This PDF file contains the front matter associated with SPIE Proceedings Volume 12717, including the Title page, Copyright information, Table of Contents, and Conference Committee lists.

Since the attitude angle between the gripper and the roller of the robot arm is not adjusted during the autonomous obstacle surmounting process of the substation inspection robot, and the obstacle surmounting route found is not optimal, the path of the substation inspection robot for autonomous obstacle surmounting is not the optimal result. Therefore, the autonomous obstacle surmounting method of the substation inspection robot based on locust optimization algorithm is proposed. After calculating the vertical distance between the overhead line and the working ground and adjusting the attitude angle between the gripper and the roller of the manipulator, a two-dimensional Gaussian scale space is constructed to extract the key feature points of the obstacle image in the space. According to the feature extraction results, combined with the locust optimization algorithm, the autonomous obstacle path of the substation inspection robot is planned to obtain the most superior obstacle path. In the experiment, the obstacle surmounting effect of the proposed method is verified. The experimental results show that the error rate of the optimal solution is low, and the path planning performance is good when the proposed method is used for robot autonomous obstacle crossing control.

As a special I/O channel in the control system, the eTPU (Enhanced Time Processing Unit) cooperates with the host computer to complete the corresponding functions. As the most foundational and important function in the eTPU, which is completed through I/O channel. This paper makes a comprehensive analysis of the thread entry and the pre-defined channel mode in the process of implementing the I/O function. Through the design of thread entry and the optimal selection of the pre-defined channel mode, the I/O function can be efficiently completed. By adopting clever design ideas and optimization, eTPU can provide more accurate and efficient functions. Especially through the implementation part of this article, the usage method of eTPU is more convenient to master.

To ensure the stable running of systems and services in a data center, O&M engineers need to collect and monitor Key Performance indicators (KPIs) generated during system and service running. Traditional performance anomaly detection methods based on thresholds and rules are suitable for simple KPI monitoring. Unfortunately, when there are significant differences between KPI values at different times of the day, or when there are significant fluctuations due to different usage patterns, it is difficult to find appropriate threshold levels that are gradually no longer appropriate for highly dynamic systems and businesses. With the popularity of Artificial Intelligence algorithms, machine learning and deep learning methods also begin to be applied in operation and maintenance scenarios, which is the emergence of Artificial Intelligence for IT Operations (AIOps), among which KPI anomaly detection is the research hotspot of AIOps. KPI anomaly detection refers to the effective data mining and analysis of KPI time series data through various anomaly detection methods, in order to quickly discover anomalies and avoid faults. Therefore, we propose a hybrid model of KPI anomaly detection based on GAN. This is an unsupervised learning method based on Generative adversarial network (GAN) and Encoder (Encoder), which can identify abnormal data in KPI. We train the GAN model with health data, and then complete the rapid mapping of new data to the potential space of the GAN model through the encoder, and then detect anomalies by combining the feature residuals of the discriminator and the data reconstruction error into an anomaly score. Experimental results show that, compared with the traditional KPI anomaly detection method and the general unsupervised method, the proposed method can capture the variable data characteristics of KPI more accurately and achieve better performance in the KPI anomaly detection task.

Photovoltaic power generation is one of the most promising forms of renewable energy generation. The forecasting of photovoltaic power is of great significance for ensuring the stability of the power grid system. In this article, a combined forecasting model based on maximum overlapping discrete wavelet transform, particle swarm optimization and bidirectional long short-term memory network is proposed to forecast photovoltaic power. Firstly, the signal component analysis of photovoltaic data is carried out using MODWT, and the frequency principal component is extracted as BiLSTM training data through signal reconstruction. Secondly, considering the fact that the number of network hidden layers and the learning rate affect the network performance, in order to obtain the most suitable network model parameters for the case data, PSO optimization is used to procure the optimal number of network hidden layers and the optimal learning rate. Finally, the parameters are assigned to BiLSTM which is then retrained to get the final forecasting results. Taking the real photovoltaic data of a certain area as an example, compared with the other eight forecasting models, it is verified that the proposed model has the best forecasting performance.

Under the influence of social development, weather and geology, the difficulty of long-distance route operation and maintenance gradually increases. In this paper, a ground station system suitable for taking off and landing in different places is developed to make the fixed wing UAV no longer fly back, improve the length of single flight operation, optimize flight efficiency, and solve the problem that inspection personnel still need to travel by bus.

