This PDF file contains the front matter associated with SPIE Proceedings Volume 12721, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

As the core equipment of the distributed photovoltaic grid-connected system to realize the functions of telemetry, telesignaling, and remote control, the secondary fusion terminal is the key target of network attacks. Based on the typical service features of secondary fusion terminals, this paper analyzes the network flow security threats of secondary fusion terminals in combination with DNP 3, which is widely used in distributed energy systems. Furthermore, the network flow security protection strategy of the secondary integration terminal based on DNP3 is proposed. The experimental results show that the proposed anomaly traffic detection model based on Attention-LSTM can effectively detect the abnormal traffic of the secondary fusion terminal.

Massive wireless debugging terminals and complex and diverse access requirements pose significant challenges to the secure access of substation terminal equipment. It is crucial to detect anomaly network traffic to ensure the security of terminal access to the substation. At present, network traffic anomaly detection based on traditional deep learning often has the problem of low computational efficiency or weak representation ability. Given the low computational efficiency of traditional deep learning, the residual network is used to extract spatial features of data, which can effectively improve convergence speed and time efficiency. Aiming at the problem of weak representation ability of traditional machine learning methods, the long short-term memory network (LSTM) is used to improve the representation ability to learn while learning the temporal characteristics of traffic and prevent the gradient from disappearing and network degradation. Experimental results show that compared with the traditional deep learning method, the accuracy of the proposed method is improved, the F1 score reaches 90.09, and the AUC is up to 0.981. By improving anomaly detection accuracy, the paper further guarantees terminal security.

With the widespread use of micro-application-based architecture for power mobility, ensuring secure authentication of users has become one of the challenges that cannot be ignored. Based on this, this paper first analyses the primary process and problems of the existing token authentication scheme. Then it proposes the overall architecture and model of secure authentication under this scheme for the security of token communication in the authentication process and uses an improved hybrid encryption algorithm for key agreement to solve the limitations of a single encryption algorithm for encrypting tokens. Finally, it can resist network attacks such as frequent replay and man-in-the-middle after security analysis. Experiments also verify that the scheme can guarantee users’ secure authentication.

At present, the time window setting of the multi-entity continuous trust evaluation model in the power mobile Internet environment lacks adaptability. The semantic retention of time series data is significant to time window setting. A trust evaluation model based on the graph model and semantic time window may provide a solution. Firstly, we analyze the entity interactions in the electric power mobile network environment and construct a global trust path diagram by combining the entropy scoring results of evidence data in this paper. Secondly, a semantic time window algorithm is proposed by combining the parameter-seeking process of the sequential data compression algorithm. Finally, update the trust value dynamically by combining the number of feature changes of entities in the global trust path graph and the semantic time window. Simulation experiments are conducted to analyze the trust value updating algorithm based on the graph model and semantic time window proposed in this paper. The experimental results show that the algorithm is feasible and accurate.

With the application of new technologies such as Internet of Things and big data in smart grid industry, new power systems based on new energy sources have emerged in response to the call of “Peak carbon, carbon neutrality.” The power grid enterprise concentrates the data assets and business access to the data external service, which needs to connect a large number of data sources. Ensuring the authenticity of the data without tampering becomes a big challenge. The power system adopts the identity authentication mechanism to resist the security attack and protect the sensitive data. However, in the process of user authentication, the sending of real identity information will lead to the reduction of system privacy, which is easy to cause the leakage of sensitive data. This paper proposes an anonymous authentication mechanism based on zero-knowledge proof for power system, which authenticates the server without revealing the identity. This mechanism uses zero-knowledge proof algorithm to design an anonymous authentication protocol framework, which consists of three stages: registration, mutual authentication and revocation. In this method, anonymous certificate and elliptic curve encryption technology are used to realize the anonymity and authenticity of users. The mechanism effectively protects the user's real identity information and maintains the sensitive data in the power system.

Gas stations are flammable and explosive hazardous places, and safety is the first priority. At present, when gas stations are reviewed for production safety standardization, they mainly rely on reviewers to score on site on the production safety standardization review form, and then conduct manual statistical analysis. The development of a standardized production safety review system for gas stations can overcome the problems of high labor intensity, low work efficiency, and general review accuracy caused by the manual mode. The system is developed by using software engineering methods, following the object-oriented development idea, adopting the development mode of separating the front and rear ends, developing the back end based on the SpringBoot framework, developing the front end using the VUE framework, and selecting MySQL 8.0 as the background database. The system mainly includes review index management module, release review task module, implementation review module, and management review result module. The practical application of the system can improve the efficiency and quality of the standardization review of gas station safety production, so as to further improve the safety management level and informatization level of the gas station.

With the development of computer network technology, network security has become an urgent problem to be solved. The network security management system studied in this paper is based on the Linux platform, the firewall, intrusion detection and other related information technology research and implementation, and through a lot of experiments to verify and optimize the software program. In this paper, we use the K-Means algorithm to solve the problem of multi-objective security policy deployment in the network security management system. Through the comparison of several related network security policy algorithms, it is found that the K-Means algorithm has a higher performance-cost ratio. The main work of this paper is as follows: 1) the data mining algorithm is studied, and the K-Means algorithm is deeply simulated; 2) Analyzed and designed the network security management system based on K-Means algorithm, and carried out detailed functional design, including the configuration of system modules, security management software and database management system; 3) Realized the network security management system based on K-Means algorithm. The system adopts B/S structure, uses K-Me algorithm to optimize the safety management system, and realizes the simulation results by MATLAB software to show that the design scheme is feasible and effective.

