This PDF file contains the front matter associated with SPIE Proceedings Volume 12636, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

A new engine for advancing college education reform and development is big data. The construction of a college education big data application is the premise and the core in order to fully utilize the value of educational big data for the reform and development of college education. An important part of the national education modernization construction is deepening the application of big data in education and supporting and leading education and teaching reform. Big data is a fundamental support platform for data collection, storage, filtering, cleaning, fusion, analysis, and specific applications in college education. The design of the higher education big data application support system is the subject of a comprehensive analysis in this paper, which begins with an examination of data management and analysis. Including indepth planning and discussion of the data exchange, data awareness, and data application layers. The big data application support system's application and realization in higher education are then discussed. It can fully fulfill its role in higher education and support the growth and progress of higher education by thoroughly discussing and analyzing the support system for big data applications.

With the rapid development of computer science and artificial intelligence, machine learning as a main method to realize artificial intelligence has developed in many fields. The purpose of this paper is to use the unsupervised algorithm of machine learning to solve the real business problem of online transactions. The traditional statistical method is replaced by machine learning method to analyze the shopping behaviors and preferences of different customers, so as to get more accurate results. The paper processed 541909 sets of original online transaction data with various technologies of data preprocessing (ETL), and used the PCA algorithm and feature selection methods to reduce data dimensions, which could yield different results due to the strengths and drawbacks of each method. The paper used the K-Means algorithm, implemented by Python, to cluster the data after dimension reduction, visualized the data sets, and drew conclusions. The experimental results demonstrate that the proposed method is effective.

The development of today's society is inseparable from the continuous renewal of science and technology. As an important indicator of scientific and technological innovation, patents reflect the core competitiveness of a country. As patent output continues to increase, the quality and value of patents become increasingly important. In this paper, the machine learning method is used to evaluate the quality of electronic information patents and identify high-quality patents from many patents, which can promote patent transfer and transformation, acquisition and custody, transaction flow, pledge financing, and other economic activities. Machine learning is more reliable and scientific than other methods. It is the task of learning corresponding rules and patterns from data and applying them to new data to make predictions.

In recent years, with the development of economic society and the improvement of people's living standards, the number and types of household rubbish are increasing. However, on one hand, due to people's weak awareness of rubbish classification, rubbish in public places cannot be collected well. On the other hand, due to the lack of rubbish classification knowledge, people can not classify some rubbish correctly, which will result in environmental pollution and waste of resources. So we design a intelligent dustbin which can carry out rubbish classification management and delivery guidance, to ensure that rubbish can be correctly placed. For those who do not know the rubbish classification knowledge or cannot determine the category of rubbish, the rubbish characteristics can be identified by scanning the rubbish and compared with the information in the system database to identify the rubbish type and display it on the screen to promote rubbish classification and various rubbish types.

3D point cloud semantic segmentation is one of the key technologies for fast 3D modeling of digital twins. To address the overfitting problem of the large-scale semantic segmentation model RandLA-Net, this paper proposes an improved RandLA-Net method based on Mix3D data augmentation. RandLA-Net is a high-volume, large-perception field model that can directly capture the contextual information of the entire 3D scene, while 3D datasets tend to be more expensive due to data sampling and labeling, often the number of scenes is small and the variance within the data is small, RandLA-Net can easily learn the overly strong contextual prior on the training set, and the model may show poor generalization ability when reasoning in realistic scenes. By introducing Mix3D to mix the two scenes to generate new training samples and implicitly place the object instances in the new contextual environment, the RandLA-Net model no longer relies solely on the scene context to infer semantic labels, but instead infer from the local structure, balancing the role of global context and local structure information in model inference and effectively reducing the overfitting of the training set context. The overfitting of the training set context is effectively reduced. Experimental results on several datasets show that our approach results in a 1.3% and 0.6% mIoU improvement of the RandLA-Net model on Semantic3D and S3DIS datasets.

The distribution of carbon emissions varies greatly among regions and industries in China. In order to master the current situation and trend of carbon emissions in different regions and industries, and formulate and adjust relevant policies in a timely manner, it is necessary to conduct real-time, continuous, accurate and digital monitoring of carbon emissions in major regions and industries. Therefore, "carbon peaking" and "carbon neutralization" put forward higher requirements for the environmental monitoring industry and also brought greater opportunities. In this paper, the spectral information data in typical atmospheric samples are collected in real time, processed and stored structurally through the atmospheric spectral detection method based on mid far infrared spectrum and terahertz spectrum, the spectral feature extraction algorithm based on deep neural network, and the atmospheric carbon emission data analysis and visual monitoring based on big data technology; Run the spectral classification algorithm with low cost and high efficiency on the big data platform, obtain the detailed data of gas sample composition in real time, and display the analysis results visually. Through simulation experiments, the results show that the proposed method can better achieve the monitoring and early warning of atmospheric carbon emissions based on big data and spectral measurement.

Hyperspectral image classification is one of the key technologies in hyperspectral image application. Since the number of bands in hyperspectral images is usually much larger than that in RGB images or multispectral images, it is necessary to reduce the dimensionality of the hyperspectral data of the target by using pre-trained models from other fields. However, this situation may lead to the loss of effective spectral information. Therefore, to overcome this problem, we propose a novel transfer learning framework based on Adaptive Spatial Spectral Relational Networks (ASSRN), which can be applied to knowledge transfer between different hyperspectral datasets. Finally, we conduct extensive experiments on several databases. The inspiring experimental results demonstrate the effectiveness of the proposed framework.

Aiming at the problem of high false detection rate and low detection rate of abnormal network traffic detection methods due to uneven data distribution of current network traffic datasets, a network abnormal traffic detection method based on CNN-GRU is proposed. The method firstly performs local feature learning on the data according to the local learning ability unique to the convolutional neural network (CNN), and replaces the fully connected layer with a global average pooling layer to reduce the feature dimension and parameters; secondly, it uses a gated recurrent unit (GRU) learns the time series features of the data to improve the nonlinear representation ability of the method; finally, for the imbalance between the data, the cross entropy loss function is optimized, and the penalty for abnormal class sample detection and normal class sample detection is passed. The adjustment of the weight increases the attention of the model to the abnormal samples, thereby improving the accuracy of the model detection. The simulation results show that the method can effectively improve the classification and detection performance.

In order to adapt to the modern complex communication environment and improve the anti-noise ability of FH-OFDM in the natural channel environment containing white noise and multipath interference, this paper proposes a data-driven double-frequency-hopping OFDM information transmission method and constructs the overall transmission scheme model. On the basis of MATLAB simulation design and preliminary implementation, the anti-noise performance of this method is further analyzed under different modulation modes and different doppler frequency shifts.

Computer chess games are one of the most significant research subjects in the field of artificial intelligence. Self-play learning depends only on the chess gaming process and the outcome. Neither knowledge nor experts are involved in the learning process. Although self-play learning based on the minimax algorithm, alpha-beta pruning algorithm and Monte Carlo search algorithm has achieved remarkable achievements, there is a lack of research on the quality of the learning samples. We consider that the shortage of convenient software tools is a primary reason why researchers have not sufficiently studied learning samples. Therefore, this paper designs and implements a Self-play Learning Sample Generator for Checkers system (SLSG system).

Federated learning is a machine learning paradigm that can protect data privacy, but the high communication cost and the arithmetic limitation of clients become one of the bottlenecks of federated learning. To address this problem, an adaptive federated learning aggregation strategy based on mobile edge computing called A-FedAvg is proposed, which first evaluates the accuracy gain metrics and cost metrics from each round of global iteration in federated learning, and then combines them into an adaptive metric by weighting parameters to indicate the number of local iterations needed for the next round of global iteration. The A-FedAvg algorithm is designed and then experimented on the MNIST and EMNIST datasets for IID and non-IID data distributions, respectively. The results show that the A-FedAvg algorithm can reduce the number of global communications compared to the FedAvg algorithm with guaranteed accuracy and without affecting the normal use of the mobile device client.

Adverse Drug Reaction (ADR) refer to harmful and irrelevant reactions that occur when normal dosage drugs are used to prevent, diagnose, treat diseases or regulate physiological functions. This definition excludes reactions caused by intentional or accidental overdose and inappropriate medication. In this paper, several models were measured and compared. The results demonstrated that base learners such as LR, SVM, RF, Adaboost, XGBoost may perform exceptionally well in some specific situations. On the other hand, if the precision of the outputs is emphasized, applying Stacking or even Multi-layer Stacking will be the most efficient tool.

Road information plays a fundamental role in many application fields. The optical remote sensing image can analyze and interpret the ground object with its high resolution, which is of great help to extract the road target. In recent years, with the support of rich sample set, the road extraction method based on deep learning has stronger applicability than the traditional road extraction method. Making full use of deep learning for accurate road extraction has become a hot and frontier issue in the field of remote sensing. In view of this, based on a large number of related literatures in recent years, this paper analyzes and evaluates the road extraction methods based on deep learning and traditional road extraction methods, and finally puts forward some suggestions and prospects for the development of road extraction from optical remote sensing images.

Image classification is an essential method to dispose the practical issues including the medical image classification, detection of objectives and process downstream tasks. However, current researchers has utilized the neural network or machine learning model to classify the images based on the image characteristic, which is extremely relies the obtained image data and the training model is a black box. Inspired by the supervised learning algorithm, we proposed a novel self-supervised structure to classify the image data-set. The model structure is consisted of three primary operations including three layers of random transformation, a main neural network layer and prediction layer. In this paper, we specifically demonstrate each components and test our model on a written number data-set. From our extensive experimental results, our proposed mechanism can identify the correct labels in image data-set with acceptable accuracy and reasonable computation cost.

Along with the boom in data rights and the increasing number of responses to protect data privacy, scholars have begun to study the "data isolation" phenomenon while ensuring data privacy and security. Users upload source data to a high computing power cloud server for centralised training required by traditional machine learning algorithms, which leads to uncontrolled data theft and sensitive data leakage problems. By removing data barriers and collaborating on modelling utilizing data from several clients, federated learning has become a successful remedy for this issue. In this presentation, the history of federated learning is discussed, federated learning’s precise defining categories is clarified, and its technological evolution is examined. How well training frameworks for federated learning, including the FedAvg, MOON, and Fedprox algorithms perform is assessed too. Finally, the current problems of federated learning algorithms and the future directions of development are also illustrated objectively.

A novel process monitoring method based on ReliefF-slow feature analysis (ReliefF-SFA) is proposed, to solve the problem of the large variable dimensions and feature selection. First, the data is mapped into a new space, and the slow characteristics of the data is obtained. The new obtained can not only extract the slowest changing components, but represent the inherent characteristics of the system. Second, the feature selection algorithm ReliefF is used to select the more weighted features, to achieve dimensionality reduction. The main subspace is established based on the obtained features, and the remaining features are put into the residual space. The corresponding statistics and control limits are calculated based on the two subspace. Finally, the process monitoring using the proposed algorithm is performed by testing the actual heat treatment furnace process data, and the results show the effectiveness of the proposed method.

In the study of UAV detection of transmission line defects, in order to improve the accuracy of detecting insulator defects, a self-explosion insulator detection method based on optimized YOLOv5 is proposed. In the training, the activation function is optimized, and the original ReLU activation function is optimized to SiLU activation function. Through the training and verification of a large number of transmission line image data collected by UAV inspection, the experimental results show that this method can effectively detect the self-explosion and falling defects of glass insulators under various complex background conditions. The mean average accuracy is used to evaluate the method. The detection rate of insulator self-explosion and flake defect is 94.7 %.

Patients with eye diseases may have a variety of pathological features. If a certain pathological feature is ignored during diagnosis, the overall treatment effect may be seriously affected. To solve this problem, a new neural network model is proposed, in which all the possible pathological features in an OCT image are trained together as the target of multi task classification. The model uses an improved residual neural network to extract the features of OCT images, and adds a spatial attention module in the residual module to facilitate the model to capture the location of lesions more accurately. MMoE model is used to classify the extracted features in multiple tasks, making full use of the potential relationship between each task to help classification, so as to improve the stability of training. In addition, the patient's personal information features and image features are fused with attention and embedded in the top of the model to improve the classification accuracy. The model is compared with VGG16, ResNet10 general single task model and other multi task learning models on 2729 human eye OCT images containing four pathological features. The experimental results show that the training time of the model is 1/4 of that of the single task model, and the average classification accuracy of the four tasks is improved by 4%.

This paper presents a competitive deep reinforcement learning (DRL) structure for the robot control, the basic method of which is to make the two actors learn from each other with competition. And the improvement of training efficiency and learning ability of the competitive DRL model is verified by the experiments under the Gym environments. For the task of robot formation, we construct a cooperation mechanism exerting robot to take the optimal action due to the state and the next reward of other robots within a certain radius around based on the Kuhn-Munkres algorithm, to avoid collisions and blocking among robots. Lastly, we conduct the simulations with the above algorithms, and the results illustrate that the trained robots are capable of self-navigation, obstacle avoidance and formation without collisions.

