Abstract:In recent years, more and more applications or micro services have been deployed to the cloud. The virtual network is the basic guarantee in a cloud environment. Virtual Network Interface Cards (NICs) and virtual network bridges are built based on physical NICs in virtualization of NIC to construct the virtual network for virtual instances such as virtual machines and containers, and these virtual network devices are uniformly configured and managed. From the perspective of the virtual NIC and the virtual network bridge, the current popular virtualization technologies of NICs are sarveyed, classified, and compared in this study. Finally, the current situation and future development in NIC virtualization are analyzed.

Abstract:In large-scale, data-intensive application scenarios, the disadvantages of row storage access to data become increasingly prominent, which is gradually unable to meet the performance requirements for high-speed data access, and data needs more efficient transmission and processing methods. Therefore, it is of great significance to expand new memory access methods compatible with row and column access at the same time to improve access efficiency, reduce overall power consumption, and save memory space. This study focuses on dynamic random storage and non-volatile storage to introduce the column-oriented memory access method in detail, highlighting the analysis of the structural design for storage units and the implementation of the column-oriented memory access process. Finally, the two different ways of memory access are compared and summarized, and row- and column-oriented in-memory databases, data mining, data encryption algorithms, and the application scenarios of real-time systems are forecasted.

Abstract:Cosmological simulation is an important research method that can help scientists understand the universe’s evolution process and verify the theoretical model. Visualization is one of the most effective ways of analyzing simulation data. Scientists can improve analysis efficiency through visualization and interactive exploration of the simulation data. With the development of supercomputers and cosmology theory, numerical simulation achieves a larger scale and higher accuracy, which brings a variety of requirements and challenges for scientific visualization. In this study, we summarize the main algorithms of visualization in cosmological simulation and demonstrate the role of visualization in analyzing the simulation data through multiple research cases. In addition, we also point out the current research hotspots and challenges.

Abstract:Edge computing can effectively solve the problems of large transmission delay, insufficient user data security, high transmission bandwidth pressure, limited computing capabilities of terminal mobile devices, and high energy consumption in traditional cloud computing. Computing offloading is a key technology in edge computing. Concerning the research status and existing deficiencies of computing offloading technology, this study first introduces the architecture of edge computing and some applications and analyzes the four main influencing factors and corresponding specific conditions. Secondly, the algorithm strategy and the role of corresponding variables in the algorithm are analyzed in three decision objectives. Finally, it summarizes the current deficiencies in computing offloading.

Abstract:Many methods for gene biomarker selection can not be directly used in clinical diagnosis because of a small number of research samples. Therefore, some scholars proposed methods of integrating different gene expression data while preserving the integrity of biological information. However, due to the batch effect, direct integration of different gene expression data may bring new systematic errors. In response to the above problems, an analysis framework integrating self-paced learning and SCAD-Net regularization is proposed. On the one hand, self-paced learning can learn the basic model from low-noise samples and then make the model more robust through high-noise samples to avoid batch effect. On the other hand, SCAD-Net regularization combines biological interaction information and gene expression data, which can achieve a better performance in feature selection. The simulation data in different cases and the results on the breast cancer cell line dataset show that the regression model based on self-paced learning and SCAD-Net regularization obtains better prediction results when dealing with high-dimensional complex network datasets.

Abstract:Mutations are induced when cells are stimulated by carcinogens. The mutational process causes a certain pattern of changes in the genome, which are called mutational signatures. Mutational signature analysis is an important task in clarifying the carcinogenic mechanism of carcinogens and driving the development of cancer research, and it will provide new insights and options for early tumor diagnosis and individualized treatment. The breakthroughs and developments of next-generation sequencing have led to the identification of massive somatic mutations, making it possible to mine mutational signatures from large-scale genomes. This study elaborates the mathematical model for mutational signature identification and introduces alternative methods and important parameters. It systematically and comprehensively compares mainstream algorithms and software and specifies the precautions for mutational signature extraction. Finally, it forecasts the future development trend of this field.

Abstract:Diversified expression of medical information and inefficient utilization of medical knowledge are encountered in online consultations. This study designs and implements an interactive intelligent diagnosis guidance system based on medical knowledge graphs. The system introduces medical knowledge graphs to provide medical knowledge and relies on entity recognition and entity linking technology to standardize the medical expression in the main condition description text. Moreover, it uses the medical entity to generate knowledge subgraphs and obtain their semantic information and merges the semantic information of the subgraphs and patient condition description to obtain department confidence. When the confidence of the recommended department is low, multi-round interactive inquiry can supplement the patient’s symptom information, and finally, the recommended department is determined. The system can provide technical support for building a fast and accurate intelligent medical system to improve the efficiency of diagnosis and alleviate the shortage of medical resources.

