This PDF file contains the front matter associated with SPIE Proceedings Volume 12923, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

With the accelerating speed of power grid construction, the number of substations is increasing, and the management of power equipment is becoming more and more complex. The traditional manual management method is time-consuming, labor-consuming and inefficient. Due to the complex scene of power equipment nameplate, the failure of standardization of nameplate size, many characters, miscellaneous types and many steel seal information in the nameplate, the recognition accuracy of power nameplate needs to be improved. On the basis of fully mining the principle of attention mechanism encoder and decoder, combined with LSTM and convolutional neural network, this paper proposes a nameplate text recognition model based on attention mechanism to solve the problem of automatic nameplate text recognition. Experiments show that the method proposed in this paper can effectively identify the substation nameplate.

The system concept of equipment modeling and simulation and the classification of equipment simulation system are introduced. According to the division of equipment simulation elements, the key technologies of equipment simulation are summarized to 6 aspects: military equipment combat theory, combat concept modeling, battlefield environment modeling, combat behavior modeling, command and information system modeling, and data simulation results visualization. The development of equipment combat simulation is prospected from four aspects: integration of equipment combat simulation, distributed simulation, big data and intelligent simulation, and the development direction of future combat simulation system are pointed out.

Wildfires, also called forest fires, are a common natural disaster that often occur in forests and are difficult to control. Detecting and suppressing them at an early stage, primarily through monitoring smoke and fires, is crucial in reducing losses. Thanks to the efforts of researchers, wildfire detection technology has advanced significantly, from traditional manual monitoring to target detection, sensor detection, and infrared detection. However, the various detection methods still have some issues, including low accuracy, high costs, slow detection speeds and susceptibility to interference. This paper presents an improved approach to identifying wildfires based on YOLOv7 with CBAM. Our experimental results indicate that our approach achieves a mean Average Precision (mAP) of 95.77%, surpassing VGG-SSD with CBAM by 0.53%, MobileNetv2-SSD with CBAM by 0.37%, and Faster R-CNN with CBAM by 2.1%. Thus, our method offers a highly accurate approach to wildfire identification.

The detection and recognition of pedestrian abnormal behavior have been currently attractive research hotspots in surveillance video content-analysis. However, due to the complexity and diversity of pedestrian abnormal behavior, the accurate and efficient behavior-recognition remains a challenging issue in the field of pedestrian safety supported by computer vision technologies. Previous studies mostly focused on detecting and locating abnormal behavior from given videos rather than recognizing pedestrian behavior, which may be incompetent in a few practical scenarios since the threat of each anomaly is needed to be evaluated. As individual pedestrian behaviors are closely related to human postures, in this study, we propose a novel experiment design approach for pedestrian abnormal-behavior recognition based on pose estimation and behavior recognition. We first classify and annotate the common abnormal behaviors in public crowds’ abnormal-behavior datasets, such as ShanghaiTech and CUHK Avenue. Then, the pose estimation toolbox, Openpose, is employed to extract skeleton sequences of behaviors annotated. Finally, the skeleton sequences are classified by the spatial-temporal graph convolutional network (ST-GCN) to implement behavior recognition. This proposed experiment design-approach can provide a technical support to validate the mathematic models or algorithms about pedestrian abnormal behavior recognition.

In recent years, pedestrian stampede accidents in cross-passages of large public places occurred frequently, as an inevitable issue in the area of public security. Abnormal behaviors is one of factors leading to pedestrian stampedes. To recognize behaviors, pedestrian pose recognition is introduced as a hot research topic in the fields of deep learning and computer vision, and has developed into a vibrant research field with various real-world applications, such as human-computer interaction, animation, 3D reconstruction, and abnormal behavior detection. This study reviews literatures related with human pose recognition to pedestrian merging areas such as cross-passage in public places. By detecting abnormal pedestrian behaviors such as falling or crowding, potential accidents in densely crowded areas, such as stampedes, can be prevented, thereby improving public safety. Then, we systematically introduce the methods of pedestrian pose recognition in cross-passage in public places from the perspective of pose estimation and behavior recognition, focusing on the research progress and advantages and disadvantages of each recognition method. Finally, the outlook on challenging issues and future development trends of the pose recognition research is drawn out.

Deep learning architectures have been widely fostered throughout the last years and used in various applications, such as object recognition, image reconstruction, and signal processing. Recognition based on radar high resolution range profiles sequences(HRRPs) is essential to robust radar target recognition. Generally, it is almost always better to recognize radar targets in low-clutter environments. Recognition methods commonly cannot extract features from HRRP sequences, especially when the environmental background is not "clean." The Restricted Boltzmann Machine (RBM) is an excellent generative learning model. By extending its parameters from real numbers to fuzzy parameters and adding a hidden layer dynamic random permutation algorithm, we developed the discriminative infinite fuzzy restricted Boltzmann machine (Dis-iFRBM). The proposed model possesses better robust recognition capability than the discriminative infinitely restricted Boltzmann machine (Dis-iRBM). In order to verify the recognition stability and robustness of Dis-iFRBM, radar HRRP sequences recognition experiments are conducted in this paper. The radar HRRPs recognition experiments were conducted to validate the identity stability and robustness of Dis-iFRBM. The experimental results on several HRRPs datasets show that the proposed method is more effective in a "noisy" environment than other models. Moreover, the proposed recognition model have some significant advantages over the traditional structures

Intention recognition task is an important subtask in intelligent question answering research. At present, the task of intention recognition is usually oriented to a single language, and there are relatively few studies on cross-language intention recognition. In this paper, the "XLM-R+TextCNN" model is proposed for cross-language intention recognition task. The text vector of multi-language representation is generated by XLM-R pre-training model, and text features are obtained by TextCNN model to realize cross-language intention recognition. The experimental data set is the self-built tourism data set. The experimental results show that the model has achieved good results in a number of comparative evaluation, which verifies the feasibility and effectiveness of the proposed method.

Artificial Neuron Network is a powerful tool help us solve various problems through its great computing capacity, which prompts its application in all fields. As a result of certain treatments of most of diseases, artificial neuron network could assist experts in analyzing state of an illness. A good prediction depends on difficulty of object and reasonable methods, and it is evaluated by appropriate metrics. this paper did a classification of the application of artificial neuron network in medicine based on Shan Yang’s work and its reference. Eventually, this paper presents some suggestions about the better directions to study in future.

The development of 5G communication and Internet of Things technology has promoted the construction of smart airports. The testing and operation of unmanned baggage transport vehicles have reduced carbon emissions and operating costs. However, vehicle batteries have a limited lifespan. Therefore, this article establishes a baggage vehicle loading and unloading scheduling model that considers battery degradation costs from the perspective of discharge depth, and plans ahead to control discharge depth and reduce battery degradation. The model adopts a detachable transport mode for baggage vehicle and designs an improved genetic algorithm for solution. In the dynamic scheduling stage, the concept of timeline and critical points, as well as a custom insert initial point, are introduced to handle flight delays or advances. Based on pre-scheduling, a re-scheduling is carried out to process dynamic information in a timely manner, and an approximate algorithm is designed for fast solution. The effectiveness of the method was verified through simulation experiments.

Most of current gait recognition models do not consider different influence of the existing gait profiles on recognition results. In addition, most of the existing gait recognition models ignore temporal information of gait profiles. We propose a temporal feature learning based on attention mechanism for gait recognition. Firstly, key feature subnet module is designed to extract key features that have great influence on recognition results, which allows the model to extract more discriminative information. Then, temporal feature extraction module is given to focus on temporal information, and extract long temporal features, short temporal features and frame level features. After that, temporal modeling module is put forward to model the relationship between temporal features based on attention mechanism. Finally, the effectiveness of the proposed method is verified by experiments.

With the rapid development of artificial intelligence technologies, federated learning (FL) is increasingly linked to the Internet of Things (IoT). Recently, researchers have found that the screening and privacy of low-quality users has a significant impact on the training of federation learning. The goal of this work is to ensure that data is trained locally, that data is not out of domain, and that the process of model aggregation is performed in ciphertext, while FL models can be generated quickly and iteratively. However, existing work is still in its infancy, with most research efforts focused on the protection of private data, and implementing their solutions is the main challenge .To address this problem, we propose an efficient and privacy-preserving federal learning scheme with the ability to screen users with low computational power. Specifically, we design a new scheme to reduce the impact of excessive time costs caused by low computational power users by iteratively executing our "Low quality user screening algorithm". Meanwhile, to achieve secure aggregation of the model, we invoke a security framework based on the threshold Palile cryptosystem [8]. Experimentally, we prove that our scheme can greatly improve the training efficiency and save the time overhead of training.

