Table of Content

24 December 2021, Volume 0 Issue 12

Previous Issue    Next Issue

Multi-objective Optimization Algorithm for Flexible Job Shop Scheduling Problem

XU Ming, ZHANG Jian-ming, CHEN Song-hang, CHEN Hao

2021, 0(12):  1-6. 

Asbtract ( 473 )   PDF (1427KB) ( 406 )  

References | Related Articles | Metrics

Flexible job shop scheduling problems have the characteristics of diversified solution sets and complex solution spaces. Traditional multi-objective optimization algorithms may fall into local optimality and lose the diversity of solutions when solving those problems. In the case of establishing a flexible job shop scheduling model with the maximum completion time, maximum energy consumption and total machine load as the optimization goals, an improved non-dominated sorting genetic algorithm (INSGA-II) was proposed to solve this problem. Firstly, the INSGA-II algorithm  combines random initialization and heuristic initialization methods to improve population diversity. Then it adopts a targeted crossover and mutation strategies for the process part and the machine part to improve the algorithm’s global searching capabilities. Finally,  adaptive crossover and mutation operators are designed to take into account the global convergence and local optimization capabilities of the algorithm. The experimental results on the mk01～mk07 standard data set show that the INSGA-II algorithm has better algorithm convergence and solution set diversity.

Operation Service Evaluation Method of Navigation Satellite Inter-satellite Link Based on Fuzzy Mathematics Theory#br#

JIA Chang, LIU Lei, SUN Jian-wei

2021, 0(12):  7-12. 

Asbtract ( 197 )   PDF (730KB) ( 129 )  

References | Related Articles | Metrics

In order to ensure the navigation satellite inter-satellite link can run normally and get its service effect, the evaluation of the service quality of the inter-satellite link operation has become a major research focus in the field of inter-satellite links. This paper constructs an inter-satellite link operation service evaluation system based on the OSI reference model, and divides the underlying indicators according to different network levels. The inter-satellite link operation services of navigation satellites are evaluated from multiple perspectives and levels including physical layer, link layer, network layer, transmission layer and application layer. According to the fuzzy mathematics theory, the weight and membership degree of each evaluation factor are obtained, which can be used to calculate the fuzzy comprehensive evaluation value of inter-satellite link network. Based on the simulation data of navigation satellite inter-satellite link, the fuzzy evaluation score of simulation data of navigation satellite inter-satellite link operation service is calculated according to the established evaluation index system and the adopted fuzzy evaluation method. In this paper, the evaluation system and methods are studied in order to provide an idea for the operation service evaluation of navigation satellite inter-satellite link.

Channel Pruning of Convolutional Neural Network Based on Transfer Learning

FENG Jing-xiang

2021, 0(12):  13-18. 

Asbtract ( 263 )   PDF (1012KB) ( 202 )  

References | Related Articles | Metrics

Convolutional neural networks are widely used in many fields like computer vision. However, large number of model parameters and huge cost make many edge devices unable to offer enough storage and computing resource. Aiming at problems above, a migration learning method is introduced to improve the sparsity proportion of the channel pruning method based on the scaling factor of the BN layer. The effects of different levels of migration on the sparsity proportion and channel pruning are compared, and  experiments based on the NAS viewpoint are designed  to explore its pruning accuracy limit and iterative structure convergence. The results show that compared with the original model, with the accuracy loss under 0.10, the parameter amount is reduced by 89.1%, and the model storage size is reduced by 89.3%. Compared with the original pruning method, the pruning threshold is increased from 0.85 to 0.97, further reducing the parametes by 42.6%. Experiments have proved that the introduction of migration method makes it easier to fully sparse the weights, increases the tolerance of the channel pruning threshold, and gets a higher compression rate. In the pruning network architecture search process, the migration provides a more efficient starting point to search, which seems easy to converge to a local optimal solution of the NAS.

BOTDA Sensing Information Extraction Based on Artificial Neural Network Using Whale Optimization Algorithm

LIU Ya-nan, GUO Nan, ZHAO Yang, YU Kuang-lu,

2021, 0(12):  19-26. 

