This PDF file contains the front matter associated with SPIE Proceedings Volume 12645 including the Title Page, Copyright information, Table of Contents, and Conference Committee Page.

According to the ubiquitous power Internet of Things strategy, the paper builds an intelligent analysis model for transformer online monitoring based on the power transmission and transformation Internet of Things architecture, summarizes the transformer’s intelligent sensing parameters and their acquisition methods, analyzes the monitoring data transmission method, and then proposes the application The algorithm model of prediction and diagnosis realizes the intelligent evaluation and early warning, and finally builds a cloud-side comprehensive evaluation model to form an efficient cloud-side collaboration system. This combination method efficiently and completely constructs a large-scale power transformer state assessment and fault warning method system based on cloud-side collaboration, which fully embodies the mechanism of rapid and timely warning at the edge, comprehensive assessment of the cloud, and collaborative operation of cloud-side information exchange. It is time-sensitive and comprehensive. It has specific guiding significance for the operation management and maintenance decision-making of the transformer under the new mode.

In view of the complex operating environment of ships and the variety of fault modes, it is difficult to achieve a unified evaluation. In this paper, the reliability of marine frequency converter system is evaluated from the fault mode to the whole system. FMECA analysis method is adopted to divide the system structure and evaluate various fault modes quantitatively. At the same time, combined with Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation method, the system reliability grade is presented comprehensively and quantitatively.

The hydrodynamic coefficient is the basis of the research on the maneuverability of underwater vehicles. In recent years, with the rapid development of underwater unmanned systems, the research on the calculation of hydrodynamic coefficient has attracted extensive attention from scholars at home and abroad. Taking an underwater vehicle as the research object to solve its hydrodynamic coefficient, three methods are used to calculate the hydrodynamic coefficient of the horizontal kinematic equation of the underwater vehicle: theoretical estimation, system identification based on actual navigation data and numerical simulation using the hydrodynamic calculation software FLUENT. In the end, the advantages, disadvantages, and applicable conditions of the three calculation methods are summarized.

Tracked vehicles and wheeled vehicles are classical weapons on the land battlefield, so their robust classification of them is an important part of radar automatic target recognition due to correct classification being of great significance to making decisions on the battlefield. To reduce the dependence of feature extraction on manual work, a novel classification method based on a Long Short-Term Memory network (LSTM) with amazing advantages in processing sequential data is proposed. The proposed method could mine the correlation inside the target, and automatically extract robust features from the target. The classification performance shows the validity of the proposed method with accuracy, precision, recall, and F1-score based on measured data respectively reaching 98.55%, 97.26%, 97.77% and 97.43% under the condition that the LSTM model is designed by four layers with the number of neurons in each hidden layer is 32. The experimental results by manually adding Gaussian white noise of different intensities to the test dataset verify the noise robustness of the proposed method. The compared results with other work indicate that the proposed method has better performance.

In order to alleviate the "power shortage panic" phenomenon caused by the lack of flexibility in the use of fixed charging stations, the introduction of mobile charging vehicles can effectively alleviate the charging problem of electric vehicles and improve the user experience. First, we divide the road network in specific areas, and then forecast the resident traffic volume (OD) data, so as to determine the initial location and distribution of mobile charging vehicles. From the global optimization of the charging time of the mobile charging vehicle at each charging position, ensure that each energy critical sensor is fully charged. At the same time, a mobile charging vehicle path planning algorithm considering the road state is proposed to determine the charging loop of the mobile charging vehicle and the charging time of the corresponding location, so as to minimize the total charging scheduling time on the loop. Finally, the performance of the proposed algorithm is evaluated through simulation experiments. Experimental results show that the proposed algorithm has low energy consumption, strong robustness and excellent performance.

Iris images captured in non-cooperative environments with visible illumination often suffer from adverse noise, which challenges many existing iris localization and segmentation methods. To address this problem, we propose a simple yet efficient anchor-free Iris Center localization and Segmentation Network, named ICSNet. Unlike existing methods that rely on post-processing on the segmented iris mask to fit the inner and outer circle, ICSNet provides an end-to-end iris localization and segmentation solution by explicitly considering the physiological characteristics of the iris. In ICSNet, an anchor-free center based double-circle iris localization network and an iris mask segmentation module are designed to directly detect the circle boundary of both pupil and iris and segment the iris region in an end-to-end framework. Extensive experiments on the challenging iris datasets (NICE-II) show that our ICSNet achieves excellent iris localization and segmentation performance. Especially, it achieves 84.02% box IoU and 89.15% mask IoU, and outperforms the state-of-the-art methods by nearly 3.8% and 1.65% respectively.

To realize the autonomous walking of the hedge trimming robot, a hedge trimming robot path tracking control system based on vision and an improved pure pursuit model is designed, and a path tracking algorithm based on genetic algorithm to dynamically optimize the foresight distance of the improved pure pursuit model is proposed. This paper first establishes the dynamics model of the robot and the improved pure pursuit model, extracts the operation navigation line, and obtains the navigation parameters by using the vision navigation system; based on the navigation parameters and the tracking error, uses the genetic algorithm to dynamically optimize the forward-looking distance in the pure tracking algorithm, and proposes a nonlinear steering control algorithm considering the heading error. The simulation test results show that when the robot driving speed is set to 2 m/s and the initial error value is 0.2 m, the proposed algorithm can converge quickly within a short time and the maximum absolute error value is much smaller than the initial value of 0.2 m, and the mean value of path tracking error is 0.022 m. The outdoor test results show that when the hedge trimming robot driving speed is 0.5 m/s, the path tracking algorithm with the improved pure pursuit model can achieve the maximum absolute error of 0.022 m. The outdoor test results show that when the hedge trimming robot travels at a speed of 0.5 m/s, the maximum absolute value of error is 0.081 m and the mean value of tracking error is 0.036 m. This paper provides technical support for the automatic navigation of the hedge-trimming robot.

A control method based on decomposed Virtual Model Control (VMC) is proposed to solve the drag teaching problem of the manipulator. According to the kinetic characteristics of the manipulator, springs and dampers are selected as virtual components, and their mathematical models are established. The motion of the end effector of the manipulator is decomposed into motion in six directions, and the corresponding virtual components are selected for movement in each direction and controlled separately. The VMC system of the manipulator is simulated, the effect of the stiffness of springs and the coefficients of dampers on the stability and phase margin of the control system is analyzed by drawing both Nyquist and Bode diagrams. An experimental platform for manipulator dragging motion control was built to conduct dragging experiments, and the dragging performance of the manipulator under VMC method and traditional torque mode was compared.

The southwestern part of Shandong province has become one of the most serious land subsidence disasters in Shandong Province. Through a long series of land subsidence monitoring in this area, the land subsidence status in this area can be fully grasped, and scientific basis for the prevention and treatment of land subsidence can be provided. In this paper, Small Baseline Subsets - Interferometry Synthetic Aperture Radar (SBAS-InSAR) technology was used to process the Sentinel-1A data from September 2019 to October 2020 in southwestern Shandong, in order to obtain the range and rate of ground subsidence in this area. At the same time, the precision is verified by the leveling results of the same period. Three feature points are selected to analyze the temporal deformation characteristics of the surface, and the influencing factors of the ground subsidence in the study area are analyzed. The results indicate that during the past 2019-2020 years, the ground subsidence in the southwestern Shandong province was mainly distributed in the east of Mudan District of Heze City, the north of Dingtao District of Heze City, Yuncheng County, Shan County, Jinxiang County of Jining City and its surrounding areas. The most severe subsidence area is located in Yuncheng County, with an annual subsidence rate of 129.1 mm/a. The ground subsidence in southwest Shandong is significantly affected by groundwater overexploitation and coal mining.

Logistics finance, as a synergistic innovation in the fields of logistics and finance, is a new way to solve the problem of difficult financing for SMEs. In this paper, a patrol process for logistics enterprises is established based on key business controls and condemnatory costs, the characteristics of patrol work are modeled, and queuing theory is used to investigate the optimal number of warehouses to be managed by a single patrol staff. On this basis, a new logistics finance business model is introduced, a many-to-many queueing theory model is used to calculate the allocation of patrol staff for a fixed number of shops in a region, and the patrol process is optimized for the new business model. The practicality of an actual logistics finance case is analyzed and verified.

This paper deeply uses the Chinese traditional pattern "falling flowers and flowing water pattern" as the research object, combs and analyzes its historical origin, analyzes its basic characteristics and composition forms, and discusses the secondary innovative design of the pattern in combination with the practice of pattern design. Using the operation demonstration method of the jacquard CAD interface, according to the actual situation of the jacquard machine, set the specifications of the gray cloth and the machine specifications, configure fabric weave, create loom layout and auxiliary needle design, generate the pattern card and simulate the pattern. The results show that according to this experimental process, the pattern can be simulated, which can provide new materials for modern clothing design.

The stability of aircraft structure is one of the important issues in aircraft structure design. This paper conducts buckling analysis on different boundary conditions of column models, conducts buckling and post-buckling analysis on different materials of rectangular plate models and conducts different radii of curved panel models in aerospace structures. The critical loads for the different models were calculated and compared by software simulation and numerical analysis. The first-order modes of 0.01 and 0.05 times as the initial defects have been introduced into the rectangular plate models in order to carry out the post-buckling studies of rectangular plate models.

Most existing urban underground sewage pipeline systems (hereinafter referred to as sewer systems) determine the existence and type of faults by manually checking videos or even entering the pipeline directly, which is a time-consuming and laborious process. Some automatic image searching based on traditional determination rules have low generalization and accuracy. Although the target recognition detection model based on the depth learning method requires a certain number of labeled images for learning and training, it improves the accuracy and strengthens the generalization. In this paper, we collected and labeled 2330 pictures as the data set and improved a deep learning model for sewer fault detection. The experiment proved that the detection accuracy of this technology for cracks, pipe deformation and dislocation, silting, corrosion, and falling off faults reached 80.5%, 91%, 94.1%, 87.5%, 88.9% respectively. The mAP value of target detection reached 88.4%.

Atomic capabilities are common elements or links supporting various applications and businesses, such as voice capabilities, SMS capabilities, RPA, computing capabilities, storage capabilities, etc. The enterprises formed through the encapsulation of internal capabilities and the introduction of external capabilities can be provided independently and have the characteristics of being integrated (standardized), reusable, pricing, extensible, authorized, and having a wide range of common needs. The current situation of China Telecom's atomic capability building is as follows: 1) The capabilities are invisible, untraceable and cannot be conveniently invoked. 2) Lack of a unified platform for capability deposition and display: various capability centers, professional companies, integration and general service systems and provincial companies in the whole group have accumulated certain capabilities during the implementation of DICT projects but failed to display and manage them in a unified way, and other units of the group cannot obtain relevant capability information. 3) The process of capability transfer is not smooth: it is difficult to transfer, long cycle, and difficult to settle, which makes it difficult to transfer and settle. The capability development unit cannot obtain benefits, and the enthusiasm of capability output is relatively low. 4) Lack of iteration of capability: the original DICT project or scientific research capability is lack of iteration and update, the project delivery and settlement are the completion of the project, and the project and capability are lack of long-term iteration and update and application mechanism. 5) The introduction of external ecological innovation capacity is slow: there is a gap between some of the current self-owned capacity and the external ecological innovation capacity, leading to weak competitiveness and development of some projects. Limited by processes and mechanisms, external ecological innovation and scientific and technological power cannot be quickly introduced and applied to internal and external projects, Lack of mechanisms and platforms for capacity exchange and introduction. In order to solve the above problems, this paper provides an open platform for building atomic capabilities, which can be checked, inquired and experienced. Provide standardized interface, real-time experience, fast delivery, 7*24 full-time service. Direct transaction between capability suppliers and front-end market demanders; Mobilize enthusiasm and enhance competitiveness. Provide an internal innovation platform to mobilize the enthusiasm of employees. The market speaks for product retention and value.

In order to solve the problems of heavy workload and low efficiency in the application of traditional CAD drawing in drilling management, a set of control software for gas drainage drilling was developed, taking the 1313 working face of Zhaozhuang Mine of Shanxi Jinneng Holding Coal Industry Group as an example, which realized three-dimensional visualization of gas drainage drilling and could analyze and evaluate the gas drainage effect. The calibration of gas geological anomalies, the automatic generation of drilling construction details and the auxiliary design of drainage drilling measures are realized. The field application effect is good. The drainage effect of gas drainage borehole is analyzed and evaluated. The drainage blank zone and geological abnormal area are calibrated, and the 3D display and details of borehole design and completion drawing are automatically output. Finally, the centralized dynamic management of drilling and drainage is realized, which guides the safe and efficient production of coal mine.

Prompt-based Pre-trained Language Models (PLMs) have demonstrated their superior performance on a wide variety of downstream tasks. In particular, the performance of prompt tuning has significantly outperformed traditional fine-tuning in the zero-shot learning and few-shot learning scenarios. The core idea of prompt-tuning is to convert different downstream tasks to mask language modeling problems through prompts, which can bridge the gap between pre-training tasks and downstream tasks for better results. Verbalizer, as an important part of prompt-tuning, can largely determine the final performance of the model, but the design of the Chinese-based verbalizer is yet to be fully explored. In this paper, we propose a method to expand the verbalizer by extracting knowledge from the training set based on a Chinese text classification task. In brief, we first segment the Chinese training set, then filter the words that can express the semantics of the labels by semantic similarity, and finally add them to the verbalizer. Extensive experimental results on multiple text classification datasets show that our approach significantly outperforms ordinary prompt-tuning and outperforms other methods for constructing the verbalizer.