Aiming at the charging difficulty of electric vehicles, this paper starts from the idea of private charging pile sharing, takes into account the needs of electric vehicles, and builds an intelligent recommendation model by adopting the collaborative filtering algorithm. By analyzing and quantifying the characteristics of electric vehicle charging demand, an evaluation model for the selection of private charging piles by electric vehicles is constructed to truly reflect the preferences of electric vehicles for charging piles. By introducing collaborative filtering recommendation algorithm, the recommendation model is proposed to obtain final recommendation results. The simulation results show that the proposed recommendation model can provide effectively charging guidance service for electric vehicles by setting the optimal recommended list length and optimal neighbor length.

Traditional dormitory management methods (such as manual signature, IC card, etc.) have a large workload and high statistical complexity, while the accuracy of fingerprint readers is very high, which is conducive to eliminating fraud, reducing costs, and improving dormitory management efficiency. This design adopts C language design, including three main functions of searching, adding and deleting fingerprints, while other modules cooperate with each other and complement each other. The main control component of the dormitory management system design based on fingerprint recognition is STC89C52 single-chip microcomputer, and the design content includes hardware circuit design and software system implementation. The hardware circuit design part contains each module, the system software part introduces the designed software environment, and finally the system simulation and hardware verification are carried out, corresponding to the main modules of the hardware circuit and equipped with a software design flow chart to realize the interconnection between the single-chip microcomputer and each module

In view of the premature convergence of genetic algorithm, poor local search ability and the strong local search ability and resistance to falling into local optimal solutions of simulated annealing algorithm, this paper combines the two algorithms and introduces adaptive crossover and mutation operators, and designs an adaptive genetic-simulated annealing algorithm. The algorithm adopts matrix encoding and through selection, crossover, and mutation operations, high fitness individuals undergo simulated annealing operations, and four different neighborhood schemes are used to obtain the optimal solution. This algorithm is applied to the project demonstration scheduling problem, and the experiment shows that the algorithm has significant advantages over other algorithms.

The application of digital twin technology in smart highway is a long-term business that covers “construction, management, maintenance, operation, and services”. As the level of digitization of highway deepens, road traffic perception information and real-time, efficient, and visual digital twin systems will provide possible for accurate data and scientific decision-making, as well as providing data foundation and platform conditions for innovative transportation applications oriented towards smart highway. This paper first analyzes the current situation, problems, and solutions of digital twin technology in smart highway. Secondly, research on the construction and light-weighting of large-scale transportation digital twin models, and simulation technology. And the digital twin platform for real-time control and decision-making in smart highway has been successfully applied to G92 section of Shanghai-HangzhouNingbo Expressway. The closed-loop effect of "real-time perception, virtual space mapping, simulation deduction, intelligent decision, real-time control, re-perception" is realized.

This paper proposes a smart security check system of urban rail transit based on the current demand for developing security checks considering the inefficient efficiency, mismatched mode, and poor performance of security checks. This paper analyzes the crucial technologies, which include differentiated security check mode, process and facility optimization, and security and ticket check integration. A simulation model of Tianruncheng Station of Nanjing Metro is established to verify the effectiveness of the key technologies. The proposed method provides a theoretical basis for developing a smart security check system.

In this paper, an attention method for person re-identification is proposed. We focus on different scales of features and strive to extract corresponding features at different fine-grained levels. By stacking attention modules with different finegrainedness to guide fine-grained and coarse-grained attention learning. The effectiveness of the proposed method is verified by experiments.

This paper proposes and designs an all-round, multi-dimensional artificial intelligence-based smart elderly care system around the problem of accompanying the elderly at home. Its main goal is to realize real-time monitoring of the life of the elderly and the capture of emotional fluctuations. When the system detects abnormal conditions such as the fall of the elderly, it will immediately send out an alarm and notify the relevant guardians. By integrating voice interaction technology, the system can recognize and understand the tone and emotions of the elderly, analyze the psychological state of the elderly and provide timely communication and comfort to help the elderly adjust their emotions and improve their quality of life. The system has set up a memo function to provide reminders for daily life, such as drinking water and exercising, so as to maintain the healthy living habits and physical condition of the elderly. The life of the elderly is relatively simple, and it is difficult to spend a lot of free time. The system has designed a VR experience hall, which has a variety of entertainment experiences such as virtual tourism, interactive games, virtual shopping , and virtual chatting . The system also uses big data visualization technology to collect and analyze the health indicators, living conditions and mental outlook of the elderly in real time, and display them in the form of trend graphs, so that the physical and mental conditions of the elderly can be fully grasped. The goal of this system is to change the traditional concept of "care for the aged" and promote the new concept of "enjoy the old".

EEG signals are better able to show changes in human emotions. However, not all channels or spatial EEG signals are valid for emotion recognition. To capture the key effective features of EEG signals that are helpful for emotion recognition, we propose to use the channel space attention mechanism in EEG emotion recognition. We used channel attention to focus on the importance of different channel signals and spatial attention to focus on the importance of different brain regions. Experimental results show that the channel spatial attention mechanism achieves good performance and that it has significant benefits for EEG emotion recognition.

With the development of electric drive and energy storage technology, internal combustion forklifts are gradually replaced by electric forklifts. Electric forklift is developing towards heavy duty. The valve control mode of the forklift fork system causes a large amount of recoverable potential energy to heat up and lose in the way of throttling loss, resulting in energy waste and system temperature rise. For heavy-duty forklift, its load is larger, more energy can be recovered, and the heating phenomenon is more prominent. In order to solve the problems, a energy recovery system with hydraulic motor-motor is proposed, and the theoretical and simulation analysis on the valve control and hydraulic motor control of the hydraulic cylinder descending process is carried out. The simulation results show that the energy recovery scheme is feasible.

A vehicle interior emotion image classification model is proposed to enhance facial expression recognition accuracy in vehicle interiors. The model is based on attention mechanism and improved ResNet. It incorporates transfer learning with a pretrained ResNet50 model as the base architecture. A Coordinate Attention (CA) mechanism module is inserted after the max pooling layer to combine feature maps and focus on channel attention and spatial relationships. CA attention modules are added after each layer to enhance feature map expressiveness. To address class imbalance in the dataset, data augmentation and weighted loss functions are used for model optimization. Experimental results show improved recognition performance on the KMU-FED dataset with a simple and implementable structure.

In recent years, although speech recognition technology has been widely used, it also faces some problems. This paper studies multimodal speech recognition in interpreting based on deep neural network. Firstly, the deep learning method and its related theoretical basis are introduced. Then, the advantages of speech corpus denoising based on acoustic expert feature extraction and training algorithm, convolution decomposition method and interpretation element analysis are described. Finally, through the experimental verification, it is proved that the recognition system can effectively improve students’ interpretation efficiency and accuracy, and the accuracy rate is more than 93%.

The collision detection technology is an important basis for the reliability of artificial intelligence planning assembly process. In the past, collision detection technology has the problem of inaccurate detection of concave parts or high occupation of computing resources, which reduces the practicability of intelligent assembly in industrial field. This article proposes a collision detection method based on stereo model stacking algorithm, which can detect concave parts effectively and occupy low resources. Combined with this method, the intelligent assembly experimental teaching platform is built, which has multiple research and teaching function modules such as assembly sequence optimization and assembly demonstration. It has achieved good results in the research of intelligent assembly field and students' practical teaching.

In urban assistance scenarios, the problem of resource allocation on the urban streets is essentially a specialized form of intelligent traffic control, where the logistics and security capabilities greatly influence efficiency of urban rescue operations. Currently, due to the lack of simulated training environments for urban assistance, solving such problems relies solely on theoretical deductions, providing limited quantitative basis for individual resource allocation and support. To simulate realistic city rescue logistics, this paper utilizes the NetLogo simulation environment to construct a two-dimensional traffic flow grid model, simplifying the urban traffic scenario. The emerging DQN reinforcement learning algorithm is employed in the software to enhance the transportation scheduling capabilities of multiple agents, addressing resource allocation challenges in various scenarios on the city rescue. This approach effectively tackles complex real-world problems and holds promising applications

The research on the distribution of passengers on the subway platform is the basis of platform traffic organization and design, which directly affects the train dwell time and density of carriages. First, the influencing factors of passengers' choice of waiting position are analyzed, in order that the attraction function of the waiting position is established. To be realistc, parameters of walking distance, the passengers’ number and queuing capacity are recalibrated, and other factors such as platform vision and rest seats is added. Secondly, the passengers’ waiting position selection model is proposed based on user equilibrium. Then, the dynamic allocation method is used to allocate the waiting position for passengers. Finally, the simulated annealing algorithm is used to calibrate the parameters of the attraction function. The proposed model is applied to the real-world data of a Guangzhou Metro station and effectively reflect passengers' behavior characteristics.

