In the process of network security situation assessment modeling,according to the characteristics of network security situation and common evaluation levels,the network security situation assessment model of PNN neural network is established,and the advantages of PNN neural network in fine-grained network security situation assessment are fully exploited.Then,in order to prevent the slow convergence caused by the fine-grained evaluation of network security situation parameters,the correction factors of PNN are left,and then the stable PNN network security situation assessment model is obtained by iterative training of PNN neural network.Experiments show that compared with the traditional PNN neural network algorithm,by using genetic algorithm to optimize the PNN network security situation assessment classification,evaluation accuracy is higher,average accuracy rate is more than 90%,and training speed is faster.

Intrusion Detection Method Based on WiFi-CSI

WANG Ying-ying, CHANG Jun, WU Hao, ZHOU Xiang, PENG Yu

Computer Science. 2021, 48 (6): 343-348. doi:10.11896/jsjkx.200700006

Abstract

PDF(2139KB) ( 964 )

References | Related Articles | Metrics

At present,Wi-Fi has been widely used in public and private fields.Deviceless human intrusion detection based on wireless technology has a broad application prospect for realizing indoor services such as asset security,emergency response and personalized services.Aiming at the problems of serious false positives and false negatives,difficult analysis of massive information,troublesome deployment,etc. of existing methods,this paper proposes an intrusion detection method based on Wi-Fi signals.Firstly,it uses the fine-grained Channel State Information(CSI) on the Wi-Fi device to capture small changes caused by human movements.Then it uses the Multiple Signal Classification(MUSIC) to sample the covariance matrix eigen decomposition.The obtained noise subspace is used to estimate the target angle(AOA).Finally,the intrusion is judged by calculating the phase difference changes of different paths caused by the movements of the human body.The difference between traditional methods and the proposed method is that the spectral peak search and phase difference are combined,with complementary advantages,the two overcome the environmental and noise interference,and solve the influence of multipath effects on the results of this paper.There are two typical indoor environments in this paper,namely the conference room and the dark room.Experimental results show that the average false negative(FN) and false positive(FP) of the method in the two indoor environments are 1.83% and 1.4%,respectively.In addition,this paper also evaluates the detection performance of the proposed method in different sports modes,and the average false negative and false positive are 2.26% and 1.46%,respectively.By comparing with other methods,the validity and stability of the proposed method are verified.It shows that this method has strong robustness and practical value,and provides a feasible scheme for the development of intrusion detection technology in the future.