The airborne system architecture describes the cross-linking and hierarchical relationships and information interactions between the abstract functions and the physical implementation of the system. This study replaces the traditional empirical and document-based design of the object search and guidance function of aircraft with a model-based design based on the system modeling language. The correlation and traceability of the model help ensure the integrity and traceability of information in the design process. In addition, the design relationships between the functional and physical architectures of the model are used to validate this modeling and design approach.

In order to improve the construction of smart home system and create a smart home environment for users. This paper studies the application of Internet of Things technology in smart home system. This paper expounds the design of smart home system, applies the Internet of Things technology to the home environment, and realizes the transformation of network transmission from wired to wireless, including the management layer, the network layer and the perception layer. Using data processing chips to work dynamically, it can efficiently collect and transmit smart home information. Achieve the control purpose of smart home system.

For the past few years, the Internet of Things (IoT), a wireless communication mode, was rising rapidly with the development of the Internet. it had been widely used in the field of greenhouse planting recently, such as "IOT + Agriculture" as the trend of new agricultural development in the future. Wherein, Dendrobiums, which were difficult to pick were demanded by more and more consumers. Although the emergence of Dendrobium greenhouse planting made the quality and yield of Dendrobium are guaranteed to a certain extent. To tackle the problem, IoT greenhouse monitoring management and control system based on Arduino for Dendrobium cultivation, is proposed in this paper, consists of four modules, data acquisition module base on ESP8266 Arduino working for data acquisition, action module for hardware controlling, wireless transition module working as the local area network for the system, and data processing module based on MQTT working as the frond-end system. Thus, a remote monitoring of greenhouse environment can control the growth conditions of Dendrobium, including air quality, and soil fertilizer factors according to the requirements of Dendrobium growth habits on the growth environment strictly. In practice, the proposed system can protect Dendrobium growth factors dynamically to assist farmers to make better planting plans and improve the quality of Dendrobium.

Given the need for managers to consider energy management of low-power devices in practical production, particularly in the field of agricultural irrigation, in order to improve the efficiency and effectiveness of water use in agricultural irrigation, there is a need to design a micro-hydropower energy saving solenoid valve system to achieve intelligent irrigation in agricultural work by controlling the opening and closing of the solenoid valve. In this paper, a micro-hydropower energy saving solenoid valve system is designed, including a generator module, solenoid valve module, microcontroller control module, power management and storage module and relay module, where the solenoid valve uses the self-locking feature, which requires 0.225W to open the valve and 0.315W to close the valve. The TP5100 chip is used to manage the charging of the 8.4V lithium battery, according to the charging The TP5100 chip is used to manage the charging of the 8.4V lithium battery, adjusting the current according to the different stages of charging. The solenoid valve system sends control signals through the host computer to establish a comprehensive agricultural IOT intelligent management system integrating self-generation, remote control and early warning analysis, which greatly facilitates the user and provides a practical solution for intelligent planting and irrigation.

This paper describes a design for a new type of Remotely Operated Vehicle (ROV) and a system to control it. The ROV and its control system are part and parcel. The control part of the ROV includes the Data Acquisition Program. The Data Acquisition Program is a GUI-based computer program that allows the operator of the ROV to visualize telemetry data from the ROV. It provides near-instantaneous communication with the ROV. The computer-aided design of the prototype is made using Autodesk’s Fusion 360. Then the design was converted to Stereolithography (STL) format and passed through a slicing software called Cura in preparation for 3D printing. An Ender 3D printer was used to manufacture the parts of the design, using Polylactic Acid Thermoplastic Polymer. Computational Fluid Dynamics is conducted to analyze the fluid hydrodynamics (water and pressure flow rates) around the hull of the remotely operated vehicle and the drag force acting on it. Several experimental results have shown the effectiveness of the remotely operated vehicle design.

In the application of machine learning to malware detection, the number of normal software samples is usually much more than that of the malicious samples, which leads to imbalance problems in training data and arises great challenges to build classification models. To improve the classification accuracy, a group-ensemble method is proposed, where normal samples are divided into several groups using sampling technology. Then each group is combined with all malicious samples to form a training data set on which a classification model is learned. Finally, all classification models are integrated into one single classification using ensemble method. The experimental results over real-world data sets indicate that the proposed method can handle the imbalance problem in malware detection with improved detection accuracy.

Although extensive research has been carried out on Human Action Recognition (HAR) and HAR has been applied in many fields of computer vision, little studies focus on the operation detection of workers in petrochemical plants. There are thousands of inflammable chemicals and large amount of toxic gas during the production in petrochemical plants, which would cause serious disaster. Thus, we argue that real-time detection of labors’ actions is required to keep their safety and the production efficiency of petrochemical plants. But most of state-of-the-art models of HAR need a lot of training time and most of them have huge model size which makes it is hard to deploy them on edge devices. To solve this problem, this paper proposes an architecture to detect labor’s operation of pipeline valves in petrochemical plant based on YOLOv5. We calculated relative positions of different objects which were detected by YOLOv5 and added an estimation layer behind output layer of YOLOv5, so that we could use our architecture to estimate labors’ actions. Finally, our experiments show that our architecture can estimate the worker’s actions with 97.5% precision. The precision, the recall, and the mAP@0.5 of object detection is 98%, the mAP@0.5:0.95 is 85%.

The out-of-distribution detection task is a hot problem in image processing, aiming to detect samples that have never been trained by a deep learning model, thus ensuring the safety of the model for use in real-world environments. Currently widely used out-of-distribution detection methods suffer from single-layer representation dependence and poor adaptability. In order to overcome these problems, this paper proposes a Multilevel Feature Detection Network (MFDNet), which detects different representations of different out-of-distribution examples with the help of intermediate layer features. At the same time, a grouping module is introduced in the final layer of MFDNet to help the model reduce the decision boundaries between categories and enhance the adaptability of the model. A comparison with several common out-of-distribution detection methods on the out-of-distribution evaluation datasets validates that MFDNet can effectively handle the out-of-distribution detection task.

As China enters an aging society, the problem of aging is becoming more and more serious. The elderly are often alone when their children are at work. Due to their poor physical fitness, they are prone to falls, injuries or even unaided situations during outdoor activities. It has been shown that more than 47% of elderly people who fall cannot stand up on their own. If they cannot be found in time, the best rescue time may be missed, and secondary damage and negative effects may even occur. In this regard, there are currently researches on fall detection technology for the elderly at home and abroad. The existing fall detection schemes are roughly divided into three types: wearable sensor detection, environmental sensor detection and vision-based detection. It can only be used in a specific environment to play a role. Due to the high cost of visual sensor detection, it is generally only suitable for indoor occasions, while wearable sensor detection technology is relatively low in cost and suitable for outdoor occasions and has more applicable scenarios. For example, Zigbee is used to transmit terminal data to the server for prediction, but Zigbee has the disadvantage of short communication distance for long-distance outdoors. The use of Bluetooth to transmit terminal data to the host device for prediction is also only applicable to short-distance indoor scenarios. This paper designs a fall detection system for the elderly based on the LiteOS operating system and 4G Cat1 module. It uses the high-precision and low-drift MPU6050 to collect human body posture information and relies on the Cat1 module of the high-coverage and high-speed 4G network to transmit the posture information. To the cloud server, the server runs the CNN model on the terminal data to make a decision, and finally feeds back the decision result to the terminal device for alarming. For a wide range of outdoor occasions, it is helpful to prevent the elderly from falling in outdoor activities and causing life injuries without rescue.

Aiming at the defect that the parameter variation affects the performance of the non-inductive observer during the long running time, the UKF is introduced into the design of sensor-less control system of permanent magnet synchronous motor. In this paper, the resistance and inductance are used as the expanded dimension state variables for state estimation to ensure the stable operation of the non-inductive observation algorithm. On this basis, the design scheme of PMSM control system is given, and the experiment is carried out to compare the improved non-inductive observation algorithm with the traditional non-inductive observation algorithm in the electrical angle estimation, and the experiment proves the excellent performance of the system for motor speed control.

We propose a fast and accurate multi-target tracking method, Twice Check. This method is a multi-target tracking method, which matches the newly appeared track and the just disappeared track and adds the low-frame quadratic verification and linear interpolation to repair the track. Our method, combined with the current SOTA method, can improve the effect of the original method on the open Multi-Target Tracking challenge (MOT) dataset. The experimental results show that our Twice Check can effectively increase the effect of the current tracking model, and our method is simple and does not increase the tracking time.

The object tracking algorithm based on the deep Siamese network has a low tracking success rate and low robustness under the conditions of target illumination changes, occlusion and deformation. Therefore, this paper proposes a deep Siamese network tracking algorithm with attention mechanism based on SiamRPN++. First, an attention mechanism is added to each layer of the feature extraction network ResNet50 to calculate the importance of each channel, so that the model can obtain more useful information. Second, since the shallow features focus on the details of the target, the deep features focus on the semantic information of the target. Therefore, a feature fusion method based on the attention mechanism is proposed to fuse the deep and shallow features to enhance the expressive ability of the features. On the OTB100 and LaSOT datasets, the success rate of our tracker is 70.1%, 51.6%. Compared with SiamRPN++, it has increased by 1.1% and 2.4%.

As an integrated supply system combining multiple forms of energy, the regional multi-energy system can realize the functions of multi-energy mutual aid and efficient consumption of renewable energy. It is expected to become an important energy supply solution to help achieve the strategic goals of carbon peaking and carbon neutrality. Existing distribution network analysis methods and tools are difficult to apply directly to regional multi-energy systems. This paper systematically combines digital twin technology with the planning and operation of regional multi-energy systems. Based on the core digital twins technologies such as panoramic digital simulation and real-time data interaction, key technologies such as dynamic planning and operation optimization of regional multi-energy systems are proposed. A set of digital twin solutions covering the whole life cycle of regional multi-energy systems is formed. At the same time, to break the bottleneck of digital twin technology in visualization and application, a demonstration and validation scheme for digital twin planning and operation technology of regional multi-energy systems is proposed. Several validation scenarios are designed to demonstrate the effectiveness and superiority of the proposed digital twin solution. And a set of regional multi-energy system digital twin auxiliary decision support software platforms integrating combined cooling, heating, and renewable energy generation facilities has been built in the demonstration area of Hongfeng Lake in Guizhou, China, in which the integration demonstration of key technologies such as digital twin-based planning and optimization has been completed. It has effectively improved the operation economy and the capacity for renewable energy consumption in the demonstration area.

Nowadays, unbounded and disordered large-scale data sets are becoming more and more common, and consumers' processing requirements for these data sets are also becoming more and more complex, such as time semantics, windows and processing delays. In order to meet the continuous development of data processing requirements on unbounded and disordered large-scale data sets, this paper analyzes the advantages of dataflow processing from the aspects of low latency, high parallelism, low synchronous memory access cost and simple on-chip logic. On the one hand, the dataflow diagram embodied in the data flow calculation model in big data processing is analyzed from the execution engine level. On the other hand, several typical dataflow processing architecture models are analyzed from the perspective of dataflow batch processing. On this basis, the problems faced in the development of current dataflow are analyzed, and the possible development direction in the future is proposed.

The power industry is developing fast and there are many power construction projects. With the growing intelligent degree of power system, the construction of power projects become more complicated. There are more than one thousand workers on the site. How to inspect and guarantee the security in construction remotely is a key subject. This paper introduces the principles and means of remote inspection system of infrastructure construction by using internet of things in power systems. Through the application of this system, violations on the construction site can be discovered in a timely and effective manner, and the planning level of construction can be better improved.

Scaling caused by long-term operation is a common fault of the heat exchanger. Fouling will reduce the heat transfer efficiency of heat exchanger, thus affecting the operation cycle of equipment, reducing product quality, and easy to induce production safety accidents. The quantitative evaluation principle and classification standard of heat exchanger fouling condition based on calculating the heat transfer coefficient as the core analysis process were proposed. The calculation methods of query temperature, pressure difference, temperature difference, effective average temperature difference, heat flow and heat transfer coefficient of fluid physical parameters were analyzed. A distributed monitoring system for fouling state of heat exchanger was developed and the system hardware design method and software operation architecture were introduced in detail. The system was successfully applied to the online monitoring of fouling state of a U-tube heat exchanger. The fouling hidden danger of the tested heat exchanger was found.

Currently, the control of civil aircraft water and waste system resides in IMA, which is a relatively advanced airborne system control mode. But due to the complex cross-linking between the control mode and the aircraft, the real environment of the aircraft needs to be restored to a greater extent through simulation in the laboratory integration verification under the premise that there is no real aircraft environment, and the aircraft signal and system failure state can be monitored and simulated. This paper proposes a set of water and waste system integrated verification model design method combining the functions of the water and waste system with resident IMA and control architecture. It simulates the real environment of the aircraft, as well as system status monitoring and fault state simulation, and completes the functional verification of the water and waste system under various working conditions at the lowest cost.