Due to the difficulty and high cost of conducting sufficient real-world road tests, it is widely accepted in the industry to use a digital twin testing system that combines virtual simulation testing with real-world road testing to test and evaluate autonomous driving systems. Digital twin testing can simulate a road test environment using technologies such as sensor simulation and vehicle dynamics simulation, advanced graphics processing, traffic flow simulation, digital simulation, and road modelling. We integrated simulation testing tools, V2X communication devices, and real testing vehicles to achieve digital twin autonomous driving testing based on virtual simulation and real environments. It generates scene data based on virtual scene simulation software, constructs simulated scenarios mapping to reality, and uses V2X communication devices to transmit simulated scene environment data in real time to the real testing vehicles. It also receives real-time status data from the vehicles and feeds them back to the scene controllers, enabling autonomous vehicles to test different scenarios on real roads. In active safety-related testing, virtual scenes can avoid accidents and save unnecessary equipment losses. The system has high repeatability, and after a large number of repetitive in-vehicle tests, it can significantly reduce functional defects in the tested system, reducing the workload of actual vehicle field testing and road testing and improving testing efficiency and cost savings.

Bearing failure is one of the common faults in general aviation piston engines, which directly affects the stability and safety of aircraft. The existing diagnostic methods are manual visual fault detection, and the diagnostic accuracy is greatly by worker experience and has low efficiency. To address the aforementioned shortcomings, an automatic diagnosis method based on deep learning networks is proposed to achieve intelligent detection of bearing failure in this paper. The proposed method adopts an improved YOLOv5 model to process photos of bearing bush during engine maintenance to automatically detect bearing faults. For specific dataset features, in order to enhance the ability of image feature extraction and further improve the ability of small target detection, the proposed DC-YOLOv5 in this paper adopts 160*160 detection layers in the network, introduces a coordinate attention mechanism in the backbone part, adds a feature multi-scale fusion structure in the neck part, and uses Alpha-CIoU loss function in the prediction part. Finally, the comparison experiments were conducted using on-site maintenance data. The experimental results show that the mAP of the proposed DC-YOLOv5 reaches 95.42%, which is 3.21% better than the original YOLOv5s. The detection accuracy is considerably improved compared with other mainstream algorithms. The demonstrated effect with the proposed method is sufficient for the maintenance of general aviation piston engines.

With the development of digital transmission and transformation systems, digital substations have become a prevalent trend in the electric power industry. However, the presence of many electrical components in substation environments with varying textures poses challenges for the reconstruction of a digital three-dimensional model and robot operation. To deal with that problem, this paper introduces the rich line features present in substation environments into the frontend feature tracking of 3D reconstruction in order to improve the accuracy of the 3D reconstruction. The optimized LSD algorithm is utilized for extracting line features in weak texture scenes. Additionally, a line segment merging algorithm is proposed to reduce the rate of mismatched line features and avoid redundancy through the combination of similar or repeated line segments. Finally, the paper presents a novel line feature matching method based on geometric constraints during the stage of online feature matching, which has the ability to improve matching accuracy while decreasing matching time. The experimental results demonstrate that the front-end visual odometer's optimization method proposed in this paper can offer better feature extraction and matching initial values compared to traditional methods for nonlinear optimization of the back-end of 3D reconstruction and loopback detection.

With the rapid development of urban modernization, the management of road ancillary facilities is more scientific and systematic. As an important part of road ancillary facilities, street lamps have the characteristics of large quantity, unified specification and discrete distribution. How to quickly 3D the large number of two-dimensional component data is a research hotspot at present. In view of the shortcomings of low efficiency, high cost and low precision of traditional manual 3D modeling, when carrying out the real 3D project of Shanghai Road, the company proposes a 3D automatic modeling method of urban road street lamp based on 3DMax. Firstly, different types of street lamp models are established in 3DMax, and then the two-dimensional vector data provided by road collection are screened through script program to determine the attribute information such as the size and direction of street lamps. Finally, the batch automatic modeling of street lamps in corresponding positions is realized. After the project test, the automatic modeling method is compared with the manual modeling method. The results show that the automatic modeling method improves the modeling efficiency, saves the cost, and can ensure the integrity of road ancillary facilities data.

With the advent of the information age, people are paying more and more attention to the security of information, and information security is already a more mature field. The advancement of artificial intelligence technology in recent years has brought about a considerable breakthrough in this field. At present, the research papers on AI in the field of information security are relatively scattered, and there are few papers analysing the hot spots and trends in this field. In this paper, we use Biblimetric, Citespace and VOSviewer to analyse 933 papers downloaded from WOS (Web of Science), summarise them and generate a visual knowledge graph to visualise and analyse the application of AI in the field of information security, through literature volume analysis, co-citation analysis and keyword co-occurrence visualisation, the current research hotspots and possible future development trends are further analysed.

Conversational humor often depends on context. Compared to one-liner humor, the task of conversational humor recognition is more complex and difficult. In addition, characters are one of the most important factors in dialogue, and most of the existing research on conversational humor recognition does not consider character information, resulting in poor results. Therefore, this paper proposes a conversational humor recognition model that combines character information and contextual feature. The main and supporting characters are set in a specific sitcom, and their gender is used as character attribute. RoBERTa, Bi-GRU, CNN and Attention are used to extract utterance feature at the word level and contextual feature at the sentence level to recognize one-liner humor and conversational humor in dialogue. This paper conducted experiments on CCL2020 Task 3 and achieved an F1-score of 53.7%, which is an improvement of 2.2% over the current best score, demonstrating the effectiveness of character information and contextual feature on the conversational humor recognition task.

The planning design of a one-way station-based hybrid fleet carsharing system (OSHFCS), which includes the planning distribution of charging infrastructure, electric vehicles (EVs), and internal combustion engine vehicles (ICEVs), poses a significant challenge for the OSHFCS operator. To address this challenge, this paper proposes an integer linear programming (ILP) solution based on the time-space network to optimize the planning process of the OSHFCS. In practice, the effective relocation of EVs can enable the operator to meet user reservations while operating under resource constraints. Therefore, we also consider vehicle relocation in the operation process. The case study in Chengdu demonstrates the applicability of the ILP model that uses Gurobi to solve this problem. Besides, we further analyze the impact of different user demands and monetary subsidization scenarios.

We observe representative information of large, medium objects and the vast background overwhelm that of small objects, which leads to poor performance of small object detection. To this end, a new module has been proposed in this paper, named Background and Foreground Attention Maps(BFAM) module, composing of three sub-modules: segmentation, background and foreground attention sub-modules. The segmentation maps positions which have the top strong semantic information in the deep layer of the backbone to the bottom layer maintaining more small object details and mask them to obtain background map and foreground map. Apply two tailored attention sub-modules on them respectively and then fuse them with different weights to detect final results. Experiments demonstrate BFAM achieves promising gains in small object detection on PASCAL VOC 2012 and Seaship datasets.

With the continuous expansion of urban and population size, the external cost of traffic congestion is increasing daily. Urban traffic prediction using multivariate data is significant in solving these problems. However, traditional prediction methods are mostly based on merely statistical theory, which only includes a single variable, resulting in the precision accuracy not being guaranteed. Therefore, based on machine learning theory, this paper collects multivariate traffic data from the main sections of Shenzhen, China. With the help of the advantages of deep learning model in data analysis, urban traffic prediction using multivariate data is proposed based on considering the characteristics of multi-parameters, which solves the problem that the traffic prediction method is mainly based on a single variable and the auxiliary information is not considered enough. To fully consider the spatial-temporal characteristics of the traffic dataset, an ABi-LSTM prediction method concerning the influence of upstream and downstream road sections is proposed, solving the problem of low prediction accuracy caused by imperfect consideration of spatial-temporal characteristics in previous traffic predictions. The results provide an important reference for traffic management departments to alleviate traffic congestion effectively, moreover, it’s convenient for residents to understand road information more clearly and make reasonable travel choices.

In the construction of urban rail transit, the space-time distribution of average daily passenger flow in the traffic network is analyzed based on important indicators such as average daily passenger flow, average passenger flow and congestion flow. Based on the data indicators, the organization and management of rail transit line connection should be planned, designed and operated more reasonably, which is to further enhance the technicality of the rail transit operation. It is one of the effective methods to predict the average daily passenger flow of the running track in some periods basing on the distribution characteristics of the average daily passenger flow and the variation model. The model established in this paper first sets the entrance and exit of the subway station as the space trajectory. Then, it regards the fixed parameters as the local function variables, and considers the real-time dynamic parameters as the external function. Considering the characteristics of the external function parameters in space, it fits with the time model accuracy, and dilutes the parameter function. Considering the driving effect of urban subway on the overall traffic environment, the genetic variation is carried out in the spatial and temporal trajectories. The local and global prediction variables are mixed with the change of time and geographical location, which has influenced the average daily commuting passenger flow. The research content of this paper can formulate rail transit travel plans suitable for passenger demand and formulate access management rules for subway stations. It also optimizes station traffic rate, thereby reducing operating costs, improving the return on investment in subway operation, and giving full play to the contribution of the urban rail transit construction.

