This PDF file contains the front matter associated with SPIE Proceedings Volume 12937, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The use of a finer-grained classroom language analysis coding form facilitates a more in-depth analysis of teachers' classroom language behavior. This study uses a finer-grained T-coding system to encode texts converted from teachers' near-field recordings, makes a preliminary exploration of multiple methods of visual teacher teaching strategies for coding sequences, and writes Python scripting software based on machine learning algorithms. This software can effectively assist teachers in self-observation and improve teaching quality.

Most of the current big data mixed attribute feature detection structures are set to be one-way, and the detection coverage is greatly limited, leading to a decline in the precision of big data feature detection. Therefore, the design, verification and analysis of big data mixed attribute feature detection method based on Kalman algorithm are proposed. According to the current feature detection, the basic feature decomposition is carried out first, and the multi-level method is adopted to break the limit of detection coverage, and the multi-level data clustering and index optimization detection structure is designed. Based on this, the Kalman measurement big data mixed attribute feature detection model is constructed, and the dual redundant data feature fuzzy processing is used to achieve feature detection. The final test results show that: after the comparison of the selected five test areas, combined with the Kalman algorithm, the precision rate of big data feature detection can reach more than 85%, indicating that this feature detection scheme has better practical application effect, stronger pertinence, more extensive detection coverage, and has practical application value.

To address a series of issues existing in the current incubators, including cumbersome data processing, limited programming capabilities, difficulties in secondary development, sharp hot and cold stress on cultures, and the inability to provide effective temperature stimulation, a smart cultivation apparatus was constructed on the base of STM32 microcontroller chip. The apparatus employed temperature and humidity sensor, JXM oxygen sensor, and MQ-135 ammonia sensor for the purpose of data collection. Additionally, the ESP-12F wireless module is utilized as a medium for custom control through a mobile application, being compatible with both of Android and iOS systems, real-time data was upload to the cloud, and real-time customized controls through mobile apps were implemented. Actual operation of this cultivation apparatus showed the proficiency to achieve functions such as automatic constant temperature and humidity control, gradient cultivation conditions setting and operation, real-time remote data uploading, and app control. The temperature fluctuation range is controlled as ±0.1 °C, and the relative humidity variation rate is of ± 1%. Additionally, the programmable and scalable performance was embedded into this apparatus by providing program burning interfaces, multiple serial ports, and relay modules, to meet the technical needs for teaching and research.

In this study, a multi-layer perceptron based identification method is proposed for the identification of wind vibration hazards of conductors in power systems. First, the background and impact of conductor wind vibration hazards are introduced and analyzed, and the related theories are discussed. Further, a data set including several feature parameters is established, and a high-quality feature vector is obtained through feature engineering, and the degree of influence of each attribute on the occurrence of wind vibration is explored. Then, a multilayer perceptron model with multiple implicit layers is designed and trained for the automatic identification of wind vibration hazards in conductors. Finally, experimental validation shows that the proposed method has high accuracy and robustness in wire wind vibration hazard identification and can be effectively applied to the safe operation of power systems.

With the rapid development of cloud computing and mobile Internet, there have been a variety of network attacks, among which distributed denial of service (DDoS) is one of the most fatal attacks. Traditional machine learning detection methods face serious challenges. This paper proposes a method for identifying DDoS attacks based on RNNs. Taking Recurrent Neural Network (RNN) as the research and improvement object, long short-term memory, bidirectional recurrent neural network and other technologies are used. Experiments on the public dataset indicate that the accuracy of the model can reach 99.97%, which is better than the traditional machine learning model and achieves a good recognition effect

With the widespread adoption of blockchain technology across diverse sectors[1], the imperative for efficient data storage and retrieval mechanisms has intensified[2]. This paper introduces an innovative approach that synergizes Zero-Knowledge Proofs (ZKP) and sharding techniques to bolster data privacy and optimize storage within blockchain systems. We present an exhaustive analysis of prevailing blockchain data storage mechanisms, pinpoint their inherent privacy vulnerabilities, and elucidate the potential of ZKP in mitigating these concerns. Furthermore, we propose a sharding-centric solution that not only fortifies data privacy but also augments the efficiency of data retrieval[3]. Experimental results, derived from real-world scenarios, validate the robustness and efficiency of our proposed methodology.

In order to promote the application of the cross-platform and plug-and-play, a development framework based on Service Oriented Architecture (SOA) is proposed. In order to achieve interface unification and service encapsulation, the concept of service template is firstly designed to standardize the generation of services instances as well as the service publishing, query and subscription so as to improve service efficiency. Besides, to fulfil the plug-and-play feature for services, a complete life cycle management is also designed to realize the dynamic service management, in which the concept of mirror service is introduced to realize the service versatility in different scenarios. The results show that SOA based framework can realize the service decoupling and merging easily, so that the service efficiency can be achieved.

With the development of information technology, storage-computing devices are widely used in the information field. Most of storage-computing devices usually have to maintain their storage and computing performance by introducing data cache technology, such compromise put the storage-computing devices in the dilemma of being challenged by the risk of losing their cache data when the power loses accidentally. In order to prevent the cache data loss caused by the accidental power loss of the system, this paper designs a method to realize power loss data protection based on Non-volatile DIMM in domestic platform, Specifically, install the Non-volatile DIMM memory in the storage-computing devices, the BMC is used to trigger the curing preservation operation of the memory in the case of power loss .When the system is suddenly power off, under the endurance of the super-capacitor, the processing unit in non-volatile memory migrates the data in DRAM to NAND Flash for curing storage, this ensures that data will not lost even in the event of a sudden power loss ,achieving the purpose of memory data non-volatile, ensuring the data security of the device.

