“This PDF file contains the front matter associated with SPIE Proceedings Volume 12787, including the Title Page, Copyright information, Table of Contents, and Committee Page.

Compliant tactile sensing has received increasing interests in human-robot interactions and soft robotic sensing. However, it is still challenging to achieve high sensing performance over an area with low wiring complexity. This paper is to implement and evaluate a soft, distributed sensor that is enabled by the technique of electrical impedance tomography (EIT), which allows areal sensing using only boundary measurements. To achieve mechanical compliance of the sensor, a functional natural rubber composites filled with exfoliated graphite (EG/NR) is fabricated by spray coating. Long-term stability of the sensing material is evaluated under cyclic tests, and a high gauge factor of 6.25 is obtained, which can be explained from the evolution of the micromorphology under strain. A continuous sensing area with a diameter of 12 cm is prepared, and 16 electrodes are attached along the periphery to implement distributed sensing. The performance of the sensor is characterized under indentation tests. Results show that the intensity sensitivity monotonically decreases with the distance to the boundary, and consistent response is obtained along all the radial directions, providing predictable performance over the sensing area.

In order to meet the characterization of circuit miniaturization, the design method was proposed on via subject structure loading open and short stub resonator object structure, which the new band-pass filter is now available. After the simulation and optimization, the physical processing test was carried out, and the volume size of the filter was 40 mm * 20 mm * 0.9 mm, and the operating pass bandwidth can reach the 1.36~2.83 GHz range, relative bandwidth was 70.2%, the insertion loss in band was less than1.5dB, the return loss was less than -15dB, the maximum suppression outside the band was less than -40dB. The measured values were in good agreement with the simulated values.

In view of the situation that using vehicle mounted sensors to detect ruts in rut detection technology can cause secondary pressure damage to the road surface ,a method of using multi rotor unmanned aerial vehicles equipped with laser radar to detect ruts is proposed ,and a scheme of multi rotor unmanned aerial vehicles equipped with laser radar acquisition system is designed .Existing rut calculation methods are analyzed, and virtual wire method is used to calculate rut depth .Laser radar data on the Zhenglan Banner section of the national highway G207 are actually collected .Study the characteristics of the road segment data, verify the rutting algorithm in this article, and provide rutting road segment results.

In order to achieve a balanced breadth and depth of knowledge graph in the energy domain and alleviate the high complexity of information sources during energy service, a comprehensive energy service knowledge graph construction and completion method based on a filter-feedback mechanism is proposed. Firstly, the domain filtering mechanism is designed to filter the information in the input text unrelated to the energy domain by calculating the entity commonality. Then, a feedback mechanism is constructed using third-party information to complete the missing entities or relations in the triplet. The method is validated on energy domain and public datasets, and experimental results demonstrate that the accuracy of domain knowledge discrimination exceeds 90%. Moreover, its performance is comparable to that of the public dataset.

This paper presents the preparation of an ITO/p-Si heterojunction on a polyimide (PI) substrate using magnetron sputtering. The microstructure and optoelectronic properties of the heterojunction were characterized using field emission scanning electron microscopy, X-ray photoelectron spectroscopy, four-probe testing, digital source meter, multimeter, UV-visible spectrophotometer, and laser confocal micro-Raman spectroscopy. Experimental results showed that the heterojunction has a resistivity of 0.82 Ω·cm and exhibits MOS capacitance characteristics in its IV curve. The heterojunction has good optical properties, with its strongest absorption occurring in the UV region at 221 nm. Under excitation by a 325 nm laser, the heterojunction exhibits an emission peak at both 586 nm and 632 nm.

A real-time all-optical switch with a wide tunable band gap is presented in this study, using photonic crystals and phase change materials to improve the performance of optical communication networks. The decisive factors of photonic band gaps are explored, and 2D photonic crystal straight waveguides are realized with an extinction ratio of 20dB. Furthermore, several types of all-optical switches with superior light manipulation capabilities are designed, and an all-optical information processing system is realized, providing a feasible solution for high-speed all-optical networks with large capacity.

Aiming at the problem that the current grid connected frequency control effect of wind turbines is poor, the grid connected frequency control method of AC excited variable speed constant frequency wind turbines is proposed, the virtual inertia control method of variable speed wind turbines is optimized, the power oscillation damping control algorithm of AC excited variable speed wind turbines is improved, and the grid connected frequency control process of variable speed constant frequency wind turbines is simplified. Finally, it is confirmed by experiments, The grid connected frequency control method of AC excited variable speed constant frequency wind turbine generator has high practicability and fully meets the research requirements.

A belt defect detection model based on unsupervised generative adversarial network is proposed for the problem of low number of defective samples in belt defect detection. The model is trained on defect-free normal samples and the detection results are obtained by calculating the outliers between the input samples and the reconstructed samples at the time of detection. Considering the information loss problem of the encoder compressing the input samples, this paper adds a selfattention mechanism to the model to enhance the extraction of useful features. In addition, the LeakyPeLU activation function in the model is replaced with PReLU to improve the fitting ability of the model, and the method is experimentally proven to achieve better image reconstruction results. The experimental results show that the accuracy of the assay using this method is 98.13%, AUC value is 99.89% and AP value is 99.75%, which was better than the other two comparative models. The method has good reconfiguration and defect detection capabilities.

Through Silicon Via (TSV) is a promising microelectronic interconnection technology in Three-dimensional integration (3D). Three-component additives are used for non-porous TSV filling. However, various experiments must be carried out to optimize the concentrations of various additives. To avoid this expensive process, Phenyl disulfide propane sodium (BSP) in the TSV filling is investigated and the correlation between the copper seed layer surface and molecular structure were simulated through molecular dynamics using materials studio software. It can be seen that the BSP concentration is 0.1 g/L, the inhibitor AESS is 1.0 g/L, the current density is 2 A/cm2 , and after 6 hours of electroplating, the TSV filling assumes a "V" growth mode, which can obtain a defect free TSV copper plating filling result.

Nowadays, with the increasing popularity of computer, network and virtual reality technology, the organic integration of virtual reality technology and digital ancient architecture plays a huge role in the fields of digital archiving of ancient architecture, cultural relics protection and online tourism. This paper mainly discusses the design process and related technical issues of the 3D roaming system of ancient architecture based on virtual reality, and analyzes the technical scheme of the system, and designs the overall design, main process and functional modules of the system. The work done in this paper can provide reference for other types of ancient architecture model and virtual tour.

This paper considers the risk assessment of railway tunnels and proposes a method based on deep learning in the context of artificial intelligence. Specifically, we use convolutional neural networks (CNN) to process and analyze sensor data inside the tunnel, and establish a supervised learning model to predict the health status of the tunnel. We also introduced a Bayesian optimization algorithm to optimize the model's super parameters, and used different evaluation indicators to evaluate the performance of the model. The experimental results show that our method is superior to traditional methods in terms of accuracy, robustness, and interpretability, and can help railway companies better manage tunnel risks.

The development of sensor technology has driven the growth of data scale and data dimensions. While the dimension of the dataset is rapidly increasing, efficiently processing high-dimensional and small-scale datasets is still a challenge for traditional machine learning methods. Currently, feature selection is an effective method used in this case. In this study, we propose a practical hybrid feature selection method called mRMR-FBSTA. The algorithm can be divided into three stages: the first stage is the filter part that provides a preliminary feature subset and necessary feature weights to avoid high computational costs caused by using only the wrapper part, the second stage called fast iteration stage which used to quickly find the acceptable solution and the third stage called global iteration stage that aims to find the global optimal solution based on the acceptable solution found by the fast iteration stage. The experimental results show that the proposed method achieves excellent performance on high-dimensional and small-scale datasets.

Aiming at lithology identification of drilling cuttings, a lithology identification method based on improved VGG16 is proposed in this paper. Firstly, the rock image data are preprocessed, and the data are expanded by image segmentation, normalization and data enhancement. Secondly, the lithology recognition model is built with VGG16 as the backbone network, finally, the convolutional neural network is good at extracting local features, but it is difficult to obtain global information, by embedding SE attention module, the network can capture long-distance features. The experimental results show that the improved VGG16 model has an accuracy of 90.32%, which is 1.65% higher than that of VGG16 network. It shows that the improved VGG16 network model used in this paper has a good effect on rock lithology identification, which proves that the study of this paper has a certain significance.