Arbitrary-shaped text detection is a challenging task where a text instance representation should be well designed to cover diverse text geometry variances. Most of current methods represent text instances as image mask or contour point sequences, but they might lead to expensive post-processing or restricted capability of shape modeling. Most recently, Zhu et al. proposed Fourier Contour Embedding (FCE) method to represent arbitrary shaped text contours in the Fourier domain, and constructed FCENet to obtain state-of-the-art (SOTA) performance. However, even the FCE signature and FCENet may suffer significant detection performance degradation on slender-shaped text instances like most other text detectors because of the mismatch between the square-shaped receptive fields and the large aspect ratio of arbitrary-shaped slender text}. To tackle this problem, we designed Structured Fourier Contour Embedding (Structured-FCE) to encode both instance-level features and local geometric features of the boundaries. Then, we constructed Structured-FCENet to predict and merge Structured-FCE signatures to construct the boundaries of arbitrary-shaped slender texts. Qualitative and quantitative results illustrate that Structured-FCENet obtain SOTA performance on CTW1500, Total-Text and DAST1500 benchmark datasets, especially for the challenging slender text instances of various shape.

The new round of scientific and technological revolution and the continuous optimization of economic structure put forward higher requirements for the prediction of logistics demand. Aiming at the problems of low prediction accuracy of traditional algorithms, we propose a logistics demand prediction method based on deep learning. Specifically, a LSTM logistics demand forecasting model is established to realize more accurate logistics demand forecasting. From the perspective of data structure of forecasting model, we compare and evaluate LSTM forecasting model with traditional statistical methods, neural network methods, support vector regression machines and other methods in terms of time series and influencing factors. The experimental results show that the logistics demand forecasting model based on LSTM has good forecasting performance, with an average absolute percentage error of about 4% and good stability.

The planning and construction of rail transit usually refer to passenger flow forecasting, so it has become an entry point for coordinating and improving rail transit operations. Based on the correlation between passenger flow and time and space, this paper conducts research on the prediction model of rail transit short-term passenger flow. This research selects Suzhou Rail Transit as the research object, processes and analyzes the data of the Automatic Fare Collection (AFC) system, and summarizes the temporal and spatial characteristics of passenger flow distribution at typical stations. On this basis, a combined model based on Radial Basis Function (RBF) neural network called parallel weighted neural network (PWNN) model is established. An example of data from Leqiao Station, a typical Suzhou Metro station, verifies the feasibility of the combined model and proves the superiority of the combined model.

Tidal prediction is of great significance to the production and practice of coastal areas. In order to improve the accuracy of tidal level prediction, a nonlinear exogenous autoregressive network tidal prediction model (DWT-NARX) based on wavelet analysis is proposed. The model decomposes the tidal data by wavelet and realizes hierarchical prediction by NARX neural network. The tidal prediction simulation experiment is carried out by using the measured tidal data of Savannah Port in the United States. The prediction results are compared with the modular NARX tidal prediction (HANARX) model. The results show that the MAE, MSE and RMSE of the prediction results of this method are the smallest. The results show that the accuracy and stability of this method are better than the modular prediction of NARX.

There are various methods of image encryption techniques, including key-based encryption, chaos theory-based encryption, wavelet transform-based encryption, etc. Although low-dimensional chaotic systems are fast, the key space is small and the security is not high. High-dimensional chaotic systems have large key space and high security, which are difficult for attackers to break, but the encryption speed is not good. In order to solve the problem of encryption efficiency of high-dimensional chaotic systems, this paper introduces a chaotic state selection system to dynamically select the state variables so as to reduce the number of iterations in the generation of the secret key, which further improves the security of the algorithm while increasing the encryption speed, making the algorithm fast and in line with the practical application scenarios.

In order to address the issues present in traditional system platform applications, this study aims to expand data enhancement through the design features of system platforms. By swiftly implementing a large amount of spatial data and effectively disseminating information about services in the information space, the goal is to improve work efficiency and achieve mutual benefits. Based on the current classification of systems, which includes desktop virtual reality (VR) systems, immersive VR systems, and shared VR systems, it was found that the overall framework of virtual reality systems is generally constructed through 3D modeling, scene creation, script logic design, and functional module development, enabling practical operational functionalities within the system platform.

This paper proposes a new criminal case recommendation method based on RoBERTa and BiLSTM for the similarity matching task of Chinese judicial documents. The method first uses the UIE model to extract entities, relations, events, sentiments and other information from the judicial documents as the semantic representation of the documents. Then, it uses the pre-trained RoBERTa model to encode the semantic representation of the documents and obtain the semantic vectors of the documents. Next, it uses the BiLSTM model to align and fuse the semantic vectors of two documents and obtain the similarity vector of the documents. Finally, it uses cosine similarity to output the most similar cases. The paper conducts experiments on a self-built dataset and shows that the proposed method outperforms the commonly used traditional models on the criminal case recommendation task, with high accuracy and recall, and can effectively improve the quality and efficiency of case recommendation.

Researchers have developed various test methods and tools to ensure the performance and security of deep neural network (DNN) applications and detect potential defects as much as possible. However, there is still a lack of a mature and comprehensive DNN test theory, leading to the uneven quality of test tools and inconsistent evaluation methods, which pose challenges for the selection and application of test tools. Test Theory (TT) has developed a series of mature theoretical approaches and has validated its usability over a long history and in a wide range of application scenarios. This paper analyses the DNN test process, defines the concept of DNN test tools, and surveys the current research status and limitations of the DNN test field. Inspired by TT, this paper systematically introduces TT into DNN testing for the first time building the theoretical model of reliability, validity, difficulty, and discrimination that conform to the characteristics of DNN, which provides a new research perspective for DNN testing and is a preliminary study for establishing DNN testing theory.

Mitigating high-speed traffic congestion and predicting traffic flow have become major challenges in urban development. Most existing prediction methods rely on graph neural networks and attention mechanisms, but their central ideas focus solely on spatial or temporal features without effectively integrating spatiotemporal features for prediction. In this paper, we propose utilizing the Maximum Information Coefficient (MIC) to explore the relationship between temporal and spatial features in traffic data. We present two approaches: Correlation Attention Mechanism (COATT) and Spatiotemporal Maximum Information Coefficient (STMIC). COATT effectively incorporates spatial features embedded in traffic flow information into the graph convolution matrix, while adding a multi-head attention component ensures attention weights are well-calibrated and the influence of noise can be minimized and its impact reduced. By leveraging STMIC, the spatiotemporal matrix is combined to establish more accurate spatiotemporal dependencies in traffic flow. Experimental results on the high-speed highway traffic flow datasets (PEMS07 and PEMS08) demonstrate that our proposed GCSTN outperforms current state-of-the-art methods in traffic flow prediction. Particularly on the PEMS08 dataset, compared to ASTGCN, our model achieved a reduction of over 7.2% in MAE, RMSE, and MAPE metrics, while compared to STGODE, a reduction of over 3.5% in the same metrics.

Object detection algorithms based on deep learning often adopt complex algorithm structures in pursuit of higher detection accuracy, resulting in a huge amount of overall model parameters and computation, making it very difficult to deploy the algorithm on mobile platforms with limited hardware computing power. This article studies the weight reduction of object detection models and proposes a lightweight algorithm YOLOv5-SNGS based on the YOLOv5 model. Firstly, the ShuffleBlock module is designed to perform lightweight transformation on the backbone of YOLOv5, significantly reducing the amount of model parameters and computation, and improving model detection speed. Then, the SimSPPF spatial pyramid pooling structure is used to extract targeted features from targets of different scales. In the bottleneck part, the GSConv convolution module is designed to fuse multi-channel feature information, and the SimAM attention mechanism is added to quantitatively evaluate the importance of neurons in the network. Key feature extraction is carried out for key areas of the image. The experimental results on the Pascal VOC dataset showed that compared to the benchmark model YOLOv5s, the parameter count of YOLOv5-SNGS decreased from 7.2million to 4.1million, and the computational load decreased from 16.6GFLOPs to 8.1GFLOPs. At the same time, the FPS during mobile platform deployment increased from 80 to 105, with an increase of 31%. While significantly reducing hardware computing power dependence and improving detection speed, the mAP of YOLOv5-SNGS remained at the same level with YOLOv5s, which can meet the target detection requirements of mobile hardware platforms

The main task of object detection is the recognition and localisation of objects in images. It is the cornerstone of image understanding and the preparation for vision tasks such as image segmentation, target tracking and pose recognition. YOLOv5 is a object detection technique that is widely used in military, medical and traffic applications. However, in complex and changing real-world scenes, there are often multiple interfering factors, so the detection effect of YOLOv5 is greatly affected. In this paper, we propose an improved YOLOv5 algorithm for complex scenes from the perspectives of background information interference and inaccurate bounding box localisation. Firstly, the SimAM attention module is added to YOLOv5 to enable the model to focus on the differences in information in the feature maps and to enhance the anti-interference ability of the network; then the EIoU loss is used in the loss function and the Focal EIoU loss function is proposed by combining the optimization idea of Focal loss to deal with the category imbalance and improve the convergence speed of the algorithm; finally ,a comparison experiment is conducted on the PASCAL VOC dataset to demonstrate the reliability of the improved algorithm.