Abstract:Modern combat System of Systems (SoS) generally consists of complex systems across various fields. Accordingly, multiple modeling and simulation tools are used for SoS development. As a result, several problems are caused. For example, heterogeneous data is hard to share among modeling tools, and the collaboration between modeling and simulation tools is difficult. To address these problems, we propose a new multi-architecture SoS modeling and simulation platform with the development capability including meta-model development, SoS modeling, as well as simulation and verification. Then, we demonstrate the platform through a man-aircraft cooperative defense scenario in the air-ground cooperative defense SoS. Specifically, the meta-models for the case scene are designed, and the operational view models are built via the proposed platform. In addition, the simulation of aircraft behavior in combat missions provides a theoretical basis for the trade-off of combat plans.

Abstract:The traditional management method for military vehicles has failed to adapt to army informatization, and it is urgent to use an efficient management mode integrating new technology. Therefore, aiming at the security authentication and authorization involved in management, this study designs a remote authorization and monitoring system for military vehicles based on artificial intelligence. The system adopts the latest face recognition technology and network technology, comprising vehicle, server, and client terminals. The vehicle terminal is deployed on the NVIDIA Jeston Nano AI development board in the driver’s cab, and it depends on a face recognition model to authenticate the driver and carries out data interaction with the server terminal. The server terminal is placed in the cloud server to manage the information of the vehicle and the client terminal, summarize the data uploaded by the vehicle, and establish data interaction between the vehicle and the client. In addition, the client terminal is deployed in the mobile intelligent terminal for remote authorization, remote monitoring, etc. The test shows that the system can realize the automatic certification, remote authorization, and remote monitoring of vehicles and prevent the illegal use of them to guarantee their safety.

Abstract:For many material companies, material design, manufacturing, and characterization are usually completed in different departments, and then the life-cycle data of materials are separately located. Also, material characterization data is presented in data files with their specific format of characterization instruments. The data of material properties and manufacturing, which is discretely located, is not appropriate for using a novel data-driven approach for material research and development. This initiates a need for a data infrastructure that can help to integrate these separately located material data to formulate a database system. However, it is challenging to develop such a data infrastructure due to diverse material characterization instruments and manufacturing processes. To address this need, we develop a general data infrastructure that can help to build a specific material characterization and manufacturing database by following the CSTM material genome engineering data principle and employing the zero-code development idea. In this study, we illustrate the design and development of the data infrastructure and use it to build a characterization and manufacturing database for liquid metal. It is demonstrated that the idea and data infrastructure for building a specific material characterization and manufacturing database are appropriate and useful.

Abstract:With the development of society, the authenticity of talent resumes is crucial for enterprise recruitment. The problems such as centralization of data management and falsification of resumes exist in the traditional resume storage based on the third-party organization. Given the above problems, a Blockchain-based searchable encryption talent resume sharing scheme is proposed. Firstly, the authenticity of resume data is guaranteed with the tamper-proof feature of Blockchain and the collaborative storage mechanism on and off the chain. Secondly, during resume data sharing, the traditional third-party organization is replaced by the decentralized feature of Blockchain, and the fairness between participants and privacy protection during data storage and update are ensured by combining searchable encryption technology. Finally, the scheme is analyzed from the perspective of security and correctness, which proves the ability of the scheme to solve the above problems.

Abstract:In response to power line inspection, a power line safety distance monitoring and early warning system based on multiple low-cost LiDARs is designed and implemented. The single-line LiDAR (RPLIDAR S1) and RoboMaster GM6020 gimbal motor are combined into three-dimensional equipment for environmental perception. Four sets of the same equipment are installed around an electric bucket for environmental perception. After a target paper is arranged, the point cloud data is collected, and the coordinates of the target paper’s global coordinate system are measured simultaneously. Then the spatial position relationship among the four LiDARs is calibrated according to the least squares principle and unified in the same coordinate system. The real-time point cloud data collected by the four sets of equipment is transmitted to the data processing platform through network communication for visual display, safety distance monitoring, and early warning, which can avoid safety hazards and improve the safety level of live line work.

Abstract:Under the pressure of rapid economic development and insufficient road construction, the urban traffic burden becomes increasingly heavier. Under the existing resource conditions, how to improve the utilization efficiency and traffic efficiency of roads is an important way to alleviate traffic problems. In this study, with an intelligent signal intersection based on a traffic flow control subsystem as the goal, we firstly analyze the related algorithms of the traffic control system and the present situation of the information system. Based on the existing system, we design the architecture of the traffic control system, comprising the intersection subsystem, regional center, as well as information and control center. The intersection subsystem realizes the real-time acquisition of traffic flow data and the local intelligent control of the intersection. The detection scheme of the intersection subsystem is designed and implemented. The data acquisition and local intelligent control of the intersection subsystem are the keys to cloud data storage and real-time release from the traffic information platform, and it is of great significance to the coordinated control of multiple intersections.