Accurate calculation of line loss in distribution networks can better guide the power system on how to optimize the power grid and how to carry out technical loss reduction work. In response to the problems of many electrical parameters, complicated steps and low accuracy of results required for the calculation of theoretical line loss values in traditional distribution networks, this paper proposes a distribution network line loss prediction method based on the light gradient boosting machine (LightGBM) model algorithm. The method uses machine learning models to model key electrical parameters and line loss values to automatically calculate the grid line loss. To address the difficulty of tuning the LightGBM model, we use a Bayesian optimization algorithm to adjust the model parameters. In this paper, grid data from the Kaggle data platform is used for analysis and validation, and the experimental results show that the model proposed in this paper has higher prediction accuracy than the traditional BP model.

In recent years, with the construction of electric energy and its related supporting infrastructure, the product iteration and update speed of the low-voltage complete switchgear industry has been gradually accelerated. Against the background of growing market demand year by year, adjusting product models according to customer requirements has gradually become the main production mode of low-voltage switchgear. In order to better improve the traditional production process, domestic and foreign research mainly revolves around the secondary development of various types of 3D CAD software. Although this can achieve a certain degree of design preview, it still lacks an effective interaction mechanism and intuitive modeling effect for the whole business flow. This paper is based on the framework of OpenCascade to parametrically model the low-voltage switchgear cabinet and key components. Combined with Qt as the front-end interaction, it provides a new solution for the customized production of low-voltage switchgear.

Aiming at the chaos control of permanent magnet synchronous motor, a dual-parameter collaborative intelligent optimal control method based on GWO-RBFNN was proposed. Starting from the perspective that the controller can automatically search for the expected motion state, the distance between two points on the Poincaré cross section is selected as the controller input. And considering the coupling effect of system parameters on the dynamic behavior of the system, a dualparameter cooperative controller is designed based on radial basis function neural network (RBFNN); Grey Wolf Optimization (GWO) algorithm is used to optimize the selection of controller parameters; By slightly adjusting the two controllable parameters of the system, the chaotic motion of the PMSM system is controlled to the expected motion state. In the simulation study, compared with the single-parameter intelligent optimization control method based on GWO-RBFNN, the results show that although both methods can achieve chaotic motion control, the control speed of the dual-parameter collaborative intelligent optimization control method based on GWO-RBFNN is faster and overshoot is smaller.

The data of smart watt-hour meter has high-dimensional characteristics, and it is difficult to represent, so it is impossible to ensure high-precision filling of missing data. Therefore, the research on missing data filling of smart watt-hour meter based on variational self-encoder is proposed. Firstly, a filling model based on variational self-encoder is constructed to denoise the measurement data of smart watt-hour meter, then the denoised smart watt-hour meter data is input into the model, and the data missing filling result of smart watt-hour meter is output. Finally, the data missing filling of smart watt-hour meter is realized according to the result. The experimental results show that, after filling, the proposed method is coincident with the original data, and the final filling operation efficiency can reach more than 90%.

The optimization of radar reconnaissance constellation is of great significance for reconnaissance timeliness. In order to solve the problems of insufficient selection pressure and poor diversity of the evolutionary algorithm based on Pareto domination in solving the many-objective optimization problem of radar reconnaissance constellation, an imporved particle swarm optimization based on fuzzy relative entropy (IFREM_PSO) is proposed. A new inertial weight strategy is proposed to enhance the convergence speed and accuracy of the algorithm, improve the external archive maintenance strategy, and enhance the convergence and diversity of the algorithm. The optimization model of radar imaging reconnaissance constellation oriented to area reconnaissance is constructed. The IFREM_PSO is used to optimize the model, and the multi-objective particle swarm optimization algorithm (MOPSO) is used as the control algorithm. Experiments show that the IFREM_PSO has better results in terms of convergence, diversity and convergence speed, and can effectively solve the problem of radar reconnaissance constellation optimization.

In recent years, with the widespread application of information network technology, information security has become increasingly important. More information leaks occur during the exchange of information flows, and opacity is a very important concept in the exchange of system information flows. It refers to the inability of external intruders to observe and infer the nature of key system behaviors through system non secret behaviors. "Opacity" is an inherent attribute of a system. The related research on opacity is widely used in various industrial control fields to meet system security and privacy requirements. This article summarizes and compares various definitions and implementation methods related to the research on opacity in discrete event systems, and analyzes their respective use scenarios and differences based on the characteristics of various types of opacity; The mathematical conditions and complexity of various opacity implementations are summarized, and the complexity of different implementations is compared. The reasons for the different complexity of different implementations are analyzed; Finally, according to the current development direction of opacity research, the future research direction of opacity is prospected.

This topic is aimed at improving the effect of chatter generated in the cutting process on the machining quality of workpieces. By controlling the servo system, the influence of lathe vibration on workpiece quality is reduced, and the cutting process is studied. According to the signal of effective vibration of workpiece, the control device of tool feed is designed to control the cutting elements in the cutting process, so as to effectively improve the machining quality and efficiency.

The current construction quality detection structure of permeable concrete pavement is mostly single processing, which reduces the detection efficiency and easily leads to an increase in the mean difference of compressive strength. Therefore, the design and verification research on the construction quality detection method of permeable concrete pavement of urban landscape engineering based on data mining is proposed. According to the change of actual measurement demand, set basic detection indicators and evaluation standards, conduct compressive strength measurement and permeability coefficient calculation, adopt the form of multi-objective, build a data mining detection structure, improve the detection efficiency, and use Dynatest secondary measurement to achieve quality detection. The final test results show that the mean difference of compressive strength of Hutuo River Ecological Environment Restoration Project 41+500~45+200 permeable concrete sections is well controlled 2.5 through comparative study of the five set detection nodes, indicating that this detection method has strong pertinence, wide detection range, controllable error, and practical application value.

In the modern power system, the safe and stable operation of the power transmission network and the special communication network depends on the healthy connection of its supporting network. In order to study and realize the above functions, a large amount of accurate data is very necessary. In this paper, the generation network G (Generator) and the discriminator network D (Discriminator) are used to continuously approach the real partial discharge voltage signal samples. Based on the WGAN model, the auxiliary classifier Wasserstein with gradient penalty is added, and the ACWGAN-GP model is proposed. Through the comparison between the data samples generated by the model and the real data samples, it can be found that there is a relatively consistent fluctuation rule between the two, but the values of the two remain different at the same time, which increases the diversity of the synthetic partial discharge voltage signal samples, is conducive to enhancing the robustness of the partial discharge detection model, and improves its generalization ability to detect faults. This provides a reference for enhancing the fault data of smart grids.

This paper proposes an improved volume method based on Monte Carlo method to address the issue of low accuracy in calculating single well controlled reserves in shale oil horizontal wells. This method first divides the calculation area into units. Then, the probability distribution function of parameters within each unit is obtained through spatial interpolation method. Finally, the Monte Carlo method and volume method are used to calculate the reserve calculation results for each unit, and the well control reserves in the area are obtained by discretizing and accumulating the calculation results of all units. This paper proposes an interpolation method that integrates XGBoost and spatial semi-variogram to perform spatial interpolation on shale porosity in a certain area. Compare the accuracy of each interpolation method using evaluation indicators such as MAE, MSE, MAPE, etc. After comparison, the spatial XGBoost interpolation method integrating semi variogram has an accuracy improvement of over 15% compared to traditional methods, and an accuracy improvement of over 4% compared to the new machine learning spatial interpolation method. This proves that the method proposed in this paper can effectively improve interpolation accuracy.