Massive botnet attacks pose a serious threat to social stability and network security. To avoid security interception, botnets mainly use Domain Generation Algorithm (DGA) to dynamically generate a large number of malicious domain names to establish communication. Therefore, it is important to study how to detect DGA domain names more effectively, and this paper proposes a method to detect DGA domain names based on multi-scale features. In the domain name feature extraction phase, extracting domain name combination features on a multi-scale convolutional neural network (CNN) based on a compressed activation model. Simultaneously combined with bi-directional gated recurrent unit (BiGRU) to extract domain name sequence features and build hybrid deep learning models to achieve the detection of DGA domain names based on lexical combination generation. The experimental results show that the method improves the F-Score evaluation metric by 7.25% in the binary classification task compared to the CNN-only model, and also has higher detection precision for lexicon-based domain names like suppobox.

Information retrieval (IR) based bug location technology is a relatively recognized lightweight bug location method at present. Most IR bug location methods solve the problem of semantic difference between natural language and code language in the bug report based on code semantic intelligibility, and use semantic similarity to construct IR model to locate source code errors through bug report. However, most IR localization studies take error report description as the guidance for code semantic generation, ignoring the difference between error report and error semantics. Due to the irregular submission of error reports and the ambiguity of error descriptions, this kind of research faces the problem of low location accuracy. We found that the code data is the data written in the specification and verified by the program compilation. Compared with the bug data submitted by the tester, the semantic ambiguity is relatively weaker. Therefore, we use code data as the semantic generation of teacher network training bug data to form SGBL method. In addition, based on the bug data set composed by Jena and other projects, we evaluated the effectiveness of our method and explained the relationship between the semantic extraction method and the bug location accuracy. The experimental results show the effectiveness of the proposed method.

In today's society, the demand for telecom packages is increasing and at the same time, there are many different types of telecom packages, so we need to recommend different types of telecom packages to different users so that each user can find the most suitable package for themselves among the many packages. We choose to build a click-through prediction model to model the interaction between user characteristics and telecom package characteristics. In this paper, we propose the FGCIN model, which generates new features by capturing important feature interactions through a convolutional neural network FGCNN, and then interacts the new features with the original features at the feature interaction layer through a compressed interaction network CIN for higher order features, followed by a deep neural network DNN for implicit feature interactions to finalize the output. In this paper we use the private dataset Telecom dataset and the public dataset Criteo for comparison experiments and ablation experiments, thus demonstrating the effectiveness and rationality of our proposed model FGCIN.

Remote sensing scene (RSS) classification is an important research topic for high-resolution HR remote sensing image understanding. Recently, many approaches have been presented for the task, including data-driven and machine learning methods. However, accurately identifying scenes from HR remote sensing images remains challenging since it is difficult to effectively extract multiscale and key features from the complex geometrical structures and spatial patterns of large-scale ground object. In this paper, we propose a novel local and global semantic relationship network (LGSRNet) for RSS classification. ConvNeXt-T with the same performance as the local vision Swin Transformer is adopted to extract feature map with powerful discriminative ability. Meanwhile, the semantic relation learning (SRL) with graph convolutional networks is presented to further learn semantic relationships between labels of RSS categories within spatial domain. Subsequently, cosine similarity is adopted to incorporate the ConvNeXt-T and SRL. Extensive experiments on two attribute-classification datasets (AID and NWPU-RESISC45) demonstrate that LGSRNet outperforms several other state-of-the-art methods.

Recently, deep learning is widely used in the field of click-through rate prediction. However, deep neural networks can only perform implicit feature interaction on features, which is unknowable. Relying solely on DNN for feature interaction may fail to uncover more information behind the data. At the same time, many click-through prediction models treat all features equally when performing feature interactions. Nevertheless, in a real production environment, some features have a significant impact on the prediction results, while others can be neglected. In this thesis, we proposed the FiDCN model (Feature Importance and Deep Cross Network). It can learn the importance of input features dynamically by giving distinct weights to different features. What’s more, the model also introduces low-rank techniques to capture higher-order feature cross efficiently. We conducted detailed comparison experiments with the classic models in the industry and ablation experiments on different datasets, and the results showed that FiDCN has good performance.

Active protection simulation is a part of the field of system-of-systems confrontation simulation, and is the content of extensive research in the field of system-of-systems confrontation simulation at home and abroad. The main objective of the research is to accurately describe the active protection process and support the rapid development of computer simulation. This paper proposes an active protection simulation method, and implements an active protection software based on this method, which consists of armor protection simulation component, incoming ammunition detection and warning simulation component, soft killing active protection simulation component, hard killing active protection simulation component, etc. It can simulate the active protection process of armored target against typical ammunition. It not only considers the relevant elements of the protection process, but also facilitates the software implementation. It effectively supports the simulation of system-of-systems confrontation.

Reducing the significant communication overhead in large-scale distributed model training is a current research hotspot. Some existing studies have reduced the communication cost from workers to parameter servers (upstream) by utilizing the in-network aggregation capabilities of programmable switches or by pruning models to reduce overall communication costs. However, none of them have investigated how to accelerate downlink communication for multiple distributed model training in data centers, given the constraints of programmable switches and links. We propose an algorithm that achieves an approximate ratio of 𝑂(logκ/β + 1). Through simulation, we demonstrate that our algorithm outperforms previous multicast methods.

Ghostnet is a lightweight image classification network proposed by Kai Han and other scholars in 2019 which was applied in terms of object detection. Ghostnet has the advantages of few parameters and high accuracy. In this paper, through the idea of transfer learning, ghost net is used as the network of feature extraction and CBAM attention mechanism is added. CBAM is not only a lightweight attention mechanism, but also a good combination of space and channels. CBAM eliminate irrelevant noise intelligences and concern to key intelligences. At the same time, the ghost convolution block is improved, and point-by point convolution is added to the ghost convolution block to obtain more image feature information. Add the squeeze and enumeration module to improve the convolution receptive field, increase the ability of the network to extract multi-scale spatial information, and introduce the residual idea to tackle the problem of information loss and gradient descent disappearance. GAUnet, designed according to the above idea, achieves better intersection and merger ratio than Unet, Unet++, DeepLabV3+and other neural networks on the multi-organ segmented chaos dataset with fewer parameters.

