This PDF file contains the front matter associated with SPIE Proceedings Volume 12087, including the Title Page, Copyright information, and Table of Contents

Large-scale wind turbine participation in the system primary frequency regulation will have a greater impact on the development and implementation of low-frequency load shedding strategies for the grid, and to explore the effect under different proportions of wind power, model of wind power participating in frequency regulation is introduced first, then the principle of under-frequency load shedding strategy in power system is expounded, and finally build large grid model in power system simulation software, and combining with the actual power grid operation are simulated, the simulation results show that considering wind power in power grid frequency regulation of under-frequency load shedding strategy, including load removal is less obvious, that wind power to participate in the frequency regulation of under-frequency reduction with a positive impact.

The output of engine speed sensor is the signal with irregular frequency, which can not be directly detected by general speed acquisition equipment. Generally, the traditional method of speed modulation is to adjust the maximum value of the signal through resistance divider or proportional amplifier circuit, and subsequently acquiring the square wave signal via hysteresis comparator. Aiming at improving traditional methods with disadvantages such as large blind area and difficult component selection, this paper develops a speed modulation method based on the characteristics of speed signal. The established method succeeded in converting irregular speed signal into square wave signal with the same frequency to meet the sampling requirements, which lay the foundation for its accurate measurement in the next step. Our work focuses on the circuit design, Multisim simulation and implementation of the developed method. The results show that the design of rotating speeds modulation is correct with concise structure, lower consumption and high stability.

In the application of small and medium power conversion such as power adapter of electronic equipment, the equipment is becoming miniaturized and lightweight, and has higher requirements on efficiency and electromagnetic compatibility. Taking the fact that high efficiency and high power density are the core index of today’s power supply. Meanwhile, the power density and efficiency of DC/DC mode power supply are improved mainly through the following technologies: advanced devices(such as forbidden broadband devices), secondary side control, soft switching technology, magnetic integration, planar transformers, etc. Thus, in this paper, a completely new topology with closed loop control based on the new power semiconductor GaN is chosen to design the power supply with high efficiency and high power density and it can work continuously for 30 minutes at 25℃. It is found out that the designed power supply is verified to be practical combined with the simulation.

With the continuous strengthening of the construction of domestic power grids, in order to efficiently use inspection robots to carry out power inspection work, to meet the needs of the development of the power industry for automation and intelligent inspection. This article analyzes the robot structure, key communication technologies and inspection routes in the power inspection system, establishes related models and design corresponding plans in the communication inspection scene, and briefly analyzes and briefly introduces the robot cluster scheduling, task control, and intelligence. The specific functions of analysis, remote real-time measurement and control, data processing, communication data link network, ground-based enhancement, and risk monitoring sub-systems provide stable technical support for the communication grid inspection.

Automatic lip-reading (ALR), also known as visual speech recognition (VSR), refers to the movement of lips to acquire the content of a speaker. In recent years, the introduction of deep learning has brought great breakthroughs to lip-reading research. Compared with traditional methods, this method can extract depth features more conveniently from large-scale data sets. We proposed an end-to-end deep learning network model containing three construction modules, STCNN, ResNet50, and Bi-GRU. STCNN was used to extract deep features and Bi-GRU was used for feature recognition. Meanwhile, we verified the effectiveness of Bi-GRU through comparative experiments. The effects of different ResNet in the whole model were compared. Finally, the accuracy of characters in the corpus reached 95.7%.

With the development of global urbanization, the requirements of traditional urban management are increasingly prominent. With the development of 5G technology, the delay of information transmission has been shortened, the speed of information transmission has been accelerated, and the large-scale wireless communication technology has become more mature. So, it has a profound impact on the construction of the smart city. But the technologies involved in the smart city still need to be solved, like the user data eavesdropping and tampering of the air interface, the DDOS (Distributed denial of service attack) attack of air interface from UE (user equipment), the malicious interference of the fake base station to air interface. This paper studies these security problems in detail and points out some solutions. These solutions include network slicing technology based on SDN (Software Defined Network) and NFV (Network Function Virtualization) technology. For these technologies, this paper also prospects some different future research directions. Finally, this paper is helpful to the research of 5G network security problems and puts forward some suggestions for future smart city construction.

In this letter, a new configuration of reflectarray antenna element is proposed. A frequency-reconfigurable element for beam-scanning reflectarray is introduced. The reconfigurable reflectarray antenna element requires two PIN diodes and one varactor to provide a reflection coefficient with continuously tunable phase over a 270 degree range which can be shifted over a 50% frequency range. Microstrip spoof surface plasmon polarition (SPP) is introduced on the side of the patch. Through this method, the reflection phase shift range of reconfigurable reflection array element increases by dozens of degrees.

With the large-scale development of renewable energy, the renewable energy represented by wind energy is greatly integrated into the power grid, and the permeability keeps increasing. On the one hand, because the wind turbine units with small inertia time constant replace the traditional generator set, the overall inertia of the system decreases; on the other hand, the frequency characteristics of the system itself change. First, taking two-feed wind turbine as an example, we review the FM control principle research of wind power turbines, the control principle and implementation of virtual synchronizer control technology and the FM control strategy of wind power farms. Finally, future research in the field of power grid regulation is discussed

Compared with visible light remote sensing images, infrared remote sensing data has a lower gray level, and its visual interpretation requires high manual experience and consumes time and energy. In this paper, a transfer learning method is proposed to train the infrared remote sensing water body extraction model based on the label information of visible light remote sensing images and a small amount of infrared remote sensing image label information. Feature-based subspace transfer learning maps different types of remote sensing image features into a shared subspace and classifies them under similar distribution. According to the analysis of features of water bodies, the nonlinear transfer learning algorithm is used to reduce the dimensionality of features. The manifold feature extraction algorithm can retain the original nonlinear distance relationship of remote sensing data after dimensionality reduction so that the water target maintains the original nonlinear features in the manifold subspace, and accurate water extraction results can be obtained. Experiments are carried out on the Band5 NIR image data set from the Lansat8 satellite. The average accuracy of the proposed method is 95.44%, which is better than other methods. The results show that the proposed isometric mapping transfer learning method based on multi-parameter optimization can accurately extract infrared water targets.

In the past year of 2020, life in the post-epidemic era continues to move towards cyberspace. The surge in magnitude and value of data has also brought more data abuse, attacks and thefts, and data security issues have become particularly prominent. Moreover, in recent years, data attacks have become more subtle and complex, making traditional security protection methods stretched. Therefore, in view of the data security threats under the new situation, it has become urgent to study new security protection methods. Due to the intuitive sensory advantages of visual analysis, security visual analysis is gradually becoming an effective means to guarantee data security and gradually arouses people's attention. As a cross-cutting research field, network security data visualization can effectively improve the perception and understanding of network security experts in the process of analyzing network security issues by providing information and interactive visualization methods. First, we review the methods of network security data visualization and analyze the basic methods and processes of visualization after investigating the current research status of network security data visualization. Then we introduce data visualization techniques suitable for network security and analyze the existing problems. Finally, we make a conclusion for future trends.

The emerging 5G communication technology has three features: enhanced mobile broadband (eMBB) based on the definition of 3GPP, Massive Machine Type Communication (mMTC), as well as Ultra-reliable and Low Latency Communication (uRLLC), which play a great role in promoting the smart city. Given the 5G network's important role in smart cities, research on the 5G network is becoming more and more important. However, there are inevitable challenges in constructing a 5G network, such as resource allocation and other issues. Based on the problem faced by the 5G network, this paper studies the following aspects: firstly, the main challenges of the network are discussed. Secondly, resource allocation is selected as the main research direction. Finally, the feasibility and effectiveness of network slicing technology for resource allocation are analyzed and discussed. At the same time, this paper also compares the more cutting-edge research results in related fields. The research in this paper is helpful to sort out the research status of resource allocation in the 5G network and look forward to the rational allocation of resources in the future to promote the development of a 5G+ smart city.