In large-scale engineering experiments, such as hydrodynamics and aerodynamics experiments, Computer Aided Software (CAE) is always used to manage a large number of parameters and experimental data to simulate the physical field through finite element analysis. To improve work efficiency, one of the most practical methods at present is to record user operations in the CAE software to script files with Domain Specific Language (DSL), such as Python, and replay those script files to reproduce user operations. In this paper, a common script module ScriptEngine supporting the recording and playback function is introduced, and it can be embedded into any CAE software to automatically complete grid division, flow field calculation, pre-processing and post-processing by running the scripts. Furthermore, a demo program embedding ScriptEngine into an opensource FastCAE was developed, and the results show that running the scripts recorded by ScriptEngine could obtain the same results as the user operations and ScriptEngine can be used as a common script module in any CAE software and other desktop software which needs the recording and playback function.

In the scenario of online monitoring of underground water in-situ remediation system, where massive real-time monitoring data is produced in a high frequency, traditional relational database cannot efficiently manage data created by the monitored system. Focusing on such issue, this paper provides a monitoring solution based on time series database. On server side this system applies InfluxDB to accept data from device side. On device side, the system chooses Programmable Logic Controller (PLC) as main controlling unit. We also use Docker during deployment for easier maintaining. Experiments show that the system can manage and store massive real-time surveillance data. This design is of some significance in terms of real-time online monitoring for automatic and IoT systems.

Aiming at the fact that traditional dynamic form development relies on a specific front-end framework and does not have portability, a visual dynamic form model generation engine for multiple front-end frameworks is proposed. First, drag and drop components and configure properties on the visual page editing platform to complete the construction of the dynamic form model. Then, through the form configuration module of the business logic layer, the business information and attribute configuration information of the dynamic form model are mapped to the relational database. Finally, the above information stored in the relational database is extracted through the front-end adapter of the business logic layer and enhanced into JSON configuration files required for dynamic form rendering with consistent structure, which are persisted to the non-relational database. The experiment shows that the dynamic form model built by the engine has good portability among various mainstream front-end frameworks after being transformed by front-end adapters.

With the increasing application of computer vision in robot systems, it is vital to improve the positioning accuracy of robot end-effectors. Hand-eye calibration is the first step to realize it. However, in the process of calibration, there are some accidental factors leading to inaccurate calibration. In this paper, a data-driven hand-eye calibration approach based on Zhang's calibration method is proposed to complete eye-in-hand calibration. Re-projection error is used as an evaluate index to evaluate the positioning error. The calibration data with errors greater than the average error were filtered to reduce the accidental errors caused by uncertain factors. Finally, taking AUBO I5 collaborative robot as an example to verify the feasibility and effectiveness of the proposed approach. It shows that our approach can improve the positioning accuracy of the robot system.

DCU (Deep Computing Unit) acceleration device is a kind of GPU-like device, which is developed by GPU technology authorized by AMD, so it is suitable to solve the parallel problems of many data elements in the same program. In this paper, a parallel immersed boundary method based on DCU acceleration device was studied to realize efficient two-phase flow simulation. The Jacobi multigrid method was used to solve the pressure Poisson equation in the simulations. Based on our serial code and combined with the hardware characteristics of DCU and HIP (Heterogeneous Interface for Portability) heterogeneous programming model, the calculation of pressure Poisson equation used in immersed boundary method was transplanted to DCU, and the parallel immersed boundary method based on multigrid method was realized. Three DCU parallel simulations were carried out: (1) solutions of Poisson equation; (2) two-phase flows of droplet deformation in shear flow and (3) bubble rising process in the static fluid. We tested them on Song Shan supercomputing platform. Compared with CPU version, the DCU parallel immersed boundary method has achieved the highest speed-up ratios of 16.888, 5.223 and 9.604 for resoling Poisson equation, drop deformation in shear flow and rising drop simulations, respectively.

The delegate power purchase mechanism in China’s electricity retail reform requires power grid companies to make accurate user load forecasting and reasonable market decisions, and the result will affect the user’s electricity price. Under this background, a monthly load forecasting method is studied for power grid company considering the impact of customer churn on the agency electricity consumption. Firstly, a customer churn model based on life cycle theory is proposed to predict the customer churn rate of a power grid company in a specific period of the development process of China’s electricity retail reform. The seasonal and trend decomposition algorithm for decomposing user monthly electricity consumption series is presented, based on which the trend, seasonal, and remainder components of user load are predicted respectively by using polynomial curve fitting methods. Then, the monthly load forecasting model for agent users is proposed according to the expected customer churn rate and the prediction of different load components. A simple electricity market is served for demonstrating the proposed method, and the simulation results show that the proposed method has higher accuracy of monthly load forecasting than the other model that does not consider the predicted customer churn rate of power grid companies.

Energy management strategy can greatly affect fuel economy of Hybrid Electrical Vehicles (HEVs). This paper proposes a Dynamic Programming (DP) algorithm to optimize the power supply ratio of Electric Motor (EM) and Internal Combustion Engine (ICE). The results show that World Light Vehicle Test Procedure (WLTP) optimized by the DP algorithm can keep the State of Charge (SOC) of the battery within the set range. Compared with the Rule-based CDCS strategy, the fuel saving rate can reach 28.4%. The method can be valuable for the develop of energy control strategy of HEVs.

As some fields can not adapt to the changes of the new situation in time, many problems are gradually exposed. For example, in modern electrical engineering, there are problems of poor overall performance and uneven structure, so we design the corresponding electrical control system based on big data technology and apply it to practical work. This paper introduces the structure and main functions of the electrical control system, including data acquisition, information transmission and electrical control. We select university students as the experimental object. The students are divided into two groups, group A students in the electrical control system design and implementation of big data technology, while group B students are still using the conventional method. The research results show that the electrical automation control system designed by group A students not only meets the requirements of intelligent dispatching, but also has a success rate of 75%, and 72% of the students in this group are satisfied with the technology applied in this experiment, and the work efficiency even reaches 95%.

Aiming at the problems of difficult track initiation and high missed detection rate in dense clutter environment, this paper presents a method of multi-target hypothetical track initiation based on motion characteristics. Most of the traditional track initiation algorithms use the sequential start method, which has a great dependence on time. In this paper, a variety of hypotheses are generated by directly judging the correlation of the point track data, and the final track information is obtained by screening the formed temporary track according to the kinematic characteristics of the target. Experiments show that this method not only greatly improves the accuracy of track initiation in dense clutter environment and reduces the missed detection rate, but also has strong adaptability to track initiation of massive data.

Temporal Action Proposal Generation (TAPG) is a challenging task at present, which aims to accurately generate temporal proposals that are likely to contain human actions in an untrimmed video. Inspired by the successful pre-training strategy applied by the transformer-based image object detection method, we propose a Self-Supervised Pre-Training transformer (SSPT) method for the TAPG task to train an action locator without any labels. As far as we know, this is the first work to explore self-supervised pre-training for transformer for TAPG. Specifically, the pre-training strategy leverages the transformer model to model the contextual information and we leverage transformer to capture contextual semantic information and global temporal dependencies of actions in the entire video can better accomplish the TAPG task. And in the pre-training process, we adapt Random Query Segments, which randomly crop multiple segments from the original video. We treat the segments as pseudo-labels and input them as queries to the transformer decoder. The pretext we designed is to locate the start time and end time of the pseudo-labels in the original video. Extensive experiments on THUMOS14 demonstrate the effectiveness of SSPT, the results show that it improves the performance of baseline and leads to a higher recall for the Temporal Action Proposal Generation task.

Recent advances of transformers have brought new trust to computer vision tasks. However, on small dataset, transformers is hard to train and have lower performance compared to convolutional neural networks. We make vision transformers as data efficient as convolutional neural networks by introducing focal attention. Inspired by the local attention networks, we constrained the self-attention of ViT to have multi-scale localized receptive field. We provide empirical evidence that proper constrain of receptive field reduce the amount of training data for vision transformers. Our best model reaches 83.16% accuracy when training from scratch on CIFAR-100 which is a significant improvement in data efficiency over the previous transformer. We also perform analysis on ImageNet to show our method does not lose accuracy on large datasets.

The repetition of test questions raises a significant challenge to test fairness. When processing the test texts, we found that the repetition of test questions is not only the similarity of context but also the context difference with the similar semantics, which is more challenging to identify. In addition, the existence of formulas in the test questions brings a greater challenge to detect the repetition of test questions. To address the above problems, we propose a semantic embedding representation model for test question duplication detection. The model incorporates knowledge point information and extracts the semantic information of test questions through a dual encoder structure and set a threshold to obtain top-n test questions with high similarity by calculating the cosine similarity between test questions. The experimental results show that this paper's proposed model greatly improved over the traditional repetitive detection model.

This paper proposes a medical image denoising method based on two-dimensional adaptive compact variational mode decomposition (2D-ACVMD). First, a binary support function is introduced into 2D-VMD in order to enhance its spatial compactness. The optimal decomposition mode number K and the penalty factor α can be determined by the moving average power spectrum method and the genetic algorithm, respectively. Then, the decomposed high-frequency noise component is removed, and the other modes are reconstructed to obtain the denoised image. According to the experimental results, this method can effectively remove the noise interference, retain the original edge and detail features, and the PSNR and SSIM obtained are higher than the previous methods.

Simulation analysis of composite structure projectile high-speed penetration thick target was carried out to investigate the characteristic law of fuse dynamics response of composite structure projectile body. ANSYS/LS-DYNA software was utilized to carry out numerical simulations of a composite structure projectile and a conventional projectile positive penetration concrete thick target to compare the dynamic response of the components of the composite structure projectile during penetration. From the results, it was concluded that during composite structure projectile body penetration, a secondary peak in the overload signals on the projectile body and fuse components occurs when the projectile was fully eroded. When the initial velocity of composite structure projectile body was higher, the erosion rate of projectile was faster. When projectile was completely eroded, the overload characteristic signals on projectile body and fuse appear earlier and more pronounced.

Deep Convolutional Network (CNN) can make the pictures generated by GAN more reasonable, but limited by the local receptive field of CNN, there are still many unreasonable places in the authenticity and semantics of the multi-object images generated from text. Therefore, a GAN-based method that incorporates a non-local self-attention mechanism is proposed. By embedding a non-local self-attention structure in the network, the network obtains global semantic information and detailed features, and uses the obtained information to perform level-by-level encoding to generate the final relatively reasonable image. The amount of parameters and calculation of the entire model is also reduced a lot. The proposed method is verified on the public COCO-stuff dataset and uses multiple indicators such as Inception Score, FID and classification accuracy score to evaluate the authenticity and diversity of the generated images. Experimental results show that the quality of the generated images is superior to that of previously proposed methods.

Task-driven Language Modeling (TLM) can reduce the amount of calculations by an order of magnitude in NLP tasks but achieve the effect of matching the BERT-like large model. It uses BM25 to retrieve relevant data subset from general datasets and combine them with task dataset for joint training, but the extraction process does not consider semantic information, and IRGAN can further extracts semantic information features in the way of adversarial training, and the combination of them can use semantic information to enhance the relevance between the auxiliary training dataset and the labeled task dataset, and then through linearly decreasing joint training in for low-resource tasks have been improved compared to TLM.

In the blind super-resolution reconstruction task, it is easy to fall into a single degradation. In order to solve this problem, we add the dropout mechanism before the last convolutional layer of the network, which can effectively prevent co-adapting, so that the network can be well restored to various degradations. To enhance the network's ability to extract high-frequency details, a HAM attention mechanism is introduced behind the residual block, which can reconstruct clearer textures. Aiming at the problem of ubiquitous artifacts in images generated in fine-grained and structured areas in the generative confrontation network, the LDL loss is introduced to make the area better restored. Experimental results show that the method proposed in this paper can achieve high performance and is closer to natural images in subjective vision.

Although generative models have made great progress in art image generation, few of them pay attention to sketch-based art image generation. Generating art images from sketch is a challenging problem suffering from two main issues: (1) how to constrain the generation of art images with sparse sketch features and (2) how to fully utilize the style information of a reference image without being influenced by their content features. To tackle these, we propose a GAN-based method for art image generation given the sketch and reference style. Specifically, a style feature enhancement module and a sketch-adaptive normalization module are constructed to enable the disentanglement of the content information from the reference style image. Experiments and comparisons demonstrate the superiority of our model over current generative models in sketch-based art image generation.

The low-light image has degradation problems such as low contrast, noise artifact, loss of details and color distortion, which cause poor visual perception quality of the image, and reduce the accuracy of subsequent image recognition, classification, and detection tasks. To solve the above problems, a low-light image enhancement method is proposed, which combines attention mechanism and Generative Adversarial Networks (GAN). A generator network model with convolutional attention module (CBAM) and residual block is designed. Three loss functions of content, color and texture are proposed for image features. The experiment is compared with the four advanced methods on LOL dataset. In this dataset, the experimental results are at least 5% better than those of other methods. The experimental results show that the saliency model proposed in this paper can effectively enhance the low-light image and obtain a good visual effect.

With the rapid development of big data, cloud computing, artificial intelligence and other technologies, the expansion of database application requirements and the change of computer hardware environment, especially the rapid development of computer network and digital communication technology, satellite communications, cellular communications, computer local area networks, wide area networks, and the proliferation of Intranet and Internet have been widely used, resulting in the emergence of distributed database systems. Especially for basic industries such as telecommunications, databases often carry thousands of business systems, and millions of queries and tens of millions of bills are carried out every day. In the face of such massive scale, high concurrency, data security, real-time analysis and other multi scenario problems, the processing capacity of traditional centralized databases is far from meeting the needs. For this reason, this paper proposes a distributed database based on MySQL transformation. Through research and implementation, it can solve the above problems, meet the requirements of distributed transaction processing, ensure data backup, high availability and high performance, and improve efficiency. This paper will also describe its modeling model and sharding algorithm in detail.