Fans are the important equipment in the manufacturing industry, which are easy to cause faults because of the complex working conditions and high speed. Various fault detection algorithms based on electrical, vibration and video have been proposed, but these algorithms are difficult to meet the detection requirements of enterprises. In this paper, a novel fan fault detection algorithm is presented based on sound signal detection technology. According to the different characteristics of the fan sound signal in the time and frequency domains, we combined the convolutional neural networks (CNN, which is sensitive to spatial information) and the gated recurrent unit (GRU, which is sensitive to timing information), and proposed a fan sound fault detection algorithm. Firstly, according to the variation characteristics of the acoustic signal in different states, the spectrogram of sound signal containing spatial and temporal characteristics is extracted, and the histogram equalization is used to enhance the spectrogram. On this basis, CNN and GRU are combined to train the Log-Fbank spectrogram respectively, so that it can better model the different characteristics of time domain and frequency domain information, so as to obtain a fan state model with high confidence and realize the sound fault detection of the fan. Experiments on MIMII dataset show that the algorithm has high fault recognition accuracy and fast convergence speed, which has good application value.

Faults inevitably occur during high-speed train operations and affect the security of the system. In order to improve train reliability, the paper proposes a fault detection (FD) framework for traction systems based on the segmental autoencoder (SAE), and within this framework, the target fault detection work is implemented in combination with a data-driven method base. The main objective of the proposed scheme is to determine the generalized kernel representation based on the knowledge learned from the autoencoder and to complete the construction of a residual generator by means of a special structure to obtain the final FD results. To verify the FD effect of the method on the traction system, the results are verified by a simulation experimental platform to ensure the effectiveness of the method on the target system.

The direction of human eye gaze is an important human behavior information that reflects the level of attention and cognitive state of the gazer towards various visual information in the environment. Eye gaze estimation has wide application value in multiple fields such as medical care, market research, and human-computer interaction. In recent years, some studies have introduced Transformer into the task of eye gaze estimation and achieved advanced performance. Although Transformer has better global modeling ability, its structural characteristics are not suitable for multi-scale feature learning in visual tasks. In addition, the global self-attention calculation for images has high complexity. This paper introduces Swin Transformer into the field of eye gaze estimation, using self-attention mechanism to perform more flexible and effective global modeling of images. The self-attention calculation uses Windows Multi-head Self-Attention(W-MSA) and Shifted Windows Multi-head Self-Attention (SW-MSA), which greatly reduces the calculation of image self-attention. The experimental results demonstrate that the Swin Transformer can obtain good results in the task of eye gaze estimation

Music source separation isolates or separates the desired accompaniment or singing voice from the song recording. This study solves this problem in the time domain using only one channel. Music contains rich information over multiple temporal scales, such as rhythm, tonality, melody, harmony, chord, etc. These different temporal scales information should be represented by features in different granularity. Inspired by this underlying nature of music, a multi-granularity solution is adopted. Furthermore, music is usually self-similar and can be treated as a fractal structure, meaning multiscale pattern repetition exists in music. We have captured these patterns in a batch of different granularity features using the multi-granularity solution. A self-attention mechanism is adopted to link these regular, predictable, and repeated features. Finally, we combine the multi-granularity and self-attention and adopt the Sandglasset model, which can utilize the information in different temporal scales and the relationship of the repeated pattern among different temporal scales. Experimental results on public datasets show that the proposed method outperforms existing systems in the single-channel music source separation task.

Neural network models are commonly used as black-box services, but they are vulnerable to model stealing attacks, where an attacker can train a substitute model with similar performance to the original model by exploiting limited information related to the target model. This can cause significant losses to the owner of the target model and pose a serious security risk. To advance our understanding of neural networks and promote the evolution of model protection mechanisms, we conducted in-depth research on neural network model stealing attacks. In this paper, we propose a black-box stealing attack framework that combines active and semi-supervised learning, even if the target black-box only provides hard-label output, an effective attack can be achieved, generating a substitute model with the same functionality as the black-box. The framework involves selectively querying the most informative samples for black-box labeling using active learning, which significantly reduces the workload of querying the black-box and enables to achieve better performance with fewer training samples. We also apply semi-supervised learning to leverage the abundance of unlabeled data and further improve model performance. We evaluated our method on various data sets and proved that the stealing ability of our method was significantly higher than 3.86%~26.64% other methods when faced with hardlabel black-box with the same number of queries, which can achieve effective black-box function stealing.

For now, cartoonization becomes more and more popular in image processing field. For example, cartoonization can be applied to image processing softwares as a filter. By now, there are some cartoonization Generative Adversarial Networks (GAN) has been proposed. However, most of these models just use RGB image as input. Different from these approaches, our cartoon landscape GAN use HSV and RGB image as input to learn cartoon images’ saturation feature and color feature. To learn cartoon images’ texture feature, we used Gramian matrix as images’ texture feature and use the texture discriminator makes generator to learn cartoon images’ texture feature. By this way, we get a great cartoon landscape generator. Our generated images have high contrast like cartoon images.

In oil drilling, the direction of oil and gas migration and the characteristics of lithological changes can be determined based on changes in formation pressure. Therefore, accurate and effective prediction of formation pressure is a prerequisite for achieving high-quality, efficient, and safe drilling. In response to the problems of high labor cost, low timeliness, and insufficient prediction accuracy in traditional methods for multi well or large area measurement, this paper proposes an oilfield water drive formation pressure prediction model based on Blending integrated model. Four models of random forest, Adaboost, XGBoost and LightGBM are selected as the base model, and the three-layer BP neural network is selected as the meta model. Compare the model in this paper with the base model separately. After comparison, the accuracy of the Blending integrated model proposed in this paper on the water injection wells test set is 94.32%, and the accuracy on the oil production wells test set is 93.87%. This model has the best performance in predicting the formation pressure of oilfield water drive, effectively improving the accuracy of predicting the formation pressure of oilfield water drive.

One of the most significant aims of natural language processing is automatic knowledge extraction. The amount of COVID-19 literature is increasing by ten thousand each month, which greatly complicates manual annotation and downstream activities. In this paper, we describe a system for biomedical multic-label topic classification. Firstly, BERT is pre-trained on PMC and PubMed biomedical corpora which helps to capture deep semantic information. Additionally, we fine-tune the pre-trained BERT using the COVID-19 literature from the LitCovid Database. Finally, we predict the topic of LitCovid scientific literature using the novel model. The experimental results of our model on the BioCreative LitCovid corpus achieves a micro F-score of 91.14%, which is 1.29 percentage points higher than BERT. The F-scores of the our model are 1.33%, 2.32%, 0.27%, 0.44%, 6.91%, 14.14% higher than BERT on Treatment, Diagnosis, Prevention, Mechanism, Transmission, Epidemic Forecasting topics respectively, which demonstrates the potential and effectiveness of the proposed framework.

The log data records the operation information of a software platform. The analysis of the log helps the manager to master the platform operation and ensure the normal operation of the platform. The rapid development of information technology has led to an explosive growth of log size. The traditional log processing architecture cannot meet the processing of massive logs. In order to process massive log data in real-time, a real-time log analysis system based on Elastic stack and Flink is designed and implemented. Based on Elastic stack, massive log data is acquired, transferred and stored in real-time. Based on Flink which is a streaming computing framework, log information is analyzed in realtime. Log analysis and statistical results are stored in MySQL database. Visualization is realized through front-end and back-end separation technology. The system has been deployed on a business platform. It can run stably against the daily average log of more than 10G, which proves the stability and reliability of the system.

As a carrier for data processing and transmission, the traditional data management methods can no longer effectively meet the requirements of high reliability and device performance. In this paper, we analyze the features and advantages of SQLite embedded database and propose the use of SQLite to manage data in embedded systems. The development process is described from various aspects such as creation, porting, compilation and verification of SQLite database, and the method of applying it to embedded system development is realized, which can be implemented into engineering practice.

Common relational graph calculation and display can adopt different calculation methods for various application scenarios to achieve the best display effect. This paper provides a multi scenario display method for the display of daily relationship graph. It can flexibly display the relationship between specified entities according to the user's needs. This technology fundamentally solves the problem of single computing and presentation mode in existing Internet relationship mapping products.

The package recommendation has always been an important issue in the marketing of mobile operators, and machine learning provides a new solution for operators. Aiming at the problem that too many training times of dirty data in the existing methods lead to the reduction of prediction accuracy and the tedious manual setting of model parameters, this paper propose a model combining grid search and XGBoost, and use the exhaustive search method to find the parameter value with the highest accuracy in the validation set within the parameter range of the given XGBoost. Compared with Random Forest and XGBoost default parameter values, experiments show that the proposed model has higher prediction accuracy and can effectively avoids the error caused by manual adjustment of parameters.