Aiming at solving the problems of insufficient feature information extraction and low accuracy in conventional anomalous sound detection methods, this paper presents a new method for detecting anomalous sound based on STgram-MFN optimization. By fusing multiple attention mechanisms for feature recalibration, it can selectively emphasize features with high informative content and suppress less useful features, thereby improving the accuracy of anomalous sound detection. Experiments on the DCASE 2020 Challenge Task two dataset show that compared with the original STgram-MFN, Its AUC has reached 94.20%, 74.29%, 88.82%, 92.86%, 99.29%, 98.06% (ToyCar, Toycar, Fan, Pump, Slider, Valve). Respectively, increased by 1.56%, 1.37%, 4.05%, 2.87%, 0.04% and 2.91%. In addition, the average AUC of our proposed method is improved by 2.13%.

Multi-Robot Multi-Target Tracking (MR-MTT) addresses the problem that a swarm of mobile robots actively detect and move to maintain surveillance of a team of dynamic targets, which is a fundamental problem in the modern robot system and has enormous potential in numerous fields. This paper investigates a series of MR-MTT algorithms, with detailed demonstrations of two distributed approaches to solving MR-MTT in finite iterations. On the one hand, a novel local algorithm is introduced to solve multi-robot multi-target assignments within limited communication rounds. On the other hand, an innovative coverage control algorithm combined with a probability hypothesis density (PHD) filter for target estimation is derived to settle detection and tracking tasks simultaneously. Both algorithms are analysed and discussed mathematically, with simulation experiments demonstrating their efficacy and performance. It is concluded that both algorithms solve the MR-MTT problem in terms of optimal coverage and optimal target estimation, respectively.

The positioning of indoor mobile machines cannot be solved by a single sensor. A positioning system based on multi-sensor data fusion is designed. The system combines the positioning results of UWB with the positioning results of low-cost MEMS inertial measurement elements to improve the positioning accuracy of the system. The system includes two Kalman filters. The primary Kalman filter is used to fuse the angular velocity information of gyroscope and the angular value of magnetometer, so as to obtain more accurate heading angle. The secondary Kalman filter combines the IMU track estimation results with the UWB positioning results to obtain accurate positioning results. The final experimental results show that the method is effective.

In this paper, the passive positioning problem of UAV formation in flight is studied, and two different formations are discussed, and two mathematical models are established. Firstly, geometric model-based exhaustive method is used to calculate with the help of Python language, which greatly saves calculation time. Secondly, the iterative merging algorithm is used, and the simulation is carried out by Matlab. The control error is in a small range, and the final conclusion is obtained, which provides a certain theoretical basis for the practical application of passive positioning of UAV formation.

Traditional unmanned logistics vehicles mainly adopt a driving strategy of running along the track, which cannot adapt to complex roads and traffic. This paper introduces a fusion decision model for the behavior of unmanned logistics vehicles in complex road conditions, which can support intersections with traffic lights and complex traffic scenarios with multiple vehicles and multiple pedestrian flows. The paper implements a multi-state machine behavior fusion decision model for unmanned logistics vehicles by separately modeling the vehicle, pedestrian, and navigation situations at intersections with finite state machines, and then implementing model fusion through a fusion decision model. The paper use test vehicles to test in an indoor open road test site, and the experimental results show that the model can effectively avoid unmanned logistics vehicle blockage and route deviation at intersections, improve unmanned logistics vehicle intersection passage efficiency, and solve the technical problem that the traditional unmanned logistics vehicle cannot adapt to complex traffic road conditions.

Deep convolutional networks have become a popular tool for image generation and restoration.In general, their outstanding performance is attributed to their ability to learn real image priors from a large number of example images.Many seemingly unrelated computer vision tasks can be viewed as special cases of image decomposition into independent layers.Image segmentation (divided into foreground layer and background layer); transparent layer separation (divided into reflection layer and transmission layer); image dehazing (separation into clear image and haze map), etc.Nowadays, in the field of image restoration using deep learning, there are many low-quality images, but the corresponding original pictures are fewer, which makes image restoration based on deep learning one of the more difficult problems in the field of computer vision.In order to break through the related problems, this paper abandons the supervised deep learning method based on a large amount of data training and then proposes an autonomous unsupervised image inpainting network, Fast-RDDIP, based on structure prior.This ability stems from the fact that the internal statistics of a mixture of layers are more complex than the statistics of each individual component.Consequently, this study constructs a matching network structure using the characteristics of various learning distribution rates of the network and additional convolution, which has characteristics like shared parameters. can perform end-to-end training.This essay explains how this technique improves image quality and dehazing. According to the experimental findings on the O-Haze dataset, Fast-RDDIP performs better than more established methods like SIHR, EIDBC, supervised neural networks like DehazeNet, NTDF, NLD, URIE, and BOPBL, as well as unsupervised neural networks like DoubleDIP models.

