Abstract:With the increasing complexity of emergencies related to the community correction object, the single fixed plan from the existing emergency plan database is unable to develop intelligent emergency response plans for different abnormal situations through dynamic data injection, which can hardly meet the demands of dynamic emergency generation. To improve the supervision quality and the informational management level of community correction object, we adopt the joint mining technology of multi-source heterogeneous data, in view of multi-source heterogeneity, complex association and dynamic evolution of abnormal situation data. On this basis, we build our judicial Knowledge Graph (KGjudicial) and crime Event Logic Graph (ELGcrime), providing data basis and auxiliary decision support for the dynamic generation of intelligent emergency plans. In addition, considering the actual business requirements for cross-regional multi-sectoral emergency coordination, we explore the multi-department information alignment method and the dynamic injection mechanism of emergency response plans. We propose the fusion technology of multiple-department emergency response plans based on our KGjudicial and ELGcrime to realize the cross-regional joint law enforcement of judicial administration departments and improve the supervision quality, while saving the management cost of community correction object. We provide technical support for the emergency response of multiple departments of judicial administration, contributing to the social security and stability.

Abstract:This study summarizes the current research on semantic-based video retrieval to help future researchers understand the technologies available in this field, and video retrieval systems are created to find the video that users want to query in a large number of video data collections on the Internet or in databases. This study introduces and discusses the semantic-based video retrieval process and also summarizes the relevant techniques to solve the main problem of a semantic gap in this process. The semantic gap is induced by the difference between the low-level features extracted from video content and the user’s cognition of these features in the real world. It is a highly concerned research topic to transform the low-level features of video content into high-level semantic concepts.

Abstract:In recent years, the quantitative investment models based on artificial intelligence algorithms have been emerging in the field of quantitative finance. These models attempt to model the financial time series through artificial intelligence methods, thereby forecasting data and developing an investment strategy. Regarding the unreliable prediction of the traditional Long Short Term Memory (LSTM) model for financial time series, we propose an improved LSTM model. The attention mechanism is added into the LSTM layer to enhance the forecasting performance of the neural network, and the Genetic Algorithm (GA) is used to optimize parameters, thus improving the model’s generalization ability. The data of China’s stock indexes and futures from the January 2019 to May 2020 is selected for the comparative experiments with state-of-the-art algorithms. The results show that the improved model performs better than other models in every indicators, proving the effect application of the model to future investment.

Abstract:Building energy-saving control is a multi-objective optimization problem considering the comfort demand. However, for the new buildings lacking operation data, it is a real conundrum to control the Heating, Ventilation and Air-Conditioning (HVAC) system to achieve both comfort and energy-saving. Aiming at this problem, this study first builds the space model of new buildings and then carries out simulation of energy consumption on the model. On this basis, it puts forward a fuzzy control algorithm based on thermal comfort of personnel to determine the optimal operation interval. Therefore, longer days of thermal comfort are enabled under the condition of lower energy consumption, achieving the goal of both energy saving and comfort. The energy-saving control based on the thermal comfort of personnel can promote the green operation of HVAC systems in buildings.

Abstract:The study of crystal structure is the basis for studying the physical and chemical properties of solid materials, and the screening of crystal structure is usually based on the principle of least energy. The use of density functional theory to calculate the structure energy requires a lot of computing resources and service time. For this reason, this research proposes a deep learning method for material structure prediction to speed up the prediction of material crystal structure. This work systematically studied and analyzed the data set optimization, training method, algorithm optimization, and so on. The network parameters and optimized algorithm of deep learning for crystal structure prediction are confirmed and coded. The optimized deep learning method is used to find out stable structure of Silicon, titanium dioxide, and perovskite CaTiO₃, the predicted structures are well agreement with the experimental results.