Federated learning empowers different gatherings to coach a machine learning model together whilst not trading their local data, and give full play to the worth of data from all parties because of the inborn deficiencies of federated learning and the security issues of information stockpiling and correspondence, it still faces a variety of security and privacy threats in practical application scenarios the security attacks and privacy attacks looked by federal learning are described. Then, at that point, the most recent security defense mechanisms and privacy protection means are summed up for such run of the typical attacks, including defense against poisoning attacks, protection against adversarial attack, defense against Free-rider attack, and defense against Sybil attack. At last, by efficiently figuring out the current risks and comparing the defense method for federal learning, the application of existing privacy protection methods to federated learning is discussed and the future exploration difficulties and advancement headings of federal learning are predicted.

Density-based clustering is famous for its ability to extract clusters of arbitrary shapes and to detect noise samples, but many existing density-based clustering algorithms suffer from high dimensional or varying density data. To address these issues, we introduce a novel density-based clustering algorithm, general ratio density (GRD). Based on k-NN graph, it combines global density estimation and local density estimation to better detect noise points in varying density situations. During the process of shifting noise points and dividing clusters, our clustering algorithm can cluster and denoise at the same time. Experiment results on real life dataset and synthetic dataset demonstrates the state-of-the-art performance of our algorithm compared to other methods.

With the widespread use of First Person View (FPV) Drone in film and television shooting, terrain reconnaissance, etc, the operator simultaneously handles the target task and obstacle avoidance trajectory planning flight task under greater pressure, which is very likely to lead to accidents, so combining monocular visual depth-of-field algorithm with Unmanned Aerial Vehicle(UAV) dynamic information to provide the controller with trajectory-assisted indication during flight becomes the focus of this research paper. In this paper, by combining the artificial potential field with deep learning, we use deep learning to regress the target motion trajectory under the action of depth-of-field potential, effectively predict multiple possible obstacle avoidance trajectories under the current position by introducing UAV positional information and UAV viewpoint images, improve the computational speed by streamlining the network, and deploy the algorithm model to Jetson TX2 for experimental validation. The experimental results show that the channel detection achieves the design requirements and realizes the effect of UAV aircraft to assist the controller in high dynamic scenes.

Federated Average algorithm (FEDAVG) is the preferred algorithm for federated learning (FL) because of its simplicity and low communication cost However, if all clients’s local data aren’t independent and equally distributed (that is, nonindependent and identically distributed), FedAvg will have the phenomenon of customer drift, which will lead to the slow convergence speed of the model, and then the efficient cooperative learning to realize the cooperative training of multiple clients will face the challenge of data heterogeneity and vulnerability to attack. This paper systematically summarizes three aspects: local model training improvement, server-side aggregation optimization, and personalized federated learning. The local model can be improved by adjusting the loss function and control variables. Servers can be optimized through asynchronous aggregation, hierarchical aggregation. Personalized federated learning can improve global model performance through both data-based and model-based approaches. This paper puts forward the future research direction of federated learning from all above aspects, and provides reference for the further research of non-IID in federated learning, so as to provide investigation and help for researchers in related fields.

In order to solve the problem that pool fire simulation is difficult to achieve real-time and realistic performance, in this paper, a pool fire rendering method based on improved particle system with collision detection algorithm is proposed. Firstly, mathematical model of pool fire is analysed, on this basis, the improved particle system model is established taking outside factors (such as wind speed, etc.) into consideration, and then via pool fire collision detection model based on adaptive binary tree, this paper uses partition algorithm based on adaptive binary tree to optimize the speed of burning liquid collision detection, this plays a crucial role for the quality and efficiency of subsequent rendering, thus realizes the pool fire rendering process optimization. Experimental results show that this method is simple to implement and the display effect of pool fire rendering is more real.

Collision-free navigation is an important research direction for multi-robot systems, in which the two core problems are navigating to the target point and avoiding other robots. Many researchers use deep reinforcement learning as navigation strategy to realize multi-robot collision avoidance navigation. However, most of them use raw sensor information or global state information of the agent as the neural network input, which is not conducive to extending the navigation strategy to a larger space. This paper proposes an improved deep reinforcement learning navigation strategy, which enables robots to learn navigation and collision avoidance strategies more accurately. This strategy converts the interactive environment state from the global coordinate representation to the relative vector representation, and attenuates the influence of the rear irrelevant agents on the collision avoidance strategy. Experimental results show that the proposed method outperforms existing learning-based methods in three indicators: success rate, additional time to reach the target, and model convergence speed.

Sentiment analysis, sometimes regarded as sentiment classification or opinion mining, can be utilized to assess public perceptions of certain products, occasions, people or ideas. There is a wide range of applications for sentiment analysis, including detecting the overall sentiment of text documents, identifying positive and negative opinions expressed in texts, and predicting the future stock market trend based on social media sentiment. In this study, the convolutional neural networks are employed to categorize the sentiment of English movie reviews, i.e., the CNNs are utilized to detect whether a review is positive or negative. Two text representation techniques, the bag-of-words model and TF-IDF, are applied to convert the movie reviews into numerical representation before transmitting the processed data to the CNN model. According to the model's prediction results on the testing set after training, the text processed with the bag-of-word model achieves a higher classification accuracy of 84.84% in the proposed CNN framework, while the TF-IDF for text processing has better precision of 88.96%.

To solve the problem of low estimation accuracy of FH data link signal parameters based on traditional time-frequency ridge method under low SNR, a time-frequency clustering estimation method based on GA optimization is proposed. Firstly, genetic algorithm is used to extract the time-frequency interval of STFT time-frequency diagram of data link signal; then the time-frequency ridge is extracted; finally, K-means clustering algorithm is used to estimate the frequency hopping frequency by classifying the time-frequency ridge with a cluster number of 6. Experimental results show that this method can accurately estimate the frequency hopping frequency under SNR =−18dB, and effectively improve the accuracy of parameter estimation under low signal-to-noise ratio compared with the traditional time-frequency ridge parameter estimation method.

In order to analyze the thermal safety of combustible cartridge case under direct thermal stimulation, a simplified three-dimensional finite element model of contact heat transfer of combustible cartridge case was established. It is assumed that the thermochemical decomposition process of combustible cartridges follows a two-step reaction kinetic model. The reliability of the thermochemical decomposition kinetic model is verified by cigarette sensitivity test. At the same time, the combustible cartridges under different thermal contact temperatures and coating thermal conductivity are numerically simulated. The results show that the simulation results of cigarette sensitivity are in good agreement with the test results, which proves the reliability of the thermochemical decomposition kinetic model of combustible cartridge. In the contact temperature range of 524 ~ 774 K, with the increase of temperature, the temperature resistance time of the cartridge decreases exponentially. In the range of 0.01 ~ 0.1W / m · K, with the decrease of the thermal conductivity of the coating, the temperature resistance time of the cartridge increases exponentially.

FedAvg is a classic algorithm for Federated Learning. The MOON algorithm is based on the idea of comparative learning. The model in the previous round is a negative sample, and the global model is a positive sample, which improves the local training of FedAvg. However, in the study of MOON, it was found that since the MOON algorithm compares models rather than pictures, the number of negative samples is difficult to expand, and in traditional comparative learning, the total number of negative samples will significantly affect the effect of comparative learning, when the model in this round is far from the model in the previous round, it may approach the earlier model, thus reducing the efficiency of contrastive learning. So, this paper studied more algorithms related to contrastive learning and federated learning. Finally, a dictionary was built based on the MoCo algorithm to save the previously generated local models. The improved MOON algorithm was named FedCOMO. With the extension of negative samples, the training accuracy on the CIFAR-10 dataset can exceed MOON.

With the increasing investment of China Telecom in technical projects, the number and scope of project applications are larger than before, which brings more difficulties to the examination and evaluation, and also increases the difficulty of management. The repeated declaration and repeated approval of projects occur from time to time. This paper is mainly based on the semantic similarity calculation method of project contents, which can find similar projects conveniently, efficiently and accurately, reduce the redundant construction of enterprise scientific research projects, improve the utilization efficiency of enterprise scientific research funds and further improve the scientific research management level of enterprises.

This paper preprocesses four common modes (T1, T1c, T2, Flair) images in clinical brain MRI examination, including image denoising, histogram equalization and offset field correction, and uses convolutional neural network model to segment the image at the pixel level. Compared with the pathologist's manual image segmentation standard, the segmentation accuracy is 95.20%, the F1 coefficient is 0.689, and the Jaccard coefficient is 0.569. Finally, an assistant diagnostic system of automatic segmentation of brain glioma images is designed, which includes a set of UI interface, which is convenient for clinicians and pathologists to use.

Extracting roads from high-resolution remote sensing images has been a popular and challenging topic, and with the rapid rise of deep learning, applying semantic segmentation methods to extract road networks is becoming increasingly popular. This paper systematically and comprehensively analyze almost all relevant studies on applying deep learning semantic segmentation methods for road extraction of high-resolution and very high-resolution remote sensing images since FCN was proposed by meta-analysis methods. As the statistical analysis progressed, we statistically analyzed the publication status, research space distribution, and citation frequency of papers in this field. And then we describe and summarize the key points that have generally received attention in papers in this area, including the model algorithm architecture, the backbone network used, the loss function, the dataset, the experimental implementation details, the evaluation metrics and quantitative comparison of models. Finally, we provide an outlook on potential future research hotspots based on the results of our analysis.

In optical remote sensing images, different degrees of cloud cover will have different degrees of obscuration to ground information. Due to the special spectral and shape characteristics of clouds, cloud extraction makes the detection and extraction of cloud cover in the process of remote sensing data processing is one of the primary problems to be solved for further analysis and utilization of image information, and cloud extraction has gradually become a hot spot and frontier of current research in recent years. In view of this, this paper categorizes and summarizes the existing deep learning theories and methods based on a large amount of relevant literature in recent years, and classifies cloud extraction into three types of methods based on convolutional neural networks, recurrent neural networks, and generative adversarial neural networks by analyzing the network architectures adopted by different methods, briefly introduces the current development trends and representative works of cloud extraction methods based on remote sensing images at home and abroad, and finally gives an overview of the current Finally, we suggest and foresee the current problems and future development trend of cloud extraction.

Microgrid is a small power supply and distribution network, which connects multiple loads and multiple power units with energy storage systems together by power electronic converter interface. Because of the inherent intermittent and random characteristics of distributed generation units in microgrid, it is very important to study the structure and control mode of microgrid system. Firstly, this paper analyzes the structure and control modes of microgrid system under different operation modes, then proposes to optimize microgrid system, and analyzes the advantages and disadvantages of traditional centralized control and distributed control modes, which provides new ideas for the future development of distributed energy and microgrid system.

Emergencies based on social platforms usually trigger high attention and extensive discussion in a short period of time, resulting in an explosive growth of information based on specific topics. If not guided, it may evolve into a serious public opinion crisis and have a negative impact on society. Taking the recent event, the Second Uncle event, as an example, we use python to write crawlers to obtain original data, then carry out preprocessing and word segmentation, and then perform statistics and data visualization to analyze regional heat and track public opinion evolution data. Further, based on the public opinion life cycle theory, a public opinion evolution analysis model is constructed to analyze the evolution process of public opinion. And a clustering algorithm is introduced to analyze the comment text and the Naive Bayes classifier is used to classify the text topic. Finally, the countermeasures and suggestions for dealing with emergencies of public opinion are discussed.

Vehicular nodes in Vehicular Ad Hoc Networks (VANETs) are featured with high speed, unpredictable sudden acceleration, and constantly changing positions. Due to the continuous mobility, the topology of VANETs becomes highly dynamic, which causes difficulties in effective data transmissions among V2V and V2I. The difficulties are originated from high communication overheads for exchanging and updating the topology information, transmission delay or interruption caused by high-speed movements. In order to boost up the performance of VANETs, we take the clustering as an efficient method to segment the moving vehicles in the traffic environment into different subgroups. In regards to the service of vehicle network, some tasks like video transmissions with long periods or intensive data interactions require a stable connection within the cluster, the longer lifetime is one of important factors for maintaining the dynamics as well as the cluster stability. To improve the life time of cluster, we develop the reinforcement learning (RL) agents into self-organized VANETs to form stable clusters and therefore increase the average lifetime, which can interact with the dynamic vehicle environment and output the optimal policy about cluster creation. A simulation is conducted to compare our RL-based scheme against other schemes with no RL factors. The results show that our method outperforms other dynamic schemes in the context of a high lifetime in comparison with the fixed policy. The average life time of our scheme has 6.5 % and 13.6 % improvement approximately in the statics of cluster change time and average vehicle’s speed.