Underwater object detection is a challenging task due to issues caused by the complex underwater environment, including high similarity between backgrounds and objects, large variance in object scales, and serious blurring of objects. These issues lead to poor performance of conventional object detection methods in underwater scenes. To address these difficulties, we propose a high-resolution underwater object detection network based on multi-scale attentional feature fusion. First, HRNet is utilized to extract high-resolution features while preserving target details. Second, a multi-scale attentional fusion module is designed to effectively integrate multi-scale information in channel and spatial dimensions. Finally, DIoU loss is adopted to improve prediction box regression accuracy, and data augmentation is used to enhance network robustness. Experimental results demonstrate that compared to other object detection methods, our approach achieves better detection performance.

Motion state detection is related to athletes' training, improvement of quality of life and daily health monitoring. The traditional motion state detection is often only analyzed by single-modal data such as video, which can't realize real-time detection of motion state, and the detection results are quite different from the actual results. In order to realize the realtime monitoring of the athlete's motion state, this study combines the knowledge of exercise physiology and other related knowledge, and uses the method of combining image and heart rate detection to realize the classification and analysis of motion state. Firstly, the image acquisition device and heart rate detection device are used to obtain the image and heart rate data of the exerciser, and the data is processed. Then feature extraction is performed on the collected data, and a motion state classification model is constructed using the key features of the extracted image and heart rate data. By setting up control experiments under different motion states, data acquisition and experimental design are carried out. The experimental results show that the proposed method is compared with a variety of latest detection models. The accuracy of this method is as high as 97.6% in the classification of motion states, and it can effectively distinguish the heart rate changes under different motion states.

Wax deposition in oil wells is a common challenge in crude oil extraction, and using hot washing to remove wax is a primary means, both in terms of cost and efficiency. However, due to the heavy workload of hot washing, traditional manual management modes face a series of problems such as inaccurate wax deposition monitoring, unclear well-washing period, and irrational resource allocation. Considering these, a novel monitoring and management system for hot washing is developed based on the Internet of Things. Through real-time data collection and transmission, the working state of oil wells can be effectively monitored. Particularly, the early warning model for wax deposition and criteria for evaluating the hot washing effect are established by data integration and mining. Following its application in the Qingcheng oilfield, the completion rate and qualification rate of hot washing increased by 17% and 21%, and the number of early warning wells decreased by 73%. Therefore, the developed system realizes data-driven management and greatly improves development efficiency.

In this article, we propose a Shapelet based adversarial attack method for GCN based traffic flow prediction models. Firstly, we extract shape patterns from the original traffic flow data to obtain a set of shape feature Shapelets with significant discrimination. Then, we use Shapelet to generate adversarial samples using a method called reverse attack. Finally, we input adversarial samples into the GCN model to estimate their impact on traffic flow prediction. We evaluated the proposed attack algorithm on two real traffic datasets. Our algorithm can mislead model predictions, reduce the speed of model predictions, and lead to incorrect congestion prediction results. This study provides a new approach to adversarial attacks against time data. It also provides a new method for improving the robustness of traffic flow prediction models. Future research can further explore more complex and covert attack methods to address evolving security challenges.

Distributed computing is a resource-centric computing paradigm. It connects all devices with computing capabilities and utilizes idle computing resources in the network to divide and complete computing tasks. Distributed computing provides services to users through collaboration and sharing of computing nodes. The nodes can be task publishers and service providers. Due to its unique architectural advantages, distributed computing plays a huge advantage in the application of the Internet of Things. There are network security issues such as identity authentication and data transmission in decentralized computing. In response to these problems, this paper studies network security issues and algorithm models in a distributed computing environment. This article analyzes in detail the factors of security threats and control costs, and builds an algorithm model that combines continuous identity authentication with security threat assessment. Based on the characteristics of network attacks, a security threat assessment model based on the transmission mechanism of infectious diseases is constructed, the network security threat index is monitored in real time, and a multi-factor continuous identity authentication algorithm model is proposed. Experimental verification results show that this algorithm model can significantly reduce the proportion of malicious nodes in the network under steady state.

There are many instances of small and medium-sized targets in drone aerial images, and existing detection algorithms are prone to missed and false detections during the detection process. To address this issue, A small target detection method based on improved YOLOv5 in UAV aerial images are proposed. Firstly, the bilinear interpolation upsampling method is used in the feature fusion section to reduce the loss of feature information during the upsampling process. Then, a small target detection layer of size 160 is added to locate and recognize small targets using shallow feature information, reducing the missed detection rate. Finally, three CBAM attention mechanism modules were added to improve the accuracy of the algorithm. On the VisDrone2019 dataset, the improved algorithm improved mAP by 2.1% compared to Algorithm YOLOv5, effectively completing small object detection tasks.

This paper mines information with potential commercial value from a large amount of commercial data, uses association rule mining algorithms to provide enterprises with operational information of various business activities, improves the management level of enterprises, and helps enterprises develop valuable value-added services solutions to make services more humane and intelligent to better meet customer needs. Logistics Information System (LIS) is an interactive system based on logistics information, combining logistics management software with computer technology, and connecting with personnel, equipment, etc. The logistics information system is a system that integrates many kinds of technologies, which involves multimedia technology, bar code technology, geographic information system, global positioning system, data mining and other technologies. Logistics information systems mainly include data warehouse systems, EDI (Electronic Data Interchange) exchange platforms, customer-oriented service platforms, and CRM (Client Relationship Management) platforms. When the minimum support in this article is 0.15, the running time of WPAR Algorithm (Weighted Partitioning Association Rule Algorithm) is 5s, and the running time of Apriori is 7s. When the minimum support is 0.95, the running time of WPAR Algorithm is 1.8s, and the running time of Apriori is 1.6s. Research shows that the use of efficient data mining and related technologies can enable enterprises to make decisions quickly and effectively in the management and strategic planning process, thereby optimizing and improving service quality.