The country attaches great importance to the safety risk of industrial control systems. It has issued more than ten national standards for industrial control safety protection and evaluation, but more of them provide guidelines for compliance safety evaluation, and less involve penetration testing. As a powerful method to discover security vulnerabilities, penetration testing can change passive defense to active, discover security weaknesses from the attacker's perspective, and verify the effectiveness of security measures. The difficulty of penetration testing for industrial control systems is that it relies more on the field experience of testers, and needs to refer to factors such as the particularity of the system industry, functional design, equipment configuration differences, and production environment. This article starts from the research on the vulnerability of industrial control systems, focuses on the penetration testing methods of industrial control systems, and analyzes and tells them in combination with actual cases.

Considering that the wireless power transmission technology is in the stage of rapid development, the magnetic coupling resonant wireless power transfer(MCRWPT) technology which is one of the most popular and important wireless power transfer methods because of its advantages in the near-field transmission is taken as an example to discuss. This article summarizes the fundamental structure and working principle of this technology, which is primarily composed of a high-frequency power supply, compensation structure, magnetic coupling structure, and rectifier-filter structure. In the end, the application prospect, the problems that need to be further studied and the development trend of this technology are discussed on the basis of the existing achievements. Despite the fact that MCRWPT is a hot topic nowadays and has a broad application prospect, including intelligence, compatibility, and other functional advantages, there are still many problems to be studied and solved by researchers.

Network slicing technology is one of the key technologies of 5G. It can enable users to access the most suitable network on demand through the end-to-end configuration of the network and increase the flexibility of network resources. Introduced the latest progress of network slicing, including slicing architecture, slicing identification, operation process, wireless network slicing, and other key standard technical solutions. At the same time, it also introduced the architecture and key technical characteristics of Software Defined Network (SDN) and Network Function Virtualization (NFV) and its algorithms. Analysis and analysis of the future of network slicing. At the same time, the challenges after network slicing and the application scenarios in the future are also proposed.

To prevent the off-grid phenomenon of wind turbines caused by voltage sag and realize the low-voltage traverse operation of doubly-fed induction wind turbine (DFIG), the elite strategy ant colony algorithm was used as a parameter optimization mechanism in the active disturbance rejection control (ADRC) system of the doubly-fed induction wind turbine. To make doubly-fed motor active disturbance rejection system have self-learning ability, an ant colony algorithm with distributed search and parallel computing ability is added to make it have automatic parameter optimization ability. After the simulation analysis of common ADRC and ant colony optimization ADRC, it is shown that the ADRC can improve the ability of low voltage crossing of the system, effectively reduce the transient current on the rotor side, and has good dynamic performance and strong robustness.

With the wide application of radio communication technology, signal source positioning technology has gradually developed into one of the key technologies in the field of radio monitoring technology. At the same time, the requirements for signal source localization accuracy are getting higher and higher. This paper introduces the basic principle of firefly algorithm, and introduces the firefly optimization algorithm to optimize the particle degradation problem of particle filtering algorithm, and further optimizes the firefly algorithm on this basis to propose the adaptive firefly algorithm to optimize the signal source localization by particle filtering, and introduces the principle and process of the proposed algorithm in detail. Finally, the proposed algorithm is analyzed and verified through simulation comparison experiments. It aims to promote the in-depth exploration of this field through this research in order to solve more realistic problems.

With the freezing and commercial use of release 16 (R16) standard in the fifth generation (5G) wireless communication network, the applications of 5G in virtual reality/augmented reality (VR/AR), ultra-high-definition video, unmanned aerial vehicle (UAV), internet of things (IoTs), and other fields have also begun to develop simultaneously. However, with the in-depth deployment of 5G, some problems such as the high cost and shortage of frequency bands have become prominent. 5G has become a new focus of competition in the communications field on a global scale, and various countries are actively deploying to seize the leading power in 5G development. In this comprehensive study, we will introduce the standardization process of 5G and the deployment of various countries and introduce the network architecture of 5G. In addition, we will analyse the various challenges of 5G deployment and give some research directions for future development.

A compact high isolation MIMO antenna is designed for 5GNR n78 band operation in mobile terminals, which consists of 8 compact monopole antenna elements. High isolation based on neutralization line decoupling technique is achieved, the working mechanism of antenna decoupling is discussed in detail in the paper. The proposed MIMO antenna is simulated by electromagnetic simulation software HFSS, the geometry size of the proposed MIMO antenna is 75×107×0.8 mm3 , etched on FR4 dielectric substrate with thickness of 0.8mm, the size of the antenna structure is in accordance with the smartphone size, the operating band covers 3.4GHz-3.6GHz of n78, and the isolation is more than 13dB at the 3.4GHz-3.6GHz band by neutralization line decoupling technique. The designed MIMO antenna has good isolation and diversity characteristics and can be used in MIMO system of 5G mobile terminal.

Aiming at the problems of the conventional beamforming in the actual scene of the towed linear array, such as broadband signal distortion and high sidelobe. The expected beam response design method of low sidelobe based on second-order cone constraint optimization is proposed. The expected beam response of the reference frequency point is designed, and the beam response of other frequency points in the processing bandwidth is transformed into the beam response of the reference frequency point. And the constant beamwidth beamforming weighted vector is obtained, which makes the broadband beamforming have frequency invariant beam response. At the same time, the norm constraint of the weighted vector is added to improve the robustness. Besides the constant beamwidth groove beamforming method is proposed to suppress fixed azimuth interference. The effectiveness of the proposed method is verified by computer simulation with a uniform linear array.

When the circuit board fails, its operating temperature will be abnormal, which will have a greater impact on the overall performance of the circuit. Therefore, it is very important to identify the fault by observing the operating temperature of the circuit board. This paper introduces a fault identification device for the electronic board card based on binocular imaging and analyzes its overall system design, system composition principles and upper computer software design results. The system uses an infrared camera and a visible light camera to simultaneously image the circuit board, and can display its thermal image with contour on the TFT LCD screen or on the upper computer software, and can intelligently identify the fault type on the upper computer. It has the advantages of low cost, small size, clear imaging, and portability, which make up for the disadvantages of current thermal imaging equipment such as high cost, poor portability, and difficulty in fault identification.

The ultra-dense network is used widely in the 5G environment. It achieves ultra-high-performance metrics that cannot be achieved by traditional mobile networks, but also brings serial interference between base station cells. There are two main deployment scenarios for ultra-dense network architecture, which are the popular macro-micro deployment scenario, and micro-micro deployment scenario, respectively. The interference coordination schemes for ultra-dense networks mainly include dynamic clustering, semi-static clustering and static clustering. In this paper, static clustering is used in the macro-micro deployment scenario. In this paper, NS3 is used to build the simulation environment with 7 macro stations and up to 150 micro stations. Two different clustering methods based on base station numbering and path loss are tested to compare their communication performance.