Abstract:To improve the racing performance of intelligent cars, this study introduces an intelligent racing car system based on IMXRT1021 with regard to selection of key components, design of hardware and circuit boards, and processing of sensor signals, as well as assembly, algorithms and control. This system consists of mechanical and hardware parts, PCB design, sensor signal processing, the recognition algorithm of racing track elements, control strategy and software design architecture. This study experimentally elaborates the recognition and control schemes of each racing elements. It compares the driving trajectories and absolute velocities of intelligent cars and analyzes the influence of different control algorithms on vital technical specifications such as finish time and stability. This design scheme shows its advantages in accurate control, sensitive steering and careful route planning, providing a sound reference for the students who are preparing for the four-wheel group in the National University Students Intelligent Car Race.

Abstract:In the era of information explosion, most users urgently need timely and effective recommendation service. As a result, the number of interactions, which a recommender system requires to recognize the drifted interests of existing users or the unknown preference of new users, would largely determine the application’s survival rate in the highly competitive consumer market. However, for the best of our knowledge, this responsiveness aspect of recommender system is far from well-studied. To bridge this gap, we propose a task-based meta learning approach towards responsive recommender system, which helps improve the recommendation quality for both existing and new users after the system only observes a limited number of incoming interactions. Basically, the leverage of meta learning contributes to the fast adaption to the optimum of the underlying model with few interactions to satisfy the responsiveness requirement. Extensive experiments on MovieLens and Netflix datasets highly demonstrate the responsiveness of the proposed method.

Abstract:Face recognition technology is widely used in security, business, finance and other fields. In view of the high cost and low ease of use of existing face recognition systems, a scheme based on Raspberry Pi for face recognition is proposed. First, the Harr cascade method is adopted in the OpenCV computer vision library to locate the faces in images. Then the improved MobileNetV2 model is employed to extract and classify the faces to obtain an optimized face recognition model. Finally, the model is ported to Raspberry Pi for face recognition. The model has an accuracy of 95% for people in the gallery while 80% for strangers. Experimental results show that the system remains stable in face recognition, with fast recognition speed and wide application scenarios.

Abstract:Meteorological cloud has become an important runtime environment of provincial meteorological systems. There is a big challenge to monitoring and maintaining the Jiangxi provincial cloud environment, because tranditional monitoring technology for server clusters cannot monitor virtual machines and cloud applications on the hand and fails to warn and automatically handle the failure. A monitoring and maintenance system for meteorological cloud based on Zabbix has been designed. It can monitor the layers of physical infrastructure, virutalization and application. Moreover, it can send the warning of failures in meteorological coud to the staff on duty and execute emergency recovery orders automatically in common failure scenarios. Through deployment and tests, it runs stably, markedly improving the operation and maintenance efficiency of the staff on duty.

Abstract:Amid the progress in cloud computing and big data, as well as the increasing scale and complexity of enterprise applications and the expanding demand for products, the traditional separate-architecture ERP system exposes disadvantages including poor scalability and low flexibility. In this study, we propose to use micro-service architecture to construct enterprise applications. Firstly, we analyze the characteristics of micro-service architecture. In light of the advantages of micro-service architecture, such as independent service, low coupling, and great scalability, we design the enterprise ERP system architecture based on micro-service and solve the problems of interface cooperation in ERP development. Then, we introduce the implementation technology, Spring Cloud, based on micro-service to reconstruct the application. Finally, we elaborate the development process of micro-service in an open source environment, including construction of Spring Boot subsystems and the service registration center, design of load balancing architecture and gateways. The system interface and performance testing are completed, demonstrating the advantages of the micro-service architecture based system, such as easy maintenance and scalability.

Abstract:In the past five years, China has carried out more than 15 000 law enforcement actions against problematic maps. However, the “problematic maps” that violate China’s territorial sovereignty, security, and maritime rights and interests have not been completely banned yet. More than 200 types of map products circulate widely in the market, and those with diversified forms and extensive uses directly or indirectly intensify the difficulty in supervision. This study designs and builds a map supervision system based on micro-service architecture of Spring Boot to realize online map supervision in a variety of application scenarios. Depending on this system, users can download standard maps online in the mapping link, review samples in the publishing link, and check the mobile supervision of published maps in the publishing link, realizing the online informationalized supervision of the whole process from the mapping link to practical applications.