Aiming at the problem of road landslide classification in Nanshui Reservoir area, a new solution was proposed by combining unmanned aerial vehicle inspection technology, deep learning and analytic hierarchy process. Firstly, the high-resolution road image data in the reservoir area was obtained by unmanned aerial vehicle inspection, and then the image segmentation was carried out by deep learning algorithm. Finally, the road landslide was graded by analytic hierarchy process, and the severity and maintenance priority of the landslide were evaluated. The experimental results show that this method can effectively realize the classification of road landslides in Nanshui Reservoir area. This study can provide scientific basis and guidance for the operation and maintenance of the Nanshui Reservoir, in order to promote traffic safety and maintenance work in the reservoir area.

Under the national strategy of "carbon peak" and "carbon neutrality", during the "14th Five-Year Plan" period, China has introduced a number of policies related to solar power generation and green PV buildings, and a variety of favorable policies to promote distributed photovoltaic in the densely populated eastern region of the rapid development. However, more problems have been exposed in the operation and maintenance of distributed photovoltaic power plants, and the research on centralized low-cost operation and maintenance of distributed photovoltaic systems and secure and efficient access to data is relatively weak. To this end, this paper takes the current situation of distributed PV data access as the background, establishes a distributed PV data acquisition system based on the Internet of Things, and focuses on the research of credible data acquisition front-end based and information security platform based on remote operation and maintenance management, which provides safe and reliable data acquisition equipment and program references for distributed PV grouped operation and maintenance.

In recent years, the external environment of enterprises has been characterized by significant uncertainty; and the increasingly complex situation has also posed numerous challenges to their supply chains. Demand forecasting, as a critical component of the enterprise supply chain, directly affects market performance. However, demand forecasting is influenced by multiple factors, leading to generally low prediction accuracy, with substantial deviations from actual outcomes. To enhance forecasting precision, this study utilizes shipment data directed towards distributors from a large domestic manufacturing enterprise during the period from September 1, 2015, to December 20, 2018. The data undergoes preliminary processing, followed by exploratory analysis and feature engineering. Conventional machine learning models are applied for prediction. By comparing the Random Forest and LGBM models, the LGBM model is selected as the best-performing one for data prediction. Forecasts are conducted at daily, weekly, and monthly granularities for the product demand in January, February, and March 2019.

The substation inspection is a necessary and important measurement to ensure the normal and safe operation of equipment, which is of great significance to reduce the occurrence of accidents and control the scope of power outages. At present, the inspection mode of substation urgently needs to change from conventional manual inspection to intelligent three-dimensional inspection. This work leverages the key technologies of intelligent three-dimensional inspection system in substation, and proposes the architecture design of three-dimensional inspection system, the substation visualization, the audible equipment design and the decision-making platform of intelligent inspection in substation. Finally, the shortcomings of the current research and the focus of the future research are discussed.

Currently, the existing human patrol and machine patrol methods lack of analysis of the surrounding environment of photovoltaic power stations, which leads to the situation that the power line inspection data are not consistent with the actual results, so that the inaccuracy of intelligent inspection results are obtained. Then, an intelligent inspection system for photovoltaic power station based on lidar point cloud data is designed. In particular, the environment and equipment information are processed centrally and the processing results are stored in the storage module. Besides, the laser radar is used to measure the distance. The laser pulse is used to collect and process the point cloud data, so as to adjust the radar parameters. By calculating the distance between the laser source and the measured point, the environmental information around the photovoltaic power station is obtained to form a three-dimensional surface model, which is used to fit the power line and achieve the best parameters. The experimental results show that the horizontal distance and vertical distance of the proposed system are consistent with the actual result, which are also accurate.

Highlighted by durability and availability, plastic product has become an indispensable part from industry to our daily lives. However, the amount of plastic waste far exceeds the amount that humans can safely dispose of. Through Least Squares, our model estimates that the maximum level for plastic waste that can be mitigated is about 11.7 million tons by fitting the data from 1980 to 2015. In the following discussion, we take the situation of each country into account, amend the model results and obtain the conjecture that the amount of disposable plastic waste reduced to the environmentally safe level is 380 million tons. The attempt to achieve the goal affects people’s quality of life and job opportunities to some extent, but it will have great benefits to the ecological environment and human health. In our models, PCA is applied to calculate the Severity Index. To improve equity, we designed a model to categorize countries by their capability to dispose plastic waste. For unsupervised clustering, t-SNE is used to divide countries into three categories. The results make sense and offer policy makers the basis to distribute the burden to every country more reasonably. We strongly recommend that countries with higher income and waste disposal capacity (Category 1) help countries with weaker integrated capacity (Category 2) by investing funds and technology. For middle-income countries (Category 0) with great potential to contribute, we should encourage them to develop sustainable economies, although the rate of economic development may decline in the process.

This paper proposes a strategy for setting charging service fees under demand response, based on the analysis of the charging behaviour of EV users, to address the problem of impact load caused by the superposition of peak electricity loads due to the disorderly charging behaviour of a large number of EVs. By guiding users to charge in an orderly manner through price signals, not only can the charging cost of users be reduced, but also the operational efficiency of grid assets and the revenue of charging stations can be improved, thus promoting a good interaction between "vehicle-pile-grid".

In order to improve the personal safety of vulnerable groups, a self-protection alarm system was designed, which integrates an STM32 microcontroller with GSM, GPS, alarm, key switch and Bluetooth modules. Develop a C language program to enable users to receive emergency support through text message alerts and alarms, providing access to their emergency contacts. Meanwhile, a stable power supply is provided by the combination of lithium batteries and solar panels. The system is easy to operate, clean, and environmentally friendly, and the user’s personal safety is guaranteed.