With the development of industrial automation technology and computer control technology, the traditional mechanical and electrical equipment can no longer meet the requirements of modern industrial production for high precision, high efficiency and intelligence. Therefore, the organic combination of advanced sensors, controllers and actuators has become one of the current research hotspots. In this paper, we design and develop a comprehensive measurement and control system integrating data acquisition and processing, process monitoring, fault diagnosis, and remote communication with "PLC-based electromechanical-hydraulic integration experiment platform" as the background. The system has the features of compact structure, complete functions, simple and convenient operation and high reliability, which can realize the real-time monitoring of the operation status information of the object under test, improve the quality of experimental teaching, and also can be used as an important means to transform the results of scientific research projects. In the current industrial control field, PLC plays an important role due to its multiple advantages, which is a common module in industrial control. PLC is a programmed control device, this is specially designed for functional control in industrial conditions, and it can realize the storage of programs.

In scientific and engineering computations, it is often necessary to solve eigenvalue problems of symmetric matrices. With the rapid development of high-performance computing, the adaptation and optimization of linear algebra solution methods, including eigenvalue problem solving, on heterogeneous platforms have become increasingly important. In this paper, we propose a divide-and-conquer method for solving symmetric matrix eigenvalue problems, which is implemented based on the SOLVER library of domestic heterogeneous platforms using the HIP programming model. We take full advantage of the multicore strengths of domestic accelerators and parallelize the solution process, such as the secular equation. Parallel reduction and merged computation optimization techniques are employed to further improve performance. Compared with implementations such as MAGMA, our optimized interface demonstrates good stability and high accuracy in various scale test cases of real-world applications, as well as a significant performance advantage. The results show that in large-scale matrix eigenvalue problems, the performance is more than twice that of the related interfaces in the MAGMA library.

Complex environments, such as dense personnel and background interference, affect the detection accuracy of whether personnel wear helmets. To solve this problem, a new detection algorithm for helmets based on YOLOv5 is studied in this paper. Firstly, RepVGGblock is used to replace the common convolution in the network to effectively utilize the computing power of GPU. Secondly, the efficient channel attention mechanism is incorporated into the C3 module of the backbone network to enhance the feature identification ability of the backbone network for the hard hat. Then, the boundary box regression function is changed to SIoU to redefine the distance loss and improve the regression accuracy. Finally, the detection performance of multi-scale targets is improved by adding detection layers. The test results on the self-made safety helmet data set show that the average accuracy of the improved YOLOv5 model reaches 94.6%, which is 4.1% higher than that of the original model, and can meet the requirements of target detection.

In order to further improve the comfort of the lighting environment in crew cabins, this paper details the design and implementation of a ship's light environment control system. The ship light environment control system consists of a dimming lamp, a lighting sub-controller and a light environment control platform. The ship light environment control system is used to realize the internal lighting of the ship's cabins. Through the application of LED light sources, regionalized lighting technology and centralized control technology, the purpose of improving the cabin lighting environment, enhancing the functionality and comfort of lighting, and upgrading the level of lighting information is achieved. The ship light environment control system not only provides lighting for the interior of the cabin, but also has a corresponding human-machine interface for remote control and management of the dimming lamps.

In order to solve the flight stability control problem of the gun-launched micro quadrotor UAV during the autonomous flight phase, this paper proposes an adaptive control law design method based on the integral backstepping method. This method is based on modeling the dynamics of the gun-launched micro-quadrotor UAV into a strict feedback form and uses the Lyapunov stability theory to construct a double loop with attitude control as the inner loop and position control as the outer loop. The control structure uses self-adaptive technology to compensate for external disturbances, and at the same time adds an integral link to the position loop and attitude loop to ensure that the UAV has good stability and robustness. The simulation experiment results verify that the adaptive integral backstepping method designed in this paper has good control performance for the trajectory tracking of the gun-launched micro quadrotor UAV and can accurately estimate the interference.

Aiming at the problems that electric bicycles are prone to fire when charged at home, such as poor effectiveness of electric bicycle blocking systems in elevator cars, and low target detection accuracy, this paper proposes an improved YOLOv5 algorithm for detecting electric bicycles in elevator cars. Firstly, an FcaNet network model is embedded in the backbone network. The FcaNet network is implemented based on discrete cosine transform, mining multiple finite frequency domain components, making full use of input channel information, and improving the network's feature extraction ability; Secondly, in the process of feature fusion, relying on the mapping of CBAM attention mechanism to channel and spatial information, the CBAM attention mechanism is fused with the C3 model to form a C3CBAM model, which optimizes the fusion process of feature information, thereby enabling the model to achieve better detection performance; Finally, replacing the activation function in the SPP structure and using ReLU to replace SiLU improves the detection accuracy and speed of the algorithm. The experimental results show that the accuracy P, the recall rate R mAP Compared with the improved YOLOv5 algorithm, the accuracy of the algorithm was improved by 3.2 percentage points, and the recall rate was improved by 2.4 percentage points, mAP Increased by 4.1 percentage points, effectively improving the detection accuracy of the electric bicycle inside the elevator car.