Abstract:OPC Unified Architecture (UA) has become the core communication interface specification in Industry 4.0. However, due to the slow update of equipment in the industrial field, the traditional communication protocol is still used by a large number of field devices, and distinct communication protocols may be adopted by different equipment manufacturers. An OPC UA gateway is designed to integrate different types of field devices into OPC UA architecture, and it can break the data exchange barriers between communication protocols. An OPC UA server is implemented in the gateway, which can connect communication protocol data to the nodes in address space of the server. The OPC UA server also enables MES and ERP management software to access data efficiently. At the same time, a graphical modeling method is designed to assist users in establishing an OPC UA information model and configuring gateway data acquisition logic. Finally, an example is given to verify the usability of the gateway for graphical online modeling.

Abstract:This study analyzes the changes of the new generation power grid dispatching and control system in architecture, human-computer interaction modes, business organization modes, etc. It sorts out the new business-oriented requirements for authority management and proposes the business-oriented authority management solution with regard to this new system. The key technologies in this system are discussed, such as path-based global controlled resource identification and definition, metadata-based controlled resource management, multi-factor access control based on a rule engine, and cross-domain access control based on upper and lower organizational relationships. This solution is verified in a prototype system and provides a multi-dimensional secure access control method of controlled resources for business scenarios in the new generation power grid dispatching and control system.

Abstract:In response to users’ difficulty in searching for what they need from massive digital book resources on the Web, this study develops and implements a personalized retrieval system of digital book resources based on Web knowledge discovery. The system uses Web knowledge discovery, intelligent agent technology, data mining, and so on to design modules such as user login model, user interest generation module, and optimized search results. This achieves the influence of user behavior on interest, the update of the personalized model, and the processing of search results, further improving the retrieval quality of digital book resources on the Web. We hope this research will provide some valuable reference for the construction of personalized retrieval services of digital book resources in the same field.

Abstract:Human behavior recognition has a strong correlation with human body poses, but many open datasets for behavior recognition do not provide relevant data of poses. As a result, few recognition methods train pose data and fuse with other modalities. Current mainstream behavior recognition methods based on deep learning fuse RGB images with optical flow. This study proposes a behavior recognition algorithm based on a multi-stream convolutional neural network, which integrates human body poses. Firstly, the pose estimation algorithm is used to generate the data of key points on the human body from the static pictures containing people, and the poses are constructed by connecting the key points. Secondly, RGB, optical flow, and pose data are respectively trained on the multi-stream convolutional neural network, and the scores are fused. Finally, substantial experimental research is conducted on ablation and recognition accuracy in UCF101 and HMDB51 datasets. The experimental results reveal that the experimental precision of the multi-stream convolutional neural network integrated with pose images increases by 2.3% and 3.1% in the UCF101 and HMDB51 datasets, respectively, proving the effectiveness of the proposed algorithm.

Abstract:To solve the problem of inefficient calculation and analysis of massive seismic data in a single machine environment, we propose a distributed architecture based method for storage, calculation, and analysis of seismic data and select the complex calculation process of a noise power spectrum as the application scenario for implementation. In light of Hadoop’s performance advantage in massive data processing, the storage and scheduling of seismic data are carried out on the Hadoop Distributed File System (HDFS). The implementation of the quality evaluation method for the noise power spectrum of seismic data in Spark distributed computing architecture is studied. The elastic dataset Spark RDD is used to automatically allocate the tasks to the computing nodes, and the seismic waveform data stored in HDFS is analyzed. In addition, the calculation results are input into the distributed database HBase in the RowKey mode, realizing the storage and extraction of the power spectra of long-period seismic noise. The calculation results show that the method based on Spark distributed architecture can support the efficient processing of massive data at the TB level in volume, which can be applied to the analysis and calculation of massive seismic data.