An enhanced YOLOv4 multi-spectral fusion pedestrian detection approach is proposed to address the problem of fusion network robustness in pedestrian detection. This method can effectively and accurately complete pedestrian detection. First, the feature extraction backbone is upgraded, and channel attention Module CAM and spatial attention Module SAM are added to the feature extraction backbone to allow for adaptive feature layer adjustment. Fusion processing based on channel and pixel direction is performed on the altered feature layer, and then the fusion layer is anticipated. The experiment is performed on the KAIST Dataset, and the capacity to generalize was assessed using the OTCBVS Benchmark Dataset. The proposed multi-spectral fusion detection approach is effective, according to the experimental results. The log-average miss rate (MR) reaches 11.03 and 8.79 throughout the full day and night when the false positive per image (FPPI) is 10-2~100 , and it also achieves good detection performance during the day. The proposed multi-spectral fusion detection approach is universal in various data sets, according to the generalization ability analysis experiment. Pedestrian detection accuracy may be accomplished adaptively regardless of whether it is daytime or nighttime detection, and speed is substantially enhanced.

In view of the rationality of the weight distribution of the evaluation indexes and the consistency of the evaluation results in the evaluation process of the power communication network operation quality, this paper selects and optimizes the evaluation indexes of the power communication network operation quality based on the operation management specification of the power communication network and the existing relevant evaluation indexes, and constructs an evaluation index system. The subjective weighting method is used to distribute the unified weight of indexes, which overcomes the irrationality of hierarchical index weighting. The feasibility of three typical index evaluation methods is verified by examples, and the consistency of evaluation results is studied by correlation coefficient simulation. The research results have reference value for improving the operation quality of the power communication network.

The fifth-generation (5G) broadband cellular standard, Blockchain and Internet of Things (IoT) are emerging technologies in the present, and they all have many advantages that benefit humanity in a vast variety of technological and scientific fields. The biggest advantages of 5G are higher speeds, low latency and the ability to support up to a million devices per square kilometer. Blockchain is a chain of records that holds data, cryptographic hash and Prove of Work (PoW) using cryptography. IoT is an overall description of devices that are connected through the internet. 5G has the potential to work together with Blockchain technology over its popularization, helping Blockchain with its problems and introducing Blockchain to IoT for more commercial use, while blockchain can benefit 5G and IoT in many ways, including optimizing some 5G enabled technologies such as cloud computing and smart cities. In this paper, discussion about the overall integration of 5G, Blockchain and IoT will be mainly focused on. Firstly, the definition and motivation of this “5G Blockchain-IoT” integration will be introduced; Secondly, the contributions of the integration on 5G enabled technologies and applications will be discussed and listed; Thirdly, challenges, issues and future perspectives of the blockchain, 5G and IoT integration will also be estimated.

Based on the practical application background, photovoltaic (PV) grid inverter single-phase and three-phase inverter is studied in detail in this paper the basic working mechanism of inverter, including topology architecture, modulation strategy, parameter design, harmonic analysis and control method, and through the Matlab/Simulink simulation platform built simulation model of single phase/three-phase photovoltaic grid inverter, The correctness and effectiveness of the proposed design method are verified.

Breaking the limitation of materials specificity for surface-enhanced Raman scattering (SERS) in the practical application has crucial scientific significance, particularly Pt nanomaterials, widely used in catalysis. This work obtains CoPt hollow Nanoparticles with sizes below 10 nm and a suitable SERS detection sensitivity substrate. Thus, it dramatically enhances Raman signals of R6G or melamine molecules by the synergistic effect of Pt and Co elements. And it illustrates the CoPt NPs substrate has the potential to be applied to the melamine molecules analysis.

As an effective means to solve the problem of spectrum resource shortage, in-band full-duplex (IBFD) communication system can theoretically double the spectrum utilization rate and provide a new solution to the problems of hiding terminals, channel congestion and end-to-end delay. The self interference (SI) of IBFD directly affects the communication quality level. Due to the uncertainty in signal propagation process, SIC with complex process is the bottleneck of IBFD development. Therefore, how to eliminate self interference (SI) and apply it to engineering practice is a hot issue in future research. In this paper, SIC technologies are classified and studied. According to three SIC methods: Antenna-Domain SIC, Analog-Domain SIC and Digital-Domain SIC, the existing SIC technologies are compared and their advantages and disadvantages and shortcomings in practical application are discussed. Finally, the key research directions and application prospects of SIC are put forward.

After analyzing and summarizing the characteristics and shortcomings of existing common time synchronization algorithms, a two-way time synchronization algorithm based on time drift compensation is proposed for the LoRa wireless ad hoc network based on TDMA mode. The algorithm improves the accuracy of time synchronization by compensating for the drift of the gateway's local clock during the time synchronization process and using symmetric transmission to eliminate the effect of channel message delay. After that, the time deviation between nodes is measured and the calculation method of the time synchronization period is given. Finally, the algorithm is analyzed and demonstrated through comparative experiments. This algorithm reduces communication costs and extends the overall service life of the network, while ensuring the accuracy of time synchronization.

Over the past few decades, ’as all kinds of network attacks have bad effects on people's daily life, cybersecurity has become an issue that is commonly concerned by society. In traditional ways, it belongs to human’s responsibility that estimates whether the traffic requests received by the network are attacks. This method is not only inefficient but also has bad timeliness. Later, with the popularization of machine learning, people attempt to train some models by constructing features to detect network attacks. However, the labor cost of feature construction is very high. In recent years, with the development of deep learning, this method has become a more popular choice because of its relatively low cost and outstanding effect on prediction. According to the above analysis, this paper firstly uses machine learning-based methods including Random Forest, Gaussian NB, XGBOOST, Decision-Tree, Logistic Regression, KNN to train the effective model. Besides, we also trained a Multi-Layer Perceptron (MLP) model to compare the differences between machine learning and deep learning in predicting network traffic attacks on the CIC-ID2017 dataset. The results show that the XGBOOST, Random Forests, KNN, Decision Tree, Logistic Regression, and Gaussian Naive Bayesian respectively achieve accuracy with 0.999, 0.998, 0.991, 0.979, 0.568, and 0.509, while MLP obtains accuracy 0.643 with loss NaN. This fact suggests that the MLP method is unsuitable for the web attack detection task compared with the machine learning based approaches.

This paper proposes a genetic algorithm-enhanced density-tapered method to design linear thinned arrays with low peak sidelobe level (PSLL). This method can reduce the randomness of PSLL of the array designed by density-tapered method and the number of independent variables when using genetic algorithm to optimize the position of array element. The numerical validation, carried out in the far-field and for narrow-band signals, points out that using the proposed method to design thinned array can obtain lower PSLL than the original density-tapered method.

Traditional UWB indoor positioning technology uses geometric algorithm to locate tags analytically, but the complex indoor environment and ranging error may lead to no solution to the equations, so it is impossible to achieve positioning. In order to solve the above problems, a UWB indoor location algorithm based on Improved BP neural network is proposed. The difference between the label coordinates and the real coordinates of the output layer is taken as the error signal. Utilize the gradient drop means to correct the each weights of the network to reduce the error. Based of BP algorithm, the studying speed is changed adaptively through the change trend of network training error to enhance the convergence speed. Comparative experiments by a location algorithm based on the traditional location technique and BP neural network. Emulation experiment outcomes indicate the iterative times and location error of the algorithm are obviously small. And the output trajectory is basically consistent with the actual motion trajectory, which has high positioning accuracy.

The recommendation system plays a strong auxiliary role in the user's handling of information overload. For products, user comments and descriptions are important reference data. Based on reinforcement learning, more accurate recommendations can be made through the interaction between users and product-related data items. However, the interactive recommendation will face some problems caused by data sparseness. For text data information, users and items can be mapped to the feature space to alleviate this problem. This paper proposed the depth deterministic strategy gradient algorithm is used to train the recommendation model. The experimental results on three real data sets show that the model has a relatively accurate recommendation effect and acceptable running time, and it can alleviate the impact of data sparsity on the recommendation effect to a certain extent.