Automatic driving is faced with complex and dynamic scenarios, especially under the influence of complex and variable factors such as longitudinal vehicle speed, road terrain, and other complex and variable factors, the vehicle's nonlinear and tire force longitudinal coupling characteristics are significantly enhanced, placing higher requirements on the performance of vehicle stability control systems. Based on this, this paper proposes a vehicle stability identification method which are able to identify vehicle stability status online based on driving style based on data driven and mechanism hybrid models. Firstly, based on vehicle stability evaluation criteria, the feature parameters are processed using unscented Kalman filtering algorithm and factor weighted analysis method to achieve quantitative stability evaluation; Secondly, using the dSPACE simulation platform, a test model for signals such as switch on, brake pedal on, steering wheel angle, and vehicle yaw rate under extreme conditions is established, and a driving stability demand identification model is established using the K-means algorithm; Thirdly, combining the demand identification model and quantitative evaluation training test data, a stability quantitative identification model is obtained; Finally, a collaborative simulation model of MATLAB/Simulink and CarSim was established and tested. The results show that the recognition method can accurately realize the reasonable classification and online recognition of the stability of autonomous vehicle, and to a greater extent, it reflects the design concept of intelligent vehicles that "cars adapt to people, not people adapt to cars".

Long Sequence Time-series Forecasting (LSTF) plays a crucial role in data-driven decision-making tasks in many fields. Transformer-based prediction methods significantly improve the LSTF challenges, but existing studies ignore the effect of noise in the series. Therefore, this paper proposes the time series forecasting model based on the wavelet transform, and it has the following main features: (1) This paper decomposes the time series with the wavelet transform and models the time series data from time and frequency domains, preserving the low-frequency components to keep the compactness of the time series; (2) This model effectively filters out high-frequency interference by combining the Fourier transform with a filter, while preserving the integrity of the effective signal; (3) Internal dependence of noise reduction sequences through the self-attentive mechanism to establish the connection between high and low-frequency signals. In this paper, experiments are conducted on four real data sets, and the performance of this model is significantly improved compared with existing advanced prediction models.

The control of oven temperature is a lagging, nonlinear system. To address the problem of poor results of traditional PID control of oven temperature, the mathematical model of oven temperature is derived by testing the temperature data of the oven. Using the traditional PID control, adding fuzzy control, particle swarm algorithm, and proposing an improved particle swarm to calculate the PID parameters in the fuzzy control, the temperature control of the oven equipment is carried out, and the weighting factor and learning factor in the particle swarm are optimized and adjusted respectively. The effectiveness is verified by MATLAB program and Simulink simulation, and compared with traditional PID and fuzzy PID, the method has fast response, small overshoot, faster stabilization time, all aspects due to the traditional method.

The traditional control method of new energy vehicle steering by wire has the disadvantage of low control performance. For this reason, the research on the control method of new energy vehicle steering by wire based on large data analysis is proposed. Based on the pre-processed data of the new energy vehicle steering by wire, the new energy vehicle steering by wire control model is established, and the control parameters of the new energy vehicle steering by wire are obtained. Based on the control parameters, the new energy vehicle steering by wire control is controlled by the control algorithm. The experimental results show that the control accuracy and efficiency of the proposed control method are 28%, 15% and 4% higher than those of the traditional method, which fully illustrates that the new energy vehicle steering by wire control method has better control performance.

In order to improve the laser ranging system in San Juan, Argentina, a design of the daytime satellite laser ranging system is proposed. With a variety of filtering technologies, the daytime laser ranging in San Juan can be realized. The theoretical results indicate that the quantity of observation data obtained by the optimized Argentine SLR system is increased to a great extent while the precision of ranging can reach up to mm level.

For the problems of modeling difficulties and signal feature extraction in inverter open-circuit fault diagnosis methods, the methods based on machine learning and deep learning still have limitations in practical applications. In this paper, a three-phase inverter fault diagnosis method based on Markov variational field and residual convolutional network is proposed. The method encodes the phase voltage signal output from the inverter by MTF to generate a two-dimensional image with timing information and state migration information, avoiding the problem of signal information loss when directly converting voltage or circuit signals into one-dimensional or two-dimensional images. At the same time, the method uses ResNet to process MTF images to extract fault features of phase voltage signals and enhance the propagation of feature information so that the feature information can be fully utilized, and accurate fault classification and identification can be achieved. The experimental results show that the proposed method can effectively identify 22 open-circuit fault types of three-phase inverters with a diagnostic accuracy of 98.10%.