Abstract:The accurate prediction of wind speed plays a vital role in the transformation of wind energy and the dispatching of electricity. However, the inherent intermittence of wind makes it a challenge to achieve high-precision wind speed prediction. Most studies consider the temporal correlation of wind speed but ignore the influence of meteorological factors with changes in space on wind speed. To obtain accurate and reliable forecasting results, this study proposes a MultiFactor Spatio-Temporal Correlation (MFSTC) model for wind speed prediction by combining a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. This paper also constructs a data representation method based on a three-dimensional matrix. For multiple sites, this model employs the improved PCA-LASSO algorithm to extract the characteristic meteorological factors. Then, it uses CNN to establish the spatial feature relationship among the sites and the LSTM network to establish the temporal feature relationship among historical time points. The final wind speed prediction results are obtained based on spatio-temporal correlation analysis. Furthermore, experimental verification is carried out on the 10 years of actual wind speed datasets from 2009 to 2018 provided by Dongying Meteorological Center. The results show that the MFSTC model is more accurate than common prediction methods, which proves the effectiveness of the proposed method.

Abstract:The shaking of Unmanned Aerial Vehicles (UAVs) is an important reason for the error caused by the visual sensors in extracting vital signs. To solve this problem, this paper proposes a method of respiration rate detection based on Variational Mode Decomposition (VMD), which is immune to UAV shaking. First, a complex controllable pyramid is used to extract the initial characteristics of respiration rates. Second, an extraction method of respiratory signals based on VMD is designed to obtain the candidate respiratory modal signals. Third, the eigenmodes with the minimum variance are selected for respiration rate detection. The experimental results show that the proposed method can effectively extract the intrinsic respiratory signals under the normal shaking condition of the UAVs. Moreover, this method can detect respiration rates in different human postures at different measured distances, and its detection accuracy is higher than that of the existing methods.

Abstract:The electromagnetic leakage signals recognized by manually extracted features are strongly subjective with feature redundancy. For this reason, different from the traditional artificial feature extraction mode based on experience, this study proposes a recognition method based on a Convolutional Neural Network (CNN), with the electromagnetic leakage signals of computer displays as the research object. This method employs the artificial intelligence-based deep learning method and applies the deep learning technology of image processing to the leakage feature recognition of electromagnetic information. Firstly, the time-frequency spectrum information of electromagnetic leakage signals is extracted as the input of the CNN model. Then, the deep-seated features are extracted by the self-learning ability of the model to recognize electromagnetic leakage signals from sources with different resolutions. Finally, the recognition accuracy reaches 98%, and the detection of a single signal only takes 40 ms, which verifies the effectiveness of CNN in the recognition of electromagnetic leakage signals. The proposed method provides an important basis for the early warning and protection of electromagnetic leakage and offers strong support to the restoration and reproduction of electromagnetic leakage video signals.

Abstract:The Mayfly Algorithm (MA), which serves as a new swarm intelligence optimization algorithm, proves to perform well in optimization. However, when it comes to complex problems related to high dimensions and linearity, MA is still prone to premature convergence. Thus, a new mayfly algorithm based on inversion variation (Inversion Variation Mayfly Algorithm, IVMA) is proposed. IVMA, which changes the operation of the original MA on mutation, stochastically selects the random dimension of an individual to approach that of the global optimal individual. In addition, it retains the evolution results with the elite strategy. The inversion operation is used to reverse the position of the optimal individual in a certain dimension segment, which enhances the ability of the algorithm to jump out of the local optimum. The results from ten test functions indicate that IVMA has high convergence accuracy and improved convergence performance.

Abstract:To improve the efficiency and intelligence of automatic cotton-picking machines and avoid false and missed picking, we propose an improved YOLOv4 target detection algorithm to detect single cotton in complex backgrounds. The K-means algorithm is used to screen the size of the clustering anchor frame and obtain the refined size suitable for the cotton data set. The attention mechanism is also introduced to the YOLOv4 algorithm, and the Squeeze-and-Excitation Networks (SENet) module is located in the network structure. During model training, the weights of pre-training are obtained by training on an open data set, and fine-tuning parameters of the cotton data set are applied to the pre-training model. Furthermore, the original data set is expanded through data enhancement and the pre-training model has been trained again. Experimental results show that the improved YOLOv4 algorithm proposed in this study can effectively realize cotton detection in the field environment.