The physical fitness test of college students mainly evaluates the physical condition and training effect through the test results of various items of students. It is a very sound and effective strategy to urge college students to actively participate in sports training. By analyzing the data of physical fitness test scores in colleges and universities over the years, the students with different physical fitness are classified and managed. According to the actual physical condition of students, different training plans are given to assist teachers in formulating more reasonable teaching programs for students. In order to reduce manual calculation and increase prediction efficiency, this paper proposes a comprehensive score prediction model based on machine learning. First, principal component analysis is used to transform multiple attributes with strong correlation into independent attributes without correlation, and the time and space for model training are reduced by eliminating redundancy. Secondly, the back propagation neural network algorithm is used to establish the physical fitness test prediction model, and the model is applied to the test data set to evaluate the performance of the model.

Deep learning-based target detection technology shows excellent performance in healthcare, industry, and transportation. The learning ability of convoluted neural networks (CNN) derives from a combination of feature extraction layers that make the most of a large amount of data. However, they usually require adequate computing and memory resources. Traditional target detection techniques have many limitations, but by using the features of deep learning self-learning, target detection techniques can reduce the complexity of artificial feature extraction. In this paper, the single-stage and two-stage target detection algorithms are analyzed from the angle of detector algorithm, and their respective characteristics are summarized. The most commonly used datasets of target detection are summarized and the evaluation indexes of target detection algorithm are analyzed. Finally, the prospect of the future is given.

In order to solve the problem of missed detection and false detection caused by light change, leaf occlusion, fruit overlap and other citrus detection factor, an improved YOLOv3 citrus detection depth learning algorithm is proposed to achieve accurate and rapid recognition of citrus in natural environment. The algorithm uses image scaling and rotation transformation, homomorphic filtering processing and other methods to process the citrus object pictures. The 1 − IOU value is the distance function, and the genetic algorithm based on K-means clustering is used to optimize the initial anchor box size, all that are enriched the citrus image data set. So the algorithm achieves effective image enhancement and improves the detection resolution and accuracy. The experimental results show that the average accuracy of the citrus detection algorithm reaches 88.7%, the average detection time of single width pixels 640×448 pixels is 131 ms and the system efficiency increases by 9.6%.

In recent years, with the rapid development of computer technology, the application of artificial intelligence has gradually expanded to various fields, including: medical care, agriculture, logistics, education, transportation and other industries. Among them, the deep integration of information technology and education teaching has become a trend, which has made the application of smart campuses develop rapidly. In classroom teaching, students' behavior status can reflect students' classroom participation and teaching effect, which has reference significance for teaching quality evaluation and improving teaching methods. Therefore, this paper studies an algorithm for object detection and student behavior recognition based on convolutional neural network, which counts students' classroom participation and analyzes classroom evaluation by identifying students' behavioral states in the classroom. The main research contents of this paper are as follows: Firstly, the research background of the application of artificial intelligence in the field of education and teaching is expounded. Secondly, the working principle and basic algorithm of image prediction technology are summarized. At the same time, the classroom behavior detection and recognition algorithm based on convolutional neural network is introduced, and the performance of the algorithm is analyzed and optimized.

In recent years, with the rapid development of the Internet, telecommunications and financial industry, credit card fraud as a non-traditional crime cases surge. People's dependence on credit card payments and mobile payments has been further deepened, which also provides a fertile soil for credit card fraud. This paper uses SVM Algorithm to build a credit card fraud prediction model, and verifies the prediction performance through data experiments, and finds that it has a high accuracy for identifying credit card fraud. If the model is applied to real life, it can avoid the loss caused by credit card fraud in time, which has a broad development prospect.

Aiming at the problem that the traditional convolutional neural network has low accuracy face detection in complex scenes, an improved YOLOv5 lightweight network algorithm is proposed. By adding Swin Transformer Block backbone network, the window scheduling algorithm is implemented by using a moving Windows, which limits the self-attention calculation mechanism to non-overlapping local Windows. Meanwhile, the sensitivity field is increased under the condition of allowing cross-window connections to achieve better feature extraction. In order to verify the superiority of Swin-Transformer-YOLOv5, the model is compared with the traditional YOLOv5 algorithm. The model uses a WIDER face in complex scenarios to expose the dataset. The simulation results show that the algorithm can guarantee the original detection efficiency. Meanwhile, the face detection model based on Swin Transformer YOLOv5 has an average accuracy (mAP) of 82.5%, 4.5 percentage points higher than the traditional YOLOv5 algorithm.

In this paper, an improved method was proposed to solve the shortcomings of SVM in the application of tomato disease and insect pest detection algorithm: extracting features of color texture features CCL (CM, CCV, LBP) for fusion, selecting the optimal parameters c and g through 9-fold cross verification, and calculating the OPTIMAL parameters by RBFSVM to obtain the CCL-SVM model. Experimental results show that the improved SVM algorithm improves the accuracy and efficiency of detection.

In this paper, aiming at the problems of early sampling difficulty and less sample data in Industrial auxiliary assembly engineering, a method of sampling by using its digital twin to enhancing samples is proposed. The system establishes the digital twin and digital twin environment space for identifying the target by using unity3D, obtains the sample data of the digital twin space randomly and forms a sample set with the actual sample data collected in the real environment. The two recognition models are trained for physical inspection through faster r-cnn. By comparing their intersection over union, it is verified that using digital twin space to sample can be applied to physical inspection.

This paper entails the mechanisms of Ethereum (ETH) blockchain in term of security aspects. Some basic definitions will be covered to help to understand the structure of mechanisms that keep the blockchain extending and increasing blocks and transactions, such as PoW, PoS, mining. The hierarchical layer model of ETH and the structure of the blockchain itself will be discussed in detail at the beginning, and the corresponding theoretical model and structure diagram will be shown. Specific mechanisms that have and would be introduced in detail, including the beginning structure of ETH while constructed, the issues caused the development of mechanisms, how to make the transition from PoW to PoS, what the shape of the mechanism after the end of the transition and after the upgrade look like. Several key concepts that have been implemented into the Ethereum blockchain are detailed as well. These concepts are “extensions” within the blockchain to improve the security, automation capability and efficiency transactions and mining and Smart Contract will be the example that be introduced in this case. In addition, some features, vulnerabilities, and countermeasures regarding forking happening in the blockchain will also be discussed at the end of the paper. The forking will be explained in two main scenarios, and officially driven hard forking and a natural occurring soft forking.

Three dimensional(3D) palmprint recognition is a prospective biometric recognition technology, and texture direction is one of its main features. Firstly, the texture direction extracted by the Shape Index(SI) is susceptible to noise interference, and to solve this problem, the noise is eliminated by convolution operation, meanwhile the texture direction is extracted using competitive coding. Secondly, the traditional Surface Type(ST) tends to ignore the small curvature feature information, while the palmprint adjacent points have the same curvature, so a Continuous Surface Type(CST) is proposed to obtain surface consistency. Finally, the method of Collaborative Representation(CR) is used to combine features to complete palmprint recognition in classification recognition. The results of the experiment on the 3D palmprint library of the Hong Kong Polytechnic University show that the average recognition rate obtained can reach 99.92%, and the average recognition time is 0.032s compared with other classification methods. The proposed method can improve the recognition accuracy of 3D palmprint while maintaining low recognition time, and the feature extraction ability and classification recognition effect are of practical value.

Restaurant recommendation systems are able to alleviate the information overload problem substantially in the catering industry and facilitate customers to make decisions. Building an effective restaurant recommendation system can bring considerable profits and improve user experience. In this paper, we propose a neural restaurant recommendation model with heterogeneous attention (Restaurant Heterogeneous Attention Model). We use heterogeneous features that users may use as references to construct both restaurant representation and user representation of our model. In our model, the main approach to obtain restaurant and user representation is a restaurant encoder and a user encoder. In the restaurant encoder, we propose a heterogeneous attention model to learn unified restaurant representation. We consider the restaurant name, longitude, latitude, and categories as the encoder input and use both word-level attention and feature-level attention to assist the model to attend to significant words and features. In the user encoder, we learn the user representation based on the restaurants that the user has been to. Also, the different restaurants that the user has been to will provide different information for user encoder learning. In order to find which restaurant is more informative for user representation learning, we also apply an attention mechanism to build our user encoder. After conducting many experiments with a dataset from Yelp, the result shows our model can greatly improve the performance of restaurant recommendations.

In this paper, a TiO₂/NiO heterogeneous composite thin film ultraviolet detector based on the dark state self-depletion effect was designed and fabricated. The N-type TiO₂ film is prepared by the sol-gel method, and P-type NiO thin films and Ni/Au alloy electrodes are simultaneously prepared by a unique one-step oxidation method. Compared with devices based on a single thin film, the dark current of heterogeneous thin film devices is greatly improved, and performances such as responsivity and response/recovery speed are also improved. The dark current of the device with the best comprehensive performance is only 0.033 nA under 6 V bias, the peak spectral response reaches 181.9 A/W at 280 nm wavelength, and the response recovery time is 717 ms and 598 ms, respectively. This research shows that the heterogeneous composite thin-film ultraviolet detector based on the self-depletion effect and other mechanisms has significantly improved comprehensive performance and has more application potential.

With the development of Unmanned Aerial Vehicle (UAV), UAV networks are beginning to exhibit more dynamic characteristics, and it plays an important role in multi-task coordination. MAC layer provides transmission quality assurance for the message transmission. In this paper, A period adaptive load statistics algorithm based on MAC layer traffic (PA-LSBM) is proposed for multi-task priority. Finally, we build a simulation model by Opnet Modeler. The experimental results show that the PA-LSBM has excellent performance in timeliness. Ensuring that the success rate of one transmission of the highest priority data is above 99%, End-to-End delay is less than 2ms.

Aiming at the problems of low accuracy and poor timeliness of complex background images processed by deep learning network models, this paper proposes to optimize the YOLOv5s model. The threshold function is used to denoise the image, and the original loss function GIOU_Loss is optimized into CIOU_Loss function and fine-tuned. The optimized model has good generalization ability and robustness, and the bird's nest detection as a case verifies the effectiveness of the method, which can be used for the detection and identification of foreign objects in high-voltage transmission lines.

The YOLOv5 used in this paper is a single-segment target detection algorithm that improves the speed and precision performance based on YOLOv4. The original anchor frame of YOLOv5 is difficult to cope with the uneven distribution of detection targets in different training sets. In this paper, the dynamic anchor frame mechanism is used to improve YOLOv5. After improvement, the accuracy of the model was 93.5 %, increased by 4.2 %, the recall rate was 92.31 %, increased by 3.64 %. Through the actual data detection, the YOLOv5s model with dynamic anchor frame mechanism can better detect the bird's nest on the high voltage tower.

Traditional detection and tracking algorithms have some problems, such as low detection accuracy and weak generalization ability. This paper introduces a detection and tracking algorithm based on YOLOv5+DeepSort. In this model, YOLOv5 algorithm completes target detection, and DeepSort algorithm is used for target tracking. After the model is trained by a certain data set, the algorithm is verified by the video sequence. The experimental results show that the algorithm can detect and track the target very well, achieving the expected goal.

In the new era of China, with the improvement of people's level day by day, people's awareness of animal protection is gradually enhanced. However, the lack of operator experience of the breeder will cause the stress response of the animals, which will lead to the physical injury of the breeder. Therefore, contactless measurements of animals will be realized in this paper. PCL point cloud plays the same important role in 3D information processing as OpenCV 2D information processing. PCL point cloud can be combined with machine learning method to solve practical application problems. In this paper, pig is taken as the research object, and the 3D reconstruction of pig point cloud is carried out through the reconstruction algorithm in PCL point cloud. Then, this paper uses the point cloud convex hull algorithm and getsurfaceArea() function to obtain the volume and surface area of the pig respectively. Finally, the point cloud reconstruction technology combined with SVR algorithm to predict the weight of pigs can control the weight prediction rate of pigs within 5%.

In order to study the effects of the wall thickness, plug block length and shell steel yield strength of the Penetrator with Enhanced Lateral Effects (PELE) on the opening size of the intrusion concrete target plate, Autodyn software was used to model and simulate the target system. Through simulation, it is found that with the reduction of PELE wall thickness, the pore diameter is gradually increasing. The results show that when the wall thickness of the shell is C-3, the plug length is D mm, and the steel strength of the shell is 900kpa, the opening diameter of PELE is 358mm, which is 40.9% higher than that of the original model.

As one of the important ways of laser imaging, Microelectromechanical system (MEMS)-based mirror imaging has the advantages of small volume and high frequency. In this paper, the relationship between key parameters of MEMS mirrors is studied. Also, the realization method of scanning imaging based on a one-dimensional MEMS mirror is studied. A method of controlling the number of scanning harnesses based on a MEMS mirror is proposed. The single-lens method is used to design and collimate the transmitting optical system with Zemax simulation software, and the effects of the different numbers of scanning wire harnesses are simulated. The result shows that the number of scanning wire harnesses can be controlled, indicating that the system's dynamic accuracy can be achieved.

The fruit picking robot based on YOLOv5 has great development and application prospect in many fields in today's society, among which the target recognition accuracy of fruit image is the key. In view of the low recognition accuracy of small target fruit, this paper improves the data enhancement in the data pre-processing stage based on the YOLOv5 algorithm. The improved image recognition rate is 0.064s and mAP@0.5 is 0.91530. The optimized algorithm has high accuracy and certain practicability, which lays a foundation for automatic picking and other services in the future.