Low-Dose CT (LDCT) scanning can greatly reduce the radiation damage to patients but would introduce serious noise and artifacts to CT images. The traditional deep learning based LDCT denoising methods are fundamentally based on the convolution operations, while their receptive fields are small and cannot capture long-range correlations. The Global Context (GC) attention modeling mechanism can solve this problem with significantly less computational burden. However, the only one global context feature used in the GC block generally cannot well describe the non-homogenous structural statistic property of LDCT images. To this issue, this paper proposes an Adaptive Global Context (AGC) modeling scheme for better representing local contextual semantic information of CT images. An AGC-based residual auto encoder-decoder network (AGC-RED) is further proposed for efficient LDCT image noise reduction. The effectiveness of the presented AGC-RED network is validated by experimental clinical studies.

Low-Dose CT (LDCT) imaging technique can reduce the radiation exposure to patients but would inevitably introduce serious noise and artifacts to the CT images. Deep learning based LDCT denoising methods with supervised learning strategy have shown promising performance. However, they heavily rely on well-aligned training image pairs that are in fact very difficult to be obtained in clinical practice. In this study, we propose a CycleGAN-based unsupervised LDCT image denoising method with a novel Selective Spatial Pyramid Attention block (SSPA). The proposed SSPA can effectively fuse the attention map with the original feature map through an attention scheme, thus better extract the representative features for fine structure restoration. Experimental results show that the proposed unsupervised method achieves state-of-the-art denoising performance for LDCT images.

In order to solve the problem of identifying mental health status among college students, firstly we extracted features from college students' status data, including physiological indicators, family income, academic performance and behavior data, then applied the Apriori algorithm improved with pre-pruning strategy to analyze the relationship between each feature and students' mental health, and finally used the XGBoost model optimized by SMOTE+ENN method to predict the mental health based on the results of the feature analysis. The model achieved good predictive performance: precision is 0.87, recall is 0.86, F1 score is 0.86 and AUC is 0.89. The experimental results on the student status data showed that the proposed model outperforms traditional machine learning and deep learning models, thus making our model able to predict the mental health of college students more effectively and accurately.

We propose a robust, accurate, and fast ultrasonic sequence target tracking method. This method is a self-supervised approach constructed using inter-frame temporal information and builds a framework on the Siamese network that can train unlabeled data. Our method has an error evaluation of 2.01 ± 3.11 mm on the publicly available Liver Ultrasound Tracking Challenge (CLUST) 2D dataset. The experimental results show that our self-supervised approach can effectively track landmarks in liver ultrasound images and that our network structure is simple enough to be competitive in terms of real-time performance.

Visible and infrared cameras can solve the practical problem of high-resolution image recognition in complex operation scenes of substations, but the prerequisite is that visible and infrared cameras need to be calibrated effectively. In this paper, a calibration method of visible light and infrared camera based on edge detection is proposed. The method adopts a visible light camera and an infrared camera to shoot two images of the same calibration plate, namely a visible light image and an infrared image, and implements preprocessing. It detects the edges of two images, extracts feature point description information by using a scale invariant feature conversion algorithm and determines four groups of determined matching point pairs. It solves a transformation model by combining an internal parameter matrix to obtain an external parameter matrix and realizes visible light and infrared camera calibration. The results verify that the parallel equipotential error is less than zero. One under the application of the method, which indicates that the calibration effect of the method is better.

By predicting flight delays, economic losses caused by delay situations can be effectively prevented and reduced. A classification algorithm for flight delay prediction based on unbalanced data is proposed to solve the problem of unbalanced number of delayed and non-delayed samples in the process of flight delay prediction. Firstly, a multi-classification model based on the CycleMLP algorithm is proposed for flight delay prediction. Secondly, the CycleMLP algorithm is improved by using the Focal Loss modified cross-entropy loss function to build the final classification model for flight delay prediction. And the experimental results show that the improved model has better prediction results for delayed flights, with a 5.54% increase in average macro precision and 1.86% increase in accuracy. Finally, the flight delay prediction results are published in the mobile APP application system to provide corresponding advice to passengers and airport departments.

Aiming at the problems of small target scale, vulnerable to background interference and insufficient feature utilization in remote sensing image target detection task, a single stage target detection algorithm based on feature enhancement and feature fusion is proposed. Based on CenterNet, a feature enhancement module is designed, which enriches and strengthens the features of small targets and solves the problem of low accuracy caused by small targets and background interference. Then BiFPN mini multi-scale feature fusion structure is used to strengthen the feature expression ability of the target and solve the problem of insufficient feature utilization. This algorithm is implemented in the average detection accuracy on UCAS_AOD dataset reaches 84.9%. The experimental results show that the improved measures in this paper effectively improve the target detection accuracy for remote sensing images.

In this article, we design this algorithm to perform the segmentation task of remote sensing images by using attention module and feature fusion to deal with the challenges of features with different dimensions, small-size objects and uncertain boundary segmentation in remote sensing images. The encoder uses a PVTv2 model to increase the feature information acquisition capability of the model, and several convolutional paths of the feature aggregation module can acquire different size information of the image. The deep network uses attention mechanisms to obtain features that are valid for the segmentation task. In the decoding phase, the feature fusion module fuses the shallow network detail information with the deep network semantic information. The experimental comparison on the ISPRS Potsdam dataset shows that pixel accuracy of the model in this paper achieves 88.59%, IoU, mIoU, and FWIoU are higher than some classical semantic segmentation networks, which effectively increases the accuracy of remote sensing image segmentation.

Overfitting is a major problem faced by deep learning, and data enhancement is an effective means to mitigate it. Erasure-based data enhancement improves model generalization by generating new training samples, but these methods lead to over-erasure and under-erasure problems due to the random nature of erasure. To alleviate these problems, this paper proposes a new attention activation map-based image erasure method: Aag-dem, selectively erase the relevant regions of the image to generate high-quality training data and finally a large number of experiments are done to prove the effectiveness and generality of the method, and the performance of the baseline is improved to some extent.

With the rapid development of society, the world has entered the digital world, which is an important part of the metaverse scene. The information processed in the digital world is discrete "0" and "1". The cornerstone of the digital world is the chip, and behind each chip is the bounce of "0" and "1". The function of the human brain gradually changed from carbon-based to silicon based. All the objects, relationships and rules of nature are finally embodied in the digital world in the arrangement and combination of binary codes "0" and "1". Flip-flop is a unit circuit that forms a sequence logic circuit. Taking D-flip-flop as an example, this paper analyzes the function characteristics of D-flip-flop and obtains the state transition truth table by simulating the changes of circuit "0" and "1" by integrated chip 74LS74 in Proteus. From the table to the Carnot diagram, the characteristic equation of D-flip-flop is obtained by simplifying the Carnot diagram. The key points of analysis of sequential logic circuit formed by D flip-flop are further discussed and verified by proteus simulation circuit. The circuit is further optimized so that the visual effect of the circuit can be optimized. Finally, through the data migration of the underlying structure of the meta-universe, the changing rules of the bit stream in the data world are obtained.

With the rapid development of computer vision, fine-grained image recognition technology has shown strong value in catering, online shopping, education, urban transportation and other fields. The fine-grained image recognition technology can be divided into strong-supervised fine-grained image recognition and weak-supervised fine-grained image recognition. Strongly supervised fine-grained image recognition requires proper manual marking of the area to be identified in the image before training. Its preparation is time-consuming and labor-intensive, and manual marking is highly professional.

Generative Adversarial Network (GAN) has achieved remarkable results in the field of image generation, but it requires a large amount of image data to participate in the training, which is not expected in real-world work. Generating more images for this category in a new task with only a small amount of data has always been a challenging task, and a few-shot image generation is proposed to accomplish such tasks. This paper proposes a novel Generative Adversarial Network for few-shot image generation. The algorithm consists of two parts: a flexible Meta-model based on inner and outer loops, and a GAN combining CNN, Multi-Head Self-Attention (MHSA), and residuals for multi-tasking. Experimental results show that the algorithm achieves satisfactory results on several datasets, exploring the potential of the optimization-based approach for few-shot image generation.

In this paper, a multi-day rolling transaction decision-making method for the wind power plant under the Shanxi power market system is formulated by researching the transaction rules of Shanxi power market, and the average transaction price of the rolling matchmaking transactions in the next two to five days has been successfully predicted through the time-delay gray model. The time-delay gray model is a model established on the basis of the traditional gray prediction model by introducing the system characteristics and time-delay items of related factors. It has been verified that the model is effective in predicting the average transaction price of daily rolling matchmaking transaction, and the predication data can play an important role in the multi-day rolling transaction decision-making of the wind power plant under the Shanxi power market system and has high application value.

Image segmentation as a crucial step in image processing and analysis, has important applications in video surveillance, medical detection and wafer detection, etc. Accurate and efficient image segmentation can bring great advantages and convenience to the realization of related tasks in these fields. In this paper, a 2.5D UNet network based on ConvNeXt is proposed to realize the image segmentation task based on the gastroscopy image dataset. The experimental results show that the proposed method has better segmentation performance than the UNet model based on ResNet50, UNet model based on EfficientNetB0, and UNET2.5D model based on EfficientNetB1.

In the task of Simultaneous Localization and Mapping (SLAM), a mobile robot should be able to recognize the location that it has previously visited. This is termed as the loop closure detection. Current loop closure detection is implemented by matching of image features extracted from present image and previous images, where good feature representation is highly required. Bag-of-visual Words (BoW) is known to be an effect feature representation for image and video matching task. However, to build the visual feature vocabulary in the unsupervised manner is still challenging. In this paper, we are aiming at evaluating the representability of the visual words and generating the discriminative vocabulary representation. We first present an entropy ranking based scheme to improve the spectral clustering algorithm by leveraging eigenvector selection and feature embedding simultaneously. We then propose to evaluate the representative ability of the visual vocabulary quantitatively based on the complexity of the decomposition of similarity matrix in the unsupervised manner. We perform experiments on the loop closure detection datasets. which contains strong perceptual aliasing in both indoor and outdoor scenarios. Results demonstrate that the proposed BoW based feature representation has strong representation ability in measuring the similarity of loop closure events and differentiating new visited places.

Fuzzy clustering is widely used in image segmentation by scholars. The existing fuzzy clustering algorithm still lacks sufficient segmentation accuracy in the face of noise-polluted images. Therefore, combining the spatial information of membership and multi-scale filtering, an improved fuzzy clustering algorithm is proposed to solve the problem that noise may occur in the homogeneous region. Firstly, the spatial and brightness information are introduced into the objective function to reflect the context information and realize the adaptive adjustment. The Gaussian kernel metric is used to replace Euclidean metric, and the low dimensional linear separable data is mapped to the high dimensional linear separable or approximately separable data. Then the membership is modified again according to the proposed spatial function to optimize the image segmentation features. Finally, the multi-scale space of the image is constructed by a Gaussian filter, which is progressively segmented layer by layer, effectively suppressing the noise interference and improving the robustness of the algorithm. The performance of the algorithm has been improved in the synthetic image and natural image, especially in the case of serious noise pollution and complex background.

In order to solve the problem of low attitude estimation accuracy of low-cost UAV without GNSS (Global Navigation Satellite System) signal, the integrated navigation algorithm was improved. Binocular vision sensor is introduced as the observation sensor, and the position and attitude angle of the image information collected by binocular camera were estimated by the SLAM theory, the position and attitude angle were entered into the EKF algorithm, and the Vision Inertial EKF (VI-EKF) algorithm is established. It realizes the high-precision navigation of low-cost UAV without GNSS signal. The biases of gyroscope and accelerometer are calculated by the relative motion increment between the two key frames of binocular camera and IMU pre-integration data, and these two biases were taken as observations to update IMU bias in the process of EKF algorithm. In this paper, the effectiveness of IMU pre-integration algorithm is verified firstly, and then the VI-EKF navigation algorithm is compared with the pure visual SLAM algorithm through dataset simulation. The simulation results show that the VI-EKF algorithm proposed in this paper has a good performance in the environment without GNSS signal.

Animal biometrics is a frontier field of computer vision, pattern recognition and cognitive science that plays a vital role in the registration, unique identification and verification of livestock (cattle). In this study, we propose a deep learning approach to cattle identification based on the characteristics of the muzzle point image (nose pattern) to solve the problem of missed or replaced animals and false insurance claims. Inspired on the state-of-the-art OSNet architecture, which was developed for person re-identification, we introduce significant modifications in order to improve accuracy in the cattle recognition problem. First, each convolution layer is replaced by a depth-wise separable convolution layer, and parametric rectified linear unit is used as a non-linear activation function. Next, we add two convolutional block attention module. Under the same experimental conditions, improved OSNet achieves significantly superior accuracy than the original OSNet, maintaining the same speed and compact storage.

This paper designs and implements a low frequency signal generator based on the interrupt and timing functions of the single chip computer. The system uses STC89C52 single chip microcomputer as the controller, outputs different types of signals through AD signal conversion, and outputs after the LM358 amplifier amplifies and potentiometer adjusts the amplitude. According to different needs, the system outputs different waveforms, frequencies, amplitudes, and phases through corresponding keys, and uses LCD to assist in displaying the generated signal information. The lower computer software of the system is implemented on Keil five software using C language, the simulation circuit is implemented on Proteus seven software, and the drawing of circuit schematic diagram and PCB circuit diagram is implemented on Altium Designer 18 software.