To ensure the stability of image guiding head control, an improved genetic and particle swarm hybrid algorithm optimized RBF fuzzy neural network tuned PID control strategy was proposed for the correction of the image guiding head stable loop. According to the kinematics and dynamics of the image guiding head, combined with the RBF fuzzy neural network PID control algorithm, a simulation model of the image guiding head stable loop was established. The simulation results showed that after the image guiding head stable loop adopted the RBF fuzzy neural network tuned PID controller optimized by the improved hybrid algorithm, its dynamic performance was better than the traditional PID controller, and the established simulation model could well eliminate the nonlinear disturbance and improve the robustness of the stable loop, which could provide certain reference in engineering application.

In view of the vulnerability analysis requirements of system targets, this paper puts forward the concept of system target vulnerability, defines system target vulnerability, analyzes the difference between system targets and single targets, and puts forward the analysis process of system target vulnerability, including preliminary determination of system target range, determination of vulnerability purpose and level, determination of key entities, construction of system damage tree, determination of damage weight and damage criteria, based on the analysis of single target vulnerability research results and other steps, this method can guide relevant work.

With the rapid development of cloud services, more and more people use cloud services to store personal image privacy. However, cloud services lack security, which leads to the leakage of personal privacy. Although some scholars have proposed that images be encrypted before being uploaded to cloud services, it will cause the images to lose their availability. Recently, Zhao et al. proposed a thumbnail preserving encryption scheme, which can ensure the security of the image while maintaining the visual usability of the image, but the encryption efficiency of the algorithm is low. Therefore, this paper proposes a thumbnail preserving encryption scheme based on bit plane scrambling. In this scheme, a new one-dimensional chaotic system is proposed for the image chaos process. An improved Zigzag transformation method is proposed to change the position of pixels. The control parameters in the new one-dimensional chaotic system have a large value range, which ensures that the key space of the encryption algorithm is large enough. The improved Zigzag transformation method adds random numbers to improve the randomness of the encryption process. Experimental results show that the scheme has high encryption efficiency, and can make the encrypted image retain the visual usability of the original image, balancing security and usability.

In recent years, with the rapid development of China's network scale, the frequency of power Internet has also increased sharply. But at the same time, in the process of power grid intelligent terminal access, all kinds of abnormal network attacks are increasingly frequent. As a key link in the power network security protection, the detection of network abnormal intrusion behavior has been paid more and more attention by researchers in recent years. In view of this, this paper proposes a network anomalous intrusion behavior detection method based on FATE-CNN. It corresponds several local intrusion detection datasets to federated learning devices one by one, and uses a dynamic local iteration method to gradually obtain the best global model. Through the comparative experiment with four intrusion detection models of LIBSVM, CNN, DNN and DBN-EGWO-K-KELM, it is verified that the model algorithm can effectively improve the value of abnormal network intrusion behavior of NSL-KDD, UNSW-NB 15 and CICIDS2017, and improve the network security of the intelligent terminal of the power grid.

The development of low-code applets is an important means for traditional enterprises to achieve their digital transformation. The low-code applets reduce the demand of clients for code development technology by providing them with a visual application development environment, and construct a digital operation mode to further accelerate the pace of enterprise refinement construction. This paper analyzed the development trend of seamless replacement of fusion templates for low-code applets, and used development technologies such as cascading style sheets (CSS) and JavaScript to implement the design of seamless replacement of templates. Based on the three-layer architecture of presentation layer, business logic layer and data access layer, as well as the “Model-View-Controller (MVC)” model, the author studied the seamless replacement of fusion templates for low-code applet site builders, and verified the builder functions through black-box testing, realizing the seamless replacement of templates.

According to the requirement of multi-network security isolation information exchange, a multi-network isolation netgap is conceived, and the exchange unit in the netgap is newly designed. A data exchange system based on the RapidIO bus is designed in the unit, so that the exchange unit can be externally connected to four network processing units to exchange four-way data, thus solving the problem of data exchange of multi-network isolation. The system realizes high-speed data transmission and exchange with the RapidIO interconnect devices with FPGA, and the devices are connected with a star topology through a RapidIO switch. The actual test results show that the system can realize the information exchange function between the four endpoint devices, and the system data transmission rate can up to 3.4GByte/s, which exceeds the exchange unit performance of most isolation netgaps and can meet the design requirements of multi-network isolation netgap.

Used in multiple industries, recommender systems enhance the user experience by creating personalized item suggestions. However, without bias awareness, the system could output unfavorable results, leading to the item under-recommendation bias. To decrease this type of bias, creating a fair top-k list (i.e., list of top-ranked items by the user) is needed. Based on the debiased personalized ranking model (DPR) proposed in the works of Zhu et al., the paper aims to reduce item under-recommendation bias in recommender systems in the Bayesian Personalized Ranking (BPR) recommendation model. This is done by applying several fairness metrics, creating fairness through an adversarial component, and using an autoencoder on the data. The advantages of the new proposed model are proven through two sets of experiments based on multiple datasets and baselines, respectively. This results in a solution that is dozens of times more efficient compared to previously proposed ones.

The core problem of point cloud registration can be explained as computing rigid transformations to align two point clouds, which is a key technology used in popular fields such as three-dimensional reconstruction, robotics, and autonomous driving. We come up with a hybrid feature-based model inspired by Deep Closest Point (DCP) and Robust Point Matching using Learned Features (RPMNET). The main innovation of this model is to combine abstract features extracted by Dynamic Graph CNN (DGCNN) with Point Pair Features (PPF) as hybrid features for point cloud registration, after that, soft matching is performed between two point clouds, and then singular value decomposition (SVD) is applied to compute the rigid transformations. Besides, we adopt the ModelNet40 dataset for training and compare the trained model with DCPV2, Iterative Closest Point (ICP) and some other ICP variants, the comparison of results indicates our model performs better than the above methods in predicting the angle of rotation when rigid transformations occur. We also test our model on clean and noise-added test sets respectively to verify the robustness of it.