Aiming at the problem of Insufficient feature extraction for data sets in video anomaly detection, an enhanced detection method combining masked convolution and channel attention is proposed. The auto-encoder structure is adapted, and a mask convolution module and a channel attention module are embedded in the memory model. The effectiveness of the proposed method is demonstrated by comparing the results of the method and the original method.

In this article, the deep convolutional neural network is applied to the recognition of diseases and insect pests of crops such as apple, grape and corn. By training the neural network model, adjusting the network layer number and introducing transfer learning, several network models such as AlexNet, ResNet, VGGNet and DenseNet are compared and combined with the network scale, Finally, ResNet is selected as the backbone network for consideration of recognition accuracy and other aspects. The server side adopts Django framework, displays the interface on PC platform, realizes the actual operation of crop disease and insect pest identification.

With the development and application of virtual reality technology, the demand for gesture recognition is increasing. Efficient gesture recognition technology is indispensable in telecommuting and learning. In the research method of this paper, the advanced hand tracking technology is used to construct the human-computer interaction system in the virtual environment. This system uses MediaPipe framework and OpenCV to identify the key points of the hand by collecting real- time image data through monocular camera, and uses Kalman filter algorithm to smooth and optimize the coordinates of the key points of the hand. Through specific gestures in Unity, the experiment of interacting with the building blocks in the virtual environment is carried out to solve the problem of accuracy and stability when taking the building blocks. This method can be applied to multi-scenario (extensible) non-contact human-computer interaction.

In order to solve the problem that the existing face key point detection methods based on depth learning do not explicitly embed the structural dependency between face key points, and most networks have high memory access costs and energy consumption, this paper proposes a depth structure face key point detection method combining depth convolution neural network and all connected conditional random field. This method uses an efficient VoVNet depth neural network to replace the previous CNN model, reduce the memory access cost and energy consumption, and capture the structural relationship changes caused by posture and deformation. The experimental results show that this method is superior to the previous methods in face key point detection, the access cost is greatly increased, and it has good generalization ability for challenging data sets.

In the process of cigarette making, loose moisture regain, as one of the most important processes, plays an indispensable role in production efficiency and product quality. In order to solve the problem of low control accuracy of export moisture content, this paper analyzes the data of influencing factors collected by the existing automation system of cigarette factory, mainly focuses on BP neural network, and uses GA genetic algorithm to optimize and adjust the initial weight of BP neural network, A high precision prediction model with global search capability is established to accurately predict the water content at the outlet of the loose water regain process. The experimental results show that the prediction model has higher prediction accuracy than the traditional BP neural network, and effectively improves the accuracy of the prediction system.

For soldier recognition in the battlefield environment, there are factors such as camouflage and object occlusion, thus leading to incomplete feature information and poor recognition effect. In this paper, we first construct a soldier target dataset conforming to the characteristics of the battlefield environment by analyzing the factors influencing the battlefield environment. Then this paper improves the yolov5 algorithm to detect soldier recognition quickly by adding a channel attention mechanism and improving the spatial pyramid pooling structure. The implementation results show that the predicted mAP value can reach 0.946 with a 3% improvement, the recall rate reaches 0.86, and the detection speed is improved by 5%. It achieves better recognition of soldiers in the battlefield environment.

Aiming at the problems that the traditional classification and recognition methods of left and right-handed motor imagery EEG signals require prior knowledge and feature extraction requires manual design, the process is cumbersome, and the recognition accuracy is not high, A one-dimensional CNN-LSTM network model that can automatically learn signal features is proposed based on the public motor imagery dataset. The CNN-LSTM network model uses a one-dimensional CNN network to automatically learn and extract the deep-level features of EEG time series, and send the feature sequence to the LSTM classifier for classification. The recognition accuracy of the proposed algorithm is 93.57%. Compared with other algorithms, the proposed algorithm can obtain higher recognition accuracy, and at the same time, it can omit the tedious data preprocessing and feature extraction steps. The proposed method is of great significance to the research on brain-computer interface recognition algorithms.

FH-OFDM system converts high-speed serial data streams into parallel transmission using frequency division multiplexing technology, which effectively improves the transmission rate and spectrum utilization of traditional frequency hopping communication system, but the reliability of communication system is difficult to guarantee due to the existence of channel fading, noise interference and other factors. In order to improve the transmission quality, the channel coding technology is usually used to improve the error detection and correction capability of the information; Turbo codes are a class of error control codes with excellent performance. The results show that the low code rate Turbo codes can improve the reliability of the transmission system more effectively, and the anti-interference performance of the system is positively related to the interleaving length.

Forecasting company's future fertilizer sales based on fertilizer sales data are important for reducing production costs for fertilizer companies. Fertilizer sales data have the feature of contextual time-series correlation, and the information of time-series correlation is not considered in the time series (fertilizer price, sales volume, etc.) prediction, which leads to low prediction accuracy. To address the above problems, a Transformer-BiGRU model is pro-posed in this paper. Transformer consists of a multi-headed self-attention, which is capable of mining the multi-feature dependencies between data, and is therefore chosen for multi-feature extraction of fertilizer sales data. However, parameters of transformer are passed in one direction, which cannot extract the temporal information well. Therefore, Bi-directional Gated Recurrent Unit (BiGRU) is introduced to extract the temporal correlation of fertilizer sales data by using its own forward and backward propagation to solve the former problem of poor ex-traction of contextual temporal information. The experimental results show that the Transformer-BiGRU model has the advantage of extracting the temporal correlation of multi-feature data and improves the accuracy of fertilizer sales prediction.

With the development of science and technology, video equipment has become popular in laboratories. Researchers using animals (e.g., macaques) as models often collect a large number of animal behavior videos for the research aims. Researchers often need to watch numerous videos to record and analyze animal movements and behaviors for subsequent analysis. Manual analysis of videos is time-consuming and inefficient. To resolve this problem, we developed an approach to detect macaques in video and then capture information about their location. Using this information, we can quickly analyze animal behaviors, such as movement preferences and behavior patterns. The method was developed based on the YOLOv5 algorithm and had excellent accuracy and robustness. In addition, this method can automatically process videos and generate statistical results and analysis charts. Furthermore, all settings are manually configurable and easy to use to meet the different requirements of researchers. We believe that this new approach will benefit many researchers.

This paper introduces perspective transformation method to improve the error caused by perspective problem in pedestrian distance detection in complex scenes. In this paper, YOLOv5 target detection algorithm is adopted, and perspective transformation algorithm is introduced in the input. It is hoped that by calibrating the video input, the camera view can be transformed into a bird's eye view after perspective transformation, so as to ensure accurate distance between detected targets. It can be seen from the experimental results that the improved YOLOv5 has a good effect on the distance accuracy of pedestrian distance detection, making the accuracy rate reach 89.63%.

By using point clouds as the only input, 3D object detection has made significant progress. However, point clouds often suffer from incomplete geometry and lack of semantic information, which makes it difficult for detectors to accurately classify and locate detected objects, in order to effectively use the rich semantic information in images to improve the performance of point-based 3D detectors, we propose a three-dimensional object detection method based on VoteNet multimodal self-attention mechanism. Firstly, the feature extraction of the data of the two modes of point cloud and image is carried out. Sencondly the three-dimensional point cloud features are mapped to the image, pseudo-3D votes are generated on the image, and then stitched with the point cloud features. Finally the stitched features are deeply fused through the self-attention mechanism. We validated our method on a challenging SUN RGB-D dataset. The results show that our model provides a significant gain (+4.8mAP@0.25) over VoteNet.

With the rapid development of the Internet, the world is gradually moving toward the information age, and the current cyber security situation is getting more and more serious. Traditional methods based on statistical analysis, feature proximity, tensor decomposition, etc. have detection limitations and low detection accuracy when addressing user network security issues. In this paper, we propose an anomaly behavior detection method based on GCN-BiLSTM. First, a graph convolutional neural network is used as a feature extractor to extract the useful graph structure information in the network as a representation vector for the entire graph. Then, a bidirectional long short-term memory network method with an integrated attention mechanism is used for training, and the abnormal behavior detection is completed by combining the extracted feature information. Experiments were conducted on the IDS2017 dataset and compared with the current typical abnormal behavior detection methods, showing that the accuracy of the GCN-BiLSTM based anomaly detection method is further improved and the overall performance is better, which verifies the effectiveness of the proposed method.

In an increasingly information-based modern society, online shopping has gradually become the first choice for people to buy goods. Therefore, more and more e-commerce platforms have emerged, such as Tmall, JD, etc., and e-commerce economy has gradually become an important part of promoting social progress. Since most e-commerce reviews are short texts, it takes a lot of time for consumers to search for useful information for themselves. Therefore, this paper clusters short texts for the first time, and takes the user's review text as the input of the model to get the clustering results. The review text with similar views will be grouped into the same cluster, At the same time, we can get the topic words in each cluster, that is, where consumers pay most attention in the consumption process. Due to the large number of comment texts, we use Dirichlet and multinomial mixed model as the first clustering model, because this model does not need to set the clusters number in advance, and has good performance on large-scale data sets. For the results of the first clustering, although consumers' concerns can be analysed, it is difficult to judge consumers' attitudes, so we take the subject words in the clustering results as input to form a new short text, use TF-IDF to extract the features of the short text constructed by the subject words, then use K-means method to judge the emotional polarity, and then use homogeneity, integrity NMI effectively evaluates the clustering results. Through the experimental results of two clusters, we can clearly find out where consumers pay most attention after purchasing a certain commodity. For example, after purchasing clothes, most consumers focus on the materials and brands of clothes. After clustering, we can get the clustering results of all consumers and the clustering subject words, that is, consumers in the same cluster have similar concerns about goods. After the second clustering, the emotional polarity of consumers can be analyzed, which is of positive significance to consumers' purchase decisions.

The solution to solve the problem of network communication security that has attracted increasing attention in the world. A type of network forum is proposed, which is based on blockchain structure and has good anonymity and decentralization characteristics. In the above case, the user's speech will not be subject to any supervision or make the user to suffer adverse consequences. At present, various cryptocurrency systems based on blockchain structure have been running stably for a long time, which provides a relatively clear technical outline and high-security guarantee for network forums based on a similar structure. However, the network forum described in this paper does not involve financial transactions, but the main content is the user's speech. To satisfy the functionality, the block addition rule based on Karma value and the central computing power allocator concept is proposed.

SCMA (sparse code multiple access) is a kind of wireless Communication technology, which is used to process the uplink portion of wireless communication and ensure a high data transmission rate and massive data connection at the age of information technology. The urgent need for SCMA technology is to address the issue that the user capacity of SCMA technology couldn’t be balanced with the complexity of the detection algorithm. As for solving this, an improved low-complexity detection algorithm I-log-MPA was analyzed, which is based on the modification of the traditional algorithm in multi-user detection. It made up for the solving of the bit error performance of MAX-log-MPA algorithm and the problem of high complexity. The simulation results revealed that the I-log-MPA detection algorithm with lower complexity and lower bit error rate could effectively improve the performance of SCMA system.

With the rise of smart city construction, the importance of vehicle re-identification based on video surveillance has become increasingly prominent. The task of vehicle re-identification mainly focuses on recognizing the same vehicle image under different cameras. In this paper, we propose a multi-label fusion framework for vehicle re-identification based on VIT network. We design a method of anti-angle distortion data augmentation to solve the problem that the performance of VIT is limited by the relative position relationship inside the vehicle structure and the angle deviation caused by the different relative positions of camera and vehicle. At the same time, the key position weight information module is inserted into the coding layer to improve the network's attention to key information. Finally, in view of the insufficient use of the location information between patches in the VIT network, we design a location code based on the Minkowski distance metric to add the relative location relationships within the vehicle individuals to the network. The experiment shows that under the same conditions, our vehicle weight recognition system is superior to most of the more advanced work in vehicle-ID and VERI776 datasets compared with the most advanced supervised vehicle weight recognition work.

Distributed Denial of Service (DDOS) attack is a typical network attack. It paralyzes the target host through a large amount of resource occupation, which poses a great threat to Internet security. Since the birth of distributed denial of service (DDOS) attacks, a large number of them have been designed for DDOS detection every year. However, in the high-speed network era, they cannot control the complex and changing network environment. Therefore, this paper proposes a dynamic threshold detection method based on machine learning. This method processes network traffic in segments, analyzes the characteristics of a segment of network traffic and converts it into four traffic attributes, predicts the threshold of the next window by sliding window, and classifies data points in the window by using similarity. When all four attributes exceed the threshold range in a certain period of time, it is considered to be subject to DDOS attacks. At the same time, a freezing backtracking mechanism is proposed to prevent dynamic parameters from being polluted by attack traffic and speed up the generation of new thresholds after the attack. Compared with the traditional fixed threshold algorithm, this method has a very low false alarm rate for complex network environment, faster processing time, and significantly improved accuracy.