Abstract:This study proposes a new cross-media retrieval model of deep hash to solve the unreasonable distribution of the hash codes of semantically similar media objects in Hamming space in the existing retrieval methods. In this model, the cross-media association loss of deep hash is improved by the Cauchy distribution in Hamming space, making the hash codes of semantically similar media objects in a short distance and those of semantically dissimilar ones far apart. Thus, the retrieval effect of the model is improved. Furthermore, an efficient model-solving method is presented in this study, and the approximate optimal solution of the model is obtained by alternating iteration. The experimental results on Flickr-25k, IAPR TC-12, and MS COCO datasets show that this method can effectively improve the performance of cross-media retrieval.

Abstract:This study proposes a new method for text detection and recognition in complex scenes to eliminate the shortcomings of a complicated text recognition process, poor adaptability, and low accuracy. This method is composed of a text area detection network and a text recognition network. The text area detection network is an improved PSENet. The backbone network of PSENet is changed to ResNeXt-101, and a differentiable binarization operation is added to optimize the segmentation network in the feature extraction process, which not only simplifies post-processing but also improves text detection. The text recognition network is formed by combining a convolutional neural network with a long short-term memory network with aggregate cross-entropy loss. The introduction of aggregate cross-entropy improves the accuracy of text recognition. Furthermore, experimental verification is carried out on two data sets, and the results show that the new method has accuracy as high as 95.6%, which is better than the previous methods. This method can effectively detect and recognize any text instances and has good practicability.

Abstract:With the development of botnets, detecting and preventing botnet attacks has become an important task of network security research. Existing studies, which rarely consider the timing patterns in botnets, are ineffective in real-time botnet detection and cannot detect unknown botnets. To tackle these problems, this study proposes a flow summary based botnet detection method. First, the network flow data is aggregated according to the source host IPs, and the flow summary records are generated in a given time window. Then, decision tree, random forest, and XGBoost machine-learning classification models are built to validate the performance of our method. The experimental results on the CTU-13 dataset show that the method we propose can effectively detect botnet traffic and detect unknown botnets. With the help of Spark technology, our method can also meet the needs of rapid detection in real applications.

Abstract:The image caption generation model uses natural language to describe the content of images and the relationship between attributes. In the existing models, there are problems of low description quality, insufficient feature extraction of important parts of images, and high complexity. Therefore, this study proposes an image caption generation model based on a Convolutional Block Attention Module (CBAM), which has an encoder-decoder structure. CBAM is added into the feature extraction network Inception-v4 and as an encoder, extracts the important feature information of the images. The information is then sent into the Long Short-Term Memory (LSTM) of the decoder to generate the caption of the corresponding pictures. The MSCOCO2014 data set is applied to training and testing, and multiple evaluation criteria are used to evaluate the accuracy of the model. The experimental results show that the improved model has a higher evaluation criterion score than other models, and Model2 can better extract image features and generate a more accurate description.

Abstract:Image de-raining is a hot issue in the low-level tasks of images, in which raindrop removal is critical. Raindrops adhering to glass or camera lenses will significantly reduce the visibility of the scenes. Therefore, removing raindrops will benefit various computer vision tasks, especially outdoor surveillance systems and intelligent driving systems. In this study, we propose a lightweight network (PRSEDNet) to remove raindrops from a single image. With recursive computation, this network uses a convolutional long short-term memory network and a feature extraction module to extract features. In combination with the original image, the final high-quality clear image without raindrops is obtained. The experimental results show that PRSEDNet, compared with the existing deep learning-based raindrop removal algorithms, possesses few parameters and high computational efficiency and can achieve efficient raindrop removal.

Abstract:Community detection task is a hotspot in data mining. In recent years, deep learning and graph data have been increasingly diverse and complex, and the task of hierarchical community detection has gradually become a focus of research. The goal of this task is to learn the hierarchical relationship between communities while gathering similar nodes in homogeneous graphs to better understand the graph data structure. The introduction of this relationship poses a higher modeling challenge to community detection algorithms. For this task, some effective heuristic methods have been proposed. However, limited by the simple assumptions of community distribution and discrete optimization learning methods, these methods cannot describe more complex graph data, nor can they be combined with other effective continuous optimization algorithms. To solve this issue, we first attempt to model a complex overlapping hierarchical community structure and propose a simple dual-task optimization model of node embedding and community detection. The relationship of nodes and overlapping hierarchical communities can be flexibly explored through gradient updates. In the learning process, we can also obtain the embedding representations of nodes and communities to apply to rich downstream tasks.

