Table of Content

31 July 2021, Volume 39 Issue 4

Previous Issue    Next Issue

Special Issue on CCF NCCA 2020

Dynamic Microservice Interaction Platform Design Based on Stream Engine

YIN Yifan, XU Kaizhou, WANG Yanhua, ZHOU Xin, CAI Hongming

2021, 39(4):  521-531.  doi:10.3969/j.issn.0255-8297.2021.04.001

Asbtract ( 2157 )   PDF (776KB) ( 437 )  

References | Related Articles | Metrics

In view of problems of traditional service design, such as high coupling, low transparency and complex change, a dynamic microservice interaction platform design based on stream engine is proposed. Service process is decomposed into fine-grained microservice models whose boundaries are defined in a unified model representation, so that a service can be implemented independently of the interfaces of other services. Microservices are connected through stream channels. Service encapsulation is carried out on the producer side based on temporal and spatial features of data to construct unified information representation. Service analysis is carried out on the consumer side to divide and reorganize data. A complete service process for business process is constructed, driven by streaming data. Based on this design, a visual microservice interaction management platform is realized and applied to spinning detection process of engine manufacturing. Compared with traditional service systems, this platform design features in lower coupling, more flexibility in service change, expansion and evolution, and improved performance in service monitoring and fault handling.

Heterogeneous Information Network Representation Learning Based on Generative Adversarial Network

LIU Xinghong, WANG Ying, WANG Xin, LAN Shumei

2021, 39(4):  532-544.  doi:10.3969/j.issn.0255-8297.2021.04.002

Asbtract ( 2118 )   PDF (837KB) ( 156 )  

References | Related Articles | Metrics

In view of the high-dimensional sparsity shortcomings of traditional heterogeneous information networks, we firstly proposed an unsupervised learning model-heterogeneous network representation learning based on generative adversarial network (HNRL-GAN) that embeds the high-dimensional vertices of heterogeneous information networks into low-dimensional vector spaces. Secondly, having analyzed the shortcomings of HNRL-GAN, we proposed an improved model, called as heterogeneous network representation learning based on generative adversarial network plus plus (HNRL-GAN++). Finally, we used HNRL-GAN and HNRL-GAN++ in three data sets, including DBLP, Yelp, and Aminer, to perform node classification and node clustering for testing the effectiveness of the two models. Experimental results show that: 1) Both HNRL-GAN and HNRL-GAN++ achieve the goal of representing high-dimensional sparse nodes in heterogeneous information networks as low-dimensional dense vectors; 2) Compared with HNRL-GAN, HNRL-GAN++ has better performance in retaining network structure information and semantic information in high-dimensional space.

Hybrid Feature Selection Algorithm Based on Mutual Information

JIANG Wenxuan, DUAN Youxiang, SUN Qifeng

2021, 39(4):  545-558.  doi:10.3969/j.issn.0255-8297.2021.04.003

Asbtract ( 2286 )   PDF (1617KB) ( 322 )  

References | Related Articles | Metrics

Traditional feature selection algorithms only focus on feature correlation and feature redundancy without considering the interaction between features. This paper proposes a hybrid feature selection based on mutual information (MIHFS) algorithm. The algorithm takes the classification accuracy of K-nearest neighbor (KNN) algorithm as evaluation index to evaluate the classification performance of selected features, effectively removes redundant and irrelevant features, and retains the interactive features. In order to evaluate the performance of the proposed algorithm, the classification accuracy, the number of selected features and the stability of the algorithm are compared with seven other feature selection algorithms such as minimal redundancy maximal relevance (mRMR) and joint mutual information (JMI) in eight datasets. Experimental results show that the MIHFS algorithm has strong stability, which not only effectively reduces the dimension of feature space, but also has better classification performance than other feature selection algorithms. Finally, in combination with grey relation analysis (GRA) method-technique for order preference by similarity to ideal solution (TOPSIS) method, MIHFS algorithm is applied to the geological evaluation of the first member of Dainan Formation at Yong’an Area, Gaoyou Sag. Experimental results show that MIHFS algorithm performs an evaluation accuracy of 80% with high reliability, and this is basically consistent with actual drilling results and proves the effectiveness of MIHFS in oil and gas geological evaluation.