The characteristics of machinery "three many and three complexes", i.e. many structures, many mechanisms, many parts, and complex types, structures, and movements, make it difficult for students to understand and digest the teachers' exciting lectures. The virtual assembly can help students to learn in a safe environment, but also to be able to get close to and assemble these complex mechanical structures to help them learn better. The collision detection technology between parts in virtual assembly is the main problem. Based on the Yixinghe engine, this paper proposes to use the depth map for subsequent collision judgments for parts that are in the virtual environment, after using the AABB enclosing box to do the initial collision judgments.

Our retinal vessel segmentation approach utilizes deep hierarchical semantic segmentation along with a closing operation. From fundus images, the retinal vessels are extracted using the supervised learning segmentation algorithm. Deep semantic segmentation that provides a hierarchical solution is adopted for rough retinal vessel segmentation. The rough segmentation results are then processed by a closing operation to refine the segmented retinal vessels. We performed experiments and comparisons with ground truths to evaluate the qualitative and quantitative effectiveness of our method. Our method effectively segments retinal vessels, as demonstrated by the experimental results.

A test platform for performance analysis and verification of anti-intrusion and dangerous approach alarm system of mining driving equipment was developed. The platform can not only simulate the radial intrusion equipment of personnel in multiple directions,but also develop a soft dummy target with the shape characteristics of Chinese human body and the infrared and radar reflection characteristics of Chinese human body for the intrusion object. The dummy target has three different poses to choose from. The platform can be used to verify the key indicators of anti- intrusion and danger approach alarm system of driving equipment, such as human posture identification, alarm boundary accuracy, alarm real-time, system false alarm rate and multi-target recognition. The test results show that the platform can effectively realize the above functions, directly avoid the occurrence of equipment extrusion and collision accidents in the mining process, and ensure the life safety of workers.

This paper proposes two improvements to address the issues of high complexity and computational burden in the Swin Transformer backbone network for feature extraction and the feature mismatch problem caused by coupled detection in the Cascade R-CNN detection network. First, a lightweight PoolFormer Block based on pooling is introduced in the third and fourth stages of the Swin-T network to reduce its complexity. Then, to improve the feature extraction capability of the lightweight Swin-T network, a coordinate attention mechanism is introduced. Second, the classification and regression tasks of objects in the Cascade R-CNN detection network are decoupled to alleviate the issue of feature mismatch in the two tasks and further improve detection performance. The experimental findings obtained from the PASCAL VOC dataset indicate that the proposed approach increased the average detection precision by 2.2% compared to the baseline model.

In order to solve the problem of single form, monotonous content and poor effect of traditional electric power training, this paper develops a power distribution training system based on virtualized digital. The system uses 3ds max and Unity engine to complete virtual scene construction, applies LOD technology to complete rendering resource allocation, and uses XML language to write and compile scripts. The operation details of electric power training are restored objectively and accurately. Based on the electric power training model of Hainan Power Grid, a VR training system based on virtual digital technology is proposed to help users better master electric power operation skills with digital means as a tool, and improve the training effect and learning results.

Low resolution is an important factor affecting the accuracy of face recognition. Existing face image recognition methods can not overcome the limitation of low resolution face image recognition, resulting in poor face recognition results. Therefore, an adaptive recognition method for low resolution face images based on high order singular value decomposition and DNN model is proposed. Integrating graph method is used to extract face image features, and high order singular value decomposition (HODVD) is used to separate face image features from identity information, image illumination information and image pose information. On the basis of big data face samples, deep neural network recognition model technology is introduced, and based on stable and open source tensorflow implementation, face image recognition is completed. The experimental results show that the recognition rate of the proposed method is 97.2% and the running time is 8.6s, which is effective in application and provides a reliable basis for related fields.

Leaf vein segmentation is of great importance for increasing food production in botany and agriculture. To further improve the performance of vein image segmentation, we propose a new model CPB-DUNet based on the method of U-Net. That is able to find the detailed features of veins from the segmented vein image itself rather than from the original input image. CPB-DUNet is composed of two horizontal U-Net structures. The second U-Net corrects and refines the original U-Net segmentation results by iteration. CPB-DUNet still uses compact bilinear pooling between each pair of decoder modules to better extract features. And we achieved IoU scores of 81.97 and 53.59, respectively, in the only two publicly available labeled vein datasets, Leaf venation networks of Bornean trees and LVD2021, the former score the best in the literature.

Flexible sensors are an important research direction in the field of smart wearable devices and human-computer interaction, and friction flexible haptic sensors are the first choice of flexible sensing materials in the field of elderly monitoring due to their high sensitivity, softness and wearability. The purpose of this study is to design a smart wearable system for heart rate monitoring of the elderly based on friction-flexible sensors. The heart rate monitoring is performed by frictional flexible sensors to achieve real-time collection of heart rate of the elderly. In the artificial intelligence dimension, the obtained heart rate data is processed to obtain risk warning and provide personalized health advice. In the software dimension, design an information feedback system that meets the needs of the elderly interface and makes it easy to operate and understand. By providing a comprehensive heart rate monitoring system, it helps the elderly to detect cardiovascular diseases and take appropriate treatment measures in a timely manner. It is expected to provide a reference for research related to improving the quality of life of the elderly and the prevention of cardiovascular diseases.

In recent years, with advancements in technology and improvements in hardware performance, virtual reality (VR) technology has been widely applied. optimizing elements within VR scenes can help ensure stable operation and enhance the user experience. Through the use of the QEM algorithm, this article optimizes the models in the scene, reducing the number of polygons and further improving the smoothness of rendering and user experience. Research results prove that the QEM-based optimization LOD technique is a highly applicable model optimization method in virtual reality application development. This method not only reduces device resource consumption but also enhances application performance and user experience, effectively lowering development costs.

Object: With the help of style conversion Network, the design of the Zhuang brocade pattern is derived to explore the new direction of inheritance and development of the Zhuang brocade pattern in the current era. Methods: The cultural connotation and design value of Zhuang brocade patterns are analyzed, and the selected Zhuang brocade patterns are used as content maps and different style maps to conduct style conversion Network experiments. The experimental results are analyzed and evaluated to find the optimal conversion solution. Investigate the effect of the conversed Zhuang brocade pattern in the derivative design. Result: By using style conversion Network, different styles of patterns can be generated quickly and can produce good derivative design effects in many products, thus achieving the purpose of integrating Zhuang brocade patterns with modern products. Conclusion: The addition of the style conversion Network algorithm makes the style simulation process significantly reduced, saving time for subsequent derivative designs and providing the possibility of mass production. The designs produced by the style conversion Network retain the original pattern characteristics of Zhuang brocade, with both auspicious symbolic meaning and national spirit, and realize the interaction between traditional crafts and the art field, in which the cultural heritage can also be shown, promoting Zhuang culture and providing new possibilities for the modern development of Zhuang brocade patterns.

Deep learning has been extensively researched for plant disease image classification, and there is a need to develop a crop disease image recognition system that provides intelligent information services to farmers. This study focuses on citrus plants and establishes a dataset comprising infected and healthy citrus leaf images. By combining diffusion models and transfer learning, four models (VGG16, ResNet18, Efficient Net, and Shuffle Net) are trained using three different approaches to evaluate their performance. The study aims to identify the most suitable model and develop an intelligent recognition system for citrus plant diseases based on image analysis. The feasibility of using deep learning for citrus disease classification is preliminarily validated using the created dataset.