Abstract:FastText is a precise and efficient text classification model, but the precision is low when it is directly applied to Chinese long text classification. Regarding this problem, this study proposes a FastText method for Chinese long text classification, which combines TextRank key clause extraction with Term Frequency-Inverse Document Frequency (TF-IDF). Firstly, TextRank is used to extract the key clauses of the text as input features. Secondly, key words of the text are extracted by TF-IDF as a feature supplement. Finally, the extracted text features are input into the FastText model, which can preserve the key features of the target text while reducing the training corpus. The experimental results show that the accuracy of the proposed method on the datasets is 86.1%, which is about 4% higher than the classic FastText model.

Abstract:In a wireless sensor network, the sensor has limited energy. If it runs out of energy, the robustness and lifespan of the network will be greatly reduced. Therefore, a data aggregation mechanism based on fuzzy reinforcement learning and fruit fly optimization is proposed to maximize the lifespan of the network and perform efficient data aggregation. First, grid clustering is applied to cluster formation and cluster head selection. Then, all possible data aggregation nodes of each grid cluster are evaluated, in which the best one is selected by fuzzy reinforcement learning. Finally, the fruit fly optimization algorithm is adopted to dynamically position the data aggregation nodes of the entire wireless sensor network. The simulation results show that the proposed scheme is better than the comparison scheme in terms of energy consumption and network robustness.

Abstract:To reduce the delay of processing intensive computing tasks by terminal equipment in pumped storage power stations , this paper proposes a task offloading algorithm based on edge computing for the Internet of Things (IoT) system of pumped storage power stations. Firstly, the computing tasks are prioritized based on the analytic hierarchy process, and an offloading model is built with the terminal energy consumption as the constraint and the processing delay of terminal computing tasks as the optimization objective. Secondly, the Q-Learning (QL) algorithm is adopted to collect the state transition information, in the hope to obtain the best offloading strategy between the terminal device and the edge node. Finally, Deep Learning (DL) is used to map the relationship between states and actions to avoid a dimensional explosion in the iterative solution of the algorithm. The simulation results show that the proposed method greatly reduces the average delay of computing tasks and can greatly improve the execution efficiency of production operations and safety monitoring associated with pumped storage power stations.

Abstract:The Automatic Identification System (AIS) can provide the time stamp, latitude and longitude, heading angle, speed, and other data information of ships. In light of multi-dimensional ship trajectories and the demand for accuracy and timeliness in trajectory prediction, this study proposes a calculation method for trajectory similarity based on image detection and matching. To be specific, the trajectory data of all fishing boats are first visualized, and then the Oriented FAST and Rotated BRIEF (ORB) algorithm and Brute-Force (BF) matching are used to calculate the similarity between trajectory pictures for classifying fishing boat trajectories. The experimental results show that the proposed method has high accuracy and can be easily implemented and is superior to the traditional methods in the processing efficiency and speed of trajectory data.

Abstract:Developers often search for appropriate APIs to complete programming tasks during software development. Many API recommendation approaches and tools have been proposed to improve the efficiency of software development, but most of these approaches do not consider user interaction information. In this study, we propose an API recommendation tool based on client/server architecture and integrate it into the VS Code IDE in the form of a plug-in. In our tool, initial API recommendation lists are generated by existing tools. Learning-to-rank and active learning techniques are used, combined with user feedback information, to re-rank the API recommendation lists, achieving personalized recommendation. Extensive experiments demonstrate that the performance of this tool has been steadily improved with the increase in feedback data.

Abstract:To deal with the inaccurate classification caused by a failure of quick and effective extraction of image features in the remote sensing image scene classification based on existing machine learning methods, we propose a remote sensing image scene classification method based on residual attention network. With the residual network as the benchmark model, attention modules are created in the dimensions of channel and space. For effective classification of the UC Merced Land-Use dataset, parameters are set reasonably and the model that optimizes the number of network layers is fine-tuned. The results show that the accuracy of our method reaches 98.1% compared with that based on the convolution neural network.