Network Intrusion Detection Based on GRU and Feature Embedding

YAN Liang, JI Shaopei, LIU Dong, XIE Jianwu

2021, 39(4):  559-568.  doi:10.3969/j.issn.0255-8297.2021.04.004

Asbtract ( 2191 )   PDF (633KB) ( 128 )  

References | Related Articles | Metrics

The existing intrusion detection methods based on neural network have not taken data classification information into consideration yet, thus, the timing information of network traffic data are not used effectively. In this paper, we propose network intrusion detection models based on gated recurrent unit (GRU) in combination with embedding technique of categorical information. Simulation experiments on the models are carried out with UNSW-NB15, which is a comprehensive network traffic dataset. Experimental results show that the proposed models not only improve the detection rate of intrusion attacks, but also provide a new way for intrusion detection in case of processing large-scale data.

Accurately Identify Zombie Enterprises Based on Decision Tree-Logistic Regression Model

WU Dongpeng, WANG Zheng, TONG Wei, YE Feng, SONG Chuqiao

2021, 39(4):  569-580.  doi:10.3969/j.issn.0255-8297.2021.04.005

Asbtract ( 2224 )   PDF (671KB) ( 458 )  

References | Related Articles | Metrics

Aiming at the problem of how to accurately identify zombie enterprises, based on the enterprise information data set published by Hunan Kechuang Information Co., LTD., a zombie enterprise identification method based on decision tree-logistic regression model is proposed. The method uses median to fill in missing numbers and outliers, analyzes data sets for feature derivation, and finally uses multiple linear regression and chi-square test to complete feature screening. In order to verify the effectiveness of the proposed method, comparative experiments are carried out between the method and the over-borrowing method, continuous loss method, random forest algorithm, BP neural network algorithm, and XGBoost algorithm in the Alibaba Cloud environment and the local environment. Each model is trained 50 times, the data selected for each training is randomly selected according to a certain proportion, and finally the average value of each index is taken as the final result. Experimental results show that the proposed decision tree-logistic regression model has the highest accuracy in the identification of zombie companies, reaching 99.98%, and the model is superior to various other integrated models in running speed with average execution time of about 1.5 s. In all scenarios, experimental results of this model show relatively small differences, verifying the effectiveness and stability of the model.

Research on Optimizing Picking Route of Multi-zone Warehouse and Multi-checking Station

YE Nan, BI Zhongqin, WEI Hengda, WU Di

2021, 39(4):  581-593.  doi:10.3969/j.issn.0255-8297.2021.04.006

Asbtract ( 2168 )   PDF (2045KB) ( 260 )  

References | Related Articles | Metrics

Different from traditional single-export and single-check stations in small warehouses, large warehouses are often equipped with multiple check stations and multiple outlets to improve the efficiency of warehouse picking and leaving. This article proposes a dynamic adjustment algorithm based on replacing review station to solve the difficulty in traversal search caused by the uncertainty of the start and end points in the scenario of multiple review stations. On this basis, dynamical adjustment strategies with path optimization and reasonable allocation of multi-task orders are provided to meet the requirement of picking operations in large-scale and complex scenarios. Compared with the example of Jingdong logistics, it is shown that the dynamic adjustment algorithm based on the replacement of review station proposed in this paper is more efficient, with shorter picking path, less picking time and more accurate warehouse picking under the same conditions.

Spatial-Temporal Weight Attitude Motion Feature Extraction Algorithm Using Convolutional Neural Network

ZHENG Changliang, PANG Ming

2021, 39(4):  594-604.  doi:10.3969/j.issn.0255-8297.2021.04.007

Asbtract ( 2333 )   PDF (1811KB) ( 539 )  

References | Related Articles | Metrics

In traditional attitude motion feature extraction process, there is the problem of low efficiency. As to this, a temporal and spatial weight attitude motion feature extraction algorithm based on convolutional neural network (CNN) algorithm is proposed in this paper. First, from selected motion spatio-temporal samples, corresponding spatio-temporal motion keyframes are extracted and output in the form of static images. Second, initial moving images are preprocessed by using moving object detection, image enhancement and other measures. Then the motion feature is vectorized by CNN, and the adaptive interpolation method of space-time weight is used to reduce the error of motion edge detection. Finally, feature extraction of attitude motion is realized from two aspects of attitude edge feature and space-time feature of attitude motion, and extraction results are output. Compared with the traditional algorithm, experimental results show that the proposed algorithm improves the number of effective features.