Abstract:Given the diverse prohibited varieties in current security inspection scenes and low-efficiency error-prone manual inspections, this study proposes network architecture, Res152-YOLO, based on the YOLOv4 optimized target detection network. Res152-YOLO uses the ResNet-152 network to replace the CSPDarknet-53 network in the original YOLOv4 and connects the improved ResNet to the YOLOv4 network to enhance the detection accuracy of dangerous goods in X-ray images. A series of networks such as YOLOv4 and Res152-YOLO are used to conduct comparative experiments on the same data set to compare the loss curves of the above-mentioned networks, the detection results for various dangerous goods, and the overall performance of each network. The results show that the improved network can improve the accuracy of security detection and eliminate potential safety hazards.

Abstract:Large-scale installation and application of monitoring equipment have brought about difficulties in operation and maintenance. This study develops a general operation and maintenance management platform. Through the front-end status collection device, the platform can obtain various state information generated during the operation of monitoring equipment. The operation and maintenance management system of the platform collects and stores status information and provides management functions such as monitoring, statistics, analysis, and alarm to support operation and maintenance. The visualization subsystem of the platform visualizes the status information and provides intelligent support for system operation and maintenance. The modular and redundant design of the status collection device makes the equipment universal, expandable, highly reliable, and easy to maintain. The status information has the characteristics of time series data. The operation and maintenance management system uses InfluxDB to store the status information, which greatly reduces the need for storage space and ensures the performance of data query and management. The general operation and maintenance management platform of monitoring equipment based on InfluxDB has been installed and tried out in multiple user units, and it is running in good conditions, with good economic and social benefits and promotion value.

Abstract:Given that the existing multi-target tracking algorithm cannot track accurately after occlusion, a multi-target tracking algorithm using the improved attention mechanism and Kalman filter is proposed. The structure of joint detection and embedding is used to extract features and accomplish object detection and identification simultaneously. A parallel-structured attention mechanism is proposed, containing both spatial and channel parts. Each part is designed into parallel branches for pooling and convolution. During tracking, the proposed velocity-prediction Kalman filter is adopted for the more accurate prediction of pedestrian movements. The CUHK-SYSU dataset is used for training, and the algorithm is verified and tested on the MOT16 dataset. The proposed algorithm can achieve 65.1% MOTA, 78.8% MOTP, and 62.3% IDF1. The experimental results show that the proposed tracking algorithm can improve the overall tracking performance and achieve continuous tracking.

Abstract:The satellite constellation which have strict format and involve query and calculation of multiple telemetry parameters. In order to solve the problem of large workload and long time consuming in preparing health management documents of satellite constellation, a method of document automatic generation is proposed. By classifying and analyzing the basic data types of the document, making the configuration file storage rules, customizing the document templates, and applying the automatic document generation algorithm to translate the document template into data-gathering documents, the knowledge reuse and general content generation in the process of document preparation are realized. The standard and effective document preparation process is established.

Abstract:In a large-scale network, the security threats faced by the host are becoming increasingly diverse. With the rapid rise of technology based on machine learning to detect malicious files, the ability to detect malware has been greatly improved, and it has also forced adversaries to change their attack strategies. Among them, the “Living off the land” strategy achieves malicious behavior by calling operating system tools or automated management programs that perform tasks. Threat detection can find suspicious behavior in the context of parent and child processes. The parent-child process chain and the related events derived from it are regarded as an undirected graph, and the supervised learning XGBoost algorithm is used for weight distribution to generate an undirected weighted graph. Finally, a community discovery algorithm is employed to identify larger attack sequences from the graph. The above algorithm is verified on the simulated attack dataset of MIRTE ATT & CK.

Abstract:Different from the traditional deep reinforcement learning method of training through transitions selected one by one from the experience replay, for the Deep Q Network (DQN) that uses the entire episode trajectory as the training sample, a method for expanding episode samples is proposed, which is based on genetic algorithm crossover operators. The episode trajectory is generated during the trial-and-error decision-making process of the interaction between the agent and the environment, in which similar key states will be encountered. With the similar state in the two episode trajectories as the intersection point, the episode trajectory that has not appeared till present can be generated to enlarge the number of episode samples and increase their diversity, thereby enhancing the agent’s exploration ability and improving sample efficiency. Compared with DQN that randomly selects samples and uses the Episodic Backward Update (EBU) algorithm, the proposed method can achieve higher rewards in the Playing Atari 2600.