Network traffic classification technology is an important means for power Internet of things to carry out network management and maintain network security. However, there are many existing traffic classification methods. Different traffic classification methods face different data sets, and the data sets used for training are limited, the update is slow, and the change of traffic characteristics is not obvious. Therefore, based on passive detection technology, this paper uses traffic analysis as a tool to collect the lossless traffic data of the target network, and then uses reinforcement learning Q-learning algorithm to classify the traffic and design the corresponding return function, and adopt ε-greedy exploration strategy and delayed return strategy to improve the learning effect of agents and improve the accuracy and efficiency of classification to a greater extent. Finally, the feasibility of the system is verified by experimental simulation. After 100 days of training, the classification accuracy has exceeded 85%, and with the increase of training time, the classification accuracy will be further improved.

In the single-type relational database scenario, using SQL statement optimization rules can effectively shorten the statement execution time. In the heterogeneous database scenario, the effectiveness of SQL statement optimization rules needs to be further explored. Connect the relational heterogeneous database through Calcite, and use predicate push down, constant transfer and sub-query de-nesting rules to optimize SQL statements. The experiment shows that the optimized SQL statement can effectively shorten the execution time in the scenario of relational heterogeneous database.

Aiming at the problem that people are afraid to choose Chinese medicines, do not know how to choose Chinese medicines and cannot use Chinese medicines correctly because of the lack of scientific understanding of Chinese medicine, a practical intelligent question and answer system of Chinese medicine knowledge was constructed. The system provides people with reliable and convenient Chinese medicine knowledge inspection and intelligent question answering service to improve the role of Chinese medicine knowledge in safeguarding people's health. Firstly, extract and clean the Chinese medicine knowledge in Chinese medicine books, set the extracted Chinese medicine knowledge as a specific format, and build a knowledge graph. Secondly, filter the natural language question, use Aho-Corasick multi-mode matching algorithm for keywords to extract, then use the cosine similarity algorithm to calculate the similarity calculation. Finally, match the question and answer template and generate the answer. Through the system test, it is proved that the system can satisfy the needs of users to obtain reliable Chinese medicine knowledge service quickly.

Aviation Support System (ASS) is a typical kind of complex system for supporting aircraft take-off and landing. The operation of ASS involves with man-machine-resource-environment system. The sensing data from ASS shows the multi-source characteristics. It is difficult to characterize the ASS status accurately and efficiently using such data source. Aiming at building intelligent ASS to improve efficiency and safety in complex environment, a multi-source data fusion method for ASS is proposed. According to the functional requirements in real-time perception, a unified architecture of multi-source data fusion for ASS is designed. The data fusion architecture is integrated with data preprocessing, association and aggregation. Finally a case study is proposed to verify the data fusion method for ASS.

Aiming at the problem that the current traditional fieldbus cannot meet the performance requirements of complex industrial environment, a solution is proposed using LAN9252 chip as EtherCat slave controller and ARM chip as microprocessor. First, the LAN9252 is used as the slave control module, responsible for data reading, writing and verification; the STM32F767 is used as the protocol processing module, responsible for protocol processing and communication of external servo devices. Secondly, realize the LAN9252 software design for data interaction between master and slave, and realize the STM32F767 software design for protocol processing of slave and protocol interaction of external servo. Finally, the slave test platform is built to test the communication function, real-time, stability and servo communication of the slave. From the test results, the EtherCAT slave of this design communicates normally with the master and external servo devices, and the slave refresh time is 15.75us, which can run stably for 14 hours. Therefore, the design can basically meet the performance requirements in complex industrial environments.

Questionnaire survey systems are a kind of computer-aided survey tool, which are helpful for the questionnaire publisher to collect the opinions and appeals of the respondents. Nevertheless, it is difficult for the questionnaire publisher to solve tricky questions such as respondents’ inactivity in answering questions, especially in the subjective questionnaire. The paper extends the privacy-friendly incentive protocols of Camenisch et al. based on the technique of signatures of knowledge and the Au-Susilo-Mu signature and put forward a subjective questionnaire system oriented to professional course teaching. The new system controls the range of participants by specific attributes. It also allows participants to remain anonymous at all phases, so they can truly feedback on their subjective evaluations. Moreover, when students are not within the scope of the survey, they are encouraged to help to review the collected subjective questionnaire answers. Furthermore, the new system introduces a token mechanism to stimulate students’ participation. Security analysis shows that the new system satisfies correctness, balance, anonymity, and unlinkability.

Side channel analysis (SCA) is an emerging password analysis method for the password chip. The physical information leaked by monitoring the password algorithm on the physical chip will be obtained by processing the obtained signal to extract the sensitive information in the encrypted device through processing analysis. In recent years, the rapid development of power consumption analysis methods in SCA has become a powerful side channel attack method. Therefore, the ability to protect password equipment from SCA's infringement has gradually become the focus of the information security industry. This article first elaborates the research status of the field of side channel analysis, introduces many problems in the field of side channel analysis, and proposes an improved power consumption analysis method for power analysis attacks under the condition of less power trace analysis. This method has successfully reduced the number of power trace points required by the recovery of the correct key in the AES-128 encryption algorithm encryption, and can still ensure a higher classification accuracy in the case of a small amount of power traces.