Biomedical event extraction aims to extract structured biomedical events from unstructured mass biomedical literature, describing fine-grained biomedical relationships between events and biological entities. Biomedical event extraction reduces human effort and provides support for constructing a relevant database of disease diagnosis. Trigger word and argument are two components of a biomedical event. The trigger word refers to the word or phrase that triggers the event, and its type determines the type of the event. In contrast, argument refers to the event's participant, which can be a biological entity or another event. The current biomedical event extraction system uses the phased approach, with trigger word identification being the first step of the phased method. In this paper, an extended version of the recurrent neural network, i.e., the bidirectional gated recurrent unit (Bi-GRU) neural network, is utilized, to conduct the biomedical event trigger identification task. Specifically, the inputted word sequence tensor first passes through an embedding module including a word embedding layer and an entity-type label embedding layer to obtain the concatenated token representations for each sentence. Then, the token representations are fed into a Bi-GRU module to acquire the contextual encoding, which is used to conduct the trigger word identification task. The experiment is based on the MLEE dataset, a commonly used biomedical event extraction dataset. The experiment result shows that the proposed model can achieve some comparable performances with Precision 79.62%, Recall 78.64%, and F-score 78.82%

To accurately identify the dynamically changing water level in the compartment, Firstly, it is needed to pre-process the image by image grayscale, image segmentation, morphological processing, apply PP-YOLO v2 algorithm to measure the water level with dynamic changing characteristics, simulate the compartment through the water tank, convert the actual level according to the proportion, and finally compare the algorithm detection result with the actual measurement result by experiment. Finally, the algorithm detection results are compared with the actual measurement results through experiments, and the relative error is derived. The experimental results show that the dynamic water level identification algorithm of compartment with PP-YOLO v2 has high accuracy, and the relative error is only 1.33%. The accuracy of this algorithm for dynamic change of water level recognition is 98.67%, which has strong application value.

Nowadays, with intruders frequently attempting to gain access to the network and harm the data, intrusion detection systems (IDS) are playing a more and more important role in providing security against malicious activities, making IDS a popular research issue. Among various intrusion detection techniques, machine learning methods have shown their advantages, such as higher detection rates, lower false alarm rates and reasonable computation and communication costs. In this paper, we describe a focused literature review of mainstream machine learning methods for intrusion detection. We divide the schemes into three categories, including classic supervised machine learning methods, ensemble learning methods and deep learning methods, and several algorithms are discussed with respect to each category.

Phishing attack is the simplest way to obtain sensitive information from innocent users. The target of phishers is to obtain key information, such as username, password, and bank account details. Cyber security officials are now looking for reliable and stable detection techniques to detect phishing sites. This paper studies the phishing websites detection technology by extracting and analyzing the characteristics of legal and forged Uniform Resource Locators (URLs) using machine learning approaches, including Logistic Regression, K-nearest Neighbor (KNN), Linear Support Vector Classifier (SVC), Random Forest, Gradient Boosting Decision Tree, and Ada-Boost, and compares their performance with respect to criterions such as accuracy, Root Mean Square Error (RMSE), precision, recall, and F1-score. The results show that ensemble methods, including Gradient Boosting Decision Tree, Random Forest, and Ada Boosting, can achieve much higher detection performance than the other algorithms in terms of all the criteria.

Video game is becoming one of the most popular ways that people spend their time. Most of the video games are processing on a local device. It’s possible to process a game on a cloud server. However, the servers are usually far away from the data centers. Therefore, it will cause latency which ultimately degrade the gaming experiences. 5G and cloud computing are being developed at an exponential rate, and it will revolutionize the gaming landscape. Meanwhile, it also has different impacts on gaming accessories through edge computing and faster-than-ever response times. In this study, the applications of 5G technology and cloud computing in gaming industry was discussed. After that, the possible impacts of 5G and cloud computing on the gaming accessories was studied.

With the continuous development of communication technology, the traditional communication network architecture can no longer meet the demand of telematics technology in the field of autonomous vehicles. The traditional wireless access network architecture needs to evolve quickly to meet the demands of huge traffic, ultra-low latency, and other telematics communication. The commercialization of 5G with its advantages of high reliability and low latency will provide powerful performance and more possibilities for Vehicle-to-Everything (V2X), an essential technology used in automotive communications. This paper introduces the four V2X functional compositions and two major V2X technology standards, namely Dedicated Short Range Communication (DSRC) and Cellular-Vehicle to Everything (C-V2X). Also, we present five remaining limitations of V2X based on the current 5G technology and analyze the specific challenges faced with their possible solutions for the enhancement. Finally, we consider the perspectives of 5G-V2X in autonomous vehicles applications and the analysis on the economy. The continued development of 5G-V2X technology in the field of autonomous vehicles could eventually lead to a significant increase in road safety, a reduction in the number of accidents, and a higher efficiency in driving.

The rapid development of artificial intelligence technology has brought new research concepts and methods to military application. Aiming to accelerate the research process of intelligent sensing technology and promote the landing application of human detection technology, this paper studies maritime military targets detection methods in infrared image sources and establishes a special detection model for this task, which utilizes sufficient visible samples to realize the hetero-source detection of infrared imaging targets. A detection model for cross-domain targets is trained through sufficient visible images, which realizes accurate target recognition in infrared images. Target information of different domain images is then fused and optimized, advantages of different-domain images are considered, and target detection features and results are combined to boost the performance of military target detection, thus laying the foundation for maritime real-time reconnaissance applications. A human detection method based on bounding boxes and a human detection method based on key points in visible images are proposed, which improves the detection accuracy of occluded and small targets in UAV aerial images. In order to bridge the gap between infrared images and visible images and improve detection accuracy in infrared images, hetero-source detection methods based on feature alignment and shared feature are put forward. Experiments are carried out on test set which is composed of randomly selected 487 images, and the proposed model achieves 84.8 AP, which is 4.8 higher than the famous Faster R-CNN, which proves the effectiveness of the proposed method.

As an important measure of the national economic strategy, the research and application of unmanned ships have promoted the development of the marine economy and are conducive to reducing marine transportation costs and labor costs. It can be specifically applied to various engineering minerals, fishery transportation, ships, and island resource replenishment, as well as ocean mapping and hydrological monitoring. This paper studies the autonomous navigation tasks of unmanned ships. The core of its technology lies in global path planning and local obstacle avoidance. Global planning specifically includes the acquisition of global sea environment information and the realization of path planning algorithms. Autonomous navigation technology is of great significance to the use of unmanned ships. This paper focuses on the analysis and implementation of global path planning algorithms, and the research and implementation of dynamic obstacle target detection and tracking algorithms in the process of local obstacle avoidance.

Biomedical event trigger extraction, as one of the sub-task of biomedical event extraction, plays an important role in biomedical research. A biomedical event trigger is a word or phrase that marks the emergence of a certain biomedical event. The recent works are usually based on the rule and the machine learning methods. However, the rule-based methods heavily rely on the concrete rules enumerated by the field expert and usually need a large amount of expert knowledge. The machine learning-based approaches usually utilize many handcraft features such as n-gram, lexicon, pose-tag and shortest dependency path. As a result, these methods based on machine learning can suffer from handcraft engineering with expensive time costs and the problem of generalization in the field transition. With the popularization of deep learning techniques, some effective frameworks in Natural Language Processing (NLP), such as adversarial training, self-attention mechanism, graph convolutional network, have been proposed to enhance the model performance for the NLP, especially the information extraction. As a task in the information extraction field, the frameworks mentioned above have been applied in the biomedical trigger identification subtask. This paper attempts to employ an external version of the recurrent neural network (RNN), i.e., bidirectional Gated Recurrent Unit (Bi-GRU) network, to extract the biomedical event trigger existing in the biological literature. Specifically, we first transform each token and entity label in the sentence to a word sequence with token index and an entity label sequence with entity label index. Subsequently, the above two sequences will be fed into the embedding layer to obtain the concatenated tensor between them. Moreover, we put the tensor into the Bi-GRU to generate the contextual encoding, which will be fed in a linear layer with an activation function to predict the probability distribution of the trigger. The final experiment on the MLEE dataset confirms that the proposed model can achieve comparable performance with an F-score of 78.82%.