Asbtract ( 219 )   PDF (1700KB) ( 150 )  

References | Related Articles | Metrics

Artificial neural networks (ANNs) have been employed to acquire Brillouin frequency shift (BFS) information measured by Brillouin optical time domain analyzer (BOTDA), however, it suffers from drawbacks such as easy entrapment in local optima and a slow convergence rate. To overcome the above shortcomings, an artificial neural network using whale optimization algorithm (WOA) for rapid BFS acquisition for Brillouin fiber sensors is proposed in this manuscript. And then a modified nonlinear WOA neural network (NWOA-NN) with a designed nonlinear convergence factor a was put forward to better extract BFS. We compared the proposed networks with ANN, particle swarm optimized neural network (PSO-NN), and genetic algorithm optimized neural network (GA-NN) models. Experimental results show that the performance of the WOA-NN model is better than the latter three, and the average RMSE of temperature obtained by WOA-NN is lower than those of ANN, PSO-NN and GA-NN by approximately 42.66%, 52.51% and 45. 93%, respectively. The average RMSE by the NWOA-NN model further outperformed the WOA-NN by 19.08%. The average time spent training the ANN, PSO-NN, GA-NN, WOA-NN and NWOA-NN networks were respectively 929.71 s, 889.49 s, 699.36 s, 580.06 s and 549.12 s, our proposed networks illustrated better performance.

Human Pose Estimation Algorithm Based on Channel Splitting

ZHOU Kun-yang, ZHAO Meng-ting, ZHANG Hai-chao, SHAO Ye-qin

2021, 0(12):  27-36. 

Asbtract ( 185 )   PDF (7759KB) ( 84 )  

References | Related Articles | Metrics

To improve the accuracy and speed of human pose estimation, a channel-split-based human pose estimation algorithm, named Channel-Split Residual Steps Network (Channel-Split RSN), is proposed. First of all, channel-split blocks are proposed to apply convolution operation for split feature in order to obtain rich feature representation. Then, feature enhancement blocks are introduced to further split feature channel and employ different strategies for different groups which can reduce similar features in feature channels. Finally, to further enhance the pose refine machine in Channel-Split RSN, combined with improved spatial attention mechanism, a pose refine machine based on feature spatial correlation, named Context-PRM, is proposed. Experimental results show that on the COCO test-dev dataset, our algorithm reaches 75.9% AP and 55.36 FPS, and the Params(M) of the model is only 18.3. Compared with the traditional RSN18 and RSN50, the AP of the model is improved by 5 and 3.4 percentage points, respectively. FPS is 12.08 faster than the traditional RSN50. On the more challenging CrowdPose dataset, our approach achieves 66.9% AP and 19.16 FPS, an AP improvement of 4.6 percentage points compared to RSN18, which effectively improves the accuracy of human pose estimation and the model has a faster recognition speed. Our source code is available at https://github.com/qdd1234/Channel-Split-RSN.

Weibo Tag Generation Algorithm Based on LDA and Word2vec

CHEN Ying, SHENG Jia-gen

2021, 0(12):  37-42. 

Asbtract ( 341 )   PDF (886KB) ( 109 )  

References | Related Articles | Metrics

Aiming at the problem that the tag generation algorithm based on the traditional LDA topic model describes the user’s interest topics incompletely, a Weibo user’s tag generation algorithm TopicERP based on the topic embedding representation is proposed. Based on the LDA model, by introducing Word2vec word embedded model, the algorithm is to conduct a comprehensive description of interest subject to the customer, and to improve the matching degree calculation method. Firstly, LDA topic model was used to analyze the topics of users’ Weibo and generate the topics of users’ interest. Then, Word2vec word embedding model was used to transform the topic text into the topic vector, which was used to calculate the matching degree. Finally, it used cosine similarity and conditional probability of topic in the document, the matching degree between topic vector and candidate tag was calculated, and Top-Q candidate tag was selected as the target user’s tag. Experimental results on MicroPCU, a public Weibo data set, show that the algorithm has better overall performance than the algorithm based on the traditional LDA topic model, and the generated user tags can describe users’ interests and preferences more accurately.