A two-degree-of-freedom redundantly driven parallel robot is proposed to address the drawback that the two-degree-of-freedom parallel robot cannot reach the boundary. Firstly, the positive and negative solutions of the two-degree-of-freedom redundantly driven parallel robot are solved by the D-H parametric method, and then the joint kinematic simulation is carried out by ADAMS and MATLAB; finally, the Monte Carlo method is used to analyze the workspace before and after adding the redundant drive. The kinematic inverse solution was found to be correct, and the redundant drive was able to increase the working space. The results of this study have important theoretical implications for the dynamics analysis and structural design optimization of redundantly driven robots.

End-to-end driving behavior decision-making is a research hotspot in the field of driverless driving. This paper studies the end-to-end lane change decision control based on the PPO (Proximal Policy Optimization) deep reinforcement learning algorithm. First, an end-to-end decision control model based on PPO algorithm is established. The model uses the information perceived from the environment as the input state and outputs the control quantity (acceleration, braking, steering). Training and verification in the driving environment under the TORCS (The Open Racing Car Simulator) platform show that the model can achieve end-to-end lane change driving behavior decision-making. Finally, compared with the DDPG (Deep Deterministic Policy Gradient) model, which is also a deep reinforcement learning method, the experimental results show that the PPO lane change decision control model has a faster convergence rate.

This article is based on the logistics drone to deliver path planning problems in the urban environment. Combined with the flying characteristics of drones and the distribution location, considering the conditions such as the customer's collection and delivery needs, the logistics drone's delivery path planning model is established. Secondly, the model is processed. For the first time, the NSGA-II algorithm is used to solve the order of the distribution path of the logistics drone, and the second will be solved by improving the traditional algorithm. In the drone execution task, if the weight coefficient of the target functions is adjusted, different transportation sequences and paths can be obtained, and different planning results can be obtained by adjusting the attention of customer satisfaction. The analysis of the algorithm is available. Compared with the traditional A*algorithm, the algorithm transportation path adopted in this article is shorter and the delivery is faster.

To satisfy the application requirements of the UAV cooperative algorithm, the current embedded implementation method of the cooperative algorithm is investigated, and an embedded system that can run complex cooperative algorithms is proposed in this paper. The system can receive and process data from other modules of UAV at high speed and provide high-performance computing power for the application of cooperative algorithms. In this paper, the feasibility analysis is carried out on the embedded system's communication ability and computing ability.

During the motion of a robotic arm, its response speed and trajectory accuracy are crucial. This paper analyzes the control mode of the traditional PID controller, combines the principle of particle swarm algorithm, and studies the improved PID control based on the PSO algorithm. Simulate the optimized control effect using simulink simulation software, and compared with the tracking effect of the traditional PID controller. The simulation results show that compared to traditional PID controllers, the optimized PID control system has shorter response time, enhanced anti-interference ability, and more stable dynamic performance.. It is shown that using an improved PID controller to control the motion of the robot arm can make the motion response speed faster, better suppress the interference of the external environment, and improve the accuracy of the robot arm's motion trajectory.

An improved Faster-RCNN defect recognition model is proposed for the recognition of hydraulic cylinder internal surface damage defects of hydraulic supports. ResNet-50 is used instead of the original feature extraction network VGG16. Meanwhile, the anchor window generated by the region suggestion network is improved using the K-means clustering algorithm, the original ROI pooling operation is replaced by RoI Align to solve the region mismatch problem caused by the quantization operation, the soft non-maximal value suppression algorithm is used for screening, and the difficult negative sample mining method to optimize the model training process to improve the generalization ability of the model. The experimental results demonstrate that the mAP of the model on the dataset is improved by 5.87% compared with the original model, reaching 81.26%.

At present, the object grasping system of robot based on vision is one of the research hotspots, and it has a broad application prospect in industry, especially in the part of sorting. The key technologies of this system include robot hand-eye calibration, grasping target detection, grasping path planning, etc. This paper provides an overview about the current status of these key technologies by collating relevant literature, including the four kinds of hand-eye calibration methods, deep learning based grasping target detection algorithms, and the advantages and disadvantages of three kinds of grasping path planning methods. In this paper, the key technologies of vision-based robot object grasping system are summarized and prospected, which can be helpful to researchers in related fields.

Through the backstepping technique, this paper designs an observer-based event-triggered adaptive trajectory tracking control algorithm for the Autonomous Vehicle (AV) platoons with motion uncertainties (e.g., unknown AV mass, air resistance, and variable conversion errors). In order to avoid the transmission of excessive multi-vehicle status information (i.e., speed, acceleration, headway, position, expected trajectory, and so on) between AVs, this adaptive platoon control algorithm proposed only uses the imprecise sampled AV positions. The event-triggered mechanism with a fixed event-triggered threshold is introduced to reduce the update frequency of AV control laws. Through the newly constructed Lyapunov function, the adaptive platoon control algorithm designed could realize the simultaneous tracking of expected position and speed trajectories, and there exists an update time interval limit of AV control laws not less than the sampling time interval of AV positions. Through numerical simulation, the adaptive platoon control algorithm designed could control the heterogeneous AV platoon in linear formation where AV speed and positions are observed precisely, and all heterogeneous AVs in the platoon track the expected position and speed trajectories simultaneously. In addition, the update time interval of AV control laws is longer than the sampling time interval of AV positions.