Steel surface defect detection has always been a research hotspot in smart factories. Aiming at the problem of insufficient recall and speed of the Yolov5 algorithm in detecting steel surface defects, the steel surface detection effect was improved by new deep learning algorithm based on Yolov5 algorithm combined with different attention mechanism algorithms. SE, ECA, and C3SE attention mechanism modules were respectively added to Yolov5's backbone in this paper, which improved the weight of important feature information, reduced the attention to other information, and improved detection recall. The experimental results indicated that the improved Yolov5 deep learning model can greatly improve recall. Compared with YOLOv5, the recall of ECA-YOLOv5 was slightly improved; the recall of SE-YOLOv5 was obvious improved, and the Inference time was reduced by 0.1ms; The recall and mAP@.5 of C3SE-YOLOv5 were significantly improved. The total parameter count of the network decreased by 13.03%, at the same time the pt file size of the network decreased by 12.72%. The improved network significantly improved the performance of the steel surface defect detection algorithm.

The contagious disease influenza that is prevalent before and during the Covid-19 era poses a substantial effect on the global health care system. Traditional computational epidemiology can simulate potential disease progress but is inefficient in gathering sophisticated and real-time data sets. Social media, the world’s most widespread data generator, provides an access to more updated epidemic surveillance. Confronted with the outburst of the Covid-19 pandemic and concurrent influenza, machine learning models are the more efficient and crucial technique in mapping the next influenza outbreak with social media data. This study utilized supervised machine learning techniques including Logistic Regression (LR), Decision Tree (DT) and Random Forest (RF) to estimate the risk of influenza outbreaks, and examined the potential of different forecasting models in predicting the outbreak. The content of the social media was modeled based on its correlations with disease outbreaks and computed through various statistical models. Comparisons of the result revealed RF as the most efficient forecasting model for influenza outbreaks during the Covid-19 pandemic and demonstrate the usefulness of this study in future disease prediction.

Intelligent applications with Internet of Things terminal as the core have penetrated into all aspects of social life. With the iterative development of Internet of Things technology, the needs of users and manufacturers may constantly change, and various software anomalies may be exposed after the deployment of Internet of Things devices in the field. How to effectively ensure the sustainable renewal of IoT devices after deployment is an urgent problem to be solved. In this paper, a lightweight firmware update method for distribution IoT terminals is proposed, which adopts differential upgrade method, only the difference between the new version and the old version needs to be upgraded, and the same upgrade effect can be achieved by transferring less data. The simulation results show that the proposed firmware update method has higher upgrade success rate and upgrade efficiency, and lower memory consumption rate.

With the increasing number of colorectal cancer patients worldwide, colorectal cancer has become one of the leading causes of cancer death. Therefore, it is particularly important to improve the detection rate of colorectal lesions. Aiming at the misdetection and missed detection of assistant physicians in lesion detection, this paper proposes a colorectal lesion detection task based on improved YOLOv8. The sample data set is constructed through the private data set (SSPH data set) and the public data set (CVC-Polyp), and the improved YOLOv8 model is trained by selecting the SGD optimizer and transfer learning method. The introduction of the GAM attention mechanism on the basis of the YOLOv8 detection model can better focus on inter-dimensional information interaction than other attention mechanisms, and improve the model's attention to lesions. The experimental results show that the method proposed in this paper has an precision of 94.9%, a recall rate of 88%, and a Map@0.5 value of 95.8%. Compared with the original YOLOv8 model, it has been effectively improved.

In the context of China's increasing efforts to build extra-high voltage transmission projects, accurate and effective project cost prediction of extra-high voltage transmission lines can not only rationally arrange the investment of grid construction funds, but also effectively improve the operating conditions of power grid enterprises and guarantee the stable development of power grid enterprises. This paper takes data mining as a means to screen out the main factors affecting the cost of ultrahigh voltage transmission project, and screens the factors through random forest, takes the screened factors as inputs and the cost of engineering units as outputs, constructs a cost prediction model based on HKELM neural network, and proves the validity and applicability of the model to the investment prediction of ultra-high voltage transmission line project through empirical research, which provides an opportunity to predict the cost of ultra-high voltage transmission line project for the grid enterprises and to improve the operation of the grid enterprises. It provides a new method for the cost prediction of ultra-high voltage transmission line project.

Most of the existing research focuses on a single MEC server scenario, and there are few studies on computing offloading and resource allocation in multi-MEC server scenarios. Therefore, this paper proposes a dynamic computation offloading and resource allocation algorithm based on hybrid decision-making deep reinforcement learning, which utilizes deep deterministic policy gradients and dueling dual deep Q-networks to enhance the Actor-Critic structure. By calling the Actor part of DDPG to deal with continuous power allocation, and then combining the Critc part of DDPG with D3QN, the discrete MID is associated with the MEC server to solve the mixed decision-making problem in multi-MEC server multiuser scenarios. From the simulation results, we can see that the proposed algorithm has faster convergence and better stability than baseline algorithms such as DQN.

Ferroelectric memristors have become a promising contender for the realization of neuromorphic computing, due to their advantages of good stability, ultra-fast switching behavior, long endurance, and low power consumption. In this work, novel lanthanum titanate (La₂Ti₂O₇) based ferroelectric memristors (Ag/La2Ti2O7/Pt) have been fabricated. Physical morphology of ferroelectric devices has been investigated by atomic force microscope, piezo-response force microscopy, and optical microscope. The Ag/La₂Ti₂O₇/Pt devices exhibit stable resistive switching phenomena. More interestingly, the conductance of devices can be modulated under continuous voltage pulse signal. Based on the results, the synaptic behaviors have been successfully mimicked including paired pulse facilitation, short-term plasticity, and long-term plasticity. In addition, the convolutional neural network based on the synaptic behaviors of the La₂Ti₂O₇ based ferroelectric crossbar arrays has been built for the application potential in image recognition. The recognition accuracy after learning is up to 96 %.