Iterative Closest Point (ICP) algorithm is usually used for registration of three-dimensional model point clouds. It is a common and mature registration algorithm. The traditional ICP algorithm has certain mismatching and the point selection condition is relatively simple, we introduce an extra RANSAC mismatching removal step into the ICP algorithm. It takes extra consideration of the spatial geometry information of the point pair selection. It improves the accuracy of the algorithm while speeding up the convergence of registration. In addition, we discuss the influence of the number of points and the similarity threshold of the algorithm under the Stanford standard point cloud data set. Finally, Gaussian curvature is introduced to improve the accuracy of the algorithm and reduce the randomness and the time of the algorithm.

Table-to-text generation is an important area of text generation, and the process of structured table-to-text generation faces problems such as the "illusion" of over-understanding table data, and the problem of selecting and ordering the content of the generated text from table data. In this paper, we present some important model mechanisms in the history of its development and a new outlook on the future development of table-to-text generation and possible technical routes. A two-stage encoder-decoder approach is analyzed as to why it is superior to its predecessors, and a new outlook is proposed for implementing table-to-text generation tasks based on graph convolutional neural networks.

Bibliometric analysis was applied to retrieve the international core journals on the subject of text generation for an explainable recommendation from 2000 to 2022. CiteSpace, a visualization-based analysis tool, was used to analyze the research status and recent development of text generation for explainable recommendations by mapping the co-occurrence of high-frequency keywords and citation bursts. The results show that in the past 22 years, the number of international articles on text generation for an explainable recommendation has been on the rise with more publications especially compared with that in China, calling for further exchange and collaboration among scholars and institutions over the world to facilitate the research progress. “Personalized” recommendation has been exerting its influence in text generation for explainable recommendations, which effectively gain the trust of users, and increase the persuasiveness and satisfaction of the recommendation system by providing recommendations with explainable texts. Text processing in Deep Learning has now been widely used for explainable recommendations and will throw its weight further in the future.

The research is based on the perspective of credible donation data for fund-raising activity, and taking a small fundraising system on campus as an example. This study is devoted to completing Agile development of the system with the technology of Scrum model and architecture of Blockchain, which describes the feasible process of design and implementation for the first prototype version. Feedback on fund-raising activities for college students on experimental campus is positively. Application result demonstrates that through the smart contract of Blockchain, it ensures the trust problem preferable of this fund-raising system.

Compared with the traditional knowledge graph-enhanced recommendation method, this paper introduces a multi-task learning module to alternately train knowledge graphs and recommendations to alleviate the data sparsity and cold start problems in traditional recommendation methods. Specifically, in the multi-task learning module, the item features and contextual content features are taken, and the features after feature interaction are obtained using the interactive attention network, as a way to learn finer-grained features, and then the gating mechanism processes the item features and entity features that fuse the contextual content, which can filter the unimportant features and obtain the important potential features, and can capture the implicit higher-order feature interaction more effectively. Optimized for multi-task learning tasks. The validity of our model was verified on three publicly available datasets.

The rapid development of information technology has advanced the progress of artistic creation. Interactive music visualization has been an important research issue in recent year, which can enhance the feeling and listening experience of the audiences without any musical knowledge. This work aims at implementing a real-time interactive music visualization approach based on Processing platform. Processing is an open source programming language and environment made for various purposes from scientific data visualization to art installations. This paper explains how to extract the music features like pitch, melody, genre to form specific visual mapping elements, such as shapes, colors and motions. Then, we have implemented the visualization by analyzing the time and frequency information of the digital music with the Processing sound libraries. The results indicate that the approach is generally successful and enriches listening experience of the audience.

With the vigorous development of the social Internet, users have paid extensive attention to recommendation technology. In the face of the cold start problem of the existing introduction methods in the Internet and the situation of the Internet information content where users are not fully considered, the personalized recommendation formula using graph entropy in the trust Internet was generated. First of all, the user's item graph is created according to the user's item and the content of information transferred between them, and the user trust graph is created by importing the trust mechanism; Then, the initial similarity between the new user and the product is obtained by using the random walk algorithm for the two images, or the new user product similarity is obtained based on the trust relationship; Then repeat the random walk process until the convergence value of similarity is stable; Then, the two groups of similarity rates can be weighted by the graph entropy on UIG and UTG, and the final selection list of target customers can be obtained accordingly. However, according to the actual sample set of Epings and Film Trust, the accuracy of RWAGE has increased by 34.7% and 19.4% respectively compared with the typical random walk algorithm, and the recall success rate has also increased by 28.9% and 21.1% respectively, which can significantly reduce the recommended cold start time and is higher than the comparison in accuracy and coverage.

With the gradual maturity of statistical machine learning, it is now time to break the traditional model of isolated learning and turn to lifelong learning, pushing machine learning to a new level. Applications such as intelligent assistants, chat robots, and physical robots are also in urgent need of this lifelong learning ability. Without the ability to accumulate knowledge and use it to gradually learn more, a system may never be truly intelligent. Continuous learning algorithms attempt to achieve the same capabilities for neural networks and solve catastrophic forgetting problems. Therefore, in essence, continuous learning performs incremental learning for new tasks.

To improve the accuracy of mixed text classification, CBAM attention mechanism is introduced in the backbone network of YOLOv7 to highlight the key features of dominant category determination. At the same time, aiming at the information loss of PANet route aggregation network, the ASFF adaptive feature fusion mechanism is embedded after it, and the feature map is adaptively weighted by pixels to better preserve the details of the feature map. Experiments show that the Macro-F1 value is improved by 1.9% compared with PSENet + SVM method which detects and classifies first.

Cardiovascular disease, especially coronary artery disease, is always a threat to human health. Myocardial infarction is a form of coronary artery disease. Cardiologists frequently use electrocardiogram (ECG) to diagnose this condition and ensure the health of their patients. Therefore, studying ECG signal classification can aid doctors in accurately identifying the disease and providing appropriate treatment. We develop structure to extract feature from ECG signals, which achieves excellent performance in classification tasks. A single lead ECG signal typically consists of P, QRS, T, and U waves, which collectively form an ECG signal beat. We utilize R-peak detection technology to obtain ECG signal beats, and extract beat features using a residual network. To avoid the loss of global information, we employ a simple onedimensional convolutional neural network (CNN) to obtain global signal features. The fully connected layer is then used to fuse the features obtained from both beats and global signal features. The classification task is completed based on the fused features. Our designed structure improves performance metrics by at least 2% when compared to the performance of a one-dimensional convolutional neural network and a residual network individually. Additionally, we also introduce the SE block into the residual network, which provides an attention mechanism to effectively suppress unnecessary features and enhance important ones. By comparing the performance of our structure with and without SE block, we prove that SE block can enhance our structure's ability to extract ECG signal characteristics.

In recent years, with the progress of neural networks, people have gradually shifted their research focus from words to images, precisely because of the emergence of convolutional neural networks, which can convert the information of images into an easy-to-understand form and make it easier for people to solve problems in the image field. Image matching is the most fundamental research direction in the field of images. With the development of technology, the algorithms applied to image matching have gradually changed from traditional sift and orb algorithms to convolutional neural network-based algorithms, such as superPoint, superGlue, D2net and other models. Although these models have certain advantages over the traditional algorithms, there are still problems that the high-level features of the image are not closely related to the underlying features. To address these problems, this paper will use the traditional algorithm to fuse the features extracted by the convolutional neural network model to improve the matching accuracy, and combine the attention mechanism in the convolutional neural network to make better extraction of the easily ignored areas in the image. The proposed method is experimented on the Hpatches dataset and Brown dataset and achieved good results with an accuracy of 91.23% and 92.36%, respectively.

To solve the problem of failing to effectively match bug report in query reduction, we propose a query reduction method based on multiple types of information, MIQR. The method divides the information in the description of bug report into three different types: stack trace, program element, and natural language. Different term weighting algorithms are used to calculate term weights for different information, and keywords are extracted based on term weights. A new query is constructed based on the summary and keywords. We conducted experiments on 5,139 bug reports. The results show that our method is superior to the state-of-the-art method.