The increasing use of computers in various fields has led to the popularization of cloud computing due to its advantages, such as large scale, high performance, low power consumption, high reliability, and low cost. However, the rapid growth of business volume and kinds of computer hardware resources, along with complex computer architecture, poses challenges in allocating hardware resources. Traditional methods struggle to meet the varying requirements of different businesses, hinder automatic tuning, and limit the improvement of overall system performance. In this regard, we propose an automatic QoS-aware resource partitioning framework that aims to maximize the overall performance of the system. Our contributions include an automatic performance tuning framework for cloud environments, a Deep Q-learning Network (DQN) based performance optimization method, and achieving a speed-up of at most 1.73 times compared to uniform partitions for CPU overload scenarios involving throughput-aware and latency-critical workloads. The proposed framework addresses the challenges faced by cloud computing by adopting different allocation methods for hardware resources that maximize performance in different application scenarios. The framework efficiently utilizes emerging hardware resource control capabilities to improve hardware and system performance. The use of DQN in performance optimization allows the framework to learn from past experiences and adapt to different situations, resulting in better resource allocation decisions. The proposed framework can significantly reduce the waste of resources and unnecessary expenditures for governments and enterprises. Our framework can serve as a guide for future research in cloud computing and related fields.

User identity authentication of computer systems can be seen everywhere in daily life. Most systems use traditional onetime authentication methods, such as password, fingerprint and face recognition based on biometrics. However, there is a risk of password being cracked or forgotten, and fingerprint and face recognition require specific hardware devices. For the one-time authentication method, once the intruder passes the one-time authentication, the operation after the intrusion of the system will no longer be restricted. Human operation behavior can be regarded as a biological behavior feature, which cannot be stolen or copied, and can be used for identity authentication. If the user's mouse click and slide is abstracted into a behavior model, the model can be used to judge the user's identity and achieve user identity authentication. This paper studies human-machine recognition model based on mouse operation behavior. This paper deeply analyzes the advanced features of mouse click and slide, and uses machine learning algorithm to train mouse operation behavior model to identify whether the mouse is operated by human or software. This method can effectively detect the malicious behavior of hackers who use software to operate the mouse to automatically log in to the system.

In view of the problem that the cost, quality and progress of the traditional prefabricated housing can not be clearly understood and grasped before the construction, BIM technology is introduced to carry out the application research of the technology. Firstly, on this basis, the site layout design and simulation of the combined building are carried out. Secondly, the lifting process of assembling component is decomposed and simulated. Finally, the decomposition and construction simulation of grouting operation are carried out. Through this process, the virtual construction can be simulated, so that before the construction, you can have a comprehensive understanding of each construction link, so as to ensure that the cost, quality and progress of the construction can be as close to the predetermined goal as possible.

The fault alarm system is to help the duty officer to check the current operation status of each equipment in the automatic hydro-meteorological observation system and whether the observation data are at normal range in real-time. The system monitors the operation status of automatic hydro-meteorological stations in real-time and customizes fault rules to check the quality of observation data. The system generates fault information via fault diagnosis and provides alarms via sound, animation, and SMS. This paper describes the hydro-meteorological automatic observation composition, data center fault alarm system design, fault diagnosis and alarm process, and demonstration applications.

A nasal pattern recognition method based on multi-feature fusion is proposed to address the problems of imperfect feature information extraction, low recognition accuracy and interference of redundant information in nasal pattern images by a single method. An improved two-channel attention mechanism (I_CBAM) is introduced in the residual network to reduce the interference of redundant information; the output feature information of Layer2, Layer3 and Layer4 in the fusion depth residual network structure is used to enrich the extracted global features of the image by using the complementarity between feature maps of different scales; meanwhile, extract the underlying local features with better matching in the nasal pattern image using the improved SURF algorithm, and fuse the extracted global features with the local features; the training of the model is supervised using the improved fusion loss function. The experimental results on the nasal pattern dataset show that the recognition accuracy is improved compared with other mainstream recognition methods.

Clothing pattern recognition on social media is a key application of Internet marketing, but it is currently implemented manually, which is very inefficient. Our goal is to solve this problem through artificial intelligence. Based on the improved Mask RCNN network, this paper introduces the attention mechanism SENet to enable the feature extractor to extract the target areas that need to be focused on. And put more weight on this part to highlight significant useful features. And this article contributes a whole new dataset of clothing versions. The comparison experiment verifies that the improved Mask RCNN network has been significantly improved in the pattern recognition of clothing.

Different from image-based pedestrian re-identification, video-based pedestrian re-identification has more available information because of more temporal information. However, previous methods are unable to extract detailed features adequately. Although the previous methods adopted the row slicing method to extract the detailed features of the image, the slicing method was too simple and the detailed features extracted were too few. We propose a slice position shifting module (SPM), which adopts a variety of ways to slice the image, to extract the feature of each image patches separately. The feature interaction is carried out by making each patches of the image generate position-shifting. How to make time clues in pedestrian re-identification data set useful is also the research focus in this field. We propose a multi-scale inter-frame proportional fusion module (MIPM) to make the features of frame fully fuse with the front and back frames, so as to extract more the temporal clues in the data set. We will form the above two modules into slice position-shifting and multi-scale inter-frame proportional fusion network (SMnet). Full experiments on the MARS dataset show that our algorithm is very useful and can reach the leading level.

In order to meet the demand of accurate attitude measurement of high-speed targets, this paper proposes a high-speed target attitude recognition method, which includes ammunition attitude information measurement, small target rapid extraction from image sequences, ammunition axis image processing, target gray level recognition, etc. On the basis of multi camera shooting of targets, this method effectively identifies the terminal trajectory characteristics of ammunition through image fusion and image processing technology, Based on this algorithm, relevant algorithm models and software can be developed to guide practical application.