Power Information Network Intrusion Detection Algorithm Based on Dueling-DDQN

WU Shui-ming, JI Zhi-yuan, WANG Zhen-yu, JING Dong-sheng

2021, 0(12):  43-47. 

Asbtract ( 201 )   PDF (1310KB) ( 106 )  

References | Related Articles | Metrics

The power information system uses smart power grids to manage power equipment. With the increase of the total amount of social electricity consumption and the promotion and development of smart power grids, the scale of power network gradually becomes larger and the management is complex. However, it is important to ensure the security of power information system. Network intrusion detection technology can effectively avoid the intrusion and attack from the network, and then ensure the security of the system. In this paper, Dueling-DDQN algorithm of deep reinforcement learning method is used to solve the problem of intrusion detection in the network. The agent obtains reward value according to the trial and error learning to train the algorithm, so as to improve the efficiency of network intrusion detection and reduce the labor cost at the same time. Finally, the NLS-KDD data set is used for comparative experiments, and the experimental results show that the network intrusion detection algorithm based on Dueling-DDQN can improve the detection efficiency, and then better protect the network security.

Fast Classification of Urban LiDAR Data Based on Regional Structure Features

HAN Jian, LI Lin, CAO Zhi-min, DUAN Chao-hui, WAN Chuan,

2021, 0(12):  48-52. 

Asbtract ( 135 )   PDF (2971KB) ( 87 )  

References | Related Articles | Metrics

Airborne LiDAR can timely and accurately obtain a large number of 3D point cloud data with accurate 3D position information. It has a wide range of applications in digital cities, forest fire prevention, intelligent transportation, etc. Among them, the 3D point cloud data of urban areas, especially urban central areas with dense plants, is often occluded by tall trees or vegetation which make it difficult to recognize data belongs to man-made objects such as buildings. This paper uses a direct quadratic polynomial fitting method to extract regional information of typical local areas of vegetation and buildings, such as tall trees, and constructs sensitive structural features of regional targets. Furthermore, through fuzzy logic, the task 3D point cloud data classification especially designed to distinguish building targets and disturbances from trees can be completed. The experimental results show that this method can quickly and effectively realize the classification of LiDAR point cloud data, and the proposed method has the good application prospect and robustness for promotion.

Detecting Electrical Circuit Elements Based on Faster RCNN

XU Jian, ZHANG Hao, XU Hang, XIE Kai

2021, 0(12):  53-57. 

Asbtract ( 164 )   PDF (1096KB) ( 96 )  

References | Related Articles | Metrics

Vectorization of engineering drawings plays a key role in the digital foundation of power grid. Due to the variety of electrical components in power grid, some image backgrounds are blurred, and the rotation angle of the electrical components is not consistent, which poses a challenge for the identification of the electrical elements in the drawings. This paper mainly studies the electrical element recognition and training with the Faster RCNN network architecture in deep learning, and performs preprocessing and feature extraction on the images to be trained, including preprocessing in smooth denoising, binarization, segmentation, etc. The feature extraction uses VGG16 network, and then uses Faster RCNN for classification. Experimental results on real-world datasets with 9 categories of electrical circuit elements show that the performance of detection and classification of electrical elements are efficient.

Video Snapshot Compressed Sensing Reconstruction Based on Multi-scale Fusion Network

CHEN Xun-hao, YANG Ying, HUANG Jun-ru, SUN Yu-bao

2021, 0(12):  58-64. 