This paper proposes a machine learning-based automatic execution algorithm, aiming to optimize the efficiency of supply chain operation. The external conditions for the robot to complete the task are set, and the objective function is constructed based on the task completion cost and time, from which the scheduling logic of the supply chain intelligent operation robot is constructed. The genetic algorithm in machine learning is used to solve the mathematical model, and the fitness function is set as the reciprocal of the objective function, executed until the minimum operation cost and time, and outputs the automatic execution scheme of the robot scheduling logic. Through experimental verification, the results show that the algorithm can effectively improve the efficiency of supply chain operation, reduce the cost, and have high adaptability and accuracy. This algorithm provides a feasible automation solution for supply chain intelligent operation robot scheduling.

This paper provides a comprehensive review of AI-assisted motion control, which plays a crucial role in industrial automation and robotics. The application of AI technology brings new opportunities and challenges to motion control. The paper begins by introducing the basic concepts of motion control and discussing the limitations of traditional methods. It then delves into the application of AI in motion control, including generative AI, AI during setup and tuning processes, AI for statistical process control, and adaptive motion control. Special attention is given to the application of AI-assisted soft motion control and its advantages and limitations. Through comprehensive investigations, the current research progress and application status of AI-assisted soft motion is summarized. Finally, the future trends and challenges of AI-assisted motion control, as well as potential research directions and application prospects, are discussed.

The power IoT network serves as a critical social function in modern society but has also become a high-value target for malicious attackers. Therefore, it is of vital importance to rapidly discover vulnerabilities in power IoT devices and locate abnormal points for repairs in order to enhance the overall security of the power system. This paper presents the implementation of a prototype system called FuzzSem-AL for automated anomaly localization in power IoT devices. By utilizing deep learning-based semantic analysis techniques, the system recovers function names from the binary programs of power IoT devices and associates these functions with the communication protocol fields of the devices. Additionally, state backtracking techniques are employed to effectively eliminate the influence of non-root cause factors and pinpoint the root protocol fields that cause device crashes. Compared to existing techniques, this paper effectively combines the processes of fuzzing and anomaly localization, enabling automated identification of the core functions that trigger anomalies. The validation of the approach is conducted using past vulnerabilities in commonly used protocols in power IoT systems, such as FTP and SMTP. Ultimately, the paper successfully locates the function positions of 11 abnormal points out of 15 1day vulnerabilities.

Wind turbine temperature is one of the indicators for determining whether the wind turbine is working properly or not. Accurate prediction of wind turbine temperature can predict the status of the wind turbine in advance and actively take preventive maintenance measures, thus reducing losses. In this paper, based on the stochastic differential equation model, a "stochastic differential equation + Markov" combination model is developed to predict the wind turbine temperature. Where the stochastic differential equation predicts the overall trend of wind turbine temperature change and Markov corrects for the effects generated by stochastic disturbances. The fitting and testing errors of the stochastic differential equation, the grey GM(1,1) model and the "stochastic differential equation + Markov" combined model were calculated by collecting data from the temperature of a wind turbine in a wind power plant in Jiuquan, respectively. The results show that the combination model "Stochastic Differential Equation + Markov" is more accurate than the single model in terms of fitting and generalization accuracy.

With country's digital transformation, the social and economic collaborative development of cities within urban agglomerations will be further promoted. It is necessary to build cross-city and over all dimensional social governance system. Since the existing social governance system is based on integration of departments and regions at different levels, different information systems cannot exchange information at any level. It is necessary to construct smart city of urban agglomerations through overall architecture design. The "One Network Management" in Shanghai is based on smart city at the municipal, district (county) and town levels. The "One Network Management" uses the three-level platform to achieve data sharing and business linkage within the domain. On the experience of "One Network Management" in Shanghai, a three-level platform for urban agglomerations is constructed, as data sharing and business linkage between cities are achieved through the municipal-level platform to realize the social governance system of urban agglomerations. After analyzing the technical difficulties of cross-city "One Network Management", the overall architecture and network topology of the "One Network Management" of urban agglomerations are introduced. Through overall architecture design, the key and difficult technical contents of "One Network Management" of urban agglomerations are introduced, including edge-cloud collaboration technology, data analysis and processing technology, detection and analysis of urban dynamic characteristics, digital twins, network sharing technology, security protection technology, maturity evaluation indicators, etc. Based on these, various problems and challenges existing in current research on smart city of urban agglomerations are discussed, and the future development direction of this field is also explored.

The Unmanned Aerial Vehicle (UAV) inspection path planning for intelligent warehouse is a typical indoor path planning technology. Through the UAV intelligent inspection of each work area in the intelligent warehouse and workshop, we can solve some issues occurred in the work area timely. In order to solve the problem of path planning efficiently, an improved hybrid optimization algorithm is proposed to minimize the inspection path, which combines intelligent optimization algorithms of particle swarm optimization and genetic algorithm. By optimizing the structure of the genetic algorithm and finding the optimal parameter combination, it is easier to converge to the shortest path for the intelligent warehouse. Simulation results show that, compared with the traditional algorithms, the convergence ability of the optimized hybrid algorithm has been greatly improved, which has also been verified.

Currently, transient stability assessment based on deep learning requires offline generation of a large amount of data to train predictive models. However, when there is insufficient training data, the accuracy of the model's predictions significantly decreases. To address this challenge, this paper proposes a method for selecting and enhancing key samples. Firstly, a quantitative analysis of the correlation between samples and the training of the model and transient stability boundaries is conducted. Based on this analysis, a sample importance index is defined, and key samples are selected based on this index. Then, these key samples are enhanced using Generative Adversarial Networks (GANs) to improve the representation capability of the sample set and help the transient stability model learn classification boundaries that better fit real transient stability boundaries. This method is applied to the New England 10-machine 39-node system and a provincial-level power grid in China. The test results show that with the enhancement of key samples to 10%, the misclassification rate of the predictive model decreased by 86.54%. This approach significantly improves the predictive capability of the model under the condition of limited training samples.