Traction network loss is an important index to evaluate the economy and safety of electrified railway operation and management and is of great significance to evaluate the transportation capacity of the line and the power supply capacity of the traction power supply system. In this paper, the equivalent circuit model for traction network parameter identification is first studied, and the formula of traction network loss measurement is derived. Next, a multidimensional impedance database model is established for storing the equivalent impedance. Based on the derived formula and the established model, the feasibility and effectiveness of applying the measurement data from the currently deployed measurement system to the method in this paper are analyzed. The results show that the loss estimation error meets the actual engineering needs.

Temperature prediction has always been a difficult point in meteorological prediction. It not only affects the global climate and ecosystem, but also has a profound impact on people 's social practice. In recent years, the temperature prediction mostly adopts the method of constructing neural network, because the method of decision tree, statistics and vector regression ignores the time series dependence, but the neural network-based models have limited its accuracy. Many scholars like to use CNN or LSTM model alone, and both of them are still lacking. Therefore, this paper proposes to combine CNN and LSTM models, which can not only give full play to the advantages of CNN model in processing data quickly, but also use the memory function of LSTM model to process temperature data with strong time series. The experimental results show that the CNN-LSTM model is more accurate than the CNN and LSTM models alone in predicting temperature.

Cloud storage has become a widespread trend with its efficient and convenient features. However, this also brings new security risks, such as identity forgery, data theft, privacy disclosure and other security problems. Based on blockchain, elliptic curve cryptography and other technologies, this paper proposes a blockchain-based cloud environment data storage scheme, which can provide users with decentralized identity authentication and data integrity verification functions, and the transaction information of data storage is stored on the chain to realize the safe storage of data.

Old photographs often suffer from a variety of imperfections, such as scratches, stains, and fading. Various techniques have been proposed for image restoration, including image inpainting, denoising, deblurring, and super-resolution. In this paper, we propose a novel approach to restoring such images by leveraging deep learning techniques. Specifically, to accurately identify scratches in images, we train a classification network on a large dataset of paired old and repaired photos. Additionally, we employ a Fourier convolution-based neural network to repair the damaged areas of the images. Our results show that our approach outperforms existing methods in terms of both objective metrics and visual quality. We believe that our work has the potential to preserve valuable memories and historical artifacts for future generations.

Out-of-Distribution (OOD) detection is critical for preventing deep learning models from making incorrect predictions, especially in safety-critical applications such as medical diagnosis and autonomous driving. However, neural networks often suffer from overconfidence, making high confidence predictions for OOD data that are never seen during training and may be irrelevant to training data. Determining the reliability of the prediction is still a difficult and challenging task. To address this challenge, we propose a new method called Uncertainty-Estimation with Normalized Logits (UE-NL) for robust learning and OOD detection. The method has three main benefits: (1) Neural networks with UE-NL treat every In Distribution (ID) sample equally by predicting the uncertainty score of input data and the uncertainty is added into SoftMax function to adjust the learning strength of easy and hard samples during training phase, making the model learn robustly and accurately. (2) UE-NL enforces a constant vector norm on the logits to decouple the effect of the increasing output’s norm from optimization process, which causes the overconfidence issue to some extent. (3) UE-NL provides a new metric, the magnitude of uncertainty score, to detect OOD data. Experiments demonstrate that UE-NL outperforms existing methods on common OOD benchmarks and is more robust to noisy ID data that may be misjudged as OOD data by other methods.

The current conventional oscillatory search algorithm for relay protection logic diagram is prone to fall into local optimum due to the late population diversity reduction, which leads to poor search results. In this regard, a sequential traversal-based oscillation search method for relay protection logic diagrams is proposed. The oscillation search algorithm is optimized by finding the second-order oscillation factor, optimizing the topology of the algorithm, and constructing the relay protection logic diagram model to obtain a library of graph elements. The experimental results show that the proposed oscillation search algorithm converges faster and has a more desirable search effect.

Nowadays, deep learning methods such as neural network models are highly effective for various tasks including image classification and natural language processing, but the increasingly high computational cost restricts the deployment of these network models in many kinds of scenarios whose resources are usually limited. Among all kinds of methods to solve these difficulties, quantization is a plausible way to reduce the storage size of these network models and accelerate their inference process by replacing the parameters such as weights with low-bit fixed numbers during the training process. This problem can be viewed as a discrete constrained optimization problem. In this work, we use Alternative Direction Methods of Multipliers (ADMM) to decouple the continuous parameters from the discrete constraints so that the original hard optimization problem is separated into several subproblems. In addition, structure-aligned quantization is also achieved, which is usually more friendly for edge computing devices to execute and accelerate. With extensive experiments on ImageNet and CIFAR10 dataset, models represented by low-bit fixed-point numbers with acceptable accuracy loss compared with original full precision models can be acquired. Compared with some previous quantization works, the quantization models obtained in this work have little classification accuracy drop compared with the original pre-trained full precision model, and a kind of hardware-friendly structure that makes the neural network easier to deploy is achieved.

The level of development of artificial intelligence in manufacturing industry directly affects the development of the country, is the centralized embodiment of the comprehensive strength of the country, directly provides technical, material and resource support for the development of the national economy and is the most basic industry that drives economic development. In this context, this paper builds an evaluation index system based on hierarchical analysis method and cloud model theory, uses fuzzy hierarchical analysis method to determine the weight values of indicators at all levels, then outputs the standard cloud diagram and evaluation cloud diagram by building a cloud model to visually reflect the fit, and finally introduces examples to verify the scientific and rationality of the model method, and through data Finally, we introduce examples to verify the scientific and reasonableness of the model method, and reflect the superiority of the model in prediction and evaluation through the comparison of data.

Based on the problem that traditional modeling is difficult to meet the predictive control of high parameters and large capacity units and to achieve the expected control effect, this paper uses BP neural network to establish a dynamic model of 600 MW supercritical once through boiler unit and concludes that neural network can effectively improve the coordination, prediction and optimization of supercritical units, and also can improve the economy and safety of supercritical units.

In this paper, an adaptive neural network control method based on state observers is proposed for the rolling motion control problem of zero-speed ships. By combining the neural network with the disturbance observer technique, the internal and external uncertainties of the rolling motion control system of zero-speed ships are overcome. Considering that the ship's rolling angular velocity is difficult to be measured accurately under the actual sea state, the state observer is introduced to estimate the unknown rolling angular velocity. Finally, the simulation results verify the effectiveness of the control scheme.

The construction industry has always been fiercely competitive, so it also has high requirements for project quality. Most of the construction period is tight, moreover, the contract cost also shows a downward trend, leading to problems in the quality, duration and cost target management of construction enterprises. This paper focuses on how to improve the management technology of construction projects, ensure that the quality, duration and cost of construction projects can be realized at the same time, and improve the competitiveness and economic benefits of construction enterprises. Based on deep learning, the assumptions of project management optimization are given, and a mathematical model for the comprehensive control of quality, cost and duration is established. The model is solved using MATLAB. Finally, the comprehensive control measures of the construction scheme are given to provide reference for relevant researchers.

In order to solve the traffic congestion problem caused by mixed traffic flow between intelligent connected vehicles and human-driven vehicles at intersections, a deep reinforcement learning-based intelligent connected vehicle car-following collaboration method is proposed to alleviate traffic pressure and improve intersection efficiency. The intelligent connected vehicle can obtain information on the status of surrounding vehicles, and the reinforcement learning can use the obtained status information to guide the vehicle in the direction that is conducive to improving traffic flow efficiency by rewarding the vehicle, and in this way alleviate the traffic congestion caused by excessive traffic pressure. The experiments were conducted using SUMO simulation and the flow project framework. The Proximal Policy Optimization Algorithms (PPO) based on deep reinforcement learning was used to improve the PPO algorithm by proposing a new reward function, which was designed in terms of traffic safety, comfort, and driving efficiency. The reward function is designed in terms of traffic safety, comfort and driving efficiency, and penalizes the vehicles when the time distance is too small, and the deceleration is too large. The experiments were carried out using SUMO simulation software to simulate a vehicle with an IDM car-following model and a vehicle with reinforcement learning to simulate an intelligent connected vehicle. The experiments demonstrate that controlling the intelligent connected vehicle by reinforcement learning method to follow the driving can effectively improve the traffic flow rate and enhance the traffic stability. It has also been demonstrated that the reinforcement learning control of an intelligent connected vehicle can effectively reduce the frequent and significant deceleration of the vehicle.

The low permeability sandstone reservoir has the characteristics of low petrophysical property and strong interbedded heterogeneity. The classification criteria of reservoir are many and different according to the region. Lithology identification, reservoir classification based on pore structure and physical property characteristics and logging phase classification are commonly used in low permeability reservoir classification. The logging phase can comprehensively characterize the logging response characteristics of the reservoir, and this classification method has great advantages. Well logging data have the advantage of well logging and complete wellbore petrophysical characteristics data and have strong continuity and comparability. According to the well logging data and petrophysical properties of formation, automatic clustering algorithm can be used to divide the logging phase. In this paper, the low-permeability sandstone reservoir in Y member of M oilfield is studied, and it is found that the four relationships of the reservoir in this member are very different, and the diagenetic characteristics and petrophysical characteristics are not clear. Conventional logging data are used to divide logging facies based on automatic clustering algorithm, and then supervised logging facies combination is carried out based on the principle of combining similar logging response characteristics. Finally, logging phases of geological significance can be divided. The results show that the method is well consistent with the actual results in logging phase division and sweet spot identification by training prediction in the whole area, which is of great significance for oilfield exploration and development.

The rectangular arrangement problem is widely used, which refers to the optimization problem of putting a finite number of small rectangles of different sizes into a large rectangle without overlapping to maximize its utilization. It belongs to the optimization problem of NP complete class. However, when the number of samples increases the difficulty and time of scheduling increases significantly. In order to solve the problem of large and medium-scale rectangular packing, deep Q-learning based on priority experience replay is used to improve the utilization rate of the board and reduce the calculation time of packing, so that the agent can effectively improve board utilization and increase efficiency in large- and medium scale two-dimensional rectangular packing.

Wind power is renewable energy and plays a critical role in the carbon emission peak and carbon neutrality plan. As wind speed and direction is unstable, the generated electricity power fluctuates, and therefore cannot directly accepted by the power grid. A common solution is to forecast the wind power generation and adjust the power generation and dispatching strategy to improve the stability and acceptance when provided to the power grid based on the forecasting results. The accurate forecasting of wind power is the key to this scheme. Common forecasting methods include physical approach and statistical approach. For physical approach, the Numerical Weather Prediction (NWP) model forecasts the wind speed and direction, so that the power is directly calculated. For statistical approach, the wind power is forecasted according to historical time series and predicted weather. The forecasting model including statistical autoregressive models, support vector machines, and neural networks, etc. However, the single model prediction lacks robustness, so we propose to ensemble multiple models to achieve robust short-term wind power forecasting. Based on supercomputing, we propose a large-scale parallel forecasting of various methods and models and a statistical method to assign reasonable weights to various models, achieving the large-scale ensemble of various models. In order to adapt to real-time data distribution changes, we update the models and corresponding weights online. The experiments are conducted on a dataset of wind power generation and meteorology from a province in China. The experimental results show that compared with the single model, the proposed large-scale ensemble model achieves better forecast accuracy and robustness.

In order to solve the problem of low detection accuracy caused by dense placement of PCB industrial equipment and occlusion overlap in industrial AR application system, an industrial equipment object detection algorithm based on improved YOLOv7 was proposed. Firstly, Coordinate attention is added into the network to strengthen the network's attention to the visible regional features of industrial equipment. Then SIoU loss function was used to improve the model loss function, which increased the convergence speed and effectively improved the regression accuracy of object position. Finally, Adaptive-NMS is introduced to adjust the threshold adaptively according to the density of objects so as to retain more correct prediction boxes. Based on the PCB industrial equipment dataset constructed by the author, the field test results verified that the proposed algorithm would be better than other 6 advanced object detection algorithms such as Fast R-CNN, and its object detection accuracy could reach 94.21%, showing significant computational advantages for the processing of occlusion equipment. The research result shows that the industrial equipment object detection algorithm based on the improved YOLOv7 is reasonable, feasible and effective, and the proposed method is more usable.

Train mixed operation refers to the trains with different speeds running on the same line, such as CTCS-3 mixed-running line, which operates both CTCS-3 and CTCS-2 trains. Existing studies have shown that train speed differences increase the tracking interval time (headway) between trains, reducing the line carrying capacity. At present, much research has been conducted on the impact of train speed differences and the ratio of trains with different speeds on the carrying capacity of the mixed-running line from the standpoints of transportation organization and traffic scheduling. As an important component of the rail transit system, the train control system has a significant impact on the rail transit system's safety and efficiency. By constructing a cellular automata model for mixed speed trains under the fixed block system, this study analyzed the effect of different block section lengths on the headway of mixed speed trains in the same line. On this basis, the dynamic block section method was proposed in this work to shorten the headway between mixed speed trains. The study quantifies the effect of block section length on the headway of mixed speed trains, which can serve as a theoretical foundation for expanding the capacity of existing high-speed railway lines.