Asbtract ( 190 )   PDF (2539KB) ( 72 )  

References | Related Articles | Metrics

Video snapshot compressed sensing is based on the theory of compressed sensing, which only projects multiple frames to a two-dimensional snapshot measurement during one exposure process to achieve high-speed imaging. In order to recover the original video signal from the two-dimensional snapshot measurement signal, the classical reconstruction algorithm is based on the sparsity of the video prior to iterative optimization solution, but the reconstruction quality is low and time-consuming. Deep learning has attracted much attention because of its excellent learning ability as well as video snapshot compression reconstruction methods that developed based on it. However, the existing deep methods lack effective expression of spatiotemporal features, and the reconstruction quality still needs to be further improved. This paper proposes a multi-scale fusion reconstruction network (MSF-Net) for compressed sensing reconstruction of video snapshots. The network expands from the two dimensions of horizontal convolution depth and vertical resolution. The resolution dimension uses three-dimensional convolution to perform different scales. In the extraction of video features, the horizontal dimension uses the pseudo three-dimensional convolution residual module to extract hierarchically the feature maps of the same resolution scale, and learns the spatiotemporal features of the video through the cross fusion of features at different scales. Experimental results show that this method can improve the reconstruction quality and reconstruction speed at the same time.

Apple Cultivar Identification Based on Convolutional Neural Network

QIU Yu, HAN Jun-ying, FENG Cheng-zhi, CHEN Yong-wei

2021, 0(12):  65-71. 

Asbtract ( 177 )   PDF (1576KB) ( 98 )  

References | Related Articles | Metrics

Against apple fruit varieties identification and classification problem, a original data set containing more than one apple fruit varieties of leaf image is provided, a new convolution model of neural network classification is built to verify the classification accuracy, generalization performance and stability, which provides theoretical basis and technical support for apple cultivars’ simple, rapid, accurate and reliable identification and classification. Taking the apple fine seedling breeding base of Jingning Fruit Research Institute of Pingliang City of Gansu Province as the experimental base, 14 apple tree varieties are selected. For each variety, about 10 fruit trees with different tree ages, tree size and growth status were selected, and about 100 mature leaves without mechanical damage are picked, and then leaf images are taken to form a data set. Then the convolutional neural network is used to train the recognition and classification model. Aiming at the recognition and classification of apple cultivars, this paper provides an original data set containing 14394 leaf images of 14 apple and fruit cultivars, and designs and implements a recognition and classification model based on convolutional neural network. The experimental results show that the model has high accuracy. The training accuracy of the training set can reach 99.88%, the verification accuracy of the verification set is 94.36%, and the test accuracy of the independent test set is 90.49%. The results of this study can help the modern apple field planting, scientific research experiments and other practical scenarios, and provide a reference for the practical application of deep convolutional neural network technology in plant variety identification and classification, and enrich the application of deep learning in agriculture.

Multi-scale Single Image Rain Removal Based on Multi-channel Separation and Integration

WANG Fan, WEI Xian, GUO Jie-long, LIANG Pei-dong

2021, 0(12):  72-78. 

Asbtract ( 207 )   PDF (3303KB) ( 118 )  

References | Related Articles | Metrics

Due to the occlusion of the image caused by the different density of rain, it has always been a challenging task to remove the rain from the images. Nowadays, image de-raining algorithms based on deep learning have become the mainstream. However, most deep learning architectures are designed by stacking convolutional layers. For the rain removal task, the images have rain streaks in different sizes. To address this issue, A convolutional neural network based on multi-channel separation and integration for single image rain removal is designed. In the first step, the separable channels and the hierarchical connection between the convolutional layers form a multi-scale module. Finally, the outputs of different channels are integrated. The proposed module can expand the receptive field and explore the spatial information between feature maps, which extracts features better. In the second step, progressive networks are exploited to repeatedly calculate and excavate contextual information, which can be well related to global features. The proposed model is easy to be implemented and can be trained end-to-end. Extensive experiments on widely used datasets and self-built rainy images dataset for autonomous driving demonstrate that the proposed method has achieved significant improvements over the existing methods.

Deep Connected Ultra-lightweight Subspace Attention Mechanism

ZHANG Chen-xiao, PAN Qing, WANG Xiao-ling

2021, 0(12):  79-84. 