Recognition Method of Human Dangerous Behavior in Multimodal Scenes Using Reinforcement Learning

ZHANG Xiaolong, WANG Qingwei, LI Shangbin

2021, 39(4):  605-614.  doi:10.3969/j.issn.0255-8297.2021.04.008

Asbtract ( 2285 )   PDF (4465KB) ( 354 )  

References | Related Articles | Metrics

In multimodal scenes, conventional human dangerous behavior recognition methods generally perform low recognition accuracy. Therefore, this paper proposes a human dangerous behavior recognition method based on reinforcement learning. Firstly, a feature extraction algorithm of reinforcement learning is used to obtain feature subsets of human dangerous behavior in multimodal scenes. Secondly, human dangerous behaviors in multimodal scenes are extracted by reinforcement learning data decision-making, and a fuzzy recognition model of human dangerous behavior is constructed. Finally, by bringing the obtained feature subsets of human dangerous behavior into the model and calculating the membership degree of dangerous behavior under different senses, the recognition of human dangerous behavior in multimodal scenes can be realized. Experimental results show that the method in this paper has a high recognition accuracy and a recognition delay of less than 300 ms.

Prediction of Dissolved Oxygen in Aquaculture Based on 3D Convolution and CLSTM Neural Network

ZHA Yukun, ZHANG Qilin, ZHAO Yongbiao, HANG Bo

2021, 39(4):  615-626.  doi:10.3969/j.issn.0255-8297.2021.04.009

Asbtract ( 2207 )   PDF (1179KB) ( 507 )  

References | Related Articles | Metrics

In this paper, a neural network based on three-dimensional (3D) convolution and convolutional long short-term memory (CLSTM) is proposed to predict the dissolved oxygen in aquaculture. Firstly, an input vector is multiplied by its transpose to form a single-channel matrix, and the single-channel matrices within a certain period of time are stacked to form a cube as the input data. Secondly, two consecutive three-dimensional convolutions are carried out on the input data to refine the characteristics of dissolved oxygen related factors, and the pooling layer is deleted for reducing calculation. Finally,the feature results of 3D convolution extraction are sent to the CLSTM model for further information extraction of time dimension, and the data is updated reversely by the gradient descent algorithm through the full connection layer. The actual data of a special aquaculture company in Xiangyang, Hubei Province were collected for experiment, and experimental results show that the proposed model has faster training convergence speed, higher prediction accuracy and better prediction stability than traditional BP neural network models, Conv3D and Conv2D, and could meet the needs of actual production.

Crack Detection of Track Slab Based on Convolutional Neural Network and Voting Mechanism

LI Wenju, HE Maoxian, ZHANG Yaoxing, CHEN Huiling, LI Peigang

2021, 39(4):  627-640.  doi:10.3969/j.issn.0255-8297.2021.04.010

Asbtract ( 2237 )   PDF (5405KB) ( 217 )  

References | Related Articles | Metrics

There are misdetections and missed detections in the crack detection of track slab or in crack pictures taken at night as using existing detection methods. For the problem, an improved method based on convolutional neural network (CNN) is proposed. In this way, high-level feature maps are divided into different groups of vectors whose feature expression would be emphasized by attention mechanism subsequently. Final confidence is accounted by aggregating the predict result of weak classifiers dynamically. With the favor of voting mechanism, predict error is reduced and robustness of model is improved effectively. Experiment results show that the improved method gains a prediction improvement of 1.6% in crack dataset and an improvement of 2.8% in CTFAR-10 dataset, in spite of the reduction of model parameters.

Environmental Sound Recognition Based on Attention Sinusoidal Representation Network

PENG Ning, CHEN Aibin, ZHOU Guoxiong, CHEN Wenjie, LIU Jing

2021, 39(4):  641-649.  doi:10.3969/j.issn.0255-8297.2021.04.011

Asbtract ( 2131 )   PDF (2605KB) ( 116 )  

References | Related Articles | Metrics

In this paper, we propose an attention sinusoidal representation network (A-SIREN). Firstly, Mel -frequency cepstral coefficient (MFCC) as an audio recognition feature is extracted from a dataset. Then, feature extraction is performed on each frame of the MFCC by using a neural network named gated recurrent unit (GRU). And audio score is calculated for each frame by using sine function and the audio is re-weighted according to the audio score of each frame. Finally, the categories of environmental sound are discriminated by using the full connection layer in combination with the Softmax classifier. In the experiments of this paper, we validated the designed model in an open-source dataset Urban Sound 8K and compared the performance of the designed model with that of other models. Experimental results show that the A-SIREN works best on the Urban Sound 8K dataset with recognition rate as high as 93.5%.