In order to predict the laser cleaning quality of composite paint layer and realize the controllable paint removal of composite paint layer, this paper proposes a machine-learning based laser cleaning quality prediction method for composite paint layer. The aluminum alloy substrate surface is uniformly coated with 20 μm green epoxy primer and 43μmwhite polyurethane topcoat as experimental samples for laser cleaning experiments; based on the experimental data combined with the three machine learning algorithms (support vector machine SVR, BP neural network, and random forest RF) to establish a prediction model between the process parameters and the cleaning quality. The experimental results show that, compared with the RF model and BP neural network model, the SVR model is more accurate in predicting the quality of laser cleaning of composite paint layers, and the coefficient of determination of the prediction model is 0.961, the root mean square error is 1.738, and the average absolute error is 1.5162.This study obtains the prediction model of the quality of removing the paint thickness with high accuracy, realizes the effective prediction of the quality of the laser removing the paint for the It lays a model foundation for further research on intelligent control of laser cleaning of composite paint layers.

With the rapid development of China's economy, the whole society's demand for electricity is becoming more and more extensive. Insulator is the guarantee of smooth operation of transmission line, and is also one of the equipment which is prone to failure. How to detect insulator damage with high precision and high efficiency and repair it is one of the key factors of unimpeded power. This paper proposes a method for detection of insulator damage based on improved YOLOv4-tiny network. This method is mainly aimed at improving the main feature extraction module and feature fusion module of traditional YOLOv4-tiny network. In order to solve the problem of missing detection of small targets, adaptive attention mechanism is introduced to improve the module of feature extraction to improve the accuracy of detection. In addition, in order to further improve the detection accuracy of insulator damage and balance the detection time, a multi-attention CSAR model is proposed to improve the performance of the feature fusion module. Finally, images collected from different weather conditions are used as test sets to verify the effect of the improved model proposed in this paper through experiments. According to the experimental results, the detection accuracy of the proposed method can reach 98%, and the detection time is controlled within 10ms, which meets the basic requirements of detection of insulator damage.

Betweenness centrality is a measure of node importance in networks, but conventional exact algorithms require a lot of time as the network size grows dramatically. This paper aims to enhance the efficiency and accuracy of betweenness centrality computation for network nodes. We propose an algorithm based on shortest path approximation and adaptive sampling. The algorithm first selects high-quality seed nodes according to degree, then approximates shortest paths, and finally chooses appropriate samples to approximate betweenness centrality. We conduct experiments on 5 different datasets, and the results show that our algorithm outperforms the baseline algorithms in terms of sample size and running time. Our algorithm not only reduces the computational cost effectively, but also guarantees the computational accuracy.

Deep clustering algorithms based on graph convolutional networks are widely used due to their strong ability to mine network structure. However, the construction of neighborhood graphs may introduce noise and affect the clustering results. Meanwhile, focusing on ordinary topology alone ignores the higher-order connections between data about attributes. To address the above problems, an unsupervised hypergraph convolutional clustering network (UHCCN) is proposed in this paper. We construct hypergraph structures using attributes and incorporate higher-order information encoding into representation learning through hypergraph convolution. Using an attribute encoder will extract node features and fuse it into the hypergraph convolution. Finally, representation learning and clustering are optimized jointly. The experiments validate the effectiveness and superiority of UHCCN.

The design of ship main dimensions is a complex problem involving multiple disciplines and objectives. Due to the high subjectivity of conventional methods that only optimize through qualitative analysis and manual trial and error, they can’t quantitatively analyze accuracy. Therefore, in this paper, the surrogate model is used for regression analysis of the complex ship principal scale problem, and then the multi-objective evolutionary algorithm is used to optimize the feasible solution generated by the surrogate model. Firstly, the regression model of each optimization objective in the ship main scale problem is established. In order to ensure the accuracy of the Surrogate model, the Root-mean-square deviation and correlation coefficient of the training set and the test set should be taken as the measurement indicators. Then, based on pareto dominance rule, the regression values generated by the surrogate model are selected. And through genetic operations such as mutation and crossover, a large number of new individuals are generated, and the outstanding individuals are screened based on elite retention strategies. Finally, experiments show that the proposed method is superior to the algorithm without surrogate model in accuracy and efficiency.

A 14×6×5 BP artificial neural network model was constructed using Matlab software. The model was trained with 400 samples, resulting in a training error of only 1.67×10^-7, which was significantly lower than the target training error 1×10^-6. Simultaneously, a test set of 100 samples was used to evaluate the model's performance, and it was found that 100% accuracy, recall, precision, and F1 score were achieved, indicating that groundwater quality could be accurately and swiftly evaluated. The results of 176 samples from 2020 to 2022 were analyzed, revealing that the area's groundwater quality was primarily categorized as Class II and Class III. During 2020-2022, the proportion of water wells evaluated as Class II and Class III was 97%, 96%, and 98%, respectively. The primary factors affecting the groundwater quality in this region were identified as the iron and manganese indicators. The BP neural network method allowed for the retention of all data information. Based on all indicators, the water quality was assessed and the non-linear relationships between data were processed to ensure the objectivity and accuracy of evaluation results.