Abstract:The model-based software testing method is one of the most effective ways to improve the quality and reliability of embedded real-time system software. However, general models usually lack the description of their real-time characteristics and software behavior for the reason that this type of software is complicated with the real-time characteristics. As a result, a wealth of professional domain knowledge is required to build them more accurately and completely. This gives modeling a rise in difficulty and cost. It is thus difficult to guarantee the adequacy and effectiveness of the test. The usage scenario is an example of the interaction between the user and the software, which describes the system behavior of the software in detail without paying attention to its internal complex structure. Therefore, to reduce the difficulty of modeling, this study builds the model based on the standardized representation of the usage scenario and uses the time-extended EFSM model to describe the real-time characteristics of this type of software; to ensure the integrity of the model, this study designs the evaluation criteria for model integrity to determine whether the model fully characterizes the behavior of the system by verifying the integrity of the constraints in the model transitions; for the incomplete model, a to-be-completed transition generation strategy is designed according to the constraints to generate the to-be-completed transition, and it is completed into the model through the execution process of the feasible transition sequence of the dynamic simulation model to enhance model integrity; finally, this study builds a time-extended EFSM model for four pieces ofembedded real-time system software and carries out a series of experiments. Experiments show that the method proposed in this study can not only build the model but also complement the generated to-be-completed transitions to the model, further improving model integrity.

Abstract:This study explores the online game flow identification based on representation learning. First of all, due to the lack of game flow in the public data set in the field of flow identification, various types of game flow are collected, and a mapping relationship between various games and process ports is established. Depending on the mapping relationship, the game flow is filtered from the collected flow to expand the public data set. Then the representation learning model in deep learning is used to automatically perform feature learning and feature selection on the pre-processed original end-to-end game flow. Finally, the game category is identified by a classifier. The convolutional neural network self-learns the features of the original information via the construction of feature space, successfully avoiding the loss of information caused by the secondary processing of the flow data set in the traditional machine learning algorithm and the dependence of the flow classification model on feature selection. The experimental results show that, compared with the classification effect of the original data set, the expanded data set has a classification accuracy improved by 5% on the neural network model. The accuracy of game flow identification reaches 92%, and the identification performance is significantly improved.

Abstract:To realize the multi-target localization based on Received Signal Strength Indicator (RSSI), this study proposes a new Distributed Multilateral Fusion Localization (DMFL) algorithm with point estimation and ellipse estimation. Firstly, the rough locations of targets are estimated by the multilateral localization algorithm. Then, in light of the interval analysis theory, the higher-order remainder bound of Taylor series expansion is obtained and the set-membership recursive algorithm is used to solve the problem of multi-target localization. Finally, the performance of the localization algorithm is verified through experiments and simulations. The results show that, under the same nodes layout conditions, the algorithm improves the accuracy of localization compared with the latest localization algorithms, with the maximum error being less than 0.3 m. Moreover, it can determine the optimal regions that contain the real locations of targets.

Abstract:Traditional collaborative filtering algorithm relies much on ratings among users, which is prone to cold start and data sparsity. In addition, the recommendation results are single. To solve the above problems, this study proposes a diversified movie recommendation algorithm based on trust factor. Firstly, the calculation method of user similarity is improved, and the trust relationship and attribute characteristic information between users are introduced. Next, clustering is conducted to divide users with the same interest into the same community. Finally, user activity, as the movie recommendation degree, is taken into consideration comprehensively in the rating. The penalty factor is introduced, so as to facilitate personalized and diversified movie recommendations for target users. Experimental results show that the proposed algorithm can improve the recommendation accuracy and diversity, achieving a good recommendation effect.

Abstract:Artificial intelligence accelerates the development of the risk control industry. Undoubtedly, risk control is the core of intelligent risk control, and a credit default prediction model is its essential means. The traditional access to risk control is based on artificial and generalized linear models. However, the data of transactions completed on the Internet are characterized by high dimensions and multiple sources, which cannot be processed by existing models. This poses a great challenge to traditional risk control. In view of this, this study proposes an interpretable credit default model based on the fusion method. To be specific, the accuracy of the prediction results is first enhanced through the fusion of base models (LightGBM, DeepFM, and CatBoost) and secondary model (CatBoost). Then, the prediction result of the fusion model is interpreted by the introduced local-based interpretability method LIME that is independent of the model. According to the experimental result of a real dataset, the satisfactory accuracy and interpretability of the model can be witnessed on the task of credit default prediction.