The basic idea of federated learning is to aggregate the local model parameters of all participants to obtain a global model, which is more universal and has better performance. However, existing federated learning has the problem of a unique model, where all participants train their local models based on the same global model. However, in different real-world scenarios, data often has different features and distributions, and there is no universal model suitable for learning data from multiple scenarios. Therefore, existing federated learning technologies cannot achieve optimal performance in multiple scenarios. To solve this problem, this paper proposes a decentralized federated learning model where each participant independently selects other models similar to their local models and aggregates them to obtain a specialized model applicable to the local scenario. The model aggregates the parameters based on their similarity with the local model. Models with higher similarity will have a greater weight in the aggregation process. The parameter aggregation is not performed at the central server but at each participant's local device. Through this mechanism, the proposed model achieves specialization of local models for participants and improves performance in different scenarios.

This article discusses a privacy preservation for federated learning based on gaussian noise scrambling. The main objective of this model is to protect user data privacy and security by aggregating model parameters from multiple parties instead of raw data sets. However, attackers may still obtain sensitive information from the model parameter information transmitted during federated learning training through certain means. To address this issue, we propose a differential privacy noise addition scheme for federated learning. By adding noise to transmitted model parameters to some extent, it prevents attackers from inferring participant information through reverse inference. We also study Gaussian noise mechanism in differential privacy protection and prove it by combining characteristics of federated learning and blockchain technology for adding Gaussian noise size according to differential privacy requirements. This article tests their proposed differential privacy protection mechanism using the NSL-KDD dataset and finds that after adding differential privacy protection, models still have good anomaly recognition performance. This study contributes to the field of federated learning by proposing a novel approach for protecting user data privacy and security during training. The proposed approach can be used in various applications such as healthcare, finance, and social media where data privacy is crucial.

The traffic scenes are complex and ever-changing, and traditional signal control methods have limitations. DRL based methods rely on a large amount of interactive data between intelligent agents and the environment, which leads to a long learning process and high computational costs. Based on this, this paper introduces imitative learning into reinforcement learning for traffic signal control. At the same time, it improves the design of reward function for the problem of frequent phase switching and too long red-light in the flow imbalance scene, and increases the cost of phase switching and penalty factors. Experiments verify the effectiveness of the algorithm in this paper.

Driverless vehicle data is an important data for remote dispatching platform to manage and maintain vehicles. Aiming at problems such as complex data, information islands and management difficulties of driverless vehicles, using the dataoriented architecture (DOA) idea, construct an automatic registration model based on the data registration center (DRC), to realize automatic registration and synchronous update of vehicle data; Analyze the basic attributes and common elements of driverless vehicle data, and on the basis of existing research, propose a registration standard for driverless vehicle data. The experimental results show that the automatic registration model can realize the automatic collection and registration of vehicle data and write it into DRC, and the remote dispatching platform can better manage and maintain remote driverless vehicles, improving dispatching efficiency.

The classical Ant Colony Optimization (ACO) has some problems such as slow convergence speed and easy to fall into local optimum, a control method for dense formation of UAVs is proposed in this thesis. Firstly, cubic mapping is used to initialize the ACO distribution for making full use of map information and avoiding falling into local optimum; Secondly, the pheromone concentration updated by reward and punishment mechanism. Adaptive pheromone volatility factor is used to balance the global search and local search, as well as accelerate the convergence of the algorithm; Then, the artificial potential field is combined with the ACO to optimize the obstacle avoidance path. Finally, the state information from the dense formation of UAVs is corrected according to the consistency control theory, and the formation consistency control model is designed to realize the dense formation control. The experimental results shows that the proposed algorithm can achieve obstacle avoidance while maintaining tight formation, the formation can converge to the desired formation quickly after the obstacle avoidance is completed.

The separation of different components in a binary mixture of CF₄ and N₂ is a challenging task. The adsorption of 603 experimental (CoRE) organic frameworks on CF₄ and N₂ binary mixture gases at room temperature and 10 different pressures was simulated using GCMC in this study. The performance of adsorption separation is described by the product of the adsorption amount of CF₄ and the logarithm of the CF₄ adsorption selectivity. Among the nine features extracted are pressure, structural features of adsorbent, and two easily accessible user-defined descriptors (AVG_SIG and AVG_SQRT_EPS). The importance of AVG_SIG and AVG_SQRT_EPS was assessed using random forest and found to play an important role in predicting adsorption separation performance. By using Harris Hawks Optimization (HHO) to explore the potential of different models, it was found that XGBoost and GangNeuron performed excellently in predicting the adsorption and separation performance of CF₄ and N₂, with prediction accuracy of 99.2% and 98.8%, respectively.

Aiming at the problem of small contrast difference between the defective region and the background region in PCB images, This Article presents an enhanced YOLOv5 algorithm with a multi-scale detection approach. To enhance the YOLOv5 algorithm, this Article upgrades its backbone network by replacing Characteristics Abstraction network with the ASPP module. This change aims to improve the network's perceptual field and Characteristics extraction capability. Secondly, in order to improve the attention of the model to other regions, this Article introduces the attention mechanism Coordinate Attention module, which embeds the location information into the channel attention and achieves multi-scale processing and Characteristics fusion. Finally, this Article uses different sizes of anchor frames for multiscale detection of defective targets. The experimental results show that the size of the improved multiscale network model is only 83% of the size of the original YOLOv5 model, and the mAP on the dataset reaches 97.2%. The algorithm proposed in this Article can effectively detect various defects in PCB images and has high detection precision and low false detection rate, which has good practical value and prospect of popularization and application.

To solve the problem of uneven energy consumption of routing protocols for wireless sensor networks (WSN), a WSN multi-strategy clustering mechanism based on improved sparrow search algorithm is proposed, which optimizes the intra-cluster task using a dual cluster head clustering strategy. Firstly, a node hierarchy strategy based on location information is designed; secondly, a multinomial constraint factor is designed to select the primary cluster head; finally, the improved sparrow search algorithm (SSA) is applied to select the sub-cluster head within the cluster. The clustering mechanism is compared with three similar clustering protocols for experiments, and the average improvement over the other three algorithms is 121.7%, 77.1%, and 12.3%, respectively. The results show that the clustering mechanism can reduce node energy consumption and extend the network survival cycle.

With the reduction of the cost of high precision 3D sensors and spoofing attacks, face presentation attack has become the main security risk of face recognition. In order to deal with presentation attacks and ensure the effective use of face recognition, presentation attack detection has gradually become a popular research topic. As advanced face presentation attacks increases deception by adding active cues, the research on face anti-spoofing ignores the local regions of the face. In this paper we propose a FPACNet to fully exploit features of face part. Extensive experiment shows that our model has achieved extraordinary performance

Since the outbreak of the new coronavirus in China in 2019, wearing masks has gained widespread attention to prevent virus transmission. However, traditional face detection models have struggled to accurately detect faces covered by masks, posing a challenge for public health and security applications. In this study, we propose a novel lightweight model for face-mask detection, called YOLO-ARGhost, which is based on YOLOv4 and incorporates an attention mechanism to enhance accuracy. Our model is designed for fast face-mask detection, overcoming the limitations of previous models. Experimental evaluations on the AIZOO dataset demonstrate that our approach achieves an impressive mean average precision (mAP) of 92.8%.

News text classification is an important subtask in natural language processing. News text is not only unstructured text, but also its structure information is relatively fuzzy. Therefore, there is less useful information that can extract directly. Previous studies often assumed that the text is a linear structure by default. In this paper, we propose a text classification model based on GCN-BiLSTM and adding a Global Average Pooling layer. The hybrid neural network can capture global text information, and solve the problems of gradient disappearance. The Global Average Pooling layer can reduce the data dimension and shortens the training time. The method has a good performance on THUCNews dataset, and the F-score reaches 89.75%.