Abstract:With the increase in global vehicles and the expansion of road surfaces, pavement crack detection has received extensive attention in recent years. Many detector models have been proposed, with some problems, though. For example, some narrow cracks may not be detected, leading to discontinuous cracks; the detailed crack edge information may be lost during filtering or pooling. On the basis of SegNet, a continuous attention mechanism is designed in the encoder layer, and a convolutional pyramid structure is added before the feature map passes through the decoder layer to reduce the fracture in crack detection and obtain more complete edge information. The Precision, Recall, and F1-measure of our approach are 2.47%, 8.21%, and 6.87% higher than those of the related method, respectively, and the Mean Intersection over Union (MIoU) of the detection results on the three open datasets, namely, Crack200, Crack500, and CrackForest is improved by 14.35%.

Abstract:Automatic method naming, as an important task in software engineering, aims to generate the target function name for an input source code to enhance the readability of program codes and accelerate software development. Existing automatic method naming approaches based on machine learning mainly encode the source code through sequence models to automatically generate the function name. However, these approaches are confronted with problems of long-term dependency and code structural encoding. To better extract structural and semantic information from programs, we propose a automatic function naming method called TrGCN based on Transformer and Graph Convolutional Network (GCN). In this method, the self-attention mechanism in Transformer is used to alleviate the long-term dependency and the Character-word attention mechanism to extract the semantic information of codes. The TrGCN introduces a GCN-based AST Encoder that enriches the eigenvector information at AST nodes and models the structural information of the source code well. Empirical studies are conducted on three Java datasets. The results show that TrGCN outperforms conventional approaches, namely code2seq and Sequence-GNNs, in automatic method naming as its F1-score is 5.2% and 2.1% higher than the values of the two approaches, respectively.

Abstract:The JavaScript malicious code detection by existing machine learning methods is complex, with large amount of calculation and difficult detection caused by maliciously confused codes. Existing approaches, therefore, fail to realize accurate and real-time detection. For this reason, a method based on Bidirectional Long Short-Term Memory (BiLSTM)-based method for JavaScript malicious code detection is proposed. Firstly, standardized data adapting to be input into the neural network is obtained by code de-obfuscation, data segmentation, and code vectorization. Secondly, the BiLSTM algorithm is used to train the vectorized data and learn the abstract features of JavaScript malicious code. Finally, the abstract features are used to assort codes. The proposed method is compared with deep learning and existing mainstream machine learning approaches, and the results show that this method exhibits a higher accuracy rate and a lower false alarm rate.

Abstract:In order to improve the test quality of aerospace embedded software and ensure the successful completion of aerospace tasks, we study code logic analysis, one of the important contents of code inspection for aerospace embedded software. We analyze the mechanism, finding, exposure, and consequences of software defects and summarize many years of engineering practice in software testing. On this basis, we put forward ten methods for code logic analysis, such as scene analysis, time sequence analysis, and imaginary fault source tracing. The application of code logic analysis methods are analyzed and code inspection is compared with other test methods, through which the engineering applicability of code inspection is given. The methods have been widely adopted in the third-party testing of aerospace software. Practical data show satisfying application effect as the defect detection rate of code inspection has increased from generally accepted 30%~70% to more than 90%. The methods and ideas can provide reference for the design of dynamic test and the research and development of automatic detection tools for software defects.

Abstract:The user profile technology is difficult to popularize in power companies and brings little effect. With regard to this, a hierarchical clustering profile recommendation model based on an improved “weighted K-means algorithm based on a firefly algorithm” is proposed. To improve the calculation speed and accuracy of user profile construction, this model designs a label model for a single business and focuses on constructing characteristic group profiles through hierarchical clustering. In profile application, the specific business and other new business are directly recommended to potential users in the target group. A simulation experiment is conducted on a sample set of high-voltage power users. The results show that the proposed model can effectively improve the calculation speed and accuracy of clustering and profile construction and application and facilitate the popularization and application of the user profile technology in power companies.