Abstract:API-related knowledge is usually scattered among multiple information sources, such as API reference documentation, Q&A forum and other unstructured texts, which is not conducive to API query and retrieval. To improve the efficiency of API retrieval, this study proposes an API knowledge graph construction method based on multi-source information fusion. API reference documentation describes the function and structure of the API from a designer’s perspective, while Stack Overflow provides the purpose and use scenarios of the API from a user’s perspective. API reference documentation and Stack Overflow complement each other, and they can provide support for API query and retrieval together. By means of analyzing API reference documentation, API and domain concepts can be extracted as entities and their relationships can be constructed. With the Stack Overflow website, Q&A and API concepts can be extracted as entities and their relationships can also be constructed. On this basis, these two kinds of information are fused to construct a multi-source API knowledge graph for the API recommendation based on the knowledge graph. To verify the proposed method, this study evaluates the effectiveness of the API knowledge graph in terms of the accuracy of knowledge extraction and the API recommendation. The experimental results show that the recommendation effectiveness and efficiency based on our knowledge graph have been improved.

Abstract:In this study, a 3D point-cloud target detection algorithm for vehicles based on attention mechanism is proposed for the recognition and positioning of the targets in autonomous driving scenarios. The algorithm first divides the sparse and disordered point cloud space into equidistant and regular voxel representations. Then, 3D sparse convolution and auxiliary network are used to synchronously extract the internal point cloud features from all voxels. Afterward, a bird’s-eye view is generated. After the internal 3D point cloud features are converted into a 2D bird’s-eye view, the spatial feature information of the target will be lost generally, which makes the final detection result and the direction prediction unsatisfactory. To further extract the feature information of the bird’s-eye view, this study also proposes an attention mechanism module, which contains two attention models and adopts a three-dimensional layout structure (front, middle, and back) to realize amplification and suppression of the feature information of the bird’s-eye view. The convolutional neural network and PS-Warp transformation mechanism are employed to perform 3D target detection on the processed bird’s-eye view. Experiments show that, under the premise of ensuring real-time detection efficiency, this algorithm has better direction prediction and higher detection accuracy than existing algorithms.

Abstract:As a hot spot of human-computer interaction, emotion recognition has been applied in many fields, such as medicine, education, safe driving and e-commerce. Emotions are mainly expressed by facial expression, voice, discourse and so on. Other characteristics such as facial muscles, mood and intonation vary when different kinds of emotions are expressed. Thus, the inaccuracy of emotions determined using a single modal feature is high. Considering that the expressed emotions are mainly perceived by vision and hearing, this study proposes a multimodal expression recognition algorithm based on an audiovisual perception system. Specifically, the emotion features of speech and image modalities are first extracted, and a plurality of classifiers are designed to perform emotion classification experiments for a single feature, from which multiple expression recognition models based on single features are obtained. In the multimodal experiments of speech and images, a late fusion strategy is put forward for feature fusion. Taking into account the weak dependence of different models, this work uses the weighted voting method for model fusion and obtains the integrated expression recognition model based on multiple single-feature models. The AFEW dataset is adopted for facial expression recognition in this study. The comparison of recognition results between the integrated model and the single-feature models for expression recognition verifies that the effect of multimodal emotion recognition based on the audiovisual perception system is better than that of single-modal emotion recognition.

Abstract:Based on the ZU Chongzhi (ZUC) algorithm, this study designs and implements a reversible watermarking algorithm for separable encrypted images. In the algorithm, the content owner first marks the image to produce a position map and then runs the ZUC encryption algorithm to encrypt the original image. After the watermark embedder obtains the encrypted image, the watermark information is embedded into the second most significant bit or the most significant bit position of the selected pixels according to the position map. Through the encryption key and the embedding key, the receiver can get the directly decrypted image, watermark information, and the recovered image. The ZUC algorithm is adopted to encrypt and decrypt the image to ensure the security of the algorithm. The image was marked before the watermark information is embedded in the selected positions. The receiver employs an adaptive difference algorithm based on the correlation between adjacent pixels to extract watermarks and restore the image, ensuring the quality of the recovered original image and the directly decrypted image. Experiments show that the proposed algorithm has high security and achieves a separable effect. At the same time, the recovered original image and the directly decrypted image also have high quality.

Abstract:Lane detection is one of the most important modules in self-driving tasks. Lane detection is a challenging task as the structure of the lane line is special, and the detection is easily affected by various environments (such as lighting transformation, obstruction, and the blur of the lane line). Considering the traditional Convolutional Neural Network (CNN) is unable to learn fine spatial features of the lane line directly, in this study, the spatial feature aggregation module is employed to enhance the features extracted by CNN in spatial dimensions, providing rich spatial features for the cascade lane predictor. The experiments show that the module learns fine global information by aggregating feature maps in horizontal and vertical directions and thus improves the performance of the lane detection algorithm in different environments without reducing the detection speed.