In brain-computer interface (BCI) systems, acquiring EEG signals is relatively easy, but labeling EEG signals is challenging. Supervised learning is widely used in existing BCI research, but its generalization ability is poor due to the influence of labeled data, which is not conducive to the promotion and application of BCI. In the fields of computer vision and natural language processing, self-supervised learning is becoming a popular research direction that can use unlabeled data for learning. In this study, we propose a self-supervised contrastive learning method for identifying motor imagery states by performing different data transformation operations on EEG data to generate positive and negative sample pairs. This allows the pre-trained network to learn general feature representations from the signals, enabling it to converge faster and achieve better classification performance when performing downstream tasks. The network's performance is significantly related to the amount of unlabeled data used in the self-supervised learning process. Through the evaluation of the collected data, we achieved an accuracy of 82.56% in classifying motor imagery states with 40% amount of labeled data, demonstrating the practical value of our method.

Online social network platforms contain massive amounts of information propagated in cascading form. It is important to predict the future growth of the information cascade, especially in early stages, to control the information propagation effectively and promptly. The existing studies don't explicitly leverage temporal information, and only use network structure and sequence feature. However, the network structure of different cascade subgraphs varies insignificantly, due to the short cascade sequences in the early stages of information propagation. Therefore, temporal feature is more important for the prediction task. CasTemporalGCN utilizes the position encoding function to encode the temporal information. Our model captures the cascade subgraphs structure information by GCN (Graph Convolutional Network), and feeds it into multi-layer self-attention encoders instead of using classical RNN (Recurrent Neural Network) model. The temporal feature is added to encoders by residual connections. Experimental results show that using temporal information significantly reduces the prediction loss by 13.8% on real Weibo dataset and 24.7% on ACM dataset respectively than state-of-the-art methods. In the ablation study, temporal information reduces the prediction loss by 19.0% for the encode structure and 10.1% for the RNN structure.

With the development of mobile online social networks, more and more attention has been paid to the task of recommending points of interest. Existing studies have proposed using user behavior sequence information and point-ofinterest information to make recommendations, but they have not fully explored the complex interaction relationships of users. Users' check-in behavior is uncertain, which increases the complexity of recommendation work. In this paper, a point of interest recommendation algorithm model based on heterogeneous graph neural network and GRUs is designed. First, this paper constructs a heterogeneous graph that includes user social relations, user check-ins, co-occurrence relations of POIs, and proximity relations between locations. The RGCN model is then applied to the constructed heterogeneous graph to obtain user and POI embeddings. We take user and POI embeddings as input to GRU. We predict where the user is likely to arrive next based on the existing check-in location. Then, based on the location score, we recommend k locations that the user might be interested in. This paper conducts experimental comparison and analysis on a real dataset Gowalla. Experiments show that this method has a better effect than the baseline model.

Deep neural networks (DNNs) had been proved to be very vulnerable to adversarial examples which generated by adding small perturbation to the input, adversarial examples will mislead DNNs to produce wrong classification results. The quality of adversarial examples should be evaluated from two aspects: attack success rate and visual imperceptibility. The traditional assumption of visual imperceptibility is that perturbation should be tight L2 norm bounds in RGB color space. In the paper, we consider the color perception factors that are more suitable for human eyes on the basis of traditional assumptions. More specifically, we expand the GANs and add color perception factor as a loss function to the training generation model, which made the perturbation more imperceptible. After that, we propose an optimization algorithm of adversarial examples based on Grad-CAM, which is used to optimize the range of perturbation while ensuring the aggressiveness of adversarial examples. Our approach shows that on CIFAR-10 and NLPS 2017 Adversarial Learning Challenge datasets, the adversarial examples created are superior to traditional attack approach in terms of L2 norm and color difference while ensuring attack success rate. The resulting adversarial examples are more natural and usually indistinguishable from the original images.

With the continuous development of the national economy, people's demand for the security of financial institutionsiscontinually increasing. Compared with human monitoring, with the help of human behavior recognition technology, collecting data from sensors to identify the behavior of criminals has the advantages of fast data processing and lowcost. This paper takes the human behavior recognition project based on RFID and deep learning as the research object, andstudies how to effectively extract behavior features and judge human actions according to the value of features. Byoptimizing the convolutional neural network, the project has both good feature construction ability and memory abilitytoextract action marks. Hopefully, all-weather, fast and accurate response to illegal acts in self-service banks andother places can be provided, and the informatization of financial institutions can be promoted.

Under the tide of economic globalization, the number of trade in financial institutions has increased rapidly, and the demand for supervision of possible criminal acts in financial institutions has consequently increased. Therefore, using computer technology to process monitoring data has become a hot topic of current research. Therefore, this paper designs and implements a human behavior identification information management system framework based on RFID technology, and provides an overall solution from RFID hardware to software to identify and alarm possible abnormal behavior in financial institutions. A replaceable deep learning algorithm is used to identify and alarm various abnormal behaviors. Through real-time Web technology, police officers can view the relevant information of criminals in real time on the web page.

The study of bug report-oriented program error location has the characteristics of strong pertinence and low cost, which is an important direction in the current research on program error location. This type of research takes bug reports and source code as input sources, and establishes a mapping relationship between the two through semantic mapping strategies to locate program errors. In the fine-grained program error location scenario, there is a problem that the location accuracy is greatly reduced. Existing empirical studies analyze the difference in location accuracy from two aspects: input source data noise and semantic mapping strategy selection, but most studies take the established location tools and methods as the evaluation object, the evaluation data type is single, and there is a lack of fine-grained analysis of constructing key variables. In order to evaluate the influence of key variables of location method on location accuracy, this paper decouples the location method through pseudo-siamese network, measures the sensitivity of location accuracy by counting the gain of location accuracy under different input source data types, and adds input source data types and a variety of semantic mapping strategies, Based on the evaluation of 23808 bug reports and corresponding source code data in 7 open source projects published on JIRA, this paper provides a more detailed empirical basis for additional data type selection and weight allocation, combined learning of multiple data types and different semantic mapping strategies in fine-grained program error location.

In modern communication systems, channel coding can be affected by noise and burst errors, reducing the error correction capability of the received decoding, and this problem is generally solved by interleaving [8]. In this paper, based on the requirements of a topic, an algorithm based on pseudo-random rank-and-file cyclic shift interleaving and deinterleaving that can satisfy multiple rates is proposed based on the commonly used packet interleaving and cyclic shift interleaving, and the algorithm is implemented on FPGA (Field Programmable Gate Array), and the correctness of the design and implementation is proved by comparing the results with MATLAB calculations

when the software implements the design of infinite impulse response filter (IIR filter), the larger the amount of filtering data and the higher the filtering order, the more time will be spent because of the data shift of each calculation.^[1] By means of parameter replication, the IIR filtering software replicates the parameters once when initializing the software. In the calculation of IIR filtering, the array subscript changes in a cycle ^[2], replacing the way of data shift for filtering calculation, thus greatly improving the execution efficiency of the software and saving the operation time. The higher the order of IIR filter, the higher the efficiency of software promotion and the more time saving.

The goal of entity relation extraction is to extract entities and the relations between entities from unstructured texts. Most of the existing research approaches are oriented towards common entity labels in general-purpose domains (e.g., time, place, person, institution, etc.) and simple texts in specialized domains (texts consisting of single sentences with low knowledge density). The existence of long-distance dependencies of entity pairs (cross-sentence entity pairs) and the phenomenon of overlapping relations (different relations sharing the same entity) in complex texts are ignored. However, complex texts are common in practical applications, especially in professional fields such as power technology standards, where the knowledge density of texts is high and the phenomenon of entity-pair cross-sentence dependency is significant. To solve the above problems, this paper proposes a novel multi-hop automatic question-and-answer-based entity relation extraction method, which combines the current well-established machine reading comprehension framework with the automatic question construction mechanism proposed in this paper, and uses the a priori knowledge provided by the question as the extraction type guide, and uses the multi-hop question-and-answer mechanism to reason about the answer span of the question, effectively alleviating the phenomena of overlapping relations and entity dependence on crosssentences in complex texts. We conducted extensive comparative experiments on the power technology standard dataset self-constructed in this paper, and the results show that the MT-auQA model proposed in this paper achieves optimal performance

Most of the existing hand data collection procedures rely on complex and bulky image acquisition systems, in fact, the data directly acquired by the acquisition systems do not have a realistic context, and the neural network models trained from these data perform not well in real-world situations, especially when people are wearing gloves in scenes with changing illumination. However, to improve the consistency of the image the traditional image fusion methods sacrifice the authenticity of the background and the supervised methods are limited by the data acquisition system which cannot generate high quality fusion image. Therefore, based on an unsupervised generative adversarial network, this paper introduced a generator based on a global attention module and proposed a histogram similarity-based image selection module to filter the input images. Our goal is to decrease the difficulty of the image migration task for the network. The model had been trained using a self-built dataset consisting of composited and real data, including images of hands wearing IMU data gloves and photographs taken in real life with similar backgrounds. Extensive experiments on image segmentation tasks demonstrated the effectiveness of the proposed model, which obtained a precision of 0.88 and a recall of 0.86 with a mean NIMA value of 4.27 compared to other state-of-the-art methods such as UEGAN and DoveNet.