In order to solve the problem of low accuracy of pedestrian target detection and recognition caused by occlusion between people when people are crowded with each other, an improved method of introducing SENet channel attention mechanism based on YOLOV5 neural network was proposed. By introducing SENet channel attention mechanism to further explore the relationship between the feature channels in the image, the feature extraction of the human target area is enhanced, and the accuracy rate of human target detection and recognition is well improved. Compared with the general YOLO V5 algorithm, the improved algorithm introduced by SENet attention mechanism proposed in this paper has a lower missed rate of target detection and a higher correct rate of target detection and recognition when people are crowded with each other, and also has the real-time performance of the general YOLO V5 algorithm.

With the widespread use of remote computing in Internet of Things (IoT) systems, security has become an important issue for IoT. The large amount of private information stored in EC devices makes them attractive to a variety of attacks. In order to address this issue, this paper proposes a security roadmap for IoT edge computing. The experimental findings indicate that the suggested approach has the potential to enhance the service-focused security that rely on IoT.

Current machine reading comprehension methods focus on modeling local or global interactions between texts. This method, which does not pay attention to both local and global structures, easily leads to problems such as insufficient machine understanding of semantics and inaccurate answering of questions. Aiming at the problem that when the traditional LSTM is used as the encoder at the encoding layer, there is a problem that the current input and the hidden state are independent of each other, resulting in the loss of context information. This paper introduces a variant of the LSTM network to enhance the information perception and interaction between text contexts and obtain richer information. Semantic representation. At the same time, considering that the attention mechanism cannot effectively extract the local structure of the text, which hinders the model from understanding the deep semantics of the text, a combination of dynamic convolutional attention and multiple attention mechanisms is used in the attention layer to capture text from different scales. Structure. Experimental verification is carried out on the DuReader2.0 data set, and the results are analyzed, which proves that the proposed model can effectively improve the ability of the machine to answer questions.

As an important branch of biometric identification, face recognition technology has a wide range of application prospects in security field, social network, autonomous driving and smart home. However, the performance of traditional face recognition methods is limited in the face of complex environments such as lighting changes, posture changes and expression changes. In recent years, the rapid development of deep learning algorithms has brought breakthrough progress to face recognition. This study aims to explore the latest advances and key challenges in face recognition technology based on deep learning algorithms. First, we introduce the basic principles of deep learning and its application in face recognition. Then, we discuss in detail the role of deep learning models such as deep convolutional neural network (CNN) and deep adversarial generative network (GAN) in face feature extraction and generation. We also analyze the importance of data sets and how data enhancement techniques can improve the performance of models. Finally, through the review of the latest research results and technical trends, this paper summarizes the potential application fields and future development direction of face recognition technology based on deep learning algorithm, and provides a useful reference for further promoting the research and application of face recognition technology.

The extensive penetration of residential electric vehicles (EVs) has brought huge challenges to the operation of power grids. Recognizing the EV charging patterns and behaviors of residential EVs can help utilities operators make reasonable scheduling strategies for balancing power supply and consumption. In this paper, recent state-of-the-art deep learning models applied to load disaggregation are used to recognize charging patterns and behaviors of residential EVs. In addition, we propose a novel sequence-to-sequence model which is constructed by seven convolutional layers and extra two dense layers. And it can provide improved EV power consumption profiles for utility and users. Performance of proposed method has been validated using real data measured from several residential charging stations. Applicability of our proposed methodology has been confirmed by assessment results owing to its satisfactory precision. Besides, a comparative analysis has been made to demonstrate its effectiveness with regard to the disaggregation performance.

This paper proposes a design solution for a distributed machine learning platform based on Apache Spark and expounds on its advantages in specific application scenarios. Through Spark's distributed computing framework, the platform realizes efficient machine learning model training and prediction. The paper details the platform's overall architecture, core components, and model implementation methods. We also validate the platform's efficiency and scalability in processing large-scale data through a product recommendation system example. Research indicates that compared to standalone systems, this Spark-based machine learning platform can significantly shorten model training time and support stream data processing. This study offers a reliable and scalable machine learning implementation solution in a big data environment.

In order to solve the problems of local optimal solution stagnation and slow convergence speed in particle swarm optimization algorithms, this paper proposes an improved particle swarm feature selection (ISNPSO) that combines simplex method and niche. The initial solution is generated using the Relief algorithm, and the particle swarm is divided into multiple subpopulations using niche technology, allowing for information exchange between subpopulations. Local optimal particles of all subpopulations are extracted for simplex optimization, which accelerates convergence speed while maintaining diversity of the population. The performance of this method was compared with 5 algorithms on 6 datasets. The experimental results show that the algorithm effectively improves classification performance while reducing the number of features.

Due to the fact that most feature selection methods based on particle swarm optimization are prone to falling into local optima, an improved transformed length particle swarm optimization algorithm (TLPSO-rr) is proposed for feature selection to address this issue. In the process of particle swarm search, mutual information is used to measure correlation and redundancy. A neighborhood search strategy is proposed to reduce redundancy, and then a transformed length strategy is used to make particles have different lengths to increase the diversity of the particle swarm and improve global search ability. To evaluate the effectiveness of this method, TLPSO-rr was compared with 7 feature selection methods on 9 different datasets. The experimental results show that this method is effective and can achieve smaller feature subsets and higher classification performance.

In recent years, knowledge graphs have made great progress in the field of agriculture. However, the traditional knowledge graph construction is mainly based on the single modality of text, but in the actual production of agriculture, it is often difficult to accurately describe the real data scene by using the single-modal data to construct knowledge graphs, thus increasing the limitations of the use of knowledge graphs. The article provides a detailed introduction to the basic concepts of multimodal knowledge graphs and the core technologies. For the application of multimodal knowledge graph in agriculture, the article focuses on summarizing the research on the application of knowledge graph in agricultural intelligent question and answer, crop monitoring, agricultural product recommendation and other fields in agriculture. It also looks forward to and analyses the challenges of constructing multimodal knowledge graph in agriculture and the prospect of multimodal knowledge graph in the field of agriculture.