In recent years, deep learning dense matching techniques have developed rapidly. Among the various training methods such as supervised, unsupervised and semi-supervised, the matching accuracy of supervised training methods is still much higher than other training methods. However, this training method requires the use of manually labelled parallax maps as samples. The accuracy of the labelling will directly affect the matching accuracy of the trained network model. Therefore, it needs to be analyzed and studied. In this paper, the noise immunity of the deep learning supervised method is studied by simulating systematic error, random error and gross error, and the experimental results show that: (i) The deep learning method has anti-noise ability within a certain range and has a better anti-noise effect on random noise, but the matching accuracy decreases at an accelerated rate as the noise increases. (ii) The pre-training method by transfer learning can effectively improve the matching accuracy, increase the noise immunity, and make the original non-converging network converge.

As a flexible demand-side response resource, the charge and discharge of electric vehicles can be reasonably and orderly controlled through the demand response strategy, so as to improve the two-way interaction between electric vehicles and the smart grid and ensure the safe and stable operation of the power grid. This paper analyzes the participation of electric vehicles in Yunnan electric demand side response application scenarios, the study of electric vehicles in public charging stations, community charging stations and under three different charging places, participate in the demand side response, including the choice of different participation strategy, cost benefit analysis, etc., and gets different application scenarios for the most appropriate electric car participation strategy.

In this paper, trajectory tracking of uncertain manipulator under arbitrary initial conditions is investigated. Firstly, the error dynamics is established according to the preset expected trajectory, the probability distribution function of random variables and the virtual error variable. Secondly, the virtual control signal and iterative learning are used to compensate the interval without operation and the iteration length with random changes. Then, the uncertainty and external disturbance of the manipulator arms are approximated by an adaptive neural network, and the practicability of the tracking algorithm is inspected by a composite energy function. Finally, the simulation example is certified to display the validity of the proposed algorithm.

The current traditional method of detecting abnormal sound of hydropower plant equipment identifies the equipment state by a single class of support vector machine to achieve the judgment of abnormal sound, which leads to poor detection effect due to the lack of extraction of audio features of abnormal sound. In this regard, the abnormal sound detection method for hydro power plant equipment based on self-encoder and weakly supervised learning is proposed. The Gaussian mixture model is used to learn the characteristics of equipment sound and purify the abnormal sound of hydropower plant equipment. The audio features are extracted from the abnormal sound of equipment, and a two-channel weakly supervised encoder detection model is constructed to realize the detection of abnormal sound of hydropower plant equipment. In the experiment, the detection effect of the proposed method is verified. The analysis of the experimental results shows that the anomalous sound detection model constructed by the proposed method has a high AUC value and a high detection performance.

Fatigue driving detection is of great significance to prevent fatigue driving and reduce the occurrence of traffic accidents. The existing fatigue driving detection method based on driver facial information has low robustness, and its speed and accuracy can still be improved. A method of fatigue driving detection based on end-to-end facial state recognition model is proposed. Based on improved YOLOv7 network, the end-to-end facial state recognition model can detect the driver’s eyes and mouth and recognize the state of these two parts directly from driver’s image. Then, according to the state of eyes and mouth, the driver's fatigue state is judged based on two criterions: the PERCLOS and the duration of yawning. Experimental results show that the proposed method can guarantee high recognition rate in complex environment and has higher real-time performance.

Cross-modality person re-identification aims to match specific pedestrian images from different modal image galleries. In order to solve the problem of excessive image modal differences, a cross-modal person re-identification method based on generative adversarial network and feature-level modal compensation is proposed. Firstly, generating corresponding cross-modal images by cycle-consistent generative adversarial networks based on knowledge distillation to solve the problem of fewer image sample pairs; secondly, designing a feature extraction network with self-attention mechanism to extract more discriminative pedestrian features; on this basis, in order to make up for the defects of low-quality generated images and more interference information, a feature-level modality compensation module is designed, firstly the single-modality features are decomposed into modality-specific features and modality-shared features, then the corresponding missing modality-specific features are generated through the modality-shared features, and finally the existing features and the generated features are fused, which enables the model to Generating modality-specific features that are relevant to people and ignoring those that are irrelevant can improve the performance of feature extraction. Through comparative experiments on the RegDB and SYSU-MM01 datasets, the results of the method on the cumulative matching characteristic and the average precision mean were analyzed. The accuracy rate has been improved to a certain extent.

The insulator is a vital component of power equipment that aids in power transmission. Aiming at transmission line faults caused by missing insulator pieces, this paper proposed to divide insulator detection and defect location into two steps. Additionally, we add Squeeze-and-Excitation (SE) module into the Faster R-CNN model at the insulator detection stage, which could enhance insulator detection accuracy under similar model complexity and also improve the detection efficiency of the YOLOv3 model on insulator defect location.

Meat freshness is an important index to measure the quality of meat, but currently, the commonly used methods of rapid inspection require specific equipment as well as the environment, and it is difficult for ordinary consumers to distinguish meat freshness quickly and accurately by using these methods. To address this problem this paper proposes a ResNet34 network based on ResNet34 network and adding channel attention mechanism to build a SE- ReseNet34 network model for meat freshness recognition is constructed, and the training method of transfer learning is used to accelerate the convergence of the model. Compared with ResNet34, MobileNetV2, DenseNet121, and ResNext50 network models, the accuracy of the model test set is improved by 1.4%, 2.5%, 0.7%, and 1.6%, respectively. The test set accuracy can reach 99.6% while having high accuracy and recall, and the results show that this is an efficient and fast method for identifying meat freshness with some practical significance.

Sequential recommendation systems exploit the user's historical item sequences to predict their next actions. Recently, dynamic graph-based methods have been studied and achieved excellent performance for recommendation. They capture dynamic collaborative signals between different user sequences by stacking multiple network layer with attention mechanism to solve insufficient interest mining problem caused by use a single user’s sequence. In online platforms, recorded user behavior data may contain noise, and stacking multiple attention network is easy to aggravate the effects of noise. In this paper, we propose Filter-enhance Temporal Graph Neural Network for Continuous-Time Sequential Recommendation (FTGRec), which connects the related interactions of different user by dynamic graph and design a module combining Fourier transform and attention mechanism to filter the noise data, to predict the order pattern of users better. Empirical results on three datasets indicate FTGRec outperforms other comparative methods.

Smoking is a common behavior in daily life, but smoking in public places not only affects public health and the health of others, but also may cause accidents such as fires. Therefore, detecting and recognizing smoking behavior is of great importance. Aiming at the problem that most current target detection networks are large in volume, have many parameters, and are difficult to deploy on low-performance device terminals, the article proposes a lightweight improved YOLOv5 smoking behavior detection algorithm model. This model replaces the backbone with a lightweight MobileNetv3 neural network, and at the same time absorbs and references the design ideas of ShuffleNet to improve the width of the neck-head part, making it smaller and more suitable for deployment on low-configuration devices. The experimental data show that after improvement, the new network has 80% less parameters and 63% less inference time, and when deployed on low-configuration device (i3-4000-m 2.4-ghz 2c4t), its inference time (50-ms) can be reduced by half compared to the original network (120-ms) and meet the requirement of near real-time. Therefore, the improved network can reduce network parameters while ensuring accuracy and can achieve near real-time performance on low-configuration devices.

The geometry of airfoils, crucial in predicting aerodynamic coefficients, can be perceived by three existing methods: manual definition of geometry parameters, polynomial and deep learning. The first two methods are capable of directly obtaining geometry-features or polynomial coefficients from airfoil coordinates, but they are insufficient to extract latent features. Current deep learning techniques are capable of extracting latent features from only Euclidean space. However, it has been proven that curves of airfoils are exists in Riemannian space. Inspired by deep learning and polynomial (especially the Bézier polynomial) and deep learning, we propose a Geometry-based Feature Extraction method (GFE) to extract airfoil geometry-features from both Euclidean space and Riemannian space. By utilizing Bézier curve, GFE can extract geometry-features (i.e., manifold metrics) of airfoils in Riemannian space. Through an internal auto-encoder, GFE encodes airfoil coordinates and manifold metrics to generate a novel feature representation. The UIUC airfoil dataset is employed to test the feasibility of GFE. Experimental results show that smooth and realistic airfoils can be rebuilt based on the geometry-features extracted by GFE, and the average MSE of rebuilt airfoils is 38.66% lower than those rebuilt by existing auto-encoders.

Compared with hand-craft features, CNN (Convolutional Neural Network) has stronger representation capabilities for images, but training CNNs with strong generalization capability is expensive. However, the froth image is quite different from the natural scene image, and the representation ability of the transfer learning model for the froth images needs to be improved. In the industrial process of flotation, the state of reagent-addition is strongly correlated, and it is difficult to accurately judge the state of reagent-addition only by relying on a single image. In this paper, the self-supervised contrastive learning pre-training of the CNN model is performed using the unlabeled froth images to improve the model's ability to represent froth images. Then we design a 1-d convolutional network integrated with the Interframe-Attention to improve the accuracy of recognizing the status of reagent-addition.

Aiming at the problem of partial occlusion and stacking in the field of robotic arm recognition and grasping, especially in non-single scenes where multiple objects are cluttered. This paper proposes a recognition and grasping method based on the improved DenseFusion algorithm network architecture to improve pose estimation’s accuracy rate and combined with the RGB information and depth information of the depth camera to estimate the 6D pose of the target, and the experimental verification was carried out on the robotic arm platform independently developed by the laboratory. The experimental results show that the proposed method can not only identify the target object accurately, but also obtain the 6D pose of the target object accurately, so as to improve the success rate of grasping by the robotic arm.

With the rapid development of the Internet, there has been a high incidence of telecom fraud cases. Among the existing methods for identifying and classifying fraudulent websites, most of them use machine learning to classify and identify URLs and static files (HTML text and image resources) of web pages. These methods cannot take into account the text features contained in the pictures in the static files of the web page, and the content to be verified is too complex to recognize and classify the website quickly and accurately. In view of this, this system proposes a method combining SVM algorithm and improved website text features to quickly and efficiently identify and classify websites. Experimental results show that the system has excellent ability to identify and classify websites.

In recent years, the requirements for strip steel surface quality have also increased. However, traditional Non-destructive testing detection is not good enough. Deep learning-based algorithms are also unable to meet the requirements of small sample data and real-time performance for strip surface defect detection. The DAG-SVM has the advantage of solving small data, non-linear and high-dimensional pattern recognition. Therefore, this paper selects the Directed Acyclic Graph (DAG) to construct the SVM multi-class classifier. And representative shape features are selected as the feature vectors of the samples. At the same time, the kernel function and parameter search are used to further improve the recognition rate. The final results show that the average recognition rate using this method is 95.5%, which can meet the practical needs. In the comparison with BP neural network, the method of this paper is also better than BP neural network in general.

Since there are many possible influencing factors of visibility, lightweight data requirements in practical applications of machine learning in visibility prediction can reduce the corresponding data observation cost and collection difficulty. By using the long-term measured data in Qingdao, this research comprehensively compares the performance of five common machine learning methods under different training parameter schemes, including XGBoost, LightGBM, Random Forest (RF), Support Vector Machine (SVM) and Multiple Linear Regression (MLR). The lightweight visibility prediction schemes based on pollutant parameter optimization are established. The seasonal training data of five machine learning models is preprocessed, and then performance evaluations of predictions are carried out. The analysis results show that in terms of models, ensemble learning models, including XGBoost, LightGBM, and RF, have significantly better seasonal visibility prediction effects than SVM and MLR models; XGBoost and LightGBM also have slightly better prediction effects than RF models. In terms of pollutant parameters, solid pollutants have a greater impact on visibility prediction than gaseous pollutants; PM2.5 is more influential than PM10 in visibility prediction. The visibility prediction scheme with six parameters using meteorological parameters and PM2.5 based on XGBoost or LightGBM model is preferably established in this research. This scheme can achieve the same prediction performance as the 11 parameter prediction scheme. The Correlation Coefficient (CC) of the results is around 0.85. The results of this study can not only be used to provide a machine learning scheme reference for practical visibility prediction applications, but also help to deepen the understanding of the factors affecting visibility.

As a new and efficient intelligent energy system, integrated energy system has been widely used. As more and more new buildings are incorporated into the system, accurate load forecasting is essential for the planning and operation of integrated energy systems. The historical data of new buildings incorporated into the energy management system is not enough to build accurate prediction models. Transfer learning, as a cross-domain learning method, has been applied in time series prediction. To solve the problem of negative transfer caused by fluctuation and randomness of load data, this paper presents a day-ahead power load forecasting model that combines transfer learning with ensemble learning. Firstly, a multivariate migration method based on data decomposition is proposed, which migrates the data of multiple buildings with high load similarity to enrich the historical data of the tar-get buildings and avoid negative transfer. Secondly, similar day and neural network integrated prediction models are presented to deal with the impact of different date types on prediction accuracy. Finally, the proposed model is validated by simulation experiments. The experimental results show that the proposed method achieves good prediction accuracy.

Accurate estimation of the State-of-Health (SOH) can provide reliable guidance information for Battery Management System (BMS) to ensure safe and efficient operation of Li-ion batteries. In this paper, we proposed a novel lithium-ion battery SOH estimation method. Firstly, we used the incremental capacity analysis method to analyze battery aging characteristics and extract the health interest, and an important smoothing step was also introduced in this process. Then a battery SOH estimation model was built based on the extreme learning machine theory. Finally, the performance of the ELM model was validated using four Li-ion batteries cyclic aging data sets. The results show that the SOH estimation error was not exceed 0.1, which indicates that our proposed method has a high estimation accuracy.