The circuit drawing software is the basis of the current electronic design system and is an important industrial software technology. At present, the more mature circuit drawing software is mostly foreign commercial closed-source software. These software usually have problems such as complicated operation, high demand for computer configuration, and difficulty for beginners. Based on Qt, this paper designs and implements a cross-platform circuit drawing software. With reference to the more mature related software on the market, the interface is concise and beautifully arranged. Based on the MVC framework, the view and the back-end model are separated, and the human-computer interaction function is realized by responding to related operations through the signal and slot mechanism.

In order to solve the problems of disordered code format and inconsistent style in software development, this paper designs and implements software that can format multiple programming languages based on the overview of code format methods. This software can format C/C++/C #/Java/Python, which is a popular programming language nowadays. It also provides a very friendly user interface, which is easy to use and suitable for team development. After testing, the software can accurately format the code and clearly display the format results. It achieves the goal of unifying software style and improving software quality.

Aiming at the problem of high-speed online spectrum analysis in engineering practice, the Fast Fourier Transform (FFT) algorithm software design was studied. Starting from the spectrum analysis mechanism, the basic concepts, decomposition algorithms and coupling methods of FFT algorithm are analyzed step by step. The application of the cycle and symmetry of rotation factors in FFT algorithm to reduce algorithm complexity and improve system performance is discussed. From the perspective of software design, the flowchart of FFT algorithm coupled layer by layer to probability density is presented. From the perspective of engineering practice, the IPC mechanism of the dual core chip DSP28379d is applied, and a scheduling mechanism for the FFT algorithm in the dual core system is designed, further reducing the runtime consumption of the FFT algorithm and weakening the impact on the main process of the system.

With the development of technology, UAV formation gradually appears in people's view. During the formation process, the UAV can obtain height information through its own perception, thus, the UAV formation problem is downscaled and the UAV is localised. In this paper, the position coordinates of the UAV are obtained through the Triposition positioning algorithm, using geometric relations, after which the UAV is distributed on a circle by adjusting the edges, and later the UAV is adjusted to the ideal position by adjusting the angle. The experimental results ultimately show that the algorithm is robust under large angular errors and can be better combined to become a signal localisation system in larger UAV formations, thus maintaining the operational capability and stability of communication transmissions in large UAV formations.

To enhance the quality of paper-based photographic images, this paper proposes an image optimization method based on Criminisi and CLAHE techniques. To eliminate the impact of plastic film and other background factors on the restoration of paper-based photographs during contrast enhancement, geometric transformations such as cropping and flipping are first applied to the image. The preprocessed image is then subjected to channel separation, and the CLAHE method is used to enhance the contrast of the R, G, and B channels separately. After channel merging, an enhanced image is obtained. The Scharr operator is used to accurately obtain the boundary of the image to be restored, and the Criminisi algorithm is employed to implement inpainting of the selected area. The experiments show that the method proposed in this paper can achieve optimization of paper-based photographic images to a certain extent, and can be further applied to inpainting tasks for old photographs and early videos.

This paper proposes a handwriting removal method based on Convolutional Neural Network (CNN). In image preprocessing, to avoid the influence of lighting and color on paper document images, illumination equalization is realized based on Contrast Limited Adaptive Histogram Equalization (CLAHE). And then the OTSU threshold segmentation method is used for adaptive threshold segmentation to obtain a binary image. In terms of datasets production, word region segmentation is realized based on edge detection, morphological processing, and contour fitting to produce test set and training set images. which include 2800 handwriting and printing. Then it is applied to the training and testing of the CNN model, the classification accuracy reached 98.25 %. Finally, the handwriting region in paper document image is fitted and removed. The experimental results show that the algorithm can better remove the handwriting content in paper document image, it is helpful to standardize the processing of paper documents.

Because the deformation of the material under load is strongly correlated with the loading path, the Temporal Convolutional Network (TCN), which can satisfy the time-series learning, is chosen to model the elastic-plastic intrinsic relationship of the material. The ABAQUS software is used to write scripts to construct the data set required for training. After training and learning the stress-strain states and their increments for different loading states, the constructed network model can well characterize the yielding process and the subsequent yielding evolution of the material. Compared with the traditional phenomenological model, it has better flexibility and computational efficiency, and provides a new method for constructing the constitutive model of materials.

Crowdsourcing is a very popular model on the Internet. It is a common task assignment method to assign multiple tasks according to workers' subjective preferences in crowdsourcing. However, workers may exchange for higher utility by providing false information. To this end, this paper proposes a truthful incentive mechanism for multi-task assignment in crowdsourcing. By introducing an incentive mechanism and designing the allocation algorithm and payment algorithm separately, workers will adopt the behavior that the platform expects to see when participating in task assignment. Experiments show that the task assignment mechanism is truthful, and satisfies individual rationality and incentive compatibility. Moreover, the mechanism can obtain results in linear time and has good computing performance.

The context of the COVID-19 pandemic has made an increasing number of people accustomed to wearing masks to avoid getting infected. This phenomenon leads to an urgent demand for face recognition with masks, which poses serious challenge for the accuracy of face recognition. This paper seeks to produce a model called AFF-Net based on an attention mechanism and feature fusion and self-attention to address the masked face recognition by designing four compact modules to achieve the optimization of the network structure to solve the problem of unstable recognition rate and the low model generalization ability. Specifically, we fuse the features of different layers in the improved full connected module to assign weights to the channels, and formulate a specific training strategy to improve the feature learning performance. Experiments on representative datasets indicate that the proposed method can outperform the competing method significantly while maintaining high real-time efficiency of masked face recognition.

In this paper, we propose an Electrocardiogram (ECG) classification model based on FFC (Fast Fourier Convolution) and ResNet. The model utilizes FFC and ResNet for feature extraction and classification. We further improve the network performance and convergence speed through batch normalization and residual concatenation. The experimental results demonstrate that the model exhibits excellent classification performance under different data distributions in the PTB-XL database and trains faster than traditional ResNet models. Additionally, we introduce a new module, FFC-R, and validate its excellent performance in ECG classification tasks. This innovation is expected to provide powerful support for the diagnosis and treatment of heart diseases.

Aiming at the problem of low accuracy in detecting small defects in insulator aerial images using the current YOLOv5s target detection network, this paper designs an insulator defect detection algorithm based on improved YOLOv5s. Firstly, in the feature extraction stage, a PST attention module is embedded to suppress redundant interfering feature information; The guidance network grasps effective features for learning during the training process. Secondly, a feature fusion module with a bilateral structure is used in the feature fusion process, making full use of contextual semantic information to improve the accuracy of small-scale target detection. The experimental results show that the average accuracy of the improved YOLOv5 algorithm is 81.82%, which is 4.71 percentage points higher than YOLOv5s algorithm and 1.2 percentage points higher than YOLOX-S. This indicates that the improved YOLOv5s network model used in this paper has good results in the detection of insulator defects and has certain application value in actual power inspection tasks.