Asbtract ( 179 )   PDF (1349KB) ( 73 )  

References | Related Articles | Metrics

In order to solve the problem of large computation or parameter overheads in deploying the existing attention mechanism of compact convolutional neural network, an improved ultra-lightweight subspace attention mechanism is proposed. Firstly, the deep connected subspace attention mechanism(DCSAM) is used to divide the feature map into several feature subspaces, and deduce different attention feature maps for each feature subspace. Secondly, the spatial calibration method of feature subspace is improved. Finally, the connection between the front and back feature subspaces is established to make the information flow between the front and back feature subspaces. The subspace attention mechanism enables multi-scale and multi-frequency feature representation, which is more suitable for fine-grained image classification. The method is orthogonal and complementary to the existing attention mechanisms used in visual models. The experimental results show that on ImageNet-1K and Stanford Cars datasets, the highest accuracy of MobileNetV2 is improved about 0.48 and 2 percent points when the number of parameters and floating-point operations are reduced by 12% and 24% respectively.

Rotation Invariant Texture Extraction of Wheel Tread Based on DT-CWT and SVM

ZHAO Cong-hui, FENG Qing-sheng

2021, 0(12):  85-90. 

Asbtract ( 128 )   PDF (1729KB) ( 66 )  

References | Related Articles | Metrics

Aiming at the problem that the rotation texture information of train wheel treads cannot be extracted accurately and effectively, a method for extracting train wheel tread features based on Radon transform and dual-tree complex wavelet transform (DT-CWT) is proposed. Firstly, the Radon transform is performed on the image of the wheel tread; then, the transformed image is decomposed by DT-CWT, and the decomposed layer of the low-frequency sub-band coefficients and the modulus of the mean and standard deviation of the high-frequency sub-band coefficients are used to construct the feature vector, and the feature is used as the basis for distinguishing whether the train wheel tread is damaged or not; finally, the classification decision is made by the support vector machine (SVM). Part of the images used in the classification test are from the automatic vehicle station, and part of the images are artificially noised. The results show that the Radon and DT-CWT algorithms used in this paper can effectively perform the rotation invariant texture extraction, and the SVM classification accuracy rate can reach 95%. It provides more accurate and convenient method support for the detection of train wheel tread conditions.

Speech Emotion Recognition Based on ARCNN-GAP Network

QIAN Jia-qi, HUANG He-ming, ZHANG Hui-yun,

2021, 0(12):  91-95. 

Asbtract ( 190 )   PDF (1036KB) ( 82 )  

References | Related Articles | Metrics

Speech emotion recognition is the most direct way for computers to understand human emotions. It is an important channel to realize the intelligentization of human-computer interaction, but the performance of the recognition model needs to be further improved. To achieve this goal, a model ARCNN-GAP based on recurrent convolutional neural network is proposed for speech emotion recognition. Where the recurrent convolution layer has elastic path, which can ensure the depth of the network and the gradient return during optimization, and extract more effective features. The global average pooling is used to reduce the computational complexity and the risk of over-fitting. And the attention mechanism is employed to focus more on emotion-related features. The fusion features of prosodic features and spectral features are studied on CASIA and EMO-DB databases, and the recognition rates are 83.29% and 75.28% respectively. The experimental results show that the proposed model ARCNN-GAP features higher recognition performance and better generalization.

Convolutional Sequence Recommendation Algorithm for RPA Softwares

HOU Cong-ying, WANG Peng, ZHU Li-xia, GUAN Xiao-ning

2021, 0(12):  96-102. 

Asbtract ( 144 )   PDF (1185KB) ( 84 )  

References | Related Articles | Metrics

In RPA (Robotic Process Automation) softwares, sequence recommendation systems are often used to complete manual processing tasks such as judgment and selection. However, the commonly used sequence recommendation system is limited by the difficulty of extracting sequence information, so it is difficult to be widely used. In order to solve this problem, this paper constructs a convolutional sequence recommendation model based on Inception. It embeds user behavior sequence information in time and latent space into an “image”, and extracts local features through dynamic and static convolutional layers. It can fully extract the user’s short-term interest preferences, and embed the user embedding matrix as the user’s long-term interest preferences into the output of the convolutional layer. They work together to build a complete set of user interest preferences and improve recommendation performance. Through experiments on three public data sets MovieLens 1M, Gowalla, and Steam, it is verified that the performance of the convolutional sequence recommendation model based on Inception is better than the latest sequence recommendation model. Among the three evaluation indicators of Top-N series (Precision@N, Recall@N, MAP), the average increase is about 10%, and the maximum increase on a single index is 14%.