Based on the power management system of a medium-sized unmanned helicopter, an automatic test system is designed and implemented. Aiming at the shortcomings of ordinary test methods, such as low efficiency, poor accuracy and poor security, this paper discusses the design and implementation of an automatic test system for power management system from both software and hardware aspects and give the specific functions and implementation methods. The embedded system based on STM32F427 is adopted, and the host computer software runs on the Windows platform and is developed in the Microsoft Visual studio integrated development environment. The automatic test system realizes the full-function automatic test of the communication of the power management system, power supply control, voltage detection, current detection, power supply abnormal handling, and power consumption abnormal handling. Compared with ordinary detection methods, this automatic test system shortens the test time, improves the test efficiency and accuracy, and ensures the safety of equipment and personnel.

The current access point load dynamic equilibrium algorithm takes user behavior as the equilibrium standard, which leads to the long processing time of the access point and the high loss of packets caused by transmission. Therefore, the load dynamic balancing algorithm based on software defined wireless network for different access points is studied. Under the wireless network architecture defined by the designed software, the association analysis of network access points is carried out. After setting the access point load balancing standard, the virtual access point is used to formulate the load balancing strategy to realize the load balancing of different access points. The simulation results show that the equilibrium waiting time of the network access point is shortened by about 50%, and the packet loss rate decreases significantly. The load balancing effect of the access point is good.

In this article, we propose an improved binary heuristic algorithm based on artificial bee colony algorithm (ABC) for the multidimensional knapsack problem. The algorithm uses binary coding and multidimensional neighborhood search strategy to improve the search strategy of ABC algorithm. In addition, cross and mutation are introduced in the employed phase and scout phase. These measures enhance the global search capability of the algorithm, avoid the algorithm from falling into the local optimum in the iteration, and further improve the efficiency and accuracy of the algorithm. The calculation results show that the improved binary artificial bee colony algorithm in this paper can obtain high-quality solutions for various characteristic problems, and it is feasible and effective for solving multidimensional backpack problems

The traditional naive bayes algorithm is a commonly used text classification algorithm, it’s attribute independence assumptions reduce the classification effect of text to some extent. In response to this problem, a weighted naive bayes text classification algorithm based on improved feature weight was proposed. First, when the TFIDF algorithm was improved from the inverse characteristic frequency, the category frequency, etc., the redundant attribute was removed, and the weight of the different feature items are used to measure the weight of different feature items, and then use the cross entropy. The feature item weight was substituted into a naive bayes formula, and the weighted naive bayes classification algorithm was constructed. Compared with several different algorithms, the experimental results show that this algorithm has significant increase in precision, recall and F1 score.

The risk and severity of network intrusion has clearly received great attention in the last decade. Meanwhile, machine learning methods have been widely employed in the area of cybersecurity. This paper introduces the network intrusion attacks and detection systems and gives an overview of literature on various machine learning models to achieve network intrusion detection, including logistic regression, k-nearest neighbors, neural networks, random forest, decision tree, and k-means clustering. We find that as the dataset gets larger, the machine learning methods yield better performance significantly. Furthermore, we discuss the prospects mentioned in the literature and put forward some key prominent future research directions in network intrusion detection systems.

Convolutional Neural Network (CNN) has recently been demonstrated as an effective model in machine learning for processing a large amount of data. In addition to its applications in natural language processing (NLP), CNN plays an important role in computer vision. This machine learning model can be applied to realize face recognition, image classification, image retrieval, object detection video classification, etc. In this paper, the author constructs an CNN model to classify videos based on the data set HMDB51. HMDB51 is a large human motion database, consisting of a total of 7,000 clips distributed in 51 action classes. Different from the 2D image, video is the three-dimensional data, including time and the width and height of each frame in the video. Therefore, it is necessary to process the video into a sequence of frames of uniform length and normalize the frames firstly. This would facilitate subsequent training and help prevent premature over-fitting as well. Besides, the impact of changing parameters in this CNN model would also be tested, such as the length of image frames, the number of sample categories, the batch size to training the data and the location of dropout layers. The results show that it is difficult to significantly improve the accuracy of recognition by adjusting these parameters alone, and those various factors need to be considered comprehensively

At present, the traditional collaborative filtering (CF) recommendation algorithm generally has the problems of data sparsity and low accuracy. In order to solving these problems, this paper proposes a multi-feature fusion collaborative filtering recommendation algorithm. Firstly, considering that different users have different preferences for different item attributes, this paper used item attribute scores and user preference information for matrix filling to alleviate the problem of data sparsity. Secondly, considering the influence of multiple features of user activity , the count of common item ratings, and item popularity on the calculation of user similarity, and used these features to improve the calculation of similarity. Finally, the matrix after data filling and the improved similarity calculation formula were used to make personalized recommendations for users. Experimental results prove that compared with traditional algorithms and other improved algorithms, the improved algorithm in this paper alleviates the problem of data sparsity to a certain extent, and at the same time has a certain improvement in accuracy

Nowadays, despite the popularity of smartphones in our daily lives, emails are still the most widespread way for people to communicate and exchange information in business and many other circumstances. However, a tremendous problem called malicious email, also known as spam, bothers people and demands constant detection and block. This paper discusses machine learning approaches to achieve malicious email detection. The data for training is more than 10,000 raw emails with Chinese text as well as features including server name, IP address, timestamp, and content. First, the contents are split into words via feature engineering. Then, the malicious email detection is carried out by multiple machines learning methods, including Naïve Bayes, Decision Tree, Random Forest, Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), and K Nearest Neighbor (KNN), respectively. The performance of these models is evaluated by criteria like precision, recall score, F1 score, and time cost. It is shown that the Naïve Bayes model yields the best results, with the F1 score being higher than 97%, which indicates that our model is promising in practice.

Convolutional Neural Network (CNN) is a feedforward neural network suitable for large-scale image processing. CNN has been largely developed these years, it is primarily used to identify displacement, and it uses training to extract features from raw data to gain learning and judgment. Age and gender classification of face images is a kind of biometric technology that identifies people by extracting their biological features. This paper combines the convolutional neural network with the classification and recognition algorithms for the gender and age of the human face, uses the Adience dataset to train the network so as to realize the function of judging gender and estimating age on the Internet. With achieved AlexNet architecture using TensorFlow, the CNN initially completed the expected goals, and analyzed and evaluated the factors that affect network performance, and discussed the improvement methods. Using the face database to train the convolutional neural network, the network finally grasps the ability of face recognition of gender and age. The main problems to be solved are: the training of convolutional neural network data sets and the adjustment of parameters; the extraction and processing of information on face images that need to be identified; the application of convolutional neural networks to face recognition and age estimation. Specifically, it includes: constructing a reasonable convolutional neural network architecture; correctly inputting data and outputting corresponding results; processing and extracting effective information of the pictures.