Abstract:This study focuses on the electric load forecasting of the core link in the power grid. On the basis of summarizing and analyzing the research results of previous researchers, a method based on the combination of pre-training GRU and LightGBM is proposed. This method first uses electrical load data to train a feature extraction network GRU, then uses the network to extract timing features, and uses LightGBM to predict the electrical load of the extracted timing features and non-sequential features. The innovation of this method is to propose a pre-training network to expand the features and fully integrate the timing features and non-timing features. And taking into account the regional differences of the power grid, the GRU network parameters were adaptively fine-tuned during the overall training process. Ensure that the extracted time series features are consistent with the current regional characteristics. Finally, it is found through simulation experiments that this method has achieved a 2% improvement in various indicators.

Abstract:With the continuous development of smart grids, diversified power service types lead to different service demands. The 5G network slicing technology can provide virtual wireless private networks for smart grids in response to challenges in security, reliability, and time delay. Considering the differentiated service characteristics of smart grids, this study aims to use Deep Reinforcement Learning (DRL) to solve the resource allocation of the Radio Access Network (RAN) slices of smart grids. This study reviews the background of smart grids and the related research on network slicing technology, then analyzes the RAN slicing model of smart grids, and proposes a slice allocation strategy based on DRL. Simulation results show that the proposed algorithm can reduce the cost and meet the resource allocation requirements of smart grids on the RAN side to the maximum extent .

Abstract:The transaction flow data of campus cards is studied to mine the interesting information to provide scientific basis for decision management. According to the consumption data of all canteens from January 2014 to February 2019, the Holt-Winters multiplication model and the ARIMA model are built with noise canceled by smoothing. The discrete time series composed of monthly data during this period is fitted and analyzed; the consumption trends in March to May 2019 are forecasted, the results of which are tested by actual values. The experiment shows that the Holt-Winters model can well fit the consumption data at higher forecast accuracy. Our method is helpful to master the consumption behavior of teachers and students in the canteens and lays the groundwork for logistics departments to optimize resource allocation and scientific decision-making.

Abstract:Nowadays, with a higher proportion of wind power generation, wind power prediction is increasingly demanding. It is, however, not satisfyingly accurate due to the intermittent and uncertain nature of wind energy. In order to reduce the complexity of the prediction model and improve accuracy, we propose a type-2 fuzzy system based on differential evolution and rule reduction in this study. This model provides a pruning strategy for the reduction in type-2 fuzzy rules, based on which and the differential evolution algorithm is adopted to optimize the parameters of the reduced type-2 fuzzy system. Finally, the prediction accuracy of our method is higher than that of the type-1 fuzzy system and the support vector regressor, according to comparison.

Abstract:The purpose of this study is to improve the recognition rate and accuracy of fraud calls. We collect the communication process data such as users’ behavior of having telephone communications and surfing the Internet by a big-data platform and conduct a comprehensive analysis combined with users’ basic attributes and mobile terminal information; also, an identification model is built by the appropriate recognition algorithm for machine learning. The proposed method can better find the internal differences between fraud calls and ordinary ones. Compared with traditional analysis based on call behavior, it can effectively improve the identification accuracy and coverage of prank and fraud calls and reduce false negatives and false positives. The proposed method performs prominently better in fraud call identification, which can be used as a new technology choice, according to actual data verification.

Abstract:Noise monitoring systems can automatically measure decibel level and process various sound environment information in real time. However, in their practical application, the noise decibel is affected by many factors such as temperature, humidity and atmospheric pressure, which leads to the errors between measured and actual values. In view of this, the correction based on relevant technologies becomes a necessity for the accuracy improvement of noise measurement. This study adopts linear regression and Back Propagation (BP) neural network to investigate the factors and coefficients of the prediction model and analyzes the correlation of factors in the model. As a result, the automatic correction model of noise monitoring is obtained. The test effect of automatic data correction by linear regression and BP neural network indicates that the fault tolerance of measurement data is optimized and the accuracy of data correction is improved. Further, the determination coefficient (R²) of the prediction model is greatly increased.