Artificial Neural Networks (ANN) are now widely used in the medical field. However, on many datasets, only using ANN does not necessarily yield high prediction accuracy to some extent. Therefore, in this paper, the authors tried combinations of different models to obtain better accuracy based on the binary classification of a heart failure dataset collected from Pakistan, Faisalabad hospital. Briefly, the author considered ANN as feature extractors and then connected some statistical binary classifiers to build models, such as ANN with Support Vector Machines, ANN with Logistic Regression, and ANN with Random Forest. In particular, the authors used two different kernels of Support Vector Machines to build models in the artificial neural network with Support Vector Machines. One kernel is the radial basis function kernel; the other is the polynomial kernel. Experimental results indicate that models combining ANN and other statistical binary classifiers all get better performance than the ANN model alone. Moreover, the model with the highest accuracy is the ANN combined with Random Forest, followed by the ANN combined with Support Vector Machines and Logistic Regression.

Compound mode I-II crack expansion is a common fracture source for the failures of mechanical components in real-world running structures. Therefore, prediction of crack extensions of mode I-II loading is a long-term research hotspot. Software FRANC3D is widely used to simulate the growth of fatigue cracks with high precision for engineering applications. However, the high computational cost for the usage of FRANC3D are obviously. Data-driven machine learning model is another strategy to predict crack expansion with low accuracy for the lack of training samples in real-world running structures. In order to fast and accuracy predict compound mode I-II crack expansion, a hybrid model of Finite Element Method (FEM) and Machine Learning (ML) is developed by interchangeably using FEM and ML. Two cases are given to validate the performance of the present hybrid model by using FEM and Support Vector Regression (SVR) and Generalized Regression Neural Network (GRNN), respectively to predict compound mode I-II crack expansions in a stress plate. Finally, to verify the high precision and efficiency of the hybrid model compared with the results of simulation and other models.

In order to better conform to the real driving scene, intelligent driving simulation test needs to generalize the test scene on the basis of the collected natural driving data. In this context, a scenario generalization method for intelligent driving simulation based on Latin hypercube sampling is proposed in this paper, which makes up for the defect of the traditional Monte Carlo scenario generalization, in which the scene parameters are not covered enough in each dimension and ensures the coverage of the generalized scene. The comparison experiment shows that the generalization method of intelligent driving simulation scenarios based on Latin hypercube sampling can be applied in a limited time. Almost all scenarios are covered within the sampling times. Under the condition that the sampling times are the same, the risk degree of simulated scenarios can be guaranteed to be consistent with the real data sources and have higher robustness.

With the continuous development of high-speed railway construction technology, intelligent construction technology supported by new generation information technology has gradually become one of the important development directions of high-speed railway technology. As an important infrastructure of high-speed railway, the construction quality of the track is directly related to the operation quality and driving safety, so it is necessary to study the intelligent construction technology of the track. The intelligent construction of the track includes three aspects, namely intelligent design, integration of design and construction, and intelligent control. Design, as the source and foundation, must first achieve intelligence. Through digital, dynamic, and visual design technology of the track, traditional blueprint delivery is transformed into digital design delivery of the track, providing a data foundation for intelligent manufacturing, construction, and operation. Based on this data foundation, for the construction process of the track panel, the digital design results are connected with the various intelligent equipment through the digital channel, realizing the automation and operation of measurement robots, track geometric position detection instruments, portable measurement and control terminals, and related tools. Establish a real-time data acquisition channel for ballastless track and construct a big data resource pool for ballastless track to form intelligent construction and control technology for the track of "quality monitoring, data access, and closed-loop management". At the same time, track intelligent construction equipment and software have been developed, which reduces the labor intensity of workers, improves the precision level of ballastless track control, and effectively improves construction efficiency.

With the rapid development of China's aerospace industry and the increasing number of tasks year by year, the traditional processing methods have been unable to meet the needs of aerospace products with multiple varieties and small batches. It is urgent to find a low-cost and efficient production method. In view of this situation, this paper puts forward a design idea of an intelligent machining unit for aerospace electronics structural parts. Centering on improving equipment startup rate, reducing personnel demand, and improving quality consistency. Following the overall idea of "unified the type of processing method, complete some work offline, improve the flexibility of machining and intelligent decision-making", taking the built intelligent FMC (Flexible Manufacturing Cell) as an example, it realizes continuous and unattended automatic processing of products, Under the premise of ensuring quality, the production efficiency has been improved and good practical results have been achieved.