For workshop workers, helmets are one of the most important protective tools. Wearing helmets during operation can further protect the lives of workers. Aiming at the situation that the current safety helmet detection model has missed detection and false detection of small targets and dense targets in complex environments such as workshops, this paper proposes a workshop operation safety helmet detection algorithm based on improved YOLOv5 in combination with practical application scenarios. The CBAM attention mechanism is added to the backbone network, so that the network pays more attention to the safety helmet target to be detected, which can improve the detection effect in complex backgrounds. The EIoU _ Loss function is used to replace the GIoU _ Loss function to improve the convergence effect of the module. By comparing the accuracy and detection speed of the improved YOLOv5 target detection algorithm with the original YOLOv5 target detection algorithm on the self-made safety helmet detection data set, the results show that the average accuracy of the improved YOLOv5 model is 1.4 % higher than that of the original model. Achieved in a complex environment for small targets and dense target detection requirements.

With the increasingly prominent problems of energy carbon emissions, environmental pollution, and climate change, China has further accelerated the construction of new energy sources to deal with a series of ecological problems. China's new energy structure mainly relies on photovoltaic and wind power generation. In addition, wind and solar power generation also has problems such as instability and strong randomness, so existing big data and artificial intelligence can be used to solve this problem and improve wind and solar forecasting. In this paper, a double-loop distributed gradient enhancement algorithm is used to improve the power prediction accuracy of photovoltaic power plants, which effectively solves the shortcomings of slow speed and poor prediction stability in photovoltaic power prediction, and ensures the power dispatching problem based on new power systems.

Value at Risk estimation and forecasting of financial asset series has always been a hot issue in economic and financial research. The RiskMetrics method is often used in the VaR estimation of traditional financial asset series. However, the RiskMetrics method cannot effectively capture the dynamic characteristics of extreme events related to the financial asset series VaR. In this paper, the generalized Pareto distribution method in extreme value theory is used to model the tail distribution characteristics of France’s CAC40 index and to analyze and predict its VaR. At the same time, to compare the performance of the models, the RiskMetrics method was selected for comparison. The calculation and test results show that the GPD method in VaR estimation, although not sensitive to risk changes, still gives good prediction performance. Conversely, the RiskMetrics method, although more responsive to risk changes, has insufficient performance for VaR estimation. Therefore, the overall performance of the GPD method is better than that of the RiskMetrics method, which is also shown by the test results of the predictions of the two models.

These days, we are bombarded with endless spam e-mails. A survey showed that spam accounts for over 70% of all emails. The harm of spam to our daily life includes: spam usually has fraud and unhealthy contents, which requires a vast waste of network bandwidth to transfer spam, and vast waste of space to store spam. As spam is usually embedded in normal e-mails, it is difficult to identify them. In this paper, main technologies to identify and block spam are analyzed including information filtering, blacklist and white list, and intention analysis. Then two experiments of different settings are conducted and analyzed to show how different settings affects the accuracy, precision, recall, and f1_score of the model: the first experiment shows how different thresholds are set to determine whether an e-mail is a spam or not effects the accuracy, precision, recall, and f1_score of the model, the second experiment shows the effects on accuracy, precision, recall, and f1_score of the model when training set is bigger than test set.

The quality of practical teaching in colleges and universities is an important evaluation index of the level of colleges and universities, but the quality of practical teaching in colleges and universities is related to a variety of factors, and the law of change is very complex. To reduce the error rate of teaching evaluation system in colleges and universities, expand the scope of actual teaching evaluation, and ensure the practical application effect of the system, this paper carries out the research of teaching quality evaluation system in colleges and universities based on the BP neural network. Firstly, the system requirements are analyzed, and the requirement structure framework is briefly described for the application requirements and annotations of the system. Then through the establishment of the BP neural network to evaluate the instruction cluster, a circular system execution process is created. Finally, in the main program, access the multi-level evaluation function module, divide the application pressure of the system, divide and verify the results of practical teaching quality, and complete the software design. The system test results show that the error rate of the final system evaluation is controlled below 2. 5% for the evaluation of the four classes, which shows that the system can control the quality of teaching evaluation in each cycle, improve the pertinence of teaching quality evaluation, and improve the final effect of teaching evaluation.

Today's wireless sensor networks are limited in the computing power, data storage and communication bandwidth of nodes, so the research on ultra-bandwidth wireless sensor network technology, media access control and routing protocol has important application value. An important application of wireless sensor networks is real-time monitoring of mine safety. In this paper, the wireless sensor network method based on real-time system is studied, and the effectiveness of the wireless sensor network technology based on real-time system is studied and verified.

In recent years, object recognition methods of satellite remote sensing images based on deep learning have developed rapidly. The deep learning method, which needs a lot of labeled data to train the network, can achieve higher performance than the traditional method. However, it is extremely time-consuming and costly to obtain a large amount of labeled remote sensing target image data. Therefore, how to get high performance remote sensing target classifier by using only a few labeled target images training is an urgent problem to be solved. Based on WGAN-GP, this paper optimizes the constraint conditions of neural networks, and proposes a depth generation model, namely CCWGAN-GP, and applies it to remote sensing image generation. The experimental results show that the image generated by CCWGAN-GP has a high similarity to the real image, and can significantly improve the performance of the classifier under the training condition with only a few tags.

The signal collected by vibration sensor is nonlinear and unstable, which is difficult to analyze in time domain or frequency domain. Therefore, the original signal is decomposed into several intrinsic mode functions (IMF) through empirical mode decomposition (EMD), and then further processing such as feature extraction is carried out. However, modal aliasing and endpoint effects exist in EMD. Therefore, complementary ensemble empirical mode decomposition (CEEMD) is adopted in this paper. CEEMD adds multiple groups of white noises with opposite signs before empirical mode decomposition, which not only reduces the problem of mode aliasing, but also improves the analysis results. The complementary ensemble empirical mode decomposition effectively deals with the collected nonlinear and unstable vibration signals.

A Chinese grammatical error correction (CGEC) method based on the BERT-BiLSTM-CRF model is proposed to address CGEC problems. Firstly, the BERT pre-training model is implemented to generate a deep bi-directional linguistic representation vector that incorporates contextual information. Secondly, the effective text information is mined by using BiLSTM to establish text dependencies. Finally, considering that the text correction process includes substitution, insertion, deletion, and permutation operations, the conditional random field (CRF) algorithm is added to constrain the output variables so they conform to certain grammatical rules in order to obtain the global optimal text sequence. In addition, considering that most characters in the input text do not need modifications and are easily predicted, the focus should be on the wrong position. Therefore, focal loss is introduced to solve this problem. Experiments on the SIGHAN15 and HybirdSet datasets indicate that our approach demonstrates its superiority in error detection and correction tasks regarding their sentence level accuracy, precision, recall, and F1 value metrics.

With the development of information technology, big data technology, and mobile Internet technology, people can quickly obtain and process a large amount of information. However, a significant quantity of detailed information makes it difficult for people to get the information they need efficiently. The emergence of recommendation systems solves this problem very well. It can actively push the information that users are interested in. Knowledge graph-based recommendation technology is mainly used to solve problems, such as the interpretability and cold start encountered by recommendation systems in analysing large data volumes. This paper introduces four aspects of recommendation technology based on knowledge graphs: recommendation algorithm, evaluation metrics, datasets, and models.

For a cognitive multi-input multiple-output (MIMO) network with a pair of primary users and multiple pairs of users, through the two-time pre-coding design of the cognitive user sender, the interference between the secondary users is reduced on the basis of eliminating the interference between the primary and secondary users, and the system capacity is effectively improved. The first pre-coding design uses the water injection method to align the cognitive user's interference with the primary user to the unused space sub-channel of the main user, and the second pre-coding design uses the maximum signal-to-noise ratio algorithm to eliminate interference between the secondary users. Theoretical analysis and simulation results show that while the algorithm effectively eliminates the interference to the main user, the throughput of the secondary user is significantly improved compared with the traditional water injection algorithm.

Modern approaches to real-time semantic segmentation algorithms often sacrifice spatial resolution for real-time inference speed, which results in poor performance. Based on this, in this paper, we propose a lightweight real-time semantic segmentation network based on global attention mechanism by improving STDCNet. First, combined with asymmetric convolution, the short-term dense stitching module is light-weighted, and the global correlation of features is enhanced through global attention. Second, the edge branch can effectively filter out the boundary-independent information in the semantic features through the edge feature fusion module, and restore the lost detail information in the decoding stage. Finally, the improved loss function ensures that the network can update parameters in a direction that is conducive to small object segmentation. Tests and analysis on the Cityscapes dataset show that the lightweight contextual attention mechanism achieves 74.9% mIoU and 118.6FPS real-time semantic segmentation inference speed.

At present, the hand rehabilitation training system mainly adopts passive rehabilitation training method, and the training mode is relatively simple, which cannot reflect the movement intention of patients. This paper has designed and produced a kind of predictive control based on the methods of electricity hand rehabilitation training system, the system can according to your hand on the multi-channel sEMG predict intentions and movement angle, and then drive the exoskeleton robot assisted hand movement, as reflected in training patients' movement intentions, to realize active rehabilitation training. In order to achieve the compliance of the control and prevent the secondary injury to patients, this paper designed the exoskeleton manipulator sliding mode control method. The simulation results and experimental results verify the correctness of the design. The sEMG acquisition and prediction system can accurately predict the motion intention of patients, and the steady-state error of the final control can be kept within 5 degrees, with good accuracy and reliability, which is expected to be applied in the hand rehabilitation training of stroke patients.

Intelligent control of wind turbine is an important part of intelligent wind farm construction, and the prediction of wind turbine operation status is the hot spot of wind turbine operation and maintenance. In this paper, the prediction research is conducted on the sensor values of wind turbine high speed brake disc temperature, stator cable temperature and high speed brake disc infrared temperature with respect to the active power, actual wind speed, generator speed and nacelle temperature of wind turbine and other status quantities. The paper compares the prediction results of multiple linear regression model and BP neural network, and proposes an optimized BP neural network prediction method based on genetic algorithm to overcome the shortcomings of lack of interaction effects and nonlinear causality in the linear regression model, and determines the optimal number of neurons in the hidden layer of BP neural network, which speeds up the training data and improves the accuracy of neural network model prediction is improved.

Apriori improved algorithm based on compression processing and hash table is given to solve the problem that the traditional Apriori (association rules) algorithm must repeatedly scan the database when performing data mining frequent itemsets. It makes full use of the method of transaction compression, firstly compressing the data, and then compressing the obtained data into the hash table for the second time through hashing technology. In this process, you only need to scan the database once, and then even if the support degree changes, there is no need to perform a second scan. The experimental results show that the improved algorithm can not only mine the association rules of the dataset correctly, but also greatly improve the operation efficiency of the algorithm.

With the continuous development of data science and financial technology, financial data visualization methods have become an essential key technology in the field of financial data analysis today. From the technical point of view, the mainstream visualization analysis takes the fusion of large screen and multiple views, and the nature of its visualization effect is more focused on the enumeration display, without fully analyzing the data characteristics from the essence. The single view visualization analysis technology is difficult to get clear and effective visualization display through correlation analysis, dimensionality reduction algorithms and principal component analysis. From the application point of view, credit card customer data, as an important part of financial data, has positive practical significance in customer profiling, product recommendation and risk prediction, and the targeted improvement research of its visualization method has an important role. The semantic feature group method combines the domain knowledge and data distribution characteristics of credit card customer churn data, composes and analyzes the semantic feature groups, and obtains explicit visualization and analysis results by combining the understanding of the actual problem and the numerical characteristics of the data itself. The accuracy and efficiency of the data representation based on the semantic feature group method are verified by comparing the data dimensionality reduction visualization methods such as multi-view fusion method, T-distribution random neighborhood embedding and principal component analysis in the experiment.

Due to the instability of power system operation, the risk of large-scale blackouts is high. The large-scale power failure accident will cause the shutdown of factories and machines, bring huge losses to the economy, and also cause serious negative impacts on people's lives. Therefore, it is a very important work to ensure the operation safety of the power grid and gradually eliminate power failure accidents. This paper studies the analysis and evaluation methods of power system security risk and vulnerability, demonstrates the necessity of power system security risk and vulnerability analysis and evaluation, explores and studies the index system of power system security risk evaluation, studies the overall framework design, evaluation process, evaluation content, etc. of power system security risk and vulnerability analysis and evaluation, and provides a basis for power security supervision.

A novel magnetic field sensing method is proposed based on the photonic crystal fiber (PCF) filled magnetic fluid (MF). Firstly, the relationships between the loss of PCF and magnetic field under different temperature (20℃, 50℃and 80℃) were analyzed theoretically by finite element method combining perfect match condition. It could be concluded that, when the external magnetic field rose from 20Gs to 300Gs, the PCF loss sensitivity was linear, and the sensitivities were respectively 0.039dBm/Gs, 0.030dBm/Gs and 0.024dBm/Gs. Then, the results were of verified by experiments. The PCF loss sensitivities were respectively 0.037dBm/Gs, 0.030dBm/Gs and 0.023dBm/Gs under temperature T=20℃, 50℃and 80℃, and the magnetic resolutions were respectively 0.024 Gs, 0.066Gs and 0.084Oe. The results indicated that theoretical calculation was consistent basically with the experimental verification. This provides a new method for magnetic field sensing.