EAST's security monitoring platform is based on the MVC monolithic application architecture, providing experimentalists with the ability to visualise the device operating parameters. With the accumulation of large amount of data and the increase of analysis requirements, the scale of the system business is expanding. The problems of strong coupling, overlap and limited scalability of the various business components have gradually emerged. This paper aims to refactor the security monitoring platform based on microservice architecture. Firstly, we build a microservice framework, divide subsystems according to user requirements and build submodules according to business links. After that, the system hierarchy is optimized: the data source proxy is encapsulated, and the data platform is built as a layer. Finally, the system was implemented and tested with Kafka and Flink as the core real-time computing modules, whose output latency is within seconds. In fact, the refactored system is scalable and stable to support the experimental monitoring of EAST.

Knowledge-based question answering(KBQA) involves using knowledge base technology to generate answers to natural language processing(NLP) questions. KBQA is one of the most challenging tasks in the field of NLP. Answer selection plays a vital role in KBQA, as it requires selecting the correct answer from a pool of candidate answers. However, knowledge bases still struggle to identify the correct answer from multiple candidate answers. To address this issue, In this paper, we propose BAMGAN, a KBQA method based on generative adversarial networks. This method uses generative adversarial networks (GANs) to improve answer selection accuracy in deep neural networks. Experimental results show the effectiveness of the BAMGAN.

Background: At present, artificial intelligence has been widely used in fields such as facial recognition. Since facial expression is the most direct feature of emotion, it can be used to predict the emotional state of police dogs by recognizing facial expression. In our study, we developed and experimentally analyzed an improved method to predict the emotions of police dogs through deep learning to recognize their expressions. Methods: We sifted through and analyzed public dog data sets on the Internet. Finally, it was decided to use the dog data set published by Stanford University, and the categories of dogs were accurately divided and marked manually. Emotion recognition and experimental analysis of police dogs were carried out by using group recombination residual network. Results: Through experimental analysis and comparison with existing methods, the accuracy of this method is 6% higher than that of the method based on residual network, and the accuracy is up to 89%. In addition, the number and size of the model parameters are reduced. Conclusion: Through experimental analysis and comparison with other mainstream identification networks, this method has a good advantage in the number of network parameters. Compared with other methods, the residual network based on group recombination can significantly improve the accuracy of emotion recognition for police dogs.

UAV navigation spoofing technology can make UAV fly according to the preset spoofing trajectory by sending navigation spoofing signals. After analysis, this paper found that the current navigation spoofing research actual transformation is not strong: one is because the current strategy is too ideal and the control is not fine enough, did not consider the UAV navigation and flight control system as the main body of the integrated navigation spoofing control model, the second is the lack of a test environment to verify the effectiveness of spoofing. Therefore, we establish a navigation spoofing control model based on flight control according to the key technologies of UAV navigation spoofing, and on this basis add the control variables of spoofing relative position to study the feasible spoofing strategy. Finally, the new information detection and hardware-in-the-loop simulation are carried out to verify the spoofing effectiveness of this strategy, which lays a good foundation for navigation spoofing of real aircraft.

For the requirements of Unmanned Aerial Vehicles (UAVs) in high payload and small space, the aircraft needs to have the characteristics of high payload, small size and stable control system. This paper designs a coaxial twin-propeller multi-rotor UAV based on the combined control of redundant flight controller and NVIDIA microcomputer. Compared with traditional quadcopter UAVs, each axis is equipped with dual motors and dual propellers in this structure, which not only has high load-bearing capacity but also greatly reduces the occupied space of the aircraft. Through actual tests, it is found that the failure rate is almost 0 with the support of redundant flight controller, and the combined control system can hover stably and locate accurately under harsh conditions. Compared with the traditional flight control system, the system has better maneuverability and lower error offset when the UAV is under high payloads, making it more suitable for the aircraft with coaxial double-propeller multi-rotor structure.

Reinforcement learning has great potential in solving complex decision-making problems in autonomous driving, especially in traffic scenarios where autonomous and human driving vehicles are mixed. This article describes the lane changing problem of mixed traffic expressway vehicles as a multi-agent reinforcement learning problem, in which autonomous vehicles learn a strategy adapted to human driving vehicles through collaboration to maximize traffic throughput. This article extends the SAC-Discrete algorithm to the multi-agent reinforcement learning framework and proposes the MASAC-Discrete algorithm. In addition, this article also proposes a motion prediction safety controller that includes a motion predictor and motion replacement module to ensure driving safety during training and testing. This article trains, evaluates, and tests the proposed method on a highway simulator under three different levels of traffic modes. The simulation results show that even in high traffic density situations, this method can significantly reduce collision rates while maintaining high efficiency. In the considered highway scenarios, its performance is superior to several state-of-theart benchmark algorithms.

In view of the current industrial programmable control equipment configuration programming platform poor autonomous ability, foreign related technology blockade and other neck problems, we take the card size raspberry pie as the embedded IOT (Internet of Things) controller, and deploy the raspberry pie to realize the whole process of control system evolution, and realize the full coverage of all aspects of automatic control online experiment. This research project is an IOT controller that can realize the whole process of control system evolution, namely M2PLab (Mathematics to practice laboratory).

In recent years, target recognition has been applied more and more in the military field, which plays a vital role in military information intelligence, so the research on target recognition of unexploded grenade has great practical significance and military value. Based on deep learning, this paper enhances the data by rotating, clipping, adjusting brightness, etc., increases the sample size and diversity of the dataset, and then verifies through experiments, and the results show that the improved method can effectively identify unexploded grenade targets in real time.

In the daily maintenance and fault diagnosis of high-speed rail on-board equipment, a large amount of equipment fault data analysis has been accumulated, but the information of these fault analysis is too scattered to be effectively used by staff. In this paper, the construction of knowledge map in the field of high-speed rail vehicle equipment fault is preliminarily explored, and the knowledge map of high-speed rail vehicle equipment fault is constructed by combining top-down and bottom-up methods. In the process of entity extraction, this paper uses BERT+BiLSTM-CRF entity recognition model to extract entities. Through comparative analysis, BERT+BiLSTM-CRF model has a good effect of entity identification, and can identify and extract fault entities. Finally, a certain scale of vehicle-mounted equipment fault knowledge map is constructed according to the fault data of high-speed rail vehicle, which can effectively realize the structured storage and visual display of fault data and facilitate the guidance of vehicle-mounted equipment fault maintenance.

This paper takes ChatGPT users as the research object and analyzes the factors affecting ChatGPT users' willingness to use by using principal component analysis. The results show that among the subjective factors, technology risk perception and expectation confirmation are important factors influencing ChatGPT users' willingness to use. Among the objective factors, social influence and usage context are the key factors that influence ChatGPT users' intention to use. Both subjective and objective factors contribute to ChatGPT users' usage behavior and can enhance users' satisfaction and willingness to use. This paper is important for enhancing ChatGPT users' intention to use and improving ChatGPT users' experience.

With the wide application of text recognition technology, the text recognition model's own limitations and environmental factors interfere with the recognition error rate is high. To address the above situation, a text error correction method based on MacBERT4CSC after text recognition is proposed. Firstly, using the single-word text recognition confidence, we obtain the suspected wrong words and positions by setting thresholds for error detection of the text to be corrected, and then iterate through the constructed confusion set for text error correction in priority, and after the traversal is completed, the MacBERT4CSC model recalls the candidate words with suspected wrong positions, and finally, after the similarity of the word code and the MacBERT4CSC model score meet the set conditions, The error correction is finally completed by sorting the candidate words that meet the conditions. By imitating the error type of word recognition to build the dataset for word recognition and calling Ali's word recognition API for word recognition to obtain the test set to be corrected, the comparison experiments show that the MacBERT4CSC recall and word bar code sorting method improves in accuracy, recall rate and F1 value compared with other deep models and traditional rule error correction methods, which verifies the effectiveness of the method.

This article proposes a sequence-to-sequence based multi-label text classification model. The model combines neural convolutional networks and self-attention mechanisms as encoders, and designs a novel decoder to decode and generate label sequences. The proposed method not only fully considers the interpretable fine-grained information in the source text, but also effectively utilizes this information to generate label sequences. When predicting labels, the global information and local information can be effectively combined to improve the accuracy of label prediction. To verify the effectiveness of the proposed model, this article conducted a large number of comparative experiments. The results show that compared with other models, the model proposed in this article has better performance in different indicators.