Due to the incompleteness of knowledge map, the existing knowledge map is usually missing, and the task of knowledge reasoning is to predict the missing facts according to a large number of information such as entities and relationships. In order to satisfy the interpretability and accuracy, this paper proposes a multi-hop reasoning method combining first-order logic and graph embedded vector, which has the mathematical rigor of logical reasoning and the high accuracy of embedded vector. For the specified query, it is transformed into a first-order logical operation representation, and the query dependency graph is instantiated, and the reverse sampling from the root node to the anchor point in the instantiation process ensures that the query must have an answer. According to the query calculation diagram combined with logical operators, reasoning is based on embedded method, which avoids explicitly traversing KG and obtains the query answer.

Automatic summarization technology uses some concise sentences to summarize the general idea of the article so that readers can understand the article's main content only by reading the abstract. Instead of simply selecting and rearranging sentences from the original text — just as the general extractive summarization does, the rewriting system model based on seq2seq applies extracted sentences as input to generate more consistent summaries with human language conventions. The BERT pre-trained model has also been applied which achieved good results. To increase coherence, we propose the Rewriting Summarization Based on Sentence Order Prediction (SOP) and Scoped-Context: after the extractive summarization, we use the SOP tasks of the ALBERT model to reorder the sentence sequence; In the abstractive summarization, we apply the group sub-tag-based attention mechanism problem to the seq2seq situation. To further reduce redundancy and irrelevance, each extracted sentence is taken as the input of the rewriter altogether, with its context within a specific scope. Our method can enhance model performance, achieve higher ROUGE scores, and maintain lower computational complexity.

A knowledge graph is a special kind of graph data, which consists of a triad. Each node in the knowledge graph has several attributes and their attribute values. The storage of the knowledge graph has been the object of academic research, and in this paper, we conduct an in-depth study on the knowledge graph data indexing and compression storage algorithm supported by the RDF graph model, and propose an optimization algorithm for the storage query after the second-level compression. The core of this paper is that after the second-level compression of the k2-tree tree, the sub-matrices are prioritized in terms of the size of data blocks, and when retrieving data, they are retrieved according to the priority, so that the blocks in front are both subject and object at the same time, which can improve the efficiency of data reading, so that the parts with more information will be retrieved first, instead of the traditional sequential retrieval, which tends to retrieve the null values or the data with less information.

The label is used to mark the object information or classification of the logo, complex label then contains more layout information than ordinary labels. These include Chinese and English characters, pictures, symbols and barcodes. When text detection is performed in complex labels, the detection capability of the algorithm is reduced and is prone to feature overlap resulting in missed and false detection. In this paper, the original YOLOv5 algorithm is optimized by an improved channel attention mechanism to form the new algorithm SA-YOLOv5. The experimental results show that the SA-YOLOv5 model can effectively improve detection efficiency and solve the problem of false and missed detections. The average detection accuracy mAP value is 92.24%, which is 0.68 percentage points higher than the original algorithm, proving the effectiveness of the algorithm.

Audio classification is widely used in our daily life. However, computing resources are limited and many audio classification datasets are generally small especially when compared to the image domain, which limits the training of large networks. Therefore, it is necessary to construct a small network with few parameters but can achieve good performance. In this paper, we propose a small Audio Classification Residual Convolutional Neural Network (AcrcNet) which is composed of two main feature extraction blocks: a time domain feature (TDF) extraction block and a high-level residual feature (HLRF) extraction block. In the TDF extraction block, the time-consuming time-frequency domain conversion is replaced with one-dimensional CNN so that time domain signals can be used as input directly. In the HLRF extraction block, we propose a Residual Convolutional (RC) module which not only deepens the depth of the network but also eliminates degradation phenomenon by using residual learning. In addition, the avgpool layer is applied at the end of the HLRF extraction block to process input of any length. In the experimental section, we use Between-Class (BC) learning to achieve good performance on the Environmental Sound Classification (ESC-50) dataset. The results of 5-fold Cross-Validation (CV) indicate that the performance of the proposed AcrcNet is better than other state-of-the-art small networks.

Superpixel segmentation is an image preprocessing technique that uses pixel blocks instead of pixels to improve the efficiency of subsequent image tasks. Existing methods are not sensitive to image texture. To solve this problem, a texture-oriented superpixel (TOS) segmentation method is proposed. Firstly, an adaptive parameter function based on pixel boundary probability is used to calculate the distance. Secondly, a gradual merging strategy is used for merging. And finally, a loop iterative framework is used to optimize the seeds. The experimental results show that TOS can effectively preserve the smaller texture information, generate seeds that are more consistent with image texture, and better detect image texture.

Inspired by the renewal process of elephant clans, elephant herding optimization (EHO) is proposed and successfully applied to many optimization problems. However, EHO is prone to fall into local optimum during the optimization process, resulting in poor global search ability. To overcome this deficiency, an improved elephant herding optimization (SCEHO) is proposed, which integrate good point set strategy, improved sine cosine search strategy, nonuniform gauss mutation strategy and greedy strategy. First, good point set is introduced to initialize the population. Secondly, sine cosine algorithm is improved and incorporated into EHO to update the position of individuals. Then, the separation operator is improved from two aspects, on the one hand, non-uniform Gauss mutation is introduced to separate individuals. On the other hand, the number of separated individuals is increased, and half of the individuals with poor fitness are selected. Finally, the greedy strategy is introduced into SCEHO to update population according to the fitness. Our SCEHO is tested on 10 benchmark functions from CEC 2019 and the results imply the superiority of SCEHO to other algorithms. SCEHO is also applied to solve WSN coverage optimization problem. The results show that coverage model based on SCEHO has higher coverage and more uniform node distribution.