In steelmaking-continuous casting production processes, the optimizing scheduling of ladles is an effective way to reduce energy consumption and improve productivity. Because the ladle scheduling must satisfy multiple conflicting goals and conflicting constraints, it is difficult to adopt the existing optimal scheduling methods. The intelligent ladle scheduling method was proposed for the ladle scheduling problem of molten steel in Baosteel in China. First, the strategy of ladle scheduling was proposed. Then, the first-order rule learning was used to select the optimal objective, and the least general generalization rule reasoning was used to extract the ladle matching rules. In addition, an optimal scheduling model for ladle was established, and an intelligent scheduling method for ladle was proposed, including the ladle matching, ladle path optimization and crane scheduling. Rule reasoning was employed to select decarbonized ladle or dephosphorization ladle. The multi-priority heuristic method was designed to determine the path from ladle to converter, refining furnace and continuous caster. The heuristic method of avoiding interference strategy was designed to decide the crane for ladle transportation in the air and the start and end time of operation. Finally, a software system for ladle scheduling was developed, and industrial validation was carried out with real data. The results show that the overall production efficiency and ladle turnover rate are improved.

In recent years, more and more attention has been paid to the study of transformer voiceprint. Therefore, this paper mainly studies the denoising technology of transformer voiceprint signal in high noise environment, compares the wavelet packet transform and the wavelet transform, respectively, and finds that the de-noising effect of wavelet packet transform is better than that of wavelet transform. The three-layer decomposition of vowels in the high-frequency part is conducive to the subsequent extraction and identification of fault signals.

The dataset used in this project is derived from the SMS spam classification dataset in the UCI Dataset Repository, and it is necessary to understand what the dataset looks like before pre-processing the data. The first step is text clean-up, and the second step is text feature extraction. This paper investigates and compares the performance of classifiers combining different dimensionality reduction methods on spam datasets to provide a reference for related classification studies. The project then uses the scikit-learn machine learning library to train the classifier, dividing the dataset into 75% training sets and 25% test sets, and introducing classifiers such as NB, IR, SVM for training. After classifier training is complete, test the result of the model on the test set. Use trained classification models to predict the category of a message (regular mail or spam) The result shows that the best performer among the various classifiers is the SVM.

As a fundamental task in computer vision, human pose estimation has undergone in-depth and thorough research, but we found that the key point labeling uncertainty is rarely paid attention to, but this uncertainty will affect the detection performance. Motivated by this finding, this paper first uses the patch-adaptive loss to address the label ambiguity problem inherent in keypoint annotation. In addition, this paper combines the advantages of convolution and Transformer architectures to inject long-range and short-range information into the network. Experimental results on publicly available datasets demonstrate that our method improves the performance of human poses outperforming state-of-the-art methods and improves the robustness of the network.

Semantic segmentation is vital for computers to process image scene parsing. Semantic segmentation requires the output of high-resolution segmentation results that classify each pixel in an image, and high-resolution feature maps consume significant computational costs in deep learning networks. To trade-off the accuracy and speed of semantic segmentation models, researchers have proposed various different approaches to extract features by semantic segmentation backbone networks. In this paper, we propose a novel semantic-guided backbone network (SG2Path) based on a multi-branch backbone architecture, in which we design a semantic-guided upsampling module (SGUM) to better fuse high-resolution spatial features with low-resolution semantic features in different branches, which effectively solves the semantic misalignment problem between feature maps of different resolutions. Experiments on the Cityscapes dataset and the Visualization analysis of the model prove the advantages of our model in the semantic segmentation task and the significant application potential.

In order to improve the accuracy of differential frequency hopping signal detection, a frequency detection method is proposed based FasterRcnn. Firstly, the received differential frequency hopping signal is converted to time-frequency by STFT. The time-frequency distribution of the signal is obtained. Then the position information of the frequency point and the frequency of differential frequency hopping signal are used as tags. The position information of the frequency points in the time-frequency distribution graph is marked with four-dimensional coordinate, and is inputted into the FasterRcnn convolutional neural network for training. The simulation results show that the frequency detection method based on FasterRcnn convolutional neural network can effectively detect the frequency of differential frequency hopping signals.

To address the problems of low segmentation accuracy and inconspicuous image foreground and background segmentation caused by inadequate target feature representation and incomplete model capture information in existing image segmentation tasks, we proposed an image segmentation algorithm based on deformable convolutional networks and self-supervised contrast learning. The existing U-Net segmentation algorithm is insensitive to the image details, and it is difficult to accurately segment the foreground and background in the image. In this paper, an improved algorithm based on U-Net network is proposed. First, deformable convolution network is incorporated into U-Net network instead of traditional convolution, and its adaptive instance target perceptual field is used to improve the ability of this paper's network to characterize foreground regions and reduce the wrong segmentation of background and fully learn the image instance details, so as to improve the segmentation accuracy of the algorithm. On this basis, inspired by self-supervised learning, a self-supervised contrastive learning module is added to learn self-supervised contrastive between instances in an image to enhance the algorithm's understanding of the image. The experimental results show that the algorithm in this paper can effectively improve the segmentation accuracy.

The infrared thermal imaging technology plays a very important role in the current power system. This paper identifies the transformer in the infrared image according to its characteristics to form an image. Build the temperature field, get the highest, lowest and average temperatures of the equipment, compare with the database, and determine whether the electrical equipment has faults, the type and severity of faults.

Object detection in real scenes has always been a research hotspot, and dual-video object detection is one of the key researches. In this paper, deep learning YOLOv5 algorithm is first proposed to detect pedestrian targets, and then the realtime stitching technology can be used to achieve image fusion, so as to achieve the dual-video target pedestrian tracking algorithm. The experimental results show that the pedestrian target obtains good continuous tracking effect in the fast moving environment.

Thermal power image segmentation is the key step of power equipment infrared diagnosis. Power equipment infrared diagnosis can timely and accurately diagnose the potential accidents and fault precursors of operating power equipment, which is an important part of power inspection. With the increasing popularity of mobile intelligent terminals, modern power inspection requires deploying semantic segmentation models on mobile devices. Considering the small memory capacity of mobile devices, we propose a new lightweight network architecture, called Edge-Assisted Context Guided Network (ECGNet), for semantic segmentation of thermal infrared electrical equipment images. ECGNet has been carefully designed to learn the context information of thermal infrared images and improve the problem of edge blur on the premise of small parameters and small memory consumption. Under the same number of parameters, a large number of experiments on LS-ETS dataset show that ECGNet can obtain better results than the most advanced method.

In order to solve the problem of co-channel interference between base station and D2D users caused by fixed path loss compensation coefficient in existing power control algorithm of device to device (D2D) communication in cellular network, a packet-based power control algorithm was proposed. Based on the traditional closed-loop power control algorithm, the proposed algorithm calculates the supplement of the path loss compensation coefficient by the distance from D2D users to base stations and cellular users, and sets three groups of interference regions to distinguish the supplement of the path loss compensation coefficient. Finally, the joint closed-loop power is used to suppress the interference of the system. Simulation results show that, compared with the traditional closed-loop power control algorithm, the signal to interference plus noise ratio (SINR) and throughput of the system can be improved to a certain extent by using this algorithm, which can significantly improve the power control performance of D2D communication in cellular networks and achieve the effect of interference suppression.

In recent years, blockchain and cryptocurrencies have become popular topics. The significant increase in the cryptocurrency market attracts more people devoted to these fields. Many people consider cryptocurrencies a revolutionary invitation or a path to wealth; some believe it is a colossal fraud. Although some research has been conducted, cryptocurrency's primary mechanism and concept are still confusing for ordinary people. Thus, in the research, we dig into blockchain technology and provide a lucid explanation of some fundamental mechanisms. The research has chosen four popular cryptocurrencies (Bitcoin, Litecoin, Dogecoin, and Ethereum) as examples, and some of their essential features will be introduced. Then, the article focuses on the consensus mechanisms and why people always talk about "mining" new coins. Finally, we analyze the usability of blockchain technology and the discussion about the future of cryptocurrencies.

To enable mobile robots to perform different tasks intelligently and autonomously in different work scenarios, we usually use SLAM to achieve simultaneous localization and map building of the robot mainly using to complete the functions. Localization and mapping are done only by detecting and matching feature points in the image, which only utilizes the geometric information in the environment and cannot obtain the linguistic information in the environment, resulting in the robot not being able to perceive the surrounding environment well. Location and mapping can only be done by detecting and matching the feature points in the image. Only the geometric information in the environment is used, and the linguistic information in the environment cannot be obtained, which makes the robot unable to perceive the environment well. In this paper, a three-dimensional semantic map is implemented based on the system design. The data uses RGBD image sequence, and the target detection algorithm YOLOv4 is applied to ORB-SLAM2 system to extract the semantic information in the environment. A three-dimensional dense point cloud map constructed by ORB-SALM2 system is segmented using an improved LCCP algorithm. Semantic information in the environment is projected to the three-dimensional point cloud according to the globally consistent camera posture, and a three-dimensional dense point cloud map with semantic information is constructed.

Ethereum is a blockchian-based technology that enables a public ledger to be created and maintained without a trusted third party. Ethereum uses elliptic curve cryptography to create accounts and signature, the consensus of PoW as well as PoS to verify transactions and create blocks, and stores data in the modified Merkle Patricia tree. This paper introduced the main three technical aspects of Ethereum and updated some concepts as its technology is updating over time. At the time of Ethereum's transition from PoW to PoS, this paper compared the two mechanism and concluded that PoS is more vulnerable to security issues and more analysis is needed to maintain its long-time security. The paper also discussed uncle block attack and proposed two possible strategies to prevent.

In this study, from the soil compaction characteristics, the Swin Transformer network in deep learning method is used to identify the soil surface rough and smooth, and compared with the traditional network method, proving that the algorithm in this paper has good results in terms of recognition accuracy and recognition speed; using Mask leveling, wavelet transform denoising and other steps in the traditional image processing method to identify whether the soil contains cracking; using the methods of physics, mathematics and soil science, fusing gravity sensors to detect soil water content, porosity, etc; using BP neural network, fusing the extracted features to achieve the purpose of real-time detection of soil compaction, and the experimental results show that the accuracy of this algorithm is 98.8%.

Image stitching is one of the important tasks of computer vision, which is used in many fields such as autonomous navigation and autonomous driving. However, traditional stitching methods rely too much on the quality of feature detection and show poor performance for images with few features or low resolution. Although existing deep learning-based methods can make up for the shortcomings of traditional methods, they are only used on mobile robots with smallbaseline or fixed perspectives. To address the above limitations, we propose an image stitching network consisting of three modules: multistage keypoint matching module, DFAST module and multistage image reconstruction module. First, we use a multistage keypoint matching module to align the reference image and the target image, and obtain deep homography estimates between reference and target images at different scales of features. After that, the DFAST module is designed to stitch images of arbitrary views and generate stitched feature maps at different scales. Finally, the multistage reconstruction network is used to reconstruct and optimize the stitched feature maps from feature level to pixel level and fuse stitched images of different scales to generate finer texture details. Experiments results show that our method surpasses previous methods including state-of-the-art traditional and CNN-based methods.

The application of high energy density explosive CL-20 is limited due to its high sensitivity and low safety. In order to reduce the sensitivity of CL-20, CL-20 and low-sensitivity RDX were combined at the molecular level to form cocrystals, so as to improve the performance of explosives. At present, CL-20/RDX cocrystal theory is not mature enough, for Studying the mechanism of CL-20/RDX cocrystal formation and its effect on CL-20 and to provide some theoretical support for the preparation of CL-20 cocrystal. Molecular dynamics simulations of CL-20/RDX cocrystal at different temperatures (200K,298K,400K) are performed. The simulation results were analyzed by radial distribution function (RDF), Hirshfeld surface analysis, and growth simulation. The results show that the CL-20/RDX cocrystal g(r) peak values of O…H at 200K, 298K, and 400K were 1.55, 1.76, and 2.00, respectively; as the temperature of the cocrystal decreases, the hydrogen bonding interaction strengthens. The g(r) peak value of the N-N bond was higher than that of the pure component CL-20, indicating that the cocrystal enhanced the strength of the N-N bond and reduced the sensitivity of CL-20. At the same temperature, the g(r) value of the N...N bond is lower than N...O, and CL-20 decomposition starts from the N-N bond. Hirshfeld surface analysis showed that the O...H contact contribution to the Hirshfeld surface area was 38.2%, which accounted for the largest proportion of interactions, and the de+di value was the smallest, that is, the interaction distance was the shortest, indicating that the driving force of CL-20/RDX cocrystal formation was derived from hydrogen bonding. The strong interaction of the cocrystal is mainly generated at the nitro group, and the hydrogen bond force of the central molecule CL-20 as a hydrogen bond donor is greater than that of a hydrogen bond acceptor. The simulated morphology is long columnar, which is similar to the experimentally prepared morphology, and its rounded shape has certain advantages in terms of sensitivity.