VASP is a widely used software, performing electronic structure calculations and quantum mechanical molecular dynamics from first principle, in the areas of material modeling and computational material science. To obtain better computing performance, VASP usually use heterogeneous architectures for acceleration. VASP6 is currently accelerated on NVIDIA GPUs with OpenACC programming model, which cannot be compiled and run on domestic heterogeneous platform. In order to use VASP on domestic heterogeneous platforms, it is necessary to design a new transplantation strategy. Following the OpenACC GPU-port, VASP was ported on the domestic heterogeneous platform with the HIP- C/C++ programming model and libraries on the platform. Hybrid density functional calculation is a hotspot of calculation in VASP. This calculation is frequently used but takes a long time. In order to provide good calculation services for researchers, it needs to be optimized. For a typical hybrid functional calculation example, this paper uses the performance analysis tool hipprof to find out performance bottlenecks. Combined with the characteristics of GPU like accelerator hardware architecture, multi-stream concurrency optimization and hardware resource allocation optimization are performed on the compute-intensive exact exchange computation. The test results shows that the calculation performance of the exact exchange is improved by 17.19% compared with the initial porting, which proves the effectiveness of the optimization.

With the improvement of people's living standards, waste production is rapidly rising, which has a substantial negative impact on human health, the environment, and economic development. Traditional garbage sorting process relies on manual sorting, which is time-consuming and ineffective. To help people quickly sort the garbage in their lives and reduce the phenomenon of misclassification and non-classification of waste in their lives, a new garbage sorting model is proposed in this paper. This model combines momentum contrast learning and a pretext task for self-supervised learning, and the trained deep neural network is used for a downstream task to classify the supervised spam image data. The combination of contrast learning and the pretext task forces the model to learn deep semantic features in the images, which enables the model to have better generalization ability, improves the robustness of the model, and achieves efficient classification of garbage. The final classification accuracy on the garbage dataset is 89.375% after finetuning the pre-trained deep neural network model, which is 1.472% better than the garbage classification accuracy with supervised learning.

With the continuous development of satellite technology, the acquisition and processing of satellite remote sensing data have become increasingly common. However, at present, satellites still use a mode of transmitting data back to the ground and utilizing ground CPU/GPU platforms for intelligent processing, which is inefficient due to the process of data transmission between the satellite and the ground. This article studies the application method and implementation of real-time remote sensing image recognition by integrating the Yulong810 on-board intelligent module on a satellite based on the YOLOv3 object detection algorithm. In response to the actual needs of satellite remote sensing image recognition, the algorithm is optimized based on the Yulong810 on-board intelligent module, which improves the recognition efficiency and reduces power consumption while ensuring the accuracy of remote sensing image recognition. By comparing with ground CPU/GPU platforms, the results show that the Yulong810 on-board intelligent module has lower power consumption and higher efficiency, and the recognition accuracy is comparable to that of ground platforms, which can completely replace ground remote sensing image detection equipment when integrated into satellites. By integrating the module into the satellite, real-time intelligent processing can be achieved in orbit, improving the efficiency of satellite remote sensing image recognition.

In modern society, violence and pornography often occur in KTV entertainment venues and on campus. It is of great significance to protect students' physical and mental health and maintain the order of KTV entertainment venues if violent and pornographic clips can be instantly identified from surveillance videos. However, pornographic video identification networks in general are different from violent video identification networks. Moreover, the traditional video classification network is not suitable for the classification of bad videos. Therefore, we propose a new deep learning network architecture for bad video classification. The network consists of a primary branch that uses optical flow information and a secondary branch that uses keyframe information. Experimental results show that this network structure is more accurate than the existing related video classification networks.

The current non-invasive load identification methods have limited recognition range and are prone to increasing time consumption. Therefore, a non-invasive load identification method based on deep learning is proposed. Based on actual recognition requirements and standards, extract recognition features, adopt multi-level processing forms, break through the limitations of recognition range, and set them as multi-level adaptive recognition nodes. We constructed a deep learning non-invasive load identification model and utilized GAGOA multi-objective recognition optimization to achieve non-invasive processing. The test results show that the time consumption of the non-invasive load identification unit obtained by applying the proposed method is controlled below 0.5 seconds, which is shorter. This indicates that the practical application effect of this method is better and has certain practical application value.

Due to the complexity of the electromagnetic environment of high-speed EMU, electromagnetic interference will be affected by many factors and malfunction, among which disturbance source in high-speed railway is one of the most important factors. As an important disturbance source in high-speed railway, the off-line discharge disturbance of pantograph and catenary is studied based on the generation mechanism of the off-line discharge of pantograph and catenary, and the radiation characteristics of the off-line discharge disturbance source of high-speed railway pantograph and catenary are studied by means of deep learning platform prediction and time-frequency analysis. Based on the generation mechanism of pantograph-catenary offline discharge, the time-frequency characteristics of pantograph-catenary offline discharge under normal train operation and iced catenary are analyzed through the equivalent circuit model of pantograph-catenary offline discharge mechanism. Based on the time-frequency signals of the whole process for the pantograph/catenary offline discharge during the experiment, the deep learning model and machine learning for the pantograph/catenary offline discharge are established, and the signal characteristics in the two-dimensional spectrum of the discharge frequency and time are mainly studied.