Today, the Internet has become more and more popular in people’s daily life. People use the Internet to get what they need. However, there are many malicious websites on the Internet. A malicious website has various unhealthy content, such as defraud content, gambling content and false information. Therefore, it is very necessary to identify malicious websites in time to prevent them from cases with various potential harm for users. Previously, the strenuous and time-consuming manual selection was the most prevalent method for malicious website detection. However, with the rise of machine learning, people could build machine learning models to train based on hand-crafted features for this task. Although being much better than manual selection, machine learning models still require a number of hand-craft features and plenty of laboring work. Nevertheless, deep learning models developed in recent ten years save a lot of work by automatically extracting features from malicious websites and produce excellent results, gaining more and more attention from researchers. This paper uses machine learning for malicious website detection, namely Random Forest [1]. And we also use a deep learning method based on LSTM (i.e., Long Short-Term Memory) to train a malicious websites detection model. Then, in order to verify the accuracy of the two models, we used malicious URLs crawled on the Internet to verify our two models. In the end, we got the results. Through comparisons in various aspects, we find that when we use the random forest method, the accuracy of our model will achieve a better result, and the model can maintain high performance than the deep learning-based method.

In this paper, a design of an intelligent container identification system is carried out. The system realizes pre-processing the video of a container, identifying the containers in the video, identifying the container number and container specification code, and determining the landing or transportation state of the container and recording the time. By using the actual video for verification, the system can well realize the above functions and can be applied in the actual scene.

The rapid development of artificial intelligence, especially deep learning technology, has made autonomous and unmanned ships become one of the directions for developing the shipping industry. As a fully automated surface robot, an unmanned ship can navigate autonomously in complex marine environments and accomplish important tasks for humans. The implementation of autonomous navigation capability relies on the ships’ accurate perception of the surrounding environment, but existing unmanned ships’ capability of perceiving the surrounding environment cannot meet the requirements of autonomous navigation for timeliness and accuracy under rough sea and high-speed navigation. In view of that, this paper focuses on local chart’s update service technology based on sensory data from unmanned ships.

Study the workshop scheduling system to optimize the allocation of resources. In order to verify the effectiveness of the frog leaping algorithm in solving the problem of deterministic flow shop scheduling, the characteristics of the deterministic Flow Shop Scheduling Problem (FSP) are analyzed, and an improved hybrid shuffled frog leaping algorithm (ISFLA) is used to solve the problem. According to the characteristics of the problem, initialize the frog population; propose the concept of subgroups, design the communication evolution model from the inside out; conduct a guided neighborhood search for frogs, thereby expanding the solution space of the algorithm and further improving the local search ability of the algorithm. The experimental results prove the practicability and reliability of the improved hybrid shuffled frog leaping algorithm, and it has good application value.

Object detection based on convolutional neural network has important applications in mobile devices, intelligent robots and other fields. However, due to the diversity of traditional models, large number of parameters and slow computing speed, it is difficult to meet the real-time detection requirements of embedded systems. To solve this problem, a mobilenet-SSD model integrating attention mechanism was proposed. Based on the lightweight network model Mobilenet-SSD, the convolution block attention module was integrated into the model, which greatly improved the accuracy of the detection model at the cost of a small increase in computation. Meanwhile, CIOU loss function is used as position loss function to optimize learning strategy and improve training effect. In addition, the processing mode of pre-selection box is improved to Soft-NMS to optimize the detection performance of multi-target overlapping. Experimental results on Pascal VOC2007 and VOC2012 data sets show that the proposed model achieves good performance in recognition accuracy and detection speed, with detection speed of 60.9 FPS and mAP accuracy of 73.6% on GTX1080.The model in this paper has the advantages of small number of parameters, fast computing speed and high identification accuracy, which can better meet the real-time detection requirements of mobile terminals and embedded systems, and provide a good target detection scheme for industrial systems and portable systems.

With the wide application of collaborative filtering recommendation systems, more and more attackers disrupt the recommendation ranking to benefit from the manipulated recommendation results. Therefore, how to effectively detect torrent attacks becomes more and more critical. To solve the problems of low extraction of user information and low utilization of user implicit features in the existing shilling attack detection algorithms, a shilling attack detection algorithm based on non-negative user embedding matrix factorization is proposed. Firstly, the user-user co-occurrence matrix is constructed by using the pointwise mutual information. Secondly, it is found that there is a linear correlation between the quality of the user-user co-occurrence matrix before dimensionality reduction and the quality of the user implicit features. The centralized similarity is used to construct the user deep similarity matrix before dimensionality reduction, to improve the quality of the user implicit features. Furthermore, non-negative matrix factorization is used to extract the features of users in the deep similarity matrix. Through the verification in movielens100k data set and Amazon data set, it is found that the improved algorithm can detect torrent attacks more effectively.

As an important part of robot technology, path planning technology of mobile robot has been widely applied with the rapid development of intelligent robot. The path planning and obstacle avoidance of mobile robots in unknown environment is a hot and difficult topic. The path planning method studied in this paper firstly uses six ultrasonic sensors to obtain the environment information around the mobile robot, then carries on the fuzzy control to the acceleration of the mobile robot, and proposes the obstacle avoidance strategy based on the improved fuzzy algorithm. and verify the feasibility of the algorithm through Matlab. The algorithm was applied in the robot Challenge Competition of 2019 China Education Robot Competition and achieved good results.

Biomedical Events, which widely exist in biomedical literature, describe dynamical biological progress. Biomedical Event extraction plays a significant role in biomedical research, including biomedical event graph construction, medicine research and base construction. Event Trigger Identification, the very first step of event extraction, is to extract the words or phrases, usually verbs or verb groups, that trigger events from sentences or whole articles. Recently, in biomedical event extraction, the main approach to this task is based on traditional Machine Learning. However, this method relies heavily on human effort and expert experience. It can be very time-consuming. Thus, a more efficient method is needed in terms of decreasing labor and time costs. In this paper, we deal with the problem mentioned above to identify the biomedical event triggers in a new way by utilizing a deep learning-based model. Specifically, we use a Bidirectional Gated Recurrent Unit (Bi-GRU) network, an external version of the original Recurrent Neural Network (RNN), to encode the context, and a linear layer is used to classify the entities and predict the triggers. Finally, a test on the Multiple Level Event Extraction (MLEE) corpus gives a satisfying result (F1-score of around 78%).

The human machine interface (HMI) provides recommended HMI conventions to use when designing satellite operations systems.Summarizes the development of aerospace product test and evaluation system and research results. Lists the different methods and systems of HMI design, points out the necessity of using aerospace product test and evaluation to deal with a lot of data generated in aerospace product testing and identification of HMI, and gives a list of principles in future HMI design.

As abuse and malicious rumor often occur in the " fans community ", which has an extremely bad impact on the society, we need to study the emotional tendency of the comment language of the " fans community " in the network, so as to identify the "loyal fans" and "black fans" and explore the language characteristics of the two types of fans. In this paper, more than 50,000 comments were extracted from common Chinese websites, and some data were pre-processed and manually annotated to construct a Chinese "fans community" comment dataset. The three supervised algorithms and one unsupervised algorithm for" fans community ". Emotional dictionary method are used to classify the " fans community " comment information. Then it is analyzed such as the content of the two types of fans comments in terms of sentences, word count, words, and so on. The experimental results show that all the methods adopted in this paper can effectively classify the comments of "loyal fans" and "black fans" by emotion dichotomy. In terms of language characteristics, the comments of "loyal fans" are characterized by multiple nouns, long sentences and regular comment time. "Black fans" comments are often verbs, short sentences and random comments.