Abstract:With the informatization of medical data, the Electronic Health Record (EHR) plays an increasingly important role in remote medical treatment and medical research. To solve the problem of EHR interaction in remote medical treatment, this study proposes a cross-organizational EHR interaction platform among hospitals. According to the business process of the platform, the activity and resource elements of each organization are extracted. According to the relationship between the activity and resource of each organization in the interaction process, the activity and resource elements of each organization are extracted from different perspectives of various organizations by combining the polychromatic set and perimeter matrix. At the same time, it studies the business process and proposes the corresponding formal description model.

Abstract:This study aims at the problem that the distillation effect decreases when the gap between the teacher network and the student network in relational knowledge distillation is too large. A stepwise neural network compression method based on relational distillation is proposed. The key point of this method is to add an intermediate network between the teacher network and the student network for relational distillation step by step. Moreover, in each distillation process, additional monomer information is added to further optimize and enhance the learning ability of the student model. The experimental results show that the classification accuracy of the proposed method on CIFAR-10 and CIFAR-100 image classification datasets is improved by about 0.2% compared with that of the original relational knowledge distillation method.

Abstract:When using the existing search engines, users often fail to construct clear and accurate query words, which leads to poor retrieval results. Traditional query recommendation methods do not fully consider the relevance of user behavior, resulting in inaccurate query recommendation results. This study builds a new query recommendation model, which is based on the click model and network embedding. Firstly, the model embeds the user’s history view behavior and click behavior through the click chain model and measures the relevance between the query and the returned documents through the attention mechanism; secondly, it uses the attribute heterogeneous network to obtain the potential semantic information in a complex heterogeneous network structure; finally, it captures the complex information in multiple spaces through multi-head attention and uses multi-task learning to make score prediction. The experimental results on the public query log provided by SogouLabs show that our model is superior to the baseline model in both discriminative and generative tasks.

Abstract:Concerning the complicated process, interdisciplinarity, and poor real-time performance of enterprise named entity recognition, a method based on concurrent subspace optimization is proposed. First, a target-constrained equation of the system is established to complete system-level optimization; secondly, a two-level model of text detection and recognition is constructed, and the model is selected, considering the advantages and disadvantages of different existing models, to optimize the discipline in parallel; then, the connection of the two-level model is constructed with the image threshold, grayscale and Hoff transform; finally, simulation experiments verify that the recognition accuracy of this method is 9% higher than that of other two-level text detection and recognition models, and the speed increases by about 20%.

Abstract:Short text matching is a core problem in the field of natural language processing, which can be applied to tasks such as information retrieval, question answering systems, and question paraphrase. Most of the past work only considered the internal information of the text when extracting text features, ignoring the interactive information between two texts, or only performed single-level interaction. Given the above problems, an Improved Short Text Matching model (ISTM) based on Transformer is constructed. The ISTM model takes DSSM as the basic architecture and uses the BERT model to vectorize the text to solve the ambiguity of Word2Vec. It relies on the Transformer encoder to extract features of the text and obtain its internal information. It considers the multi-level interactive information between the two texts and finally infers and computes the degree of semantic matching between two texts by the concatenated vector. Experiments show that compared with the classic deep short text matching model, the ISTM model proposed in this study shows better results on the LCQMC Chinese dataset.

Abstract:The traditional KNN algorithm has shortcomings such as low classification efficiency. This study proposes an efficient weighted KNN algorithm that combines the idea of multiple representative points. It uses the concept of the upper and lower approximate regions of the variable precision rough set and integrates the clustering algorithm to generate a representative point set and construct a classification model. Then it adopts the structural risk minimization theory to optimize the classification model and analyze the factors that affect the classification model. During the classification process, the relative position of the test sample is obtained according to the similarity between the test sample and each representative point. Moreover, the category of the test sample in the lower approximate region can be directly determined. If the test sample is in other areas, the sample within the coverage of each representative point is weighted according to the relative position of the test sample and each representative point to determine the type of the test sample. Experiments on the data set in the field of text classification show that the algorithm can improve the performance of the classification model.

Abstract:Crew scheduling is an important part of an airline operation plan. Reasonable crew scheduling can help airlines save great crew costs and increase their revenue. Since the crew scheduling process involves massive complex constraints, it is an NP-hard problem; thus it is difficult to optimize the solution. This study proposes an optimal solution to crew scheduling problems according to Satisfiability Modulo Theories (SMT), which converts various constraints in the crew scheduling process into first-order logic formulas and sets the solution goal as minimum cost and maximum crew utilization. In addition, it transforms the problem into an optimal solution under the satisfiable condition of the given logic formula and uses the SMT solver, Z3. Experiments show that the algorithm in this study can solve the crew scheduling problem of a certain-scale flight plan, bringing benefits to airlines.