Application of Edge Computing in Intelligent Transportation Systems

WU Jian-bo, ZHU Wen-xia, JU Liang, XU Zhi-fang

2021, 0(12):  103-109. 

Asbtract ( 383 )   PDF (1839KB) ( 217 )  

References | Related Articles | Metrics

With the popularization of automobiles, traffic congestion is becoming increasingly serious. Although the cloud-based intelligent transportation systems (ITS) can relieve traffic pressure, it can no longer meet the demand of transmission bandwidth and delay requirements of new on-board applications such as assisted driving and autonomous driving. In order to realize the data real-time processing, ensure public information and traffic safety, and improve the transportation system efficiency, edge computing (EC) is applied to ITS. The development of ITS is described and the overall architecture of edge-based ITS is proposed, which makes full use of the characteristics of edge computing such as physical proximity, high bandwidth, low latency and location recognition to solve the problems of information transmission delay, data processing delay and large transmission load. Next, the key technologies of edge computing in ITS are discussed from the aspects of wireless transmission, information perception, computing offloading and collaborative processing. Finally, the future opportunities and challenges for the application of EC in ITS are pointed out.

Microblog Hot Topic Discovery Based on Text Dual Representation Model

LIU Meng-ying, WANG Yong

2021, 0(12):  110-115. 

Asbtract ( 250 )   PDF (1054KB) ( 103 )  

References | Related Articles | Metrics

Microblog is an important platform for information dissemination in contemporary life, mining hot topics on microblog has become one of the important research directions nowadays. In view of the problems of traditional hot topic discovery methods in dealing with microblog text, such as lack of semantic information in text representation, poor effect of mining hot topics and so on, this paper proposes a text dual representation model based on frequent word sets and BERT semantics(FWS-BERT), which calculates the weighted text similarity to perform spectral clustering on microblog text, further, microblog topic mining is carried out based on affinity propagation (AP) clustering algorithm with improved similarity measurement. Finally, a topic heat evaluation method is proposed by introducing the H index in bibliometrics. Experiments show that the proposed method is higher than the single text representation method based on frequent word set and K-means method in contour coefficient and Calinski-Harabasz (CH) index value, and can accurately represent the topic and Evaluate-the popularity of microblog data.

A Feature Filtering and Instance Transfer Framework for Cross-project Defect Prediction

DIAO Xu-yang, LIU Xiao-yang, XU Li, CHEN Tian-qun, XU Ya-zhou

2021, 0(12):  116-122. 

Asbtract ( 144 )   PDF (836KB) ( 106 )  

References | Related Articles | Metrics

In cross-project software defect prediction, the feature correlation and the difference in instance distribution between the source project and the target project are the main factors that affect the performance of the prediction model. From the perspective of feature filtering and instance transfer, we propose a framework for cross-project defect prediction called KCF-KMM. Specifically, during the feature filtering phase, it uses K-medoids clustering algorithm to select features, filtering out features that have low relevance to the target project. During the instance transfer phase, the KMM algorithm is used to calculate the distribution difference between the source project and the target project instance, so as to assign the influence weight of each training instance. Finally, it combines a small amount of labeled data in the target project to establish a mixed defect prediction model. To verify the effectiveness of KCF-KMM, it is compared with the classic cross-project software defect prediction methods such as TCA+, TNB and NNFilter from the perspective of accuracy and F1 value. The prediction performance of KCF-KMM can be improved by 34.1%, 0.8%, 21.1% and 14.4%, 3.7%, 10.6% on the Apache data set, respectively.