In recent years, energy storage safety has received more and more attention. What is energy storage? According to scientific and technical personnel, energy storage mainly refers to the storage of electrical energy, which is a technology that stores different forms of energy in different ways through a specific device or physical medium, so that it can be used later when needed. Energy storage is an emerging industry that spans multiple disciplines, involving electrical, materials, thermal physics, machinery, control, and information. Thermal runaway, leakage, overcurrent, overcurrent, pressure relief, secondary disasters, etc. may cause fire accidents in energy storage power stations. This paper starts with the twin technology to realize a cloud platform system that monitors the safety factors that affect energy storage equipment: battery SOC (battery state of charge) and SOH (state of health estimation method).

In this paper, we propose a 4-tags RFID system that utilizes passive ultrahigh-frequency RFID tags. This system is capable of both auto deduction and positioning of vehicles at roadside parking and localization of vehicles for collision warning on the road. In this work, beamforming technique is utilized to send radio waves to the proposed system, beam switching is applied to change the angle of the main beam. An analysis is conducted to evaluate the limitations of the system, possible improvements are proposed. The location of vehicles with the proposed system can be captured with angle feedback from the RFID tag, safe driving and reducing human labor can be achieved with the proposed system.

Aiming at the problem that automatic code generation technology requires a lot of human intervention and cannot produce a lot of source code, a random code generation method based on syntax tree hierarchical model was proposed. In this technique, the syntax elements in the code abstract syntax tree are divided into five layers according to different degrees of refinement by constructing a hierarchical syntax tree model, and the element composition expression is proposed to formally describe the structure of the code at the text level. Based on this approach, an automatic code generation framework is proposed. Supported by a corpus, the framework can generate multiple nested element formation expressions which are finally converted into usable source code by translators. Experiments from the three dimensions of code complexity, control flow and semantic similarity prove that this method can generate a large number of source code randomly, and the generated code has low similarity in control flow diagram and semantics.

Non-intrusive load monitoring (NILM) is a practical method to provide equipment-level power consumption information, which can be used to improve a variety of application scenarios in smart grids. This paper proposes a CNN-LSTM-based NILM decomposition method, which overcomes the problem of insufficient feature extraction in the existing methods for power decomposition. First, a convolutional neural network (CNN) is used to extract the local features of the aggregated power data. Then, the long short term memory (LSTM) network is introduced to perform global feature extraction on the basis of extracting local features to achieve the fusion of local features and global features. In this way, the proposed method can refer to more comprehensive features when performing power decomposition, which facilitate the decomposition of appliances with different power level. In the simulation experiment on the public data set UKDALE, the average accuracy, recall, and F1 values of the proposed method on multiple electrical appliances of different power levels reached 0.81, 0.94, and 0.86, respectively. At the same time, the MSE and SAE indexes of appliances with simple state were reduced to 1.67 and 1.24, respectively, which fully verifies the effectiveness and advancement of the method proposed in this paper.

Computer simulation technology is a new technology, which is developing rapidly at home and abroad in recent years, and has expanded this technology to many fields. From the perspective of practical application, computer simulation technology has made good achievements in the field of sports. Computer simulation technology can simulate human motion through high-speed photography and motion capture, and find the problems existing in the process of athletes through the calculation of system algorithm on the computer, so as to better improve sports performance. In the motion simulation system, the emerging new algorithm theory can carry out multi-dimensional data acquisition and input, and analyze the stress condition of human bones in detail, which is a good tool for athletes' training. Firstly, this paper summarizes the current research of computer simulation technology and motion simulation system, and then puts forward the problems and corresponding solutions of motion simulation technology in China, hoping to provide some references for motion simulation technology in China.

Aiming at the lack of dynamic interaction in the application of computer database technology in information management, this paper completes the data processing and calculation of information controlled by remote intelligent technology. Based on computer database technology, information management forms a series of system models and personalized processing patterns. Database technology studies and solves the problem of the effective organization and storage of a large amount of data in the process of computer information processing, reduces the redundancy of data storage in the database system, realizes data sharing, guarantees data security and efficiently retrieves data and processes data. And computer database is also divided into distributed database, parallel database, engineering database, deductive database, knowledge base, fuzzy database and so on. Database technology is an important part of modern information science and technology, and also the core of computer data processing and information management system

While training predict models to assess new data, relative researchers always incline to design some strategies to reduce amount of selected sampling points to save cost and time, especially when information about these sampling points is hard to get. Among these strategies, active learning is a popular and useful tool to enhance the efficiency of training predict model by improving the quality of sampling points selected. In biological field, making specific experiments to get result could cost some time and budgets, which fits the situation exactly. While when the information on each data points is limited, active learning method is relatively hard to apply. Thus, in order to relieve this situation, this paper attempts to solve the problem by optimizing the structure of datasets. In the paper, a specific dataset is selected to test the performance of several traditional active learning methods. Meanwhile, a special trick which aims to optimize the configuration of data space is proposed to enhance the performance of both predict models and active learning methods. According to experiments, it turns out that the optimization on data space could let the predict model fit better to the datasets and could help enhance the effect on active learning methods, which has performance enhancement of 5%~22% during the process (5%~20%) of training predict model. By combining with traditional active learning method, the increment could be risen up to 9%~32% under the same progress (5%, 10%, 15%, 20%), which stands for the percentage of data used to train predict model in all the dataset.

In order to make a trade-off between the ratio of compressed images and the sparsity of high-frequency coefficients, and optimize the wavelet transform image compression method, this paper proposes a blocked graded wavelet image compression method based on image complexity detection. Based on the image characteristics, the block complexity detection is defined. The row and column complexity map of the compressed image is calculated by color difference, and the decision of wavelet classification is made according to the content of the block. Through the system test, the blocked graded wavelet image compression has made a good choice among compression ratio, compressed image quality and wavelet high frequency coefficient sparsity.

We propose a Dual Attention-guided Residual U-Net (DARUNet) for Image denoising. This learning is based on the U-Net with several dual attention residual modules embedded in the network to realize the image denoising of Gaussian synthesis noise. Specifically, our method combines the advantages of the reconstruction and transformation of U-Net, the ability of strengthening model training of residual structure and the guiding role of attention mechanism. Among them, the attention mechanism adopts the dual attention parallel structure including spatial attention and channel attention, which effectively guides the model to reduce noise. The model can retain the detail features of the original image, and has excellent ability for image denoising. The resulting images are more natural and have higher quality. A large number of experimental results show that our method achieves advanced performance qualitatively and quantitatively.

The image caption task aims to generate a corresponding semantic description for an image. The current algorithm of generation captioning has these problems, that is insufficient information understanding in the process of generation image description and insufficient research of the relationship between image features. In order to solve these problems, we propose an image description method using the relevance attention mechanism and ITEM encoding. This model uses ITEM encoding to obtain text features containing contextual information and image information, so it can obtain richer semantic information. Then the relevant attention module is used to capture the relevant features which is a correlation between image features and text semantic information. Finally, the LSTM is used for decoding the relevant features to generate image caption. This paper conducts experiments to verify the effectiveness of the model on the MS COCO data set and Flickr 30k data set. The experimental results show that compared with the baseline model using visual attention, various evaluation indicators of our model have improved significantly.

This paper proposes an algorithm for extracting the instantaneous frequency of the incomplete signal under low SNR based on the Viterbi algorithm. We first set the corresponding parameters, and use STFT to project the signal onto the time frequency plane. Locate the maximum value of the plane, thereby positioning the signal component. Taking the time point as the axis, the time-frequency plane is divided into two parts before and after. The Viterbi algorithm with threshold detection is performed in the front and back parts to extract the instantaneous frequency of the signal, and connect them into a complete instantaneous frequency of the signal component. After that, zeroing is performed near the extracted IF path, and then the maximum value in the full plane is re-searched to extract the instantaneous frequency of the next component. Experiments show that the algorithm has achieved good results under the low signal-to-noise ratio condition.