Online courses in colleges and universities need professional equipment and settings for students to experience the operating environment, but the real equipment is expensive and the operating experimental conditions are limited, which results in the knowledge expression defects of professional courses. The development of virtual simulation laboratory can help the laboratory to realize the practice environment of 3D virtual simulation, which requires the integration of new technologies such as virtual reality into teaching. This paper adopts the latest GL3D Web design technology and Three. js related open source frameworks to design the system. With the help of the virtual laboratory, the whole experimental process and experimental results can be fed back to the virtual laboratory users in real time through three-dimensional visualization, which further improves the effect of experimental teaching and reduces the pressure of experimental work. Through virtual simulation technology, this paper builds a flexible teaching platform to promote the teaching function of controlling virtual equipment with the help of physical control system.

A two-stage hotspot detection method for aerial infrared images is proposed to address the issues of high cost, low efficiency, and low accuracy in traditional photovoltaic power plant detection technology. This method achieves component level localization and fine classification diagnosis of hotspot defects in infrared images. The method proposed in this paper is to integrate deep learning algorithms and traditional image algorithms to better identify defects. Firstly, this paper uses the edge detection algorithm to segment the target contour for the image itself at different scene gray values; Secondly, considering the differentiation of related factors, this paper carefully classifies the correlation factors based on the EfficientNet network. In order to ensure the rapid detection of the model. Experimental results show that the accuracy of the proposed model reaches 97.1%, and the speed is also faster. This shows the superiority of the algorithm model proposed in this paper.

Lodging is a key factor, which affect yield and quality of wheat. Timely and accurate acquisition of wheat lodging information is conducive to lodging identification and breeding of improved varieties. In this study, based on the UAV visible light image obtained during the wheat filling period, the wheat lodging image data set was constructed, and the ACSNet network model was selected as the main model of segmentation task. The model can effectively extract local context information features. The result show that the prediction accuracy of ACSNet model was 87.5 %, which could accurately and efficiently complete the automatic classification of wheat lodging. It had high accuracy and applicability, and provided an important basis for agricultural insurance companies to identify wheat lodging and determine agricultural losses.

Aiming at the problems of low detection accuracy and poor robustness of electric vehicle helmets in complex traffic scenarios, an improved YOLOv5-based electric vehicle helmet detection model is proposed. First, a shallow feature detection layer is added to enable the model to better detect small targets. Then, the attention mechanism is introduced in the backbone network, which optimizes the feature extraction ability of the network. Finally, replacing the CIoU loss function with WIoU improves the convergence speed and regression accuracy. The improved YOLOv5s algorithm was trained and tested on a homemade electric vehicle helmet dataset, and the improved YOLOv5s algorithm improved the average accuracy mean, accuracy rate, and recall rate by 6%, 4%, and 6.7%, respectively, over the original algorithm. The experimental results show that the electric vehicle helmet detection model with improved YOLOv5s algorithm has high helmet detection accuracy, excellent detection of small targets, and strong generalization capability.

[Objectives] In order to improve the emergency response speed and emergency response capability of aviation emergency rescue, and solve the pain points of long emergency preparation time, single emergency training scene and high cost of actual combat training, a virtual training system for aviation emergency rescue is created by using virtual simulation engine. [Process and Method] The three-dimensional modeling technology was used to truly restore the disaster accident scene and rescue disposal environment. The A* algorithm was improved by the Bessel curve principle to carry out the path planning simulation of the rescue equipment handling process. Based on the integrity of the virtual operation, the correctness of the operation steps, and the rationality of the operation time, a scoring evaluation system was established. Users can choose the learning mode or training mode independently according to the knowledge, select the airborne rescue equipment and complete the simulation of the assembly process and assembly mode of the rescue equipment. [Results] Through the operational test and functional test, the results show that the system can effectively simulate the installation method and emergency treatment method of airborne rescue equipment, and improve the response speed of emergency rescue.

To address the current problems of combining single domain-specific knowledge and poor fusion in suicidal ideation detection tasks, this paper proposes a Multi-Head Knowledge Attention Mechanism model that fuses domain knowledge (DK-MHKA) to fully integrate the suicide risk severity lexicon and the user's neurotic personality traits. The model involves integrating suicidal tendencies attributes into the semantic domain that encompasses the user's social media content, with the aim of enhancing the model's linguistic representations. Furthermore, the method employs a multi-head knowledge attention mechanism to effectively combine various sources of features, resulting in an enhanced predictive capability of the model. The experimental findings indicate that the suggested DK-MHKA model outperforms alternative baseline models in terms of forecasting precision. Additionally, the ablation experiments confirm the individual contributions of each module to the overall performance of the model.

Exploring the cerebral structure-function correlation advances our understanding of the brain's underlying principles. The brain structure is static, while brain function is dynamically changing; thus the relationship has the unmatched issue of the temporal and spatial dimensions. It makes the structure-function relationship a challenging task. Partial least squares correlation (PLSC) is applicable to examining the structure-function association. Most of the studies explored how structural traits were linked to functional ones. However, if the relationship is explored at a voxel-wise level, perhaps more interesting results will be found. Actually, the ordinary PLSC cannot effectively handle the unmatched temporal dimension issue at the voxel-wise level, failing to establish suitable linear models. We propose a novel variety of PLSC, termed constant-block PLSC, to handle this issue. The performance of our method is compared with the ordinary PLSC by applying it to simulated structural magnetic resonance imaging (sMRI) along with resting-state functional MRI (fMRI) information. As indicated by the results, the superiority of our approach to the common PLSC is that it extracts proper quantity of significant and realer spatial patterns, enhancing the cross-block covariance to 36.02%from 0%. Our constant-block PLSC technique is promising for assessing the relationship between cerebral architecture and functionality

Brain structure-function relationships are becoming an important topic in neuroscience. The structure-function correlation can be performed at the scales of voxels, connections, networks, and so on. Most of the studies have been carried out based on brain structural and functional connections. Actually, if the studies are performed based on whole-brain voxels, which maybe lead to more interesting findings. The present work analyzed the relation of structural magnetic resonance imaging (sMRI) data-based whole-brain mean gray matter volume with resting-state functional magnetic resonance imaging (fMRI) data-based whole-brain voxel activity by adopting constant-block partial least squares correlation (PLSC) across the lifespan. Additionally, the resting-state networks, consisting of default mode, auditory, dorsal attention and sensorimotor networks have been identified by brain structure-functional relation analysis. According to our results, the whole-brain mean gray matter volume lays the foundation for brain function, which also contributes to the implementation of different activities.

To forecast the condition of traffic networks in the future, it is crucial to model the spatial and temporal correlation of traffic series. The majority of current research has been on creating complicated graph neural networks that can capture common patterns using preconfigured graphs. In this paper, we claim that predefined graphs may be avoided and that adaptive graphs can be used to capture spatial correlations between traffic series and improve the performance of graph neural networks. In order to capture the temporal relationships of sequences, we also aggregated gated recurrent neural networks. Then, we encode the relative time position of the sequence in order to fully extract the characteristics of the traffic sequence. Finally, we add the values of the previous day and the same day of the previous week as a reference in the final prediction to improve the accuracy of our prediction. Experimental results on two sets of real-word traffic data (PeMSD4 and PeMSD8) demonstrate that our method is better than the existing methods.

Federated learning has rapidly become a research hotspot in the field of distributed deep learning because raw data does not need to be shared. However, it is still characterized by problems in terms of privacy security and computing power. Split learning is a new distributed deep learning technology. By dividing the complete deep learning model, the client and server can participate in the training together to overcome some problems related to computing power and privacy and security. In this paper, the concept of split learning is introduced. This article introduces the use of multi-user and parallel training methods to improve the learning efficiency of split learning.

This paper presents a visual traceability system for the Farm-to-Customer (F2C) IoT agricultural model, aimed at improving product traceability and quality. The system integrates intelligent hardware such as sensors, RFID chips, and smart labels, and employs blockchain, big data, cloud computing, and AI technologies for comprehensive product traceability services. The system uses a data-driven approach, leveraging data mining and analysis to provide consumers with intuitive and personalized traceability information. Its visual design, encompassing product identification, traceability information display, and data monitoring interfaces, is key to its success, offering a seamless user experience. This innovative system significantly enhances product quality and safety management, promoting a safe consumption environment. Its success relies on the integration of multiple technologies and continuous design refinement. Applicable across various industries, it ensures product quality and safety, boosting consumer confidence.