In order to mine the medication rules of traditional Chinese medicine in treating thyroid diseases and analyze the performance of association rule algorithm in traditional Chinese medicine data, the improved Apiori algorithm based on transaction compression technology and hash technology is introduced. Furthermore, an improved FP-growth algorithm is proposed in view of the characteristics of more transaction duplicates and longer single transaction in the TCM data transaction database. Combined with these improved association rule algorithms, the data mining analysis of traditional Chinese medicine is carried out, and the algorithm time efficiency diagram is drawn to analyze the algorithm efficiency. The results show that the mining results of association rules of all algorithms are consistent, which reflects the basic treatment principles of TCM in treating thyroid diseases by regulating phlegm and eliminating galls.

In this paper, we present a blind signal separation method for time-frequency overlap communication signals. The method is derived in the general context of independent component analysis, which we refer to as the signal minimum projection algorithm (SMP). Unlike conventional blind signal separation algorithms that use signal independence, we determine whether the signal is independent by observing the length of its projection on each axis of the scatter plot, and give a detailed proof. Simulation results show that this algorithm runs fast while maintaining good separation performance compared to other algorithms.

In this paper, we propose a crop planting recommendation algorithm based on ensemble learning to recommend the most suitable crops for farmers based on environmental characteristics and soil element content, which can achieve scientific planting and crop yield increase. Firstly, we adjust the scaling ratio of N (nitrogen), P (phosphor) and K (potassium) elements, which play an important role in crop growth, and use KNN (K-Nearest Neighbor), XGBoost and RF (Random Forest) as weak learners, and use GA (Genetic Algorithm) to optimize the important parameters for KNN, XGBoost and RF, and combine these three weak learners to obtain the ensemble learning model through a soft voting mechanism. After training and testing on the Kaggle public dataset, the accuracy of the crop planting recommendation algorithm based on ensemble learning can reach 94.36%.

Due to the high-dimensional structure parameters and time-consuming of numerical simulations, it is hard to obtained the optimized structure parameters of metasurface structures. To address this issue, in this paper, in order to maximize the performance of the proposed asymmetric polarization converter, a deep learning with heuristic algorithm approach is applied to search for the optimal set of structure parameters. With optimized structure parameters, the average transmittance is larger than 0.5 for the wavelength range from 447 to 475nm and a maximum transmittance of 0.623 at 457.5nm. The bandwidth of transmittance larger than 0.5 is 6nm wider and the maximum transmittance is 0.02 higher than those of the reinforcement learning model we used for the parameters’ optimization of the same structure in our previously published papers respectively. The optimization time is about 10 minutes, which is only 50% of that used in the reinforcement learning algorithm.

Video summarization is used widely in the field of fast browsing and retrieval of videos by generating keyframes or segments to achieve video compression. Existing methods mostly explore based on image content, ignoring the temporal characteristics of videos, resulting in summaries lacking temporal coherence and representativeness. We propose a video summarization network based on an encoder-decoder framework. Specifically, the encoder part extracts features using a convolutional neural network and enhances the weight of key features through self-attention mechanism. The decoder part consists of a bidirectional long short-term memory network fused with a random forest, and adjusts the proportion of the random forest and bidirectional long short-term memory network in the loss function to make the model more stable and accurate in prediction. Experiments were conducted on two datasets and compared with seven other methods, and the comprehensive experimental results prove the effectiveness and feasibility of the proposed method.

The transmission success rate of the traditional dynamic allocation method of communication spectrum resources is easily affected by the environment. Therefore, mobile edge computing technology is introduced to conduct a comprehensive study on the dynamic allocation of communication spectrum resources of high-speed railway automatic control system, establish the communication model of automatic control system, and realize the dynamic allocation of communication spectrum resources based on mobile edge computing, realizing the frequency compatibility between networks and efficient allocation of spectrum resources. The test results show that the link transmission success rate of this method can reach more than 97%.

In systems where unmanned navigation and manned behavior of ships coexist, the intelligence-driven unmanned ships lack the adaptability to complex environments that humans exhibit when faced with realistic navigation environments and will affect the decision judgment of manned ships. To solve the problem, a knowledge distillation-based multi-intelligent body maritime collision avoidance navigation method for human-like behavior is proposed. The method first preprocesses the AIS data, and the core technique is to use the knowledge distillation method to combine the expert strategy with the reinforcement learning method, which introduces the human piloting ship habits and makes the ship’s automatic navigation exhibit human characteristics. The test results show that the unmanned ship trained by the method can learn human driving ship habits outside the reinforcement learning reward function setting, and enhance the training efficiency of reinforcement learning.

The popularization of the Internet of Things and cloud computing has provided powerful conditions for the construction of agricultural digital management systems. Among them, the Internet of Things has advantages such as efficient resource utilization, convenient data collection, and security; Cloud computing stands out for its secure, inexpensive, and stable computing and storage capabilities. Therefore, design agricultural digital management systems under the Internet of Things and cloud computing, build a network environment based on the Internet of Things, manage software computing processes through distributed processing means, and maximize the data attributes of agricultural digital management systems. The scientific design system framework includes a management system application layer, a cloud computing network layer, a big data access layer, and a big data awareness layer. This article focuses on designing three basic frameworks for the cloud computing network layer, and designs the logic and functional entity structure of the system's functional modules based on the construction requirements of the Internet of Things and cloud computing for agricultural digital management systems. The research results show that the designed agricultural digital management system has a high technical level and a high utilization rate of system resources, which is optimistic.

In order to quickly and accurately identify forest fires, a forest fire detection method based on CenterNet without anchor frame was proposed. Firstly, lightweight backbone CSP-VoVNet is used to improve the backbone network and improve the detection speed. Secondly, the multi-scale fusion was carried out by introducing the weighted bidirectional characteristic pyramid BiFPN of ECA to improve the detection accuracy of small fireworks. Finally, Smooth L1 was used to regression center point offset to improve model robustness. The experimental results show that the method can accurately identify the small size forest fireworks and realize the early warning effect under the premise of high detection speed.