Lightning ignition source protection design of aircraft fuel system is an important factor affecting aircraft flight safety. There are many factors related to ignition source, and there exist optimization possibilities in the design process. Therefore, it is difficult to confirm design only by means of tests - on one hand, it is difficult to evaluate the design systematically, on the other hand, the cost is high, and the time is long. Simulation analysis is more and more widely used in aircraft lightning protection design. This paper introduces a simulation application based on EMA3D software in aircraft fuel system lighting protection design evaluation, through the analysis of the key factors affecting the ignition source to evaluate the rationality of lightning protection design.

Malicious language identification has been a critical issue in current network environment, automatically investigating malicious sentences has been concerned by security protection of neural language processing However, existing research about malicious investigation is almost based on the keywords detection and the system cannot detect the cryptic information in networks, which needs human to sign the malicious keyword and cost lots of computation costs. In this paper, we describe the usage of a neural network for language identification of malicious information. The neural network model was trained on a dataset of malicious and non-malicious information samples and then simulated on a set of unseen malicious text samples. From our extensive experimental results, we can conclude that the proposed model was able to precisely identify malicious language and achieve a high score approximately about 0.874. The results indicate that neural network models can be a useful tool for detecting malicious language.

Graph Neural Networks (GNNs) extend traditional deep neural networks to graph-structured data and have greatly succeeded in graph representation learning scenarios. However, the current mainstream GNNs based on message passing mechanisms are challenging to scale to large graph data due to their severe scalability and over smoothing issues. In this paper, we propose a novel scheme called the Message Gain Aggregation Architecture (MGAA) to alleviate the over smoothing problem while maintaining a scalable architecture and good expressiveness. Our MGAA network considers the nodes themselves separately from their neighbors and treats the neighbors as gains to be imposed on the nodes. A theoretical complexity analysis demonstrates that our proposed MGAA network maintains linear temporal and spatial complexity. Furthermore, the experimental results show that our model achieves sufficiently competitive performance in a comparison with the current popular state-of-the-art GNNs.

Both sides in contemporary warfare must thoroughly consider their strategies and implement effective tactics to increase their offensive or defensive capabilities. It is critical to emphasize the strengths and weaknesses of one’s own side to establish a balanced and stable situation on the battlefield. This paper presented a suggested method that employed Markov networks to define the weight of nodes in a map, resulting in an optimal predetermined layout. Furthermore, we proposed adopting BP neural network algorithms to give the army autonomy in decision-making. That enabled each unit to develop its own decision-making process and enhance the overall war strategy via repeated simulations.

In order to solve the problem of insufficient feature expression ability in gesture recognition in emergency rescue, this paper proposes a multi-feature fusion dynamic gesture recognition network (MGN-LSTM). The model first extracts the global information and local information of each granularity through a Multiple Granularity Network (MGN). Then, a layer of Long Short-Term Memory network (LSTM) is used to extract the time information of gestures. Finally, in the model prediction stage, the candy algorithm is added to detect the edge feature information of the gesture, and the recognition result is obtained by combining the global, local and time information. The experimental results show that the multi-feature fusion method proposed in this paper has good feature expression ability and has been verified in both the self-built emergency rescue gesture dataset and the public dataset IsoGD, which improves the accuracy of gesture recognition.

In order to solve the problem that the recognition accuracy of a single mode model depends on the emotion type, this paper proposes a multimodal emotion recognition model based on the Di-directional Gated Recurrent Unit (BiGRU) network and attention mechanism. In this paper, BiGRU neural networks are used to extract high level features from the original features of text, audio and video, and then high-level features are carried out. When high level features are fused, the attention mechanism is used to adjust the weight of emotional features, and finally emotional recognition is carried out to obtain emotional classification. The model is trained and tested on the CH-SIMS dataset, with an accuracy of 75.15%. Experimental results show that the proposed model has lower computational complexity and higher recognition accuracy.

The accuracy of traditional convolutional neural networks in expression recognition is relatively low, this paper makes improvements on the basis of ResNet network. On the one hand, the two 3*3 convolutional layers of ResNet residual block are replaced by two 1*1 convolutional layers plus a 3*3 convolutional layer, which not only maintains the accuracy of the network model but also reduces the amount of computation at the same time, the ReLU activation function is replaced with Mish. On the other hand, the convolutional attention mechanism is added to the network model to suppress the useless feature information, so as to improve the recognition accuracy of facial expressions. The experimental results show that the improved network model has significant improvement over the original network in RAF-DB and FERPlus data sets respectively.

Because the deep learning model is highly dependent on data, and the extraction effect of data sample features will directly affect the prediction accuracy. Based on this, in order to improve the accuracy of short-term power load forecasting, a load forecasting method based on time series generation antagonism network TimeGAN and short-term memory network LSTM is proposed for data enhancement. First, in order to optimize the effect of model feature extraction, the sample six-dimensional feature data is reconstructed into nine-dimensional feature data according to the date and weather characteristics. Then, analyze the correlation between historical data and sample distance, use TimeGAN model to enhance the data, and then reconstruct the data set. Finally, the prediction model of long and short-term memory network is created to import the reconstructed data to predict the electric load in the next 24 hours. The experimental results show that this method is superior to the prediction methods of CNN-LSTM, CNN-BiLSTM and LSTM models, and TimeGan-LSTM has higher prediction accuracy.

Aiming at the problem that web front-end dynamic form rendering often depends on a specific framework and has poor portability under traditional development methods, this paper proposes a dynamic form rendering implementation method for multiple front-end frameworks. First, the dynamic form configuration file is split into corresponding form element areas according to the keywords of the form elements in the dynamic form configuration file, and then native JS is used to encapsulate the split form elements in a componentized manner. Then, for different front-end pages, the frontend frame assembly adapter completes the adaptation and assembly of the encapsulated form element components, and finally completes the rendering of the dynamic form under different front-end frame pages.

Cantilever beams with defects in different locations are studied in the finite element software. The acceleration response of the cantilever beams is processed by the shock-waveform decomposition method in which the characteristic parameters of the acceleration are extracted, and the dataset is formed. Then, the dataset is trained by BP (Back Propagation) neural network to identify the location of defect. It is shown that the defect detection method based on the shock-waveform decomposition method and BP neural network has high defect detection accuracy and efficiency.

In order to efficiently transmit various types of spatial data in Advanced on-Orbit System (AOS), an AOS virtual channel scheduling algorithm based on Q learning is proposed. This algorithm comprehensively considers the service priority, scheduling delay and frame residual, establishes a reward function model, and evaluates the advantages and disadvantages of virtual channel sequencing according to the obtained Q value, and obtains the optimal virtual channel scheduling order after repeated iterative training. The simulation results show that the virtual channel scheduling algorithm based on Q learning has smaller average scheduling delay and frame residual, improves information transmission efficiency and optimizes system performance.

Traditional anomaly detection methods cannot take into account the influence of external factors. Therefore, it is difficult to obtain an accurate detection effect when performing gas anomaly detection. In this paper, we propose an anomaly detection model based on LSTM-WGAN. The multi-layer LSTM network captures temporal dependencies and introduces a self-attention layer to integrate embedded weather, holidays and other influencing factor data. Embed different features into the WGAN framework and make the detection results more accurate by fused models. This paper conducts experiments on the real gas load data set. The experimental results show that the model in this paper has obvious performance advantages in precision, recall and F1 value compared with traditional time series anomaly detection methods.

The Bilayered Neural Network (BNN) prediction model uses to predict the State of Health (SOH) of lead-acid batteries used in aerial work platforms. After the cycle life test of lead-acid batteries, a multi-layer structure prediction model builds to improve the accuracy of the prediction results. Compared with the linear model, tree model, Support Vector Regression (SVR) model and Gaussian Process Regression (GPR) model, the training and test prediction results of BNN are better, and the test RMSE is 0.0562.

This paper seeks to improve the intelligence, real time and security of carbon trading and to reduce the default frequency of carbon trading. For these purposes, a carbon transaction model for the power industry was proposed based on blockchain. Firstly, a blockchain-based carbon trading framework is constructed, based on BIM carbon emission calculation to suit the general transaction scenarios. Then, using credit value evaluation and smart contracts to ensure the transparency, openness, and non-tampering of credit scores. On this basis, the transaction priority reward mechanism promotes the trustworthiness of each transaction node and maintains transaction security. Finally, the direct transaction based on credit value was designed to improve the transaction efficiency and security. Through case analysis, the carbon transaction model for the power industry, which is based on the credit value of the blockchain, supports fast and frequent energy transactions, as it overcomes the confirmation delays of energy blockchain transactions. The proposed model improves the efficiency of carbon transactions, effectively suppresses default frequency, and maintains the order of the carbon trading market.

Automotive connection function is an important part of the Intelligent and Connected Vehicle (ICV) industry and an important support for the realization of automated driving. This paper summarizes the automotive connection function technology standards and regulations and compares and analyzes the typical connection function application scenarios related to driving safety developed by leading automotive industrial counties and international organizations. Combined with correlation between technology and typical working conditions, in terms of technology maturity, technology cost, product market plan, etc., use statistical analysis, get the result of priority ranking of automotive connection function application scenarios related to driving safety. Based on the research, proposes suggestions for standard system construction.

With the rapid development of neural network, significant breakthroughs have been made in many aspects of object detection which has been widely used in many fields such as intelligent transportation and industrial detection. The State Grid is also focusing on object detection and related technologies to advance the intelligence of China's power grid in operations and maintenance and other fields. Therefore, in this paper we propose a fire detection oriented in server room. Since edge computing ensures responsiveness and reduces server load, object detection is better suited for deployment under edge computing. However, embedded devices in edge computing often have significant limitations in terms of storage and energy consumption. Traditional object detection algorithms are implemented by neural networks, which require large computational and storage resources, which conflict with edge computing. Therefore, we implement lightweight fire detection algorithms for edge computing via MobileNet to reduce parameter size and increase speed, and enhance multi-scale detection capabilities with the YOLOF algorithm.

Impedance control is hard and important for high frequency Printed Circuit Board (PCB). The paper introduces a BP-GA algorithm that applies the optimization on the iteration process of weight parameters during the training process. The results shows that the model has a good performance in impedance prediction of PCB. The RMSE and SSE of this model are 4.083 and 7.003*e3, respectively.

China's production of carbon for aluminum electrolysis has ranked first in the world for 19 consecutive years and has become the world's largest exporter of carbon for aluminum electrolysis production. In the process of carbon mixing and kneading, the uniformity of paste is an important factor affecting the quality of carbon anodes. In this paper, an intelligent sensing system is designed for modelling and analysis of the paste uniformity index. The system collects the three-phase voltage/current, stirrer speed and stirrer rotation angle of the carbon kneader in real time, calculates the active power, stirrer torque and stirrer rotation angle according to a specific cycle, and establishes a soft measurement model of the relationship between stirrer torque and time and rotation angle. During the kneading process, 72 torque time series are generated at 5° intervals, and the torque variation pattern at the same angle is analyzed in relation to the homogeneity of the paste, thus forming a criterion for the end point of the kneading process. A comparison with actual data from the field shows that the system is stable and reliable and can accurately determine the end point of kneading, avoiding manual errors in judgement and improving the level of automation and quality of kneading, while improving the performance of the carbon anode and achieving energy saving and consumption reduction.

In this paper, we construct a multi-index evaluation model based on Entropy Weight-TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) and 0-1 programming algorithm to complete the evaluation and selection of suppliers. We discuss the various capabilities of suppliers and analyzed their impact on supplier scores. At the same time, through the collected data of the past suppliers of the enterprise, we select the appropriate indicators and use the entropy weight method to calculate the index weight. Then the TOPSIS method is used to evaluate the suppliers, and then the 0-1 planning algorithm is employed according to the evaluation results. Finally, using the evaluation model, combined with the historical data of multiple suppliers, we establish five evaluation intervals to evaluate suppliers. Effective evaluation of suppliers can help enterprises to better select suppliers in order to facilitate long-term stable development of enterprises.

After four generations of development, aviation ammunition has become the most widely used and most widely used ground strike weapon in the world. Aiming at the requirements of realizing the weapon's precision counterattack capability in the complex and changeable war environment in the future, research the application of artificial intelligence technology in weapons and equipment. Based on the technical characteristics of artificial intelligence, this paper proposes the specific application of artificial intelligence technology in the field of aviation ammunition and analyzes the development prospect of aviation ammunition intelligent technology in combination with the future development trend.

Waste engineering pre-demolition glass is often retained as recycled material to be recycled in subsequent engineering construction. In order to achieve effective recycling of different types of glass, we take the study of high potassium glass and Pb-Ba glass as an example and proposes a hybrid engineering glass clustering model with subjective and objective verification. This model firstly determines the two-dimensional classification index system through cardinality analysis, and then constructs a two-level classification model based on K-Means cluster analysis and hierarchical cluster analysis to qualitatively determine the classification result of glass with a class number of three. Subsequently, the optimal class number for both high potassium glass and Pb-Ba glass is quantitatively determined to be three based on the distance evaluation function. The sensitivity of the hybrid clustering model was found to be high.

In order to effectively replace the daily fee urging work of the grid manager, realize the lean fee urging ability, give full play to the channel efficiency, and improve the efficiency of multi-channel collaborative fee urging and customer experience, an intelligent decision-making model for multi-channel collaborative fee urging was built. The sliding window and the anti k-nearest neighbor method are used to obtain multi-channel collaborative fee urging samples. According to the availability, integrity and business relevance of the data, based on incremental principal component analysis, combined with entropy method, the feature of multi-channel collaborative fee urging is selected. By using Logistic regression algorithm, an intelligent decision model of multi-channel collaborative fee urging is built to realize intelligent decision oriented to multi-channel collaborative fee urging. The experimental results show that the proposed method has a strong multi-channel collaborative fee urging capability and can effectively improve the efficiency of multi-channel collaborative fee urging.