The interrogative understanding task is an important subtask in question-and-answer systems, and currently, most of the existing interrogative understanding methods are oriented to single-intent scenarios, which do not match the actual scenarios. Based on this, we propose a joint modeling model PBIU (Prompting Based Interrogative Understanding) based on cueing learning for the interrogative sentence understanding task. Instead of the traditional pre-training-fine-tuning approach, a pre-training-prompting-prediction approach is used to formalize two subtasks into the same form using a generic pre-training model, and different prompts are used to complete different subtasks. At the same time, the model introduces auxiliary subtask slot prediction and learns the relationship between labels, thus improving the effectiveness and generalization of the model. The experimental results on the publicly available datasets MixATIS and MixSnips and the self-built multi-intentional interrogative dataset MixTFQD in the tourism domain validate that the model has better performance in the interrogative sentence comprehension task in the tourism domain, which is of more importance for applying to realistic real-world scenarios of interrogative sentence comprehension tasks.

Text correction aims to determine whether natural language text contains grammatical errors, text errors, etc., and correct sentences. Previous work usually adopts byte pair encoding (BPE), which may lead to semantically related Chinese characters being separated. In addition, the previous models can only extract the superficial semantic features, but cannot capture the global deep semantic relationships. In this paper, we introduce the time convolutional network (TCN) to capture multi-scale semantic information, so as to promote the development of global semantic information. The CTC_BERT model uses the synonym masking strategy to reduce the semantic segmentation of related words and adds a fully connection layer as the error detection layer. In order to verify the performance of the CTC_BERT model, a comparison experiment has been carried out on SIGHAN2015+Wang271K Chinese error correction dataset. The results show that the accuracy of this model can reach 81.4%, which is better than that of BERT, BART, ConvSeq2Seq and other conventional models, and effectively improves the performance of text error correction.

The development of the Internet makes online reviews tend to be multilingual. In the study of aspect sentiment analysis tasks, traditional monolingual methods is unable to fully extract the reviews sentiment information in multilingual situations. To solve the limitations of traditional monolingual sentiment analysis in multilingual situations, a ngram-based multilingual gated convolutional neural network model was proposed. The model divided the review text into the left context and the right context, and uses convolutional neural network to obtain ngram features of different size between the left and right contexts. After max pooling, the gating mechanism is used to realize the interaction between ngram features. Finally, the features extracted from the left context and the right context is used to detect aspect categories. In this paper, several experiments have been conducted for the proposed model. The experimental results show that the performance of the proposed model is significantly better than other comparison models, which can more effectively extract the emotional information of user reviews more effectively and improve the accuracy of aspect category detection.

In a large metro network, passenger path selection is a key component of the study of passenger flow performance in complex metro networks. The Logit model has been widely welcomed by scholars due to its simple and practical characteristics. However, existing studies often ignore the impact of congestion, transfer on passenger path selection. Therefore, this paper proposes a passenger path selection model considering multiple factors. Firstly, this paper analyzes the main factors that affect passenger path selection behavior by questionnaire. Then, the discrete selection model considering the ride time, transfer time, transfer times, congestion degree is established, and the calibration method of parameters is given. Finally, the validity of the model is verified by taking Suzhou rail transit as an example. The results show that the multi-factor model proposed in this paper can estimate the passenger path selection behavior more accurately.

The PBFT consensus algorithm commonly used in the consortium chain has the disadvantages of three-stage consensus process, resulting in high communication costs. This paper proposes an enhanced PBFT algorithm that utilizes a credit mechanism to select master nodes that can improve system availability. The three-stage consensus process of PBFT is optimized, and the master node collects the confirmation messages broadcast by the whole network during the confirmation stage, effectively reducing communication overhead. At the same time, the block height synchronization mechanism is used to ensure the high consistency of consensus node blocks, restore the block height of failed nodes, and achieve fast consensus. Empirical results show that the improved P-PBFT consensus algorithm effectively reduces resource consumption and latency.

Early detection and treatment of esophageal cancer may improve the survival rate of patients, despite its high incidence and mortality. The use of computer technology can assist in the diagnosis of esophageal cancer. RNA-Seq gene expression data can be used for the diagnosis of esophageal cancer, but it is difficult to analyze directly because of its high dimension and small sample size. Applying computer technology to this data can solve these problems. In our work, we used the RNA-Seq gene expression dataset and considered the specificity of the sample, proposed an artificial intelligence approach for esophageal cancer classification through selecting the comprehensive features of RNA-Seq gene expression data using mutual information feature selection and obtaining a set of sample specific features by generating adversarial examples using one-pixel attack method to reduce the dimensionality of the dataset. Finally, the deep learning method is used to construct a deep neural network as the classifier. The experimental results reveal that this method outperforms other state-of-the-art algorithms in terms of accuracy and other metrics.

This paper takes geographical indication product data of Yunnan province as the research object to build a visualization platform for geographical indication products of Yunnan Province. Aiming at the current problems of geographical indication products in Yunnan Province, this platform has designed and implemented a comprehensive visual analysis and display platform, which includes five functional modules: Visualization analysis of geographical indication product data, WebGIS of geographical indication products, Geographical indication tracing, Geographical indication product 3D tilt photography, Reporting and statistics of geographic indication product information. It can provide technical support for geographical indication related managers, users and marking enterprises on visual visualization of geographical indication products.

In view of the instability of the existing cigarette quality control level, there are still large fluctuations in the physical characteristics of tobacco, physical characteristics of tobacco and cigarette smoke quality control indicators. The radial basis function neural network (RBFNN) has the advantages of simple structure, fast convergence and overcoming the local minimum problem. The prediction model of RBFNN cigarette silk process quality control index was established, the parameters of cigarette silk process were optimized, and the optimized results were compared with the traditional orthogonal test results. The results show that the prediction model of RBFNN technology quality control index for cigarette silk production has high accuracy, while the traditional orthogonal test method has differences in parameter setting. Therefore, RBFNN is more accurate in predicting the optimal technological parameters of cigarette silk making and improves the production efficiency of cigarette silk making.

With the development of information technology and big data, large amount of multimodal data with opinion biases has exploded. Compared to unimodal data, multimodal data can accurately identify users' emotional tendencies and enrich emotional attitudes from a variety of perspectives. To effectively identify the consistency of sentiment representations between image and text semantics, as well as the variability in the degree of contribution in sentiment analysis, an imagetext fusion sentiment analysis method based on text attention is proposed. The method constructs attention-based models for extracting text and images' features, emphasizes the words containing sentiment information, and builds a sentiment classification model using multimodal feature fusion. Experimental results show fusion method proposed has achieved better performance in both accuracy and F1 values compared with baseline models.

Taking KCS ship model as an example, using numerical tank technology and commercial CFD software, the free heave and trim motion of ship model sailing in infinite depth of still water is simulated in the time domain. Based on the coupled solution of ship motion equation and fluid flow equation, the VOF method is used to deal with the nonlinear free liquid surface, and the overlapping grid technique is used to deal with the grid. Finally, the relevant conditions of the ship model such as resistance, heave and trim are obtained, and the detailed information of the flow field including free surface wave, ship surface wave height and ship surface dynamic pressure are obtained at the same time, which is helpful to understand the mechanism of ship heave and trim motion. The calculated results are compared with the relevant experimental data, and the agreement with the experimental data is good.

Ptychography is a powerful technique that combines scanning transmission X-ray microscopy (STXM) and coherent diffraction imaging (CDI). Thanks to the fluctuation of light and the development of phase retrieval algorithms, ptychography has achieved ultra-high imaging resolution. However, iterative solutions are time consuming. A cascaded residual network is used to recover the amplitude and phase information of the sample directly from the received diffraction patterns. With the powerful coding and decoding capabilities of the residual blocks, the correspondence between diffraction patterns and sample distribution information can be learned by the network. Experimental results show that the use of cascaded residual blocks enhances the reconstructive capability of the network.

QR codes have become the most prevalent means of information transmission in today's world, finding widespread usage across various domains. Nonetheless, being a publicly available technology, QR codes entail inherent risks of exposing sensitive information during the transmission process. To address this vulnerability, this paper introduces a QR code secret embedding scheme based on an enhanced EMD method, aiming to safeguard sensitive information within QR codes. In this scheme, secret messages are transformed into binary data streams, and an improved EMD strategy is employed to precisely adjust the pixel values of the cover QR code, facilitating the embedding of confidential data. Despite the modifications, the resulting marked QR code remains legible and meaningful, thereby diminishing the curiosity of potential attackers. With the aid of a customized scanner, the secret information can be swiftly extracted. Compared to existing approaches, this scheme boasts a higher capacity for secret payload and superior embedding efficiency.

As ROV technology gradually matures, underwater robots are being applied in a large number of deep-sea operations, which requires more professional operators to be trained. However, due to the complex and dangerous underwater environment, long sailing operation time, and expensive cost of ROV, to training a pilot is always low safety, long cycles, and high costs by using real equipment. Therefore, this paper provides a method for ROV training in a virtual environment, which can be used to train new operators and provide an environment for researchers to conduct edge testing. This system is developed based on Unity3D and implements functions such as deep-sea scenes, multi-camera screen display, simulation of three ocean currents, real-time multi-beam sonar simulation. In order to achieve realistic effects, visual and control simulation, this system uses advanced features in Unity, including HDRP, physics engine, and image processing technology. Through this system, operators can be trained more safely, efficiently, and lower cost.