Abstract:Aiming at the discrete job-shop logistics scheduling problem with Automated Guided Vehicles (AGVs), a multi-objective discrete job-shop logistics scheduling optimization model is constructed to minimize the time penalty cost of the transfer task and the total travel distance of the vehicle. A multi-objective hybrid Variable Neighborhood Search Genetic Algorithm (VNSGA-II) based on Pareto optimization is designed. Based on the genetic algorithm, the Pareto stratification and crowding-degree calculation method of NSGA-II are used to evaluate the population to achieve multi-objective optimization. The elite individuals are protected by adding the optimal memory to improve the optimization ability of the algorithm and avoid falling into the local optimum. Moreover, the local optimization ability of the variable neighborhood search algorithm is used to design six random neighborhood structures in light of the model features in this paper, thereby solving the optimal value. For a lower cost, the insertion neighborhood based on the critical AGV and the exchange neighborhood adjustment based on the critical transfer task are proposed. Finally, with a discrete job-shop logistics scheduling problem as an example, VNSGA-II, Nondominated Sorting Genetic Algorithm II (NSGA-II), and Strong Pareto Evolutionary Algorithm 2 (SPEA2) are adopted respectively. The results show that VNSGA-II can get a better Pareto solution set, which verifies the effectiveness and feasibility of the algorithm.

Abstract:A dynamic data partition system based on node load mainly considers the load of CPU, memory, and bandwidth of nodes. It first uses the quadratic smoothing method to predict the load of nodes, then combines AHP and entropy index weight to get the processing capacity of each node, and finally dynamically adjusts the load balance of the system for different application scenarios to improve the response speed of applications. It includes the modules of load monitoring and collection, prediction, data pre-partition, and data migration. Given the heterogeneity of node resources in a distributed environment, it aims to reduce the data transmission between nodes in the process of data analysis and calculation, make full use of node computing resources, and improve the parallel computing speed of application analysis through load balancing. Therefore, this study proposes a dynamic partition data mechanism based on node load to improve the system load balance and application response speed and assist the relevant staff in making the decision. This study combines data analysis application scenarios integrating Spark and Elasticsearch for testing.

Abstract:The release of social network data may lead to the disclosure of user privacy; for example, the user identity may be recognized by malicious attackers by analyzing the degree of nodes in the network. Concerning this problem, a k-degree anonymous privacy protection scheme based on the average degree of nodes is proposed. The scheme first depends on the greedy algorithm based on the average degree to divide social network nodes, so that the degrees of nodes in the same group are modified to the average degree, thus generating k-degree anonymous sequences; then the graph structure modification method with priority to retain important edges is used to modify the graph, thus achieving k-degree anonymity of the graph. In this scheme, the average degree is introduced when k-degree anonymous sequences are generated, which improves clustering accuracy and reduces the cost of graph structure modification. At the same time, because the indicator-neighborhood centrality, which measures the importance of edges, is considered in the graph structure modification, important edges are retained in preference, and a stable network structure is maintained. The experimental results show that this scheme improves the network resistance to degree attacks, greatly reduces information loss, and improves the utility of published data while protecting user privacy.

Abstract:To simplify the processing steps of VAT invoices and improve recognition accuracy, we propose a method based on HRNet and YOLOv4 to extract structural information of VAT invoices. Firstly, we detect predefined keypoints in the VAT invoice with the HRNet method to align the invoice to a standard template. Then detect the structural information cell in the invoice by YOLOv4. And lastly use CRNN to recognize the cell block image to obtain structural data. The experimental results on real business VAT invoices show that the proposed method gets a detection accuracy of 75.7 at 0.5 mAP, reaches a detection speed at 12.85 fps, and achieves an Element Correct Ratio (ECR) at 69.30%. The results indicate that the proposed method can simplify the process and improve the accuracy of recognition, and it can apply to the scene where requires high real-time performance and needs to deal with complicated noise situation.

Abstract:For quick construction of a large-scale and high-quality Chinese face recognition dataset, a semi-automatic construction method is proposed in this study. Compared with the existing dataset construction strategies, this method can quickly build a large-scale Chinese celebrity face dataset, which is named CCFace (Chinese Celebrities Face). The dataset contains 506874 face images of 431 persons, with an average of 1176 images of different ages and postures per person. This method makes up for the shortage of available Chinese face image datasets in the face recognition community. In the experimental section, the effectiveness of the dataset is tested on various models and the results show that it can be used as the training set of the State Of The Art (SOTA) model. It is believed that this method and the dataset will attract more people to join the research team of face recognition and promote the face recognition applications in China.