In order to optimize the spatial layout of jet fans in the tunnel, the influence of the location of jet fans on the air flow and vehicle exhaust in the tunnel are studied in this paper. Firstly, the three-dimensional physical model of tunnel and mathematical model of jet fan ventilation and vehicle exhaust emissions are established. Then the effects of fan longitudinal spacing and transverse spacing on the velocity distribution, pressure-rise coefficient of jet fans and vehicle exhaust in the tunnel are studied through CFD numerical simulation, so as to obtain the optimal layout parameters. The results show that when the fan longitudinal spacing is 110-m to approximately 140-m, the pressure-rise coefficients of fans are 0.880 to approximately 0.891, which increase first and then decrease as the longitudinal spacing increasing, and the optimal longitudinal spacing is 130-m. When the transverse spacing of fans are 1.68-m, 2.24-m, 2.8m and 3.36-m, the pressure-rise coefficients of fans are 0.918 to approximately 0.936, which increase first and then decrease as the transverse spacing increasing. The transverse spacing of fans should be set between 2.24-m and 3.36-m, and the pressure-rise coefficients can reach above 9.2, and the optimal transverse spacing is 2.8-m. When the jet fans are arranged horizontally and longitudinally, the concentration of CO in the tunnel presents zigzag fluctuations. When the transverse spacing of fans are 1.68-m to approximately 3.36-m and the longitudinal spacing are 110-m approximately 130-m, the concentration of CO in the tunnel is far lower than the safety concentration. The research results can be used for reference and guidance in the layout and installation of jet fans in tunnels.

This paper proposes a method of semantic bias analysis and discrimination based on SVM. This paper adopts the word frequency coefficient analysis method based on a single word. Through the analysis of function words (auxiliary words, conjunctions, etc.), it studies the language differences between different forum users. After training the SVM classifier, carry out the test. The accuracy of the semantic bias test results reaches 77.5%.

Logisim is a cross-platform educational tool for designing and simulating digital logic circuits, and it is widely used in experiments for courses such as computer organization in universities. As Logisim is widely used, tasks such as plagiarism detection require the calculation of Logisim circuit similarity, and the Logisim circuit similarity algorithm becomes very important. Logisim can represent circuits as text in XML format or render circuits as RGB images, which means that Logisim circuit similarity can be calculated using both text similarity algorithms and image similarity algorithms. However, due to the characteristics of Logisim circuits, text similarity algorithms and image similarity algorithms cannot achieve good results. In this paper, we propose a Logisim circuit similarity algorithm based on the maximum connected subgraph. We first represent the Logisim circuit as a collection of maximum connected subgraphs, then calculate the proportion of nodes with the same type, similar location and isomorphic local subgraphs in maximum connected subgraphs as the similarity of maximum connected subgraphs, and finally calculate the proportion of maximum connected subgraphs whose maximum similarity exceeds a threshold as the similarity of the collection of maximum connected subgraphs (i.e., Logisim circuits). We construct two Logisim similar circuit datasets based on real-world data. The experimental results on these datasets show that our method outperforms well-known text similarity algorithms and image similarity algorithms.

In order to explore the problem of text analysis of power customer complaint work order, a text analysis of power customer complaint work order based on data mining is proposed. This method explores the research on the text analysis of power customer complaints work order through the key technical problems and solutions of information recommendation represented by data mining. The research shows that the combined curve is closer to the actual situation, and all prediction indicators have improved, indicating that the combination effect is good. At present, the data platform is still in the process of improvement. Therefore, enterprises cannot be complacent and stay at the current analysis stage. They should keep pace with the times, strive to improve the big data platform, and further improve their data mining and analysis capabilities to adapt to the increasingly fierce market competition.

The new Coronavirus epidemic has had a huge impact on the economy, politics, and culture worldwide. However, it is very difficult to obtain accurate data on the new crown epidemic due to various uncertainties, such as the difficulty of detection. In this paper, we use objective and real Baidu search indexes as the basic data set and use factor analysis with the improvement of entropy method to reduce the dimensionality of Baidu search index data to solve the problem of fixed parameters caused by its excessive dimensionality. After that, the WOA algorithm is used to optimize the parameters of the conventional BP neural network, thus making the fit and accuracy greatly improved, which is of great practical significance for the prediction of epidemic data.

The rapid development of the power grid brings massive business data and operation data to the field of power dispatch and control, which puts forward new requirements for fast response and intelligence of data retrieval in the dispatch system. Considering the characteristics and business needs of power dispatch and control, and based on cloud computing and big data technology, this paper studies the construction of an intelligent search engine for power dispatch and establishes a lexicon and knowledge graph in the field of power dispatch and control, covering various types of power data such as dispatch and control model, operation, and grid information. This provides a convenient and unified data search portal and provides comprehensive, accurate, fast, and intelligent search services for dispatch and control operation and management personnel.

Aiming at the problems of traditional Grey Wolf Optimization algorithm (GWO), such as easy to fall into local optimization, low precision of optimization, and slow convergence speed, a fast grey wolf optimization algorithm with exponential iterative search is proposed. The tent chaotic mapping method is introduced to realize the diversification of the initial grey wolf population. To balance the global search scope and search speed, an iterative search method for exponential nonlinear control parameters is designed to avoid the algorithm falling into local optimization and accelerate the convergence speed. Finally, an information-sharing method based on the inertia weight coefficient is proposed for the current optimal position of individuals, which dynamically updates the individual step size of grey wolves and improves the speed and accuracy of optimization. Through a variety of standard test functions, compared with the traditional GWO algorithm and three improved algorithms, the performance of the proposed algorithm with fast convergence and strong optimization ability is verified.

With the increase of the elderly population, the number of patients with Alzheimer's disease is increasing year by year. The economic burden of Alzheimer's disease is heavy, and it is urgent to strengthen the prevention and treatment of Alzheimer's disease. This paper selects the relevant research documents that have been published in domestic and foreign journals and conferences and applies the rough set attribute reduction algorithm to the prediction and analysis of Alzheimer's disease, by analyzing the prevalence of Alzheimer's disease in a certain region from 1990 to 2019, the incidence of Alzheimer's disease in the future is predicted, and effective prevention and treatment measures for Alzheimer's disease are proposed.

Machining simulation is often used to verify the rationality of NC (Numerical Control) machining toolpath and is the core function of CAM software. In order to improve computational efficiency, machining simulation algorithm is usually designed based on the discrete geometric model of workpiece, which is widely used in engineering. With the discrete geometric model as input, machining simulation algorithm is difficult to generate real-time changing 3D rendering visualization for workpiece. This paper proposed a GPU-driven mesh generation method for machining simulation. This method improves the traditional conversion method based on Marching Cube algorithm and uses the parallel computing capability of GPU to improve the efficiency of mesh generation. The method proposed in this paper supports real-time user interaction and may also be applied to virtual engraving modeling and digital twin for NC machining.

With the rapid development of information equipment, people's demand for teaching aids is increasing. Among them, the construction of writing aid evaluation system is a challenging task, its difficulty lies in the diversity of writing forms and the diversity of context. In this paper, we propose a Hybrid Model based on an ensemble module, which has the characteristics and advantages of both Roberta and Deberta. Our model includes enhanced mask decoder and disentangled attention transformer layers. At the same time, we used FGM adversarial training to improve model performance. We carried out tests on the Predicting Effective Arguments data set, and experiments showed that the performance of our model far exceeded that of other single models, such as Bert model, Roberta model and Deberta model.

Grassroots governance staff are submitting increasing amounts of text data to governance system due to encouragement of intelligent governance work. Misclassified work order data results in data loss, leading to wasted labor and material resources. Hence, effective text classification techniques are needed for correction. Training results are insufficient to suit industry needs due to limited work order data and difficult data conditions. For difficulties above, this research presents a mobileBERT-based preset class fusion text categorization algorithm (Preset class fusion BERT, PCFBERT). Templated preprocessing of address information, label representations produced by algorithm encoder are separated from input data representations to compute contrastive loss function and improve model performance on one hand and used as classifier input to produce final prediction results on the other. Traditional text classification algorithms cannot guarantee high speed and accuracy in face of tiny samples of actual data and complex data conditions. Suggested technique solves this problem. Templated preprocessing of address names reduces noise and makes dataset information more reliable. Label fusion and contrast learning techniques based on mobileBERT improve model data representation in this paper. Triple mapping module fuses pre-defined label class information into classifier and better utilize global information to improve decision-making. Combining strategies above improves structure robustness and performance without increasing model size. This research evaluates recommended strategy using simulated data from desensitized work orders and compares it to many popular text classification methods. Suggested method outperforms popular text classification algorithms in classifying work order text data. Same higher performance is achieved in two open Chinese multiclassification text datasets.

With the continuous development of distributed energy, the distribution network's demands for flexibility and operation efficiency are improving rapidly. Intelligent distributed network has the advantages of less pollution, high reliability, and high energy utilization, and is widely used in power systems. However, massive devices connected to the distribution network make the service data of the distribution network grow explosively and bring difficulties to the processing and decision-making. This paper proposes a Priority-driven Price Matching algorithm (PP-Matching) based on digital twins, making the cost of establishing matching between the device and the server with more conflicts higher by employing server bidding. Data simulation results show that the proposed algorithm can well solve the problem that the device is difficult to process and obtain service decisions in real time and improve the operation and maintenance management level of distribution network effectively.

Aiming at the problem that the double-neighborhood gradient algorithm has good real-time performance but general detection accuracy in complex environments, an improved double-neighborhood gradient infrared small target detection algorithm was proposed. Firstly, a single-scale three-layer window is used to traverse the input original image for saliency calculation, and the candidate targets are obtained after threshold processing; then the candidate targets are pixel-completed by the region growing method; finally, a final detection result is obtained by filtering the candidate targets using the gradient features of the targets. The results show that the method improves the background rejection factor and contrast gain by an average of 1.3 times and 1.6 times, respectively, compared to the double-neighborhood gradient algorithm, which can effectively suppress background clutter and enhance the target. The method in this paper has advantages in detecting infrared weak and small targets, and the effect is more significant in complex backgrounds.

In the power system, load regulation refers to maintaining the stability and optimal operation of the system by adjusting the power consumption. Traditional load control methods usually rely on manual analysis and decision-making and have many limitations and shortcomings. This paper proposes a load regulation method based on AI agent and constructs the implementation framework of AI agent and cloud platform cooperation. In this paper, the concept and background of AI under intelligent power consumption are briefly described. Secondly, a stochastic forest algorithm based on time-domain feature expansion is proposed to realize the evaluation and analysis of load operation status and provide a basis for decision-making of load regulation. Carry out data analysis and calculation on the cloud platform based on the evaluation results to achieve accurate load control. Finally, the simulation experiment verifies that the distributed load regulation optimization method based on artificial intelligence can effectively solve the problem of power system load regulation.

Although there are a large number of human behavior activity identification data and related wearable technologies, there are still many challenges to realizing the understanding of human behavior. Thus, it is required not only perceive and effectively process the information data of human behavior activities, but also accurately and effectively classify the information data. The core of understanding human behavior is the identification and monitoring of human activities. In this paper, a classification model based on the k-nearest neighbor method was established, and a s-fold cross validation method was proposed. The results showed that the precision of the as-proposed model was about 82%, and the generalization ability was good. Besides, weights were added to the predicted cosine distance, and heuristics were used to search for the best weights, making it was more inclined to fall into the neighborhood of the test set samples. As a result, the overall accuracy could be gradually improved. The prediction accuracy showed an upward trend with the increase of iterations, indicating that this method is indeed helpful to overcome the overfitting problem of human activity classification strategy.

Aiming at the privacy protection problem in the cloud computing environment, a signature algorithm based on the Decision Diffie-Hellman (DDH) assumption and the discrete logarithm problem is designed when studying the Certificateless Homomorphic Signature (CLHS) technology. Our algorithm can resist pollution attacks from cloud computing users, collusion attacks from semi-honest participants, and malicious attacks from key generation centers. While protecting user privacy, our algorithm also ensures the efficient development of cloud computation server services.

The conservation of modern architectural heritage is different from that of ancient or traditional buildings, and there are special features in the development process from the creation of conservation issues to the implementation of conservation and utilization measures. As the national task of energy conservation and emission reduction increases, the demand for retrofitting the performance of existing buildings increases simultaneously, and China's modern architectural heritage cannot meet the requirements of modern buildings in terms of building energy consumption and indoor comfort. Traditional retrofitting strategies lack quantitative analysis of the retrofitting target and have no clear objectives, making it impossible to achieve the expected results after retrofitting. Therefore, this paper proposes an intelligent algorithm-based multi-objective optimization strategy for retrofitting architectural heritage in order to solve this problem. A neural network is introduced into the genetic algorithm to realize a multi-objective optimization algorithm, which breaks through the limitation of the number of optimization objectives while enhancing the optimization search speed and develops a multi-objective optimization transformation decision process for the reuse of modern architectural heritage. The empirical results prove that the multi-objective optimization transformation decision method proposed in this study can improve the transformation decision efficiency.