With the rapid progress of information technology, the application of visualization is becoming more and more popular, and more and more learning in the future will use various visualization means. There is no doubt that visual organs will play an unprecedented role. Visualization may become the mainstream way of learning, and human learning efficiency will be greatly improved. In order to adapt to such a process, it is necessary to analyze in detail how to perform visual analysis. This paper will describe in detail the applications and functions of CiteSpace software. It is hoped that the fans of science knowledge map are able to know, master and use CiteSpace software to draw science map efficiently in the shortest time, and effectively guide the study and research practice. Therefore, we take the 2012-2022 Master's degree thesis in International Chinese Education as an example to show how to use CiteSpace software for visual analysis. In this paper, CiteSpace (visual analysis software) is used to carry out quantitative research and knowledge graph analysis on the master's thesis of CNKI (China National Knowledge Network database), revealing the evolution path and new development trend of research focus in this field in the past 13 years, and expanding the perspective of International Chinese Education research in multiple dimensions. The diachronic and synchronic analysis of the literature in this field is carried out, and the research focus in different stages and different periods is studied in multiple dimensions and perspective, which further helps us demonstrate the application and research significance of CiteSpace software.

Digitization and visualization are inevitable trends for future equipment. Nowadays, with the deepening of new military changes and the iteration of military equipment, the requirement for digital life cycle management of equipment is becoming more and more urgent. In order to better guide the development of digitalization and visualization of future equipment, Visual Simulation Technology comes into being and can provide an effective method for it. Combined with the current situation and application of modeling and simulation, with a focus on Visual Simulation, the basic concepts, application processes and implementation software are described, and this area of equipment is introduced and combined with a comprehensive analysis to summarize the latest application prospects of the relevant technologies, in order to provide reference for future equipment visualization modeling and simulation.

Transparent objects have two unique optical properties of reflection and light transmission. Current depth cameras cannot perceive them effectively, and the obtained depth images contain a lot of holes and noise information. Robotic grasping and 3D reconstruction of transparent objects becomes difficult. Recent methods mainly focus on depth completion for nontransparent objects, while transparent objects are rarely studied. This paper presents a deep learning approach to obtain a complete depth image from raw RGB-D images of transparent objects. Specifically, our algorithm uses an encoder-decoder architecture, the encoder is used to extract the feature information of transparent objects, and the decoder uses deconvolution operations to complete the depth of transparent objects. In the encoder stage, this paper adopts the Joint Convolutional Attention and Transformer block (JCAT). It consists of two parts: the convolutional attention layer and the Vision Transformer, which is beneficial to extract local and global features. In the decoder stage, Convolutional Block Attention Module (CBAM) is added to enhance feature fusion in channel and spatial dimensions. We conduct a large number of experiments on the ClearGrasp dataset. The experimental results show that our method is reliable.

The exploration of stable diffusion model for generating traditional Chinese painting-style line drawings is a promising field. However, the results directly generated by fine-tuning the stable diffusion model using the line drawing dataset have flaws in details, especially on the face. These flaws include unclear faces, facial feature aliasing, and inability to show expression. Therefore, we propose a method of line drawing generation and face optimization based on stable diffusion and matting technology. The dataset we used for fine-tuning is divided into two parts: the whole character line drawing and the face part of the line drawing, The face mask is obtained by matting after generating line drawing using a fine-tuned stable diffusion model. Finally, the face part of the figure is redrawn in the mask area by inpainting to achieve the face optimization effect. Through the above steps, we can not only generate line drawing but also solve the problems of unclear faces, facial feature aliasing, and inability to show expression.

Knowledge graph can be used to clarify the curriculum knowledge system and describe the knowledge point and its structural relationship. In order to improve the learning efficiency and level of learners in the process of programming language learning, the typical examples related to knowledge can be incorporated into the knowledge graph. In this study, the knowledge graph of C# programming language is constructed as an example: first, the data is obtained; Second, the ontology is constructed and the data is processed. Then use the Neo4j graph database to store knowledge; Finally, knowledge retrieval and visualization are realized, and knowledge graph can be used to assist teaching and personalized learning.

To reasonably allocate the heterogeneous network resources of Internet of Things (IoT) to meet the various needs of users in a cost-effective manner. The study proposes an improved genetic algorithm (IGA) method for finding the optimal solution of heterogeneous network resource allocation. The method first transforms the IoT heterogeneous network resource allocation optimization problem into an objective function to find the optimal solution problem, then uses the improved genetic algorithm to find the optimal solution under the constraints, and finally analyses and simulates the performance of the model for application. The results show that the algorithm stabilises after 10 iterations in the IoT-23 dataset and 14 iterations in the IoT-Healthcare dataset, with stability values of approximately 3.16 and 3.14 respectively, and that the computation time of the algorithm is 8.8m and 3.6m in the two datasets respectively when the number of users is 100, which is much less than the other two algorithms. In the simulation experiments on the accuracy of the calculation results, the study achieved an accuracy rate of over 92% and 94% in A and B respectively. Combining the above data, it can be shown that the method can accurately achieve the efficient allocation and utilisation of resources under the constructed network environment and demand.(Abstract)

The massive data in the power grid system is stored in various databases. However, when users want to query the contents of the databases, they need to write structured query language (SQL), which is difficult for non-professionals. This paper focuses on the research of Text2SQL technology in the power domain based on the Seq2Seq model. On one hand, the BERT model is introduced to classify the queries of power equipment defect texts, solving the problem of inaccurate prediction of aggregate functions in SQL statements. On the other hand, the decoding strategy with fused rule-constrained attention mechanism is employed to mine the data features of the input text, addressing the issue of information loss in input sequences. Finally, the improved Seq2Seq model is proposed by integrating the above strategies. Experimental results show that compared to the Seq2Seq model, the proposed algorithm achieves a 24.88% improvement in accuracy.

From the perspective of increasing the number of objects in the object detection dataset and ensuring their quality, this paper proposes contextual copy-paste data augmentation for object detection, which generates new images that match the objective situation by copying and pasting objects in the dataset. We train a object background classifier based on the background features of the same class objects. Use the trained classifier to classify the image background and paste the object that matches the background based on the classification results. The experimental results on PASCAL VOC 2012 showed that the algorithm we proposed effectively increased the number of objects in the dataset. The mean average precision of the four selected object detectors on the PASCAL VOC 2012 test set increased by about 0.9%, proving the effectiveness of our proposed data augmentation algorithm for object detection tasks.

In order to understand the research hotspots and frontiers in the field of smart healthcare, this study utilized COOC14.1 and VOSviewer1.6.13 software to conduct a visual analysis of 689 core journal articles obtained from the CNKI, VIP, and Wanfang databases. The results of the network diagram indicate that the research interest in the field of smart healthcare has been high in the past decade. The academic community primarily focuses on the application of internet technology in the construction of smart healthcare platforms and the development direction of the smart healthcare field. In recent years, research on smart hospitals and deep learning has become a hotspot and frontier topic. Promoting the development of smart healthcare not only requires policy support but also the integration of advanced intelligent technologies.

In order to solve the problem of database security intrusion identification and rationally plan the query path of database index, a method of database security intrusion identification based on mapping deep learning is proposed. A database security intrusion identification framework is constructed. Authorized users are grouped and calibrated according to the security level. The weights of plaintext keywords are obtained by mapping deep learning algorithm, and the keys of ciphertext are obtained. The encrypted results are uploaded to the main server, and the index of ciphertext is obtained by mapping deep learning algorithm, so as to realize database security intrusion identification. The experimental results show that the proposed method is more real-time for encrypting and decrypting a large number of data in the database, and the maximum recognition rate is 95%, which shows that the proposed method has good recognition effect.

In order to solve the limitations of SLAM system when using RGB-D camera to achieve dense mapping of large outdoor scenes, this paper proposes a visual SLAM dense mapping system based on Pseudo-LiDAR. First, this method does not require a depth sensor and generates Pseudo-LiDAR by a separate module to provide geometric information to the system. The input stereo images are used to generate depth map by employing depth estimation network PSMNet, and the coordinates of each pixel point in the 3D world are calculated by combining the a priori camera parameters, which is Pseudo-LiDAR. Next, the Pseudo-LiDAR and stereo images are input to SLAM for pose estimation using the surfel representation to further achieve dense mapping. The ideas in this paper are experimentally validated on a large outdoor KITTI dataset and demonstrate a more robust dense mapping effect.