Abstract:As a highly condensed high-level semantic information, the text information of Wujin style Tibetan scripts in natural scenes not only has great research and practical value, but also can be used to assist researchers with text understanding in Tibetan scenes. At present, there are few related studies on the detection and recognition of Wujin style Tibetan scripts in natural scenes. Based on the manually collected image data set of Wujin style Tibetan scripts in natural scenes, this study compares the detection performance of common text detection algorithms on such scripts. The recognition accuracy of the sequence-based text recognition algorithm, CRNN, under different feature extraction networks is also compared on the image data set collected. Examples of recognition failure during the recognition of Wujin style Tibetan scripts in 314 real natural scenes are analyzed as well. Experiments show that the differentiable binary network, DBNet, used in the text detection stage has better detection performance on the test set. The accuracy, recall, and F1 value of this method on the test set reach 0.89, 0.59, and 0.71, respectively; when MobileNetV3 Large is used as the feature extraction network in the text recognition stage, the CRNN algorithm has the highest recognition accuracy of 0.4365 on the test set.

Abstract:The Autonomic Nerve Wreath (ANW) is an important diagnostic feature in iris diagnostics. However, how to extract it against the interference of light spots, pigment spots, and eyelashes is still a problem. This study presents an ANW extraction method based on a genetic algorithm. It takes the local point density as the fitness, selects the optimal individual through roulette, and compares the genetic fitness between two adjacent chromosomes to determine the paternal parent. Experimental results show that the proposed method can avoid the interference of light spots, pigment spots, and eyelashes and thus improve the search efficiency and accuracy of the algorithm. The extracted rolling wheel is close to the actual rolling wheel.

Abstract:In a Radio-Frequency IDentification (RFID) system, the interaction data between an RFID tag and the reader is in a wireless communication mode. Due to the inherent openness of the mode, the interaction data can be easily obtained by a third party. For the sake of data security, a lightweight authentication protocol is designed in this study. A lightweight pseudo-random function is selected as the data encryption algorithm, which can reduce the overall calculation amount of the RFID system and ensure the security of the interaction data. Meanwhile, it can operate on an input parameter of any length and give an output of a fixed length. A comprehensive analysis of the protocol from the perspectives of security, calculation amount, and gate circuit shows that the protocol has high security requirements and the overall calculation amount is also less than those of other protocols in comparison.

Abstract:For the reliable acquisition of mixed signals and the effective separation and recognition of physiological parameters in a non-contact sleep monitoring system, a piezoelectric film sensor is used to obtain the pressure signal of the human body in the sleep state. Meanwhile, a charge amplification circuit and a signal conditioning circuit are adopted to collect the pressure signal in real time. During signal processing, empirical wavelet transform is employed to separate single modal components, such as the BallistoCardioGram (BCG) signal and the respiratory signal. Then, the K-means algorithm is used to cluster different types of peaks in the BCG signal so that the heart rate can be calculated through the average heartbeat cycle. The experimental results show that the designed monitoring system has strong adaptability and it can extract respiratory and heartbeat signals.

Abstract:In view of the complicated cabin structure of the Three Gorges ship lift and the difficulty in selecting the equipment inspection route, the inspection route in the ship lift cabin was taken as the research object and the planning for the route was converted into a Traveling Salesman Problem (TSP). Through the weighted undirected graph of the inspection route and the spatial coordinates of the inspection points, a spatial structure model of the inspection points of the ship lift was built. The ant colony algorithm was applied to calculate the optimal inspection route for day shift and swing shift, respectively, via the Matlab software. The experimental results show that the optimal inspection route calculated by the ant colony algorithm meets the equipment inspection requirements of the Three Gorges ship lift.

Abstract:Sewage treatment processes are a set of complete solutions for the treatment of urban domestic sewage and industrial wastewater, and they are widely used in various fields. According to the treatment scale, the characteristics of water quality, the environmental functions of the receiving water, and the actual situation and requirements of the local area, the treatment processes of urban domestic sewage should be optimized and determined after the measurement of technological characteristics and economic costs. Thus, the design of urban domestic sewage treatment processes can be regarded as a special form of multi-parameter optimization. Firstly, the process methods of sewage treatment should be sorted out, and the knowledge base of sewage treatment processes should be designed. Secondly, with the parameters and environmental information of each process as inputs, a scheme composed of the processes from the knowledge base is automatically generated based on the sewage treatment process database and the set intelligent algorithm. Specifically, the scheme includes the sequence of each process module, the size of internal components, the operation cost prediction, and the treatment effect. This study adopts the messy genetic algorithm to recommend the sewage treatment processes. The multiplicative inverse of the total cost of the process sequence is taken as the fitness, and the optimized scheme with the lowest cost is automatically generated when multiple pollutant indexes reach the standard. Experimental results show that the messy genetic algorithm can efficiently and accurately recommend the scheme when the process sequence length changes.