The existing knowledge could not portray the interrelation among the factors affecting cross-border e-commerce capability, the objective of this study is to explore the interrelation between the influencing factors of cross-border e-commerce capability using the RBF-DEMATEL model. Firstly, this paper constructs an index system about the capability of cross-border e-commerce based on the TOE (technology-organization-environment) theory, and then use RBF neural network to calculate the weights between the target index and the influencing factor index to get the direct correlation matrix to identify the influencing factors of cross-border e-commerce capabilities based on the traditional DEMATEL method. The study employs empirical data of 249 enterprises, it shows that international marketing situation, international e-commerce logistics technology, and support conditions of international electronic payment are the three most important three aspects of the cross-border e-commerce capability of enterprises. The result indicates that this method is reliable and effective. The study provides theoretical support for improving the capability of cross-border e-commerce. In addition, this paper opens new horizons for future research in similar areas

With the widespread application of cloud service platforms, the operational efficiency of data operators has been improved. However, if people upload data containing user privacy to the cloud, you will inevitably face the risk of data leakage. Yang et al.[1] first proposed a public key encryption system that supports equality testing in 2010, but this system lacks an authorization mechanism and has certificate management problems. This article proposes CLEET-DBA for the first time, which is a complete encryption system because it adds a timestamp authorization mechanism, solves the certificate management problem, and improves computing efficiency.

In this paper, the sufficient and necessary conditions for the convergence of the MKSOR method are given for a class of block two-by-two linear systems when the eigenvalues of the iterative matrix of Jacobi method are real or pure imaginary. And new results expanding the range of the convergence for the MKSOR method. Numerical experiments show that the effectiveness of new theoretical results.

Aiming at the problems of using mobile device sensors to recognize human behaviors such as low accuracy, few types, and no consideration of transition actions, this paper proposes a hierarchical Bagging-SVM human behavior recognition method. This method first applies the moving average filtering algorithm to the collected sensor data for noise reduction and smoothing; then calculates the amplitude of the acceleration vector sum to characterize human motion behavior, extracts time domain features to build a layered model; finally uses the Bagging-SVM algorithm Perform hierarchical recognition of behaviors in hierarchical models. Experimental results show that compared with other recognition methods, this method can recognize human behavior more accurately and has higher robustness.

Target recognition is a hot issue in the field of high resolution synthetic aperture radar (SAR). Extracting aircraft targets quickly and effectively from SAR images is one of the important tasks of battlefield intelligence reconnaissance. This paper introduces the three main stages of typical target detection and recognition in SAR Image: image preprocessing, feature extraction and classifier, and analyzes the scattering characteristics of aircraft targets in SAR image. This paper combs and summarizes the relevant literature in recent years, introduces the research progress of SAR aircraft target recognition, and finally prospects the existing problems and further development of SAR image target recognition technology.

In recent years, crowdsourcing has been widely used in many actual businesses. Foreign websites such as Uber and Upwork, and domestic websites such as Zhubajie have all adopted the crowdsourcing model to achieve success. However, traditional crowdsourcing uses a three-party centralized model, including task publishers, workers, and crowdsourcing platforms. This scheme has problems such as a single point of failure, lack of trust in all parties, high user risks and costs, and privacy protection. This article proposes a crowdsourcing scheme based on blockchain technology to build a decentralized crowdsourcing model for task publishers and workers. Blockchain technology is used to establish transactions between task publishers and workers. And combined with smart contract technology to support crowdsourcing mechanisms and task management. At the same time, this paper will also propose a decentralized crowdsourcing operation mechanism based on blockchain to ensure that the blockchain crowdsourcing platform does not require the participation of third parties. And can complete the reasonable transaction between the task publisher and the workers through the credit mechanism, transaction mechanism, privacy protection mechanism. Finally, the feasibility of the scheme is verified through smart contract experiments on Ethereum.

The hypoelliptic operators with drift appears in many research fields, for instance, mathematical finance theory, kinetics and models of human vision. In this paper, we establish the Lorentz boundedness of fractional integral operators on homogeneous group, then applying this result and the relevant properties of the fundamental solutions, we obtain global Lorentz estimates of hypoelliptic operators with drift. These estimates expand the regularity of operators generated by vector fields.

In order to ensure the safety of ATC (Air Traffic Control), air traffic controllers need to artificially compare the ATC instructions between the original ones and repetitions from the pilot, which may bring errors and flight conflicts. Hence, an automatic verification method of ATC command repetition based on the Siamese network is proposed in this paper. Firstly, the instruction texts of the original ones and repetitions are converted into word vector sequences, respectively. The attention mechanism is introduced to generate word vectors containing interactive information. Secondly, the two generated word vectors are input into BiGRU for semantic extraction. Finally, the text similarity between the two generated word vectors is measured by Manhattan distance. The experiment results show that the method is effective in the automatic verification task of ATC instruction repetition, and its average test accuracy reaches 94.81%.

In the process of construction, it is not uncommon for workers to enter the construction area without wearing safety helmet correctly. In order to reduce the occurrence of such accidents, this paper uses the target detection algorithm to quickly check the situation of construction workers wearing safety helmets. Firstly, the Stochastic Gradient Descent algorithm is used to improve the speed of target detection. Secondly, for different sizes of helmets, the receptive field is increased by dilated convolution to achieve multi-scale target detection. Thirdly, in order to improve the positioning and classification accuracy of the model, the feature fusion method of skip connection is used to improve the feature extraction ability of the network. Finally, the identity of construction personnel is determined according to the color of safety helmet to improve the expression ability of detection results. The experimental results show that the detection speed of this algorithm reaches 32fps, and the MAP reaches 88%, which can meet the detection requirements in the actual engineering scene. At the same time, the matching of helmet type and worker identity is realized.

A keyframe is a crucial image frame used to describe a shot, and the use of keyframe technology can significantly reduce the amount of data for video retrieval. For example,video-on-demand, face recognition under the camera, key lens retrieval of medical images, etc. Aiming at the problems in the current video keyframe extraction process that the extraction accuracy is low and cannot meet the real-time performance, this paper proposes a real-time video keyframe extraction algorithm CTM-NN based on the inter-frame difference method combined with clustering and neural network. The algorithm uses the inter-frame difference method based on the set threshold, HOG plus HSV first-order moment feature extraction algorithm, and uses the K-means++ clustering algorithm to finally train its own ResNet-50 model, aiming to accurately and efficiently extract real-time video Keyframes. In order to verify the algorithm proposed in this paper, experiments were carried out in the finished news video, landscape video, and real-time concrete mixing video. The experimental results show that the method proposed in this paper can meet the extraction accuracy and meet the keyframe extraction speed of the real-time video so that it can save the keyframes, automatically label while maintaining the time sequence. All in all, the CTM-NN algorithm proposed in this paper has achieved good results in the extraction and storage of real-time video keyframes

Aiming at the high computational complexity of radon ambiguity transform, this paper proposes a sortie resolution method of radon ambiguity transform based on golden section iteration. Firstly, the high computational complexity of radon ambiguity transform is analyzed, and then the golden section iterative algorithm is used to improve radon ambiguity transform. This method greatly reduces the algorithm complexity. The Doppler frequency modulation slope of the formation target is estimated by an iterative algorithm, and then the echo signal is phase compensated. Finally, the formation target sorties can be distinguished in the frequency domain.