Small Target Detection Algorithm of UAV High Resolution Image Based on Center Point and Dual Attention Mechanism

WANG Shengke, REN Pengfei, Lü Xin, ZHUANG Xinfa

2021, 39(4):  650-659.  doi:10.3969/j.issn.0255-8297.2021.04.012

Asbtract ( 2365 )   PDF (6345KB) ( 269 )  

References | Related Articles | Metrics

Unmanned aerial vehicle (UAV) images have characteristics of high resolution, large field of vision and small target. However, existing object detection methods are generally insufficient in extracting the features of these small targets. Aiming at this problem, a small target detection algorithm is proposed in this paper. First, in order to improve the ability of feature expression for small targets, CenterNet, a detection network which uses center points to represent small targets, is adopted, and a deformable dual attention mechanism is induced. Then on this basis, for the problem of deficiency of original nonmaximum suppression (NMS) in dealing with nested redundant frames, we propose to use a generalized non-maximum suppression (G-NMS) in the process of redundancy detection elimination. Finally, LegoNet convolution unit is introduced to reduce convolution parameters and achieve balance between precision and velocity. The main validation data sets used in this paper are Visdrone 2019 and UAV_ OUC. Images in UAV_OUC have higher resolution than those in VisDrone2019. Compared with CenterNet, the detection accuracies of UAV_OUC and VisDrone2019 are improved by about 10% and 2% respectively.

Three-Dimensional Point Cloud Segmentation for Plants

LAI Yibin, LU Shenglian, QIAN Tingting, SONG Zhen, CHEN Ming

2021, 39(4):  660-671.  doi:10.3969/j.issn.0255-8297.2021.04.013

Asbtract ( 2524 )   PDF (5604KB) ( 614 )  

References | Related Articles | Metrics

Aiming at the irregular shape and uneven density of plant point clouds, a three-dimensional point cloud segmentation method applied to plants is proposed. Three plants of tobacco, corn, and cucumber are used as sample data, in which outliers and background points are removed by filtering and other preprocessing methods. Plant population is segmented by the Euclidean clustering algorithm, and leaf organs are segmented by region growing algorithm, edge extraction algorithm, super voxel clustering algorithm, and segmentation algorithm based on unevenness. The proposed method is used to segment three-dimensional point clouds of tobacco and corn, and the coverage rates are 87.5% and 96.9%, respectively. This verifies the feasibility and effectiveness of the method and provides clues for the automatic extraction of plant leaf organ phenotypes.

MAC Protocol Design for RF-Powered Wireless Sensor Networks

LIU Leyu, WANG Zumin, ZHENG Zupeng, QIN Jing, JI Changqing

2021, 39(4):  672-684.  doi:10.3969/j.issn.0255-8297.2021.04.014

Asbtract ( 2089 )   PDF (722KB) ( 130 )  

References | Related Articles | Metrics

To solve the problem of power supply in traditional wireless sensor network (WSN), a medium access control (MAC) protocol based on RF energy capture is proposed. Firstly, the time division multiple access (TDMA) technology is used to allocate channel according to time slot between adjacent nodes, so that data can be transmitted uncontested between source nodes and sink nodes. At the same time, the energy consumed by nodes in each cycle is controlled, and the duty cycle is adjusted indirectly. Then a load threshold is set in the buffer as the basis of each node’s role in the communication transformation to realize time synchronization between nodes. Finally, experimental simulations of various network topologies are carried out to evaluate delay rate and data loss rate. Experimental results show that, compared with the adaptive TDMA-based MAC (AT-MAC) protocol, the proposed protocol not only has lower latency rate, but also has a network simulation structure closer to the real network scenario, which can meet the performance requirements of wireless sensor networks.

Automatic Classification of Bamboo Flute Playing Skills Based on Deep Learning

GUO Yubo, LU Jun, DUAN Pengqi

2021, 39(4):  685-694.  doi:10.3969/j.issn.0255-8297.2021.04.015

Asbtract ( 2059 )   PDF (1123KB) ( 214 )  

References | Related Articles | Metrics

A dataset named Breath and two neural network reference models named Breath1d and Breath2d respectively are proposed for bamboo flute skill classification, and the optimal method is achieved for different classification tasks on this dataset. This paper divides the Breath dataset into subsets, and takes the multi-layer perceptron as the benchmark method of performance evaluation. First, the subsets are trained and predicted by the breath1d and breath2d models, and then the long short-term memory (LSTM) network model is used for auxiliary testing. Finally, the most suitable classification reference model for subtasks is obtained. When the whole dataset is classified, the breath2d and breath1d models are fused, and the data enhancement method is used. All of these make the classification accuracy of the whole dataset reach 91.3%. Compared with traditional audio classification tasks, this work expands the research field of music classification, and has a great effect on the modernization of national music.