Communication network is an important part of social life, but it is extremely vulnerable to various network attacks, and the energy of data nodes in the network is limited, and abnormal behavior in the network cannot be detected in the highdimensional traffic data mode. Therefore, an intrusion behavior detection method based on machine learning algorithm was proposed to detect attack behavior in communication network in real time. Firstly, through the restricted Boltzmann machine algorithm model of machine learning, high-dimensional traffic data features are extracted and dimension reduction operation is carried out. Then, combined with the characteristics of communication network cluster structure, a multi-layer support vector machine classification model is established to detect network abnormal state. Finally, the public intrusion detection data set is used to verify the proposed intrusion behavior detection model. The experimental results show that the detection accuracy of the model for network traffic can reach 99.63%, in order to provide reference for further research on intrusion behavior detection methods in the future.

Conventional UAV multi-type patrol task allocation method mainly uses MEC(mobile edge computing) technology to unload UAV data, which is easily affected by the change of dynamic unloading mechanism, resulting in a high delay in patrol task allocation. Therefore, it is necessary to design a new UAV multi-type patrol task allocation method based on deep reinforcement learning. That is, the deep reinforcement learning technology is used to construct the multi-type inspection task allocation model of UAV, and the multi-type inspection task allocation algorithm of UAV is designed, thus realizing the multi-type inspection task allocation of UAV. The experimental results show that the design of UAV deep reinforcement learning multi-type inspection task allocation method has good distribution effect, reliability and certain application value, and has made certain contributions to improving inspection reliability and reducing comprehensive inspection cost.

The rapid growth of e-commerce platforms has resulted in an overwhelming abundance of product information, leading to choice paralysis and information overload for users. To address these challenges and enhance the user experience, an intelligent recommendation system based on machine learning is proposed. This paper aims to design and implement such a system for e-commerce platform product information. Firstly, we review the relevant theories and technologies of recommendation systems, including the application of Kafka message queues and the Lambda architecture, as well as two common recommendation system types: content-based recommendation and collaborative filtering. Then, we propose a Bayesian pairwise recommendation algorithm based on one-class collaborative filtering and Bayesian pairwise ranking and optimize the model. The performance and effectiveness of the algorithm are validated through experimental evaluation and comparison with other benchmark algorithms. Regarding recommendation system design, we propose the corresponding architecture and modules. Importantly, we conduct functional and non-functional testing to verify the usability and performance of the system. Through this research, we provide a design and implementation solution for an intelligent recommendation system for e-commerce platforms, offering references for improving the accuracy of product recommendations and user experience. Future research can continue optimizing the algorithms and expanding the recommendation system's functionality and scope.

WiFi sensing is a hot research field of the Internet of Things (IoT). Utilizing channel state information (CSI) to locate or recognize posture, gesture, gait and so on has attracted the interest of many research scientists. Although previous studies have achieved some achievements, they have stuck to traditional machine learning methods due to data structure and some other limitations. With the development of hardware technology, there are more sub-carrier sequences in the CSI that can be collected, which makes CSI matrix more like an image rather than a time series. So our team believes that we can learn from deep learning methods in the field of image processing and computer vision to further improve the accuracy of the model. In fact, using gait recognition as an example, we collected more than 1,000 pieces of data from three volunteers, added attention mechanisms to a residual neural network which was pre-trained with ImageNet, and got a better model after multiple epochs of training. Specifically, our deep neural network achieved 94.6% recognition accuracy in multi-object classification.

A self-supervised mixed comparison recommendation method based on meta-learning (MSHCL) is proposed to address the problem of poor recommendation accuracy in social recommendation algorithms. Collaborative filtering based on graph neural networks can model the inter-node interactions between users and items, and make effective use of higherorder neighbor information. However, its representation is very susceptible to interaction noise, and thus the great potential of node-level information is not well utilized. To address the above issues, We first learn the embedding of nodes through the network view and meta-path view to fully capture the heterogeneous network structure. In addition, self-supervised learning as a novel learning method using unlabeled data effectively mitigates the data sparsity problem, which in turn fuses self-supervised learning into hybrid contrast learning model training. We have done empirical and ablation studies on a real dataset to demonstrate that the MSHCL model outperforms the current mainstream methods.

Deep learning is an essential part of machine learning, and neural network is an important technology in machine learning all the time. Neural networks can simulate the neural net and synaptic structure of the biological brain to complete data analysis and problem processing and can summarize the law from a time experience like human beings. For the human brain, the more layers and synapses the human brain neural network has, the deeper the level of abstraction of the input features established for some unknown things, so the deeper the accuracy of people's understanding is relative. Then, can the same logic be applied to neural networks in deep learning? Or, to put it differently, the deeper the neural network is and the more nodes it has, the better its performance must be.

In the contemporary retail sector, deciphering customer behavior is crucial for businesses vying for a competitive advantage. While customer loyalty has conventionally been gauged through parameters such as repeat purchases and promotional response, the merging of technology, media, and commerce has necessitated a reevaluation of these benchmarks. This study recognizes the inadequacy of solely relying on traditional transactional data to evaluate loyalty. Utilizing Machine Learning (ML), we crafted a novel model to predict customer loyalty, taking into account a plethora of modern factors, including online behavior and social media interactions. This research leveraged a diverse dataset and sophisticated ML algorithms to identify subtle patterns and relationships. Key findings include the significance of online behavior in loyalty prediction, the emerging role of social media interactions, and the nuanced influence of device preferences on brand loyalty. Notably, this manuscript offers a pioneering ML model that accurately reflects the complexities of the modern retail landscape, providing a valuable tool for businesses seeking to foster enduring customer relationships.