We study the multi-output probability regression prediction and uncertainty quantification problem in multi-output regression prediction with gradient boosting models, and propose an uncertainty estimation method based on Gaussian Graph Model (GMM) constraint and natural gradient. Under the assumption of Gaussian distribution, we employ GMM to model the conditional independence relationship of the outputs, and propose a mixed parameterization to represent the subfamily of Gaussian distributions defined by the graph model. We set this mixed parameters as the learning target of the gradient boosting model, and use natural gradient in Riemann space that based on Fisher information matrix (FIM) instead of the vanilla gradient to be the steepest descent direction. The formula to compute FIM and natural gradient for the subfamily defined by GMM is deduced. Experiments show that the proposed method is able to incorporate the priors on the outputs to optimize the search space, therefore, lead to better uncertainty estimation results on multi-output regression for gradient boosting models.

The typical deep learning detection workflow of sonar images uses a transfer-learning paradigm, i.e., to select an external optical image pre-trained classification network as the backbone for feature extraction, and then to fine-tune the entire R-CNN network on detection datasets. Though transfer learning can effectively avoid overfitting and accelerate network training, transferred models are usually too redundant and cannot achieve the optimal generalization. This paper uses Automatic Deep Learning (AutoDL) to improve the automatic target detection performance of sonar images. An improved neural architecture search algorithm is applied to automatically design the architectures of R-CNN detectors. The automatically designed architectures don’t require external optical images to pre-train and have lower requirements for R-CNN parameters and computations. Experiments on an actual measured sonar image dataset composed of small targets show that, the automatic detectors designed by the neural architecture search algorithm are better than the transfer learning detectors in terms of recall, detection precision, computations, and parameters.

In order to solve the real-time control problem of hand rehabilitation exoskeleton robot, a motion angle decoding model was proposed based on surface EMG signal and synchronous motion angle value. The long short-term memory neural network was used to construct the hand motion angle decoding model. During recognition, EMG signal and synchronous angle signal are sent to the model for decoding, and the output of the model is the angle prediction value after 200ms. The experimental results show that the combination of motion angle signal and EMG signal can significantly improve the decoding ability of the model.

In view of the existing BOC modulated signal synchronization methods have problems such as small application range, limited sidespeak suppression ability and low synchronization probability, and the large code tracking error of BOC modulated signal synchronization in multipath environment, an improved BOC signal multipath suppression method based on code-related reference waveform was proposed. The improved method makes the correlation peak of the cross-correlation function between the BOC modulated signal and the local reference waveform significantly smaller than the correlation peak of the autocorrelation function of the BOC signal by designing the local reference waveform, so as to effectively suppress the error influence caused by the multipath effect. The simulation analysis of BOC signal using narrow correlation method, Double delta method and the new method shows that the new method has better multi-path suppression performance than narrow correlation method and Double delta method, and can almost completely suppress the code loop tracking error caused by multi-path effect in the case of long delay. This study can provide a theoretical reference for the high-precision reception of Beidou navigation signals based on BOC signals, and thus provide a foundation for the enhanced application of satellite navigation system.

As a key part of autonomous driving technology, 3D object detection in Bird-Eye-View(BEV) view is designed to accurately detect road conditions based on multiple sensor inputs. However, current 3D detection methods are usually biased towards the acquisition of depth information. There is a strong correlation between 2D object detection and 3D object detection in BEV, especially the latest 2D detection has high robustness and accuracy in complex environments. To the end, this paper proposes a 3D object detection algorithm with object feature enhancement , which reasonably uses mask generated by 2D object detection as prior information and facilitates object extraction in BEV. In particular, in order to amplify the difference between the target and the background, we first designed the Data Preparation Module(DPM) to convert the 2D detection results into a binary target mask which completely separates the target from the background. Then, an Object Enhancement Module (OEM) is designed to enhance the contrast and texture details between the object and background. Finally, we propose Channel Interaction Module(CIM) considering the correlation between multi-channel features, which fully allocates the weights between channels. In this way, MaskBEV-Tiny scores 34.0% mAP and 41.1% NDS on the nuScenes val set which surpasses BEVDet-Tiny by a largin of 2.9% mAP and 2.0% NDS.

Images taken in hazy weather are susceptible to the effects of haze, resulting in blurred images and low contrast to the extent that important information is lost in the image. Therefore, this is necessary to dehaze haze images, process the image information and ensure the normal operation of other computer vision tasks. Traditional deep learning-based image dehazing methods often suffer from uneven haze removal, colour bias and loss of detail. To solve this problem, this paper proposes a single image dehazing method (IMNet) based on pyramidal input for image dehazing. The network is divided into three modules: an intensive feature extraction module, a pyramid input branch and a detail deepening module. This paper uses two loss functions in combination, which can help preserve texture details more effectively. Experimental results have shown that IMNet outperforms other dehazing algorithms in terms of metrics and visual effects.

In modern war, the continuous development of missile technology makes the air combat environment more complex. With the emergence of intelligent mobile targets, the interception and guidance of missiles are facing severe challenges. It not only requires the continuous improvement of the precision guidance performance of missiles, but also requires multiple missiles to work together to maximize the effectiveness of group warfare. In this paper, a new method of missile interception guidance strategy based on game theory is proposed, which is based on dynamic game and matrix game in game theory. By establishing a many to one missile target game model, the strategy set of missile and target is determined. Then, according to the many to one missile target game model, the system equation is determined, and the cooperative guidance strategy is obtained through the calculation results. Finally, the Nash equilibrium solution of the dynamic game model is obtained by using the model prediction algorithm. And a cooperative guidance strategy suitable for many to one interception is obtained. The simulation results verify the effectiveness of the guidance law.

Sewage treatment technology is developing day by day, a systematic, objective, complete evaluation method to evaluate the sewage treatment process is imminent. Sewage treatment AAO process as a domestic sewage treatment plant land is relatively more process, we in comprehensive evaluation, the AAO process evaluation index system in the role in the whole sewage treatment process is different, to the final treatment of water quality weight is different, which requires us to design a scientific and reasonable sewage treatment AAO process performance evaluation index weight determination model. This paper combines AAO process performance evaluation index system with BP neural network model, design and study it, and the evaluation index weight is determined through BP neural network model, which improves the effectiveness and objectivity of process performance index weight.

Wellbore pressure field prediction plays an important role in the safe production of gas well. Fracturing sliding sleeve is a part of typical downhole operation tools in wellbore, and the pressure loss of fluid through fracturing sliding sleeve is an important parameter affecting the prediction of the entire wellbore pressure field. In order to study the pressure loss law of typical fracturing sliding sleeve, a simulation model of typical fracturing sliding sleeve flow was established in this paper. In order to improve the accuracy of calculating pressure loss in sliding sleeve, the Eulerian multiphase flow model was selected. The effects of different displacement, liquid density and ball seat diameter on the pressure loss of fracturing sleeve were studied. The results show that with the increase of displacement and liquid density, the pressure loss of the fluid through the sliding sleeve increases. However, with the increase of the ball seat diameter of the fracturing sleeve, it is the opposite, among them, the displacement and the ball seat diameter bigger influence on the pressure loss. Finally, based on the simulation results, the empirical formula of flow pressure loss of acid fluid, fracturing fluid and sand-containing fluid in fracturing sliding sleeve was fitted, and the overall average error was less than 5.71%, which can be used for pressure field prediction and calculation. The research results can provide guidance for the evaluation and diagnosis of wellbore pressure field.

Oil is a crucial strategic resource and a guarantee for the steady operation of a nation. Lack of oil supply has a significant impact on both national security and social, economic, and social development. Given that the PRD region is one of China's most significant economic zones, we took nine indicators from both supply and demand (pipeline transportation of refined oil products, oil and by-product production, electricity consumption, car and motor vessel ownership, population, GDP, output value of primary, secondary, and tertiary industries), and obtained the oil consumption of the PRD for a lengthy period of 15 years. We have also designed and implemented an oil consumption forecasting model based on a long and short-term memory network algorithm, with a loss rate of 0.1, which verifies the validity of the model. The suggested model is a useful tool for predicting regional oil consumption and can serve as a foundation for study of future oil supply and demand in the PRD region.

Starting with the demand of greenhouse gas detection, this article designs the harmful gas detection and alarm system based on 51 single-chip microcomputers. The hardware circuit takes STC89C52 as the control core, MQ-7 gas detector as the sensor, and ADC0809 is the digital-to-analog conversion chip, supplemented by power supply module, display module, button module, short message module, memory module, etc. Realize the concentration display of harmful gas in the greenhouse, and give the alarm when it exceeds the standard. At the same time, the GSM module can also send short messages to the mobile phone, playing the role of remote prompt.

In the environment of rapid development of technology, scientific research, processing, manufacturing and other fields are also developing in the direction of richer functions and more sophisticated design, and the requirements for the properties of various parts are becoming increasingly strict. In the field of electronic information, the small resistance of a part also deeply affects the quality of the product. In this paper, a set of micro resistance measurement system based on MCU is completed by combining the relevant knowledge of MCU with resistance measurement technology. The micro resistance measurement system based on STC89C52 single chip microcomputer is designed to provide better experience for users, and intelligent products are designed around customer needs. A micro resistance measurement system including LCD module and resistance measurement module is designed. The micro resistance measurement system based on SCM designed in this paper effectively reduces the system volume, improves the portability of the system, effectively reduces the production cost, has high measurement accuracy, and makes the system easy to use. It is a convenient and fast integrated design of SCM system that can adapt to various workplaces.

In order to achieve high efficiency and energy saving electromechanical integration zero transmission structure, a permanent magnet synchronous motor using an external rotor is adopted (PMSM) direct drive mode of coal mine belt conveyor. Through the research and analysis of the special application environment and control strategy of this kind of motor, vector control strategy method based on active disturbance rejection technology and sensorless control is adopted. To improve the speed regulation performance of belt conveyor when the load frequency changes conduct modeling and simulation analysis The improved sliding mode observer can effectively improve the accuracy of rotor position estimation and tracking performance, the effectiveness of the algorithm is verified.

In recent years, with the development of industrial technology gradually mature, modern industry has carried on the thorough research to the measurement field, put forward higher requirements for accurate measurement and performance optimization in other directions. Therefore, the focus of this paper is to test the networking system based on laser diameter measurement to achieve accurate measurement of cable radius. The measured experimental data can show that the system can show high precision, high efficiency and real-time control for accurate diameter measurement. Through the development of the corresponding measurement data host computer display interface, verify its feasibility, thereby reducing the product development cycle, with strong practicability.

The transient heat transfer finite element model of a 30mm caliber gun with a local barrel is established for the temperature distribution characteristics of the barrel during continuous high frequency firing of small caliber guns. Combining the classical internal ballistics theory and heat transfer principle, the heat flux of the high temperature gas of gunpowder in the barrel is calculated based on the inverse heat conduction problem model. In this paper, the heat flux distribution is used as the heat load condition of barrel heat transfer, and the temperature distribution characteristics of barrel at different flame temperatures during continuous firing are analyzed based on LS-DYNA. The results show that during the continuous firing of the gun, the heat flux of the inner chamber of the barrel changes exponentially, and the higher the flame temperature, the greater the peak heat flux. The temperature of the inner wall of the barrel is characterized by sawtooth-shaped pulse change. The higher the flame temperature, the more heat is accumulated in the barrel, which makes the temperature of the barrel rise more quickly.

Web service specifications of WSRF and WSN provide a new method to solve the problem of stateful middleware services. At present, cloud native technology has been widely used in the field of stateful middleware management platform. Based on stateful Web services, a middleware SOA framework for the control of physical examination population has been designed, and its stateful resource encapsulation method and state, message notification mechanism when resource changes and other key technologies have been described in detail. Finally, the middleware has been implemented through Globus Toolkit 4.0, and good application results have been achieved on the medical examination appointment platform.

The earthquake emergency assessment model proposed in this paper uses the Python Programming Language to achieve automatic output. The overall framework includes data stacking and analysis and automatic output of the epicenter, casualties, and economic loss. When an earthquake occurs or an earthquake drill is simulated, the program can, by inputting the three elements of the earthquake, automatically output the economic loss and casualties caused by the earthquake, calculate disaster-related information in a timely and rapid manner, and quickly generate an earthquake emergency assessment report. This is essential for formulating a proper emergency rescue strategy in a timely manner and improving the efficiency and quality of rescue efforts.