The evaluation of street view perception plays a crucial role in the design and improvement of urban street layouts by urban planners. However, the most commonly used method for evaluating street views by planners and researchers is through questionnaires, which may not effectively capture users' psychological feelings. Currently, there is a lack of research on providing suggestions for improving street views that score low in perception evaluation. To address this issue, we propose a method for highlighting areas where street views are insufficient in six aspects which are beautiful, safety, wealthy, lively, boring and depressing to provide suggestions for improvement. Reference pictures and corresponding improvements suggested by the proposed system are necessary as they can provide urban planners with intuitive suggestions for street improvement. In this study, we recruit volunteers to rate street view maps in the six aspects mentioned above. We then use the scene graph generation method to characterize the relationship between objects in the street view images. Finally, we apply the graph-matching algorithm SimGNN to identify three pictures that are highly similar to the graph structure of the high-score street views as reference images. This approach effectively provides suggestions for street view images with low-score, while also offering an intuitive way to improve street view perception for urban planners. Overall, our proposed method provides a more comprehensive and effective way to evaluate and improve street views, which can contribute to better urban planning and design.

To address the problem of low efficiency in test case generation, an Elite Opposition-Learning Particle Swarm Optimization Based on Selection and Mutation Strategy (SM-EOLPSO) is proposed in this paper. Firstly, nonlinear decreasing inertia weight with random offset is set so that the search ability can be adaptively adjusted to the situation. Secondly, opposition-based learning is performed to enhance global detection ability and improve population diversity; meanwhile, selection and mutation operations in genetic algorithm are introduced to speed up convergence and prevent falling into local optimal solutions. Finally, the branch distance is used to construct the fitness function and evaluate test cases. Experimental results show that the algorithm is competitive in terms of the number of iterations and generation time for automatic test case generation.

Based on cadaver data on non-lethal kinetic projectiles impacting the human abdomen, the effectiveness of the THUMS AM50 Dummy Finite Element Model has been established. Using the Abdominal Injury Criterion (AIC) and Abdominal Peak Force (APF) to assess the abdominal injury, it has been determined that frontal impact of projectiles with a mass of 45g and velocity of 60m/s on the human abdomen results in a 50% probability of liver injury at the AIS2 lever or above. After comparing two different abdominal injury assessment criteria, it has been observed that the existing criteria show significantly different levels of adaptability under varying impact conditions.

In this paper, we propose a new MRMRKA-SVR-GARCH model to predict the volatility of time series. This model is a combination of the MRMRKA-SVR model and the GARCH model. Since SVR-GARCH model has been successfully employed to forecast volatility. Empirical results show that the multi-kernel support vector regression based on minimal redundancy maximal correlation criteria and kernel target alignment (MRMRKA-SVR) tend to beat single-kernel SVR models in terms of forecasting accuracy. First, we use the MRMRKA algorithm to select a set of kernels to establish a multi-kernel, and establish the MRMRKA-SVR model. Then we use MRMRKA-SVR to estimate the volatility equation of GARCH model, and establish MRMRKA-SVR-GARCH model. Furthermore, the models’ predictive ability was evaluated using three empirical analyses on China Unicom's stock price data and stock price data of ICBC. Under different basic kernel numbers m, compare the prediction errors between the proposed model and the SVR-GARCH model based on the mixed Gaussian kernels in the two stock data, the results show that the prediction accuracy of the MRMRKA-SVR-GARCH model fluctuates slightly with the change of m , and the MAE and RMSE of the MRMRKA-SVR-GARCH model are lower than the mixed-Gaussian-kernel SVR-GARCH model. Description MRMRKA-SVR-GARCH model managed to outperform the mixed-Gaussian-kernel SVR-GARCH model. Therefore, the MRMRKA-SVR-GARCH model can provide a certain degree of reference in analyzing the trend and prediction of volatility, and has high feasibility and effectiveness.

This paper proposes an algorithm based on an improved version of YOLOv8l, which is designed for small target detection on flapping wing drones. By adding a small object detection layer and introducing Multi-head self-attention (MHSA), the algorithm effectively reduces interference from irrelevant backgrounds and enhances the network's feature extraction performance. Experimental results on both the flapping wing dataset and the VisDrone dataset demonstrate that compared with the baseline YOLOv8l algorithm, the improved algorithm shows a 4.1% and 3.3% improvement in the mAP@.5 and mAP@.5:.95 indicators, respectively, and a 5.9% and 4% improvement on the VisDrone dataset. Particularly noteworthy is the improved algorithm's performance on the mAP@.5 index, which achieved 50.1% on the VisDrone dataset, proving its robustness and exceptional performance in small target detection. These results illustrate the algorithm's effectiveness and practicality, making it a valuable tool for flapping wing UAV vision applications.

There are few uniform models for networks in the real world. This article selects the communication network for study. Usually, when the communication base station exceeds its capacity, a common measure is to transfer the overloaded part of the node to relieve the communication system. This paper introduces the logistic function as the probability function of node failure and develops a cascading failure model with the features of the communication network. Based on the wireless mesh network (WIN) also known as multi-hop mesh network (MHMN) technique to transfer the overloaded part, a new measure which is expanding the node capacity to impact the network vulnerability is discussed. Case study demonstrates the existence of a threshold value for node capacity expansion factor keeping a good balance between cost and network vulnerability and proves the threshold value in the communication network in the real world.

Flexible job-shop scheduling problem (FJSP) is an extension of classical job-shop scheduling problem, which is a typical NP difficult problem with complex modeling and solving difficulties. In order to solve the problems of long coding length and low efficiency of solving FJSP by two-layer coding genetic algorithm, a FJSP solving algorithm based on automata and genetic algorithm is proposed in this paper. For machining automaton model is established, and then according to priority machine automata model generation chromosomes, through genetic algorithm for scheduling policy. Simulation results show that the proposed algorithm is faster and more efficient than the traditional two-layer coding genetic algorithm.