In the OS Kernal, how to allocate disk space is a difficult question. Buddy system is one of important way in various allocating disk space way. We can reduce the fragmentation by Buddy System. However, Buddy System cannot avoid the fragmentation created because the logic in algorithm. How to improve the algorithm is a important way to reduce the fargmentation created and increase the life about system. Because of the Buddy System always merge and split the memory page, fragmenation is created in the process and accumulate in system. With the passage of time, the space which can be used by system will less and less in disk. If we can reduce the number of times of the memory page’s merged and spilted in the process of system in used, the fargmentation will be reduce in the system used.

Wireless body area network (WBAN), as a low-power and miniaturized network attached to the human body, is extensively used in health and medical fields. Therefore, it is completely necessary to take into account the impact of human motion on the network in a mobile environment. The channel characteristics of WBAN are different due to its different human body characteristics, motion characteristics, and environment, but the characteristics of human motion are sometimes regular and sometimes irregular, making the changes in the channel of WBAN unpredictable. In this paper, a relay transmission and channel fading reserve control strategy was designed by considering the residual energy of nodes and channel changes, aiming to solve the problem of shadow fading caused by human motion and unstable links caused by channel changes. The experimental results indicated that compared with the traditional IEEE 802.15.6 protocol, this protocol reduced the network energy consumption and packet loss rate, so it was suitable for reliable data transmission under human motion.

The failure of power system will cause potential safety problems of power supply in specific areas. A large number of power tests show that the security of power system is one of the important foundations to ensure the stable operation of power grid. With the rapid progress of computer technology, the research on new technology of substation high voltage test has become a hot topic for scholars. This paper mainly analyzes the detection technology of high voltage test and the application of new technology.

In order to study the influence of the structure of prefabricated fragments warhead on the killing power, a projectile structure scheme with two prefabricated fragments installation modes, which are external and internal, is designed based on a certain grenade, the simulation process of warhead explosion and fragmentation dispersion is accomplished by AUTODYN, and the characteristics of fragmentation parameters under the two structures are obtained. The results show that the fragment velocity of the prefabricated warhead is 4.7% higher than that of the external warhead. The built-in prefabricated fragment warhead structure is more conducive to the formation of effective fragmentation. The reduction of the cylinder wall thickness of the prefabricated fragment warhead will lead to the premature rupture of the projectile body, and lead to the decrease of the initial velocity of the whole fragment.

In this paper, a TiO₂/NiO heterogeneous composite thin film ultraviolet detector based on the dark state self-depletion effect was designed and fabricated. The N-type TiO₂ film is prepared by the sol-gel method, and P-type NiO thin films and Ni/Au alloy electrodes are simultaneously prepared by a unique one-step oxidation method. Compared with devices based on a single thin film, the dark current of heterogeneous thin film devices is greatly improved, and performances such as responsivity and response/recovery speed are also improved. The dark current of the device with the best comprehensive performance is only 0.033 nA under 6 V bias, the peak spectral response reaches 181.9 A/W at 280 nm wavelength, and the response recovery time is 717 ms and 598 ms, respectively. This research shows that the heterogeneous composite thin-film ultraviolet detector based on the self-depletion effect and other mechanisms has significantly improved comprehensive performance and has more application potential.

Nowadays, with the increasing diversity of Internet programs, more and more programs have launched more functions. However, many system vulnerabilities have gradually surfaced, and the process of finding and repairing vulnerabilities has become an imminent problem for developers. Code snippets related to dynamic memory allocation and resource preemption operations are a major source of operating system failures. In this paper, a new fault detection technique based on dynamic kernel tracing is proposed. The fault source and fault type are determined by collecting kernel call stack and global state transition information. The fault detection technology is implemented as a loadable kernel module in Linux system, which can effectively collect kernel information without adding hardware or modifying the original system code. The experimental results of fault injection show that the proposed method can detect faults effectively, and the detection delay is smaller than the method based on timeout time and performance index. How to deal with these vulnerabilities is the subject of this article.

In this article, a 126 kV single break vacuum circuit breaker is taken as the research object to analyze the high-power density permanent magnet synchronous motor mechanism and repulsion mechanism. Also, a combined operatingmechanism is designed. Based on the different requirements of vacuum circuit breaker for opening and closing, a scheme of combined operating mechanisms has been proposed. The combination of repulsive opening and permanent magnet motor closing ensures the rapidity of opening and the stability of closing. An equivalent excitation circuit of the electromagnetic repulsive opening mechanism has been established, and the motion characteristics of the designed highspeed permanent magnet operating mechanisms have been obtained in the opening process. Also, the cogging torque of the drive motor is designed and optimized to carry outthe dynamic simulation of torque output of the high-power permanent magnet motor driven by the high voltage circuit breaker. A suitable high-speed buffer device is designed, which can not only effectively buffer the kinetic energy of switch motion but also keep the moving guide rod at the opening position without any rebound. The testing machine platform of variable stroke and load reaction are also established. Eventually, the prototype is developed and tested.

This paper introduced the development of a hand rehabilitation therapy system based on virtual reality technology aimed at encouraging stroke patients with hand movement disorders to undertake rehabilitation exercises. The system uses Leap Motion controller, Unity3D development platform and Visual Studio2019 integrated development environment to design, including three modules: game module, interaction module, adaptive module. The game module includes physical simulation, collision detection, audio-visual feedback and other functions, which can better immerse patients in the virtual environment. The interaction module realizes the function of collecting and transmitting the hand movement information and interacting with the game module. According to the patient's performance, the adaptive module, such as game time and object size, can adjust the internal parameters of the system reasonably and appropriately, so as to match the task with the patient's ability. The system has the characteristics of self-adaptation, strong immersion and simple deployment, which can improve the fun of training and the sense of gain of players, and lay the foundation for stroke patients to carry out hand function rehabilitation training at home.

Artificial intelligence technique based on deep learning is gradually improving. Strategic power grid technique with a new generation of transformation capability is becoming the focus of all countries. The intervention of new energy yield electricity has made a great contribution to the planning of power grid at this stage. Therefore, the new energy grid connected to the power grid information platform has become the driving force for the development of new energy industry. This paper explains the impact of new energy grid connection on power grid quality, and also explains the application of artificial intelligence technique connected to power grid information platform.

In the last few years, mechanical fault diagnosis (FD) based on deep learning has been systematically applied to various industrial fields. However, most methods rely on enough labeled data, but collecting enough sample of failure is timeconsuming and labor-intensive. It remains challenging to train a model for FD with numbered training data and to function correctly under intricate operating scenario. In response to this problem, the FD method based on Siamese neural network (SNN) has shown promising results in recent years, especially in the case of insufficient training samples; the fault diagnosis work has achieved good results and has been widely concerned by researchers at domestic and foreign. To better combine the SNN for fault diagnosis and give full play to the advantages of the SNN in fault diagnosis, this paper summarizes and analyzes the fault diagnosis methods based on SNN. Firstly, the basic framework of the SNN method is introduced, and the advantages and application scenarios of the framework are pointed out. Secondly, the existing fault diagnosis methods based on SNN are classified, including the principle of the technique, improved methods, and shortcomings. Finally, the FD methods based on SNN are summarized, and the development of the SNN in FD is in prospect.

The PMSM based full speed control method is an effective way to realize sensorless control of permanent magnet synchronous motor (PMSM), but because of the complex structure of PMSM, high performance requirements, and the existence of nonlinear, strong coupling and other problems, this method can not achieve high-speed operation in the full speed range. To solve these problems, this paper proposes a sensorless control method for PMSM in the full speed range based on adaptive sliding mode control. Firstly, an adaptive sliding mode filter is loaded on the stator resistance to reduce the influence of current component in the sliding mode controller; Then, the stator resistance is estimated according to the stator resistance, torque and speed variation; Finally, the nonlinear mathematical model of the motor is established to estimate the variation of torque and speed. The PMSM sensorless full speed control model is built in Matlab/Simulink. The simulation results show that this method can effectively realize the PMSM sensorless control in the full speed range.

At present, one of the problems of Chinese word segmentation is the low efficiency of Out-Of-Vocabulary (OOV) detection in the field of expertise. Due to restrictions on the characteristics of the words of the profession itself, the word segmentation of architectural texts is not very effective in identifying OOV. This paper proposes a new method to recognize OOV, which is an unsupervised method based on improved algorithm and entropy. This paper uses algorithms to identify strings with relatively large interdependencies between texts, filters through the stop-words vocabulary and corpus to obtain candidate dictionaries, calculates the entropy between candidate dictionaries, and determine the final OOV by setting an accurate threshold, Add the recognized OOV as a professional dictionary for word segmentation. Experiments show that by using the algorithm proposed in this paper, the recognition effect of OOV in architectural text has been significantly improved. Compared with the algorithm, P ( precision ) increased by 15.92 %, R ( recall ) increased by 7.61 %. Therefore, the final word segmentation precision can reach 82.15% and recall can reach 80.45%.

Aiming at the problems such as illumination variation, occlusion interference, complex background and diversity of objects in the grasping process of space manipulator, a grasping perception method of space manipulator which can be used for real-time calculation was proposed. A hybrid algorithm strategy of long-distance gray feature search, lightweight detection network classification and lightweight segmentation network was adopted to obtain the centroid/position/attitude. First, when the end is far away from the target, the simple gray distribution feature is used in the field of view to detect whether the target is there or not, so as to quickly judge the area where the target is concerned. Then, the target area is extracted, and the lightweight deep convolution detection and recognition network is used in the process of close movement to carry out continuous recognition and classification of the target, determine the key parts to be captured, determine whether the target is an effective target, and constantly adjust to ensure that the target stays in the middle of the camera. Finally, when a relatively short distance is reached, the lightweight segmentation convolutional network is used to conduct semantic segmentation of the target, and the shape binarization data of multiple angles are obtained. In this way, the position and attitude information of the target are obtained according to the binocular stereo vision, and the feedback control computer is used to complete the fetching task. Compared with previous designs, the identification accuracy of the proposed method is increased by 1.4%, and the calculation speed is reduced by 58%. The simulation results show the effectiveness and adaptability of the proposed method.

The fault simulation of satellite control subsystem is an important part of the satellite fault simulation verification platform. It adopts the international standard modeling and simulation universal interface FMI, and through the full digital simulation mode, uses advanced simulation technology and virtual instrument technology to build a high fidelity user monitoring interface to achieve visual simulation flight, equipment fault simulation, user operation, information monitoring and other functions. The system level key failure modes of the satellite are simulated and verified in the development stage, and the ground recurrence and analysis of the failure phenomena are assisted in the satellite on orbit operation stage. It supports the satellite test and stable operation in orbit through satellite fault verification, satellite electrical performance test and other work.

In order to study the damage characteristics of typical tank targets under shock wave load, a full-scale finite element model of tank targets was established by using AUTODYN software. By changing the charge and the position of detonation point, eight simulated working conditions were set up to obtain the damage law of tank armor and internal key components during the process of explosion damage to typical tank targets. The research results show that when the charge explodes on the right side of the tank, the displacement deformation of tank armor occurs at the position corresponding to the explosion center level, and the deformation at the turret joint decreases with the increase of detonation distance. Judging from the impact vibration of the people in the car, all of them are severely damaged, and the impact vibration of the driver is the largest. Because the gunner and the captain are in the turret, the impact vibration is smaller than that of the driver. The fire control computer has great impact vibration response under different working conditions.

Abstract text. In view of the low monitoring accuracy of drivers' dangerous behaviors, tired driving and distracted driving in traffic safety, and the inability to accurately and timely give early warning, in order to make-up for the shortcomings of existing detection methods and improve the accuracy of identification, in this paper, a novel distracted driving behavior detection method (CF-Net) based on CNN and multi-model fusion is proposed for intelligent cockpit environment. Firstly, a distracted driving detection model was built through the classification network, and the influences of different parameter values and classification networks on the model were analyzed on the open source data set of distracted driving. The XGBoost learning tool is added to the Stacking training mode to optimize the model. By comparing various models, the multi-model fusion network (CF-Net) distraction detection model is further applied to the driving environment, and the model's overall generalization ability and effectiveness are verified. The experimental results show that the detection performance based on a multi-model fusion network (CF-Net) is better, the test accuracy reaches 93.45%, the verification loss is reduced to the minimum, for the current intelligent cockpit environment, the driver distraction detection has a significant advantage.

With the development background of compressed air launcher for single weapon cartridge, the compressed air launch system and internal ballistic calculation simulation model design with different drain diameter is established by using fluid-solid coupling AUTODYN software, relying on the theory related to the effective energy of compressed air, the mathematical model built is calculated and simulated. The simulation results show that: at 70mm drain diameter, the acceleration obtained by the launcher is 8.2m/s²and reaches the maximum value, the initial velocity of the launcher, which meets the requirements of weapon combat performance. The application of cold energy compressed air in the field of man-portable weapons firing has some feasibility.

This paper introduces the wireless power transfer system (WPT system) of application prospect, introduces the WPT system of basic components, put forward the LLC series resonance circuit topology structure, put forward the new modes to completely symmetrical resonant compensation network apply to the primary sides ; the simulation results show the effectiveness of the compensation circuits.