As a key part in the water environmental protection chain, the safe and stable operation of wastewater treatment plant is the prerequisite to ensure the quality of effluent. A major requirement for achieving this goal relies on the availability of real-time monitoring of key or primary process indicators. These indicators reflect the important information of wastewater treatment processes, but most of them are hard-to-measure or not easy to measure. Their real-time availability is often associated with expensive capital and maintenance costs, as well as being characterized by time-delayed responses that are often unsuitable for real-time monitoring. In order to deal with the online real-time acquisition of these variables, the soft-sensing techniques provide a feasible and effective solution. This dissertation surveys and discusses the modeling of data-driven soft-sensing techniques in municipal wastewater treatment plants. The main contributions are described as follows. A deep neural network soft-sensor model for wastewater treatment is proposed. Supervised neural network methods are one of the most popular data-driven soft-sensor modeling methods in wastewater treatment. However, most of them are shallow architecture, which are incapable of performing tasks effectively when got stuck in extremely complex situations, such as severe weather conditions. One of the potential solutions is resort to the neural networks with multi-layer structure, i.e., the deep neural networks. However, deep architecture neural networks suffered the difficulty of training, poor generalization ability, and so on, and thus be few successful application cases in early period. Furthermore, wastewater treatment has its unique practical matters, leading to the difficulty of collecting adequate data. Therefore, few literatures devoted deep neural networks to the soft sensor modeling for wastewater treatment. In response to this situation, a deep neural network soft-sensor model for wastewater treatment is constructed based on a deep learning model: stacked auto-encoder. At last, the main research work and corresponding results of this dissertation are briefly concluded, and the future research are discussed.

In this paper, we consider the group consensus of first-order hybrid Multiagent System (MAS). The hybrid MAS contains both continuous-time and discrete-time agents. We propose four consensus protocols for different cases. The consensus analysis is mainly based on algebraic graph theory and matrix theory. Sufficient and necessary conditions are provided for different cases to solve the group consensus of the hybrid MASs. Finally, numerical simulations are designed to illustrate the effectiveness of the proposed protocols.

The purpose of the Question Generation (QG) task is to automatically generate a reasonable and relevant question for the given document, the goal of this paper is to improve the performance of the question generation model and thus generate higher quality questions. This paper mainly focuses on the neural network model of Seq2Seq architecture and aims to two problems existed: exposure bias caused by the difference between the training and testing processes of the current Seq2Seq architecture model, and insufficient utilization of semantic information of the input documents. This paper proposes a question generation method based on semantic enhancement and reinforcement learning, which enhances the document representation through semantic graph; Add the reinforcement learning framework, use three special indicators for question generation tasks to fine-tune the model parameters during the training process to improve the performance of the model. The experiments’ results on show that this method effectively improves the quality of generating questions.

To address the issue of unknown initial position and lost localization for mobile robots, this paper proposed a global localization algorithm based on the fusion of vision and laser. In the mapping stage, a feature point map and a laser grid map are established respectively based on visual SLAM and laser SLAM. In global localization stage, the bag-of-words model is used for quick visual localization. Priori localization information in the grid map is obtained through aligning the maps. Precise localization is achieved using an adaptive Monte Carlo localization (AMCL) algorithm, realizing fast and accurate global localization for mobile robots.

The correctness of the secondary circuit protection plate for the on/off operation is a prerequisite for the correct action of various types of protection. In the context of the intelligence of various power system equipment, the intelligent operation and management of the secondary protection plate in substations is still a blind spot. This article designs an intelligent operation and "dual confirmation" error prevention and control system for the secondary voltage plate in substations based on machine vision artificial intelligence technology, creating a new intelligent operation mode for the secondary system, achieving on-site precise identification, error prevention interlocking, and digital intelligent verification of the on/off status of the secondary hard plate, completely eliminating problems such as false on/off, missed on/off, and effectively improving the verification efficiency of the secondary protection plate in substations to ensure the correctness of the on/off operation.

Distribution network transformers have the function of controlling power equipment, but faults and anomalies often occur during practical applications. Therefore, a fault diagnosis method for distribution network transformers based on artificial intelligence is designed. Select characteristic fault diagnosis parameters, design an intelligent closed-loop transfer fault diagnosis model, and use cost matrix to complete the fault diagnosis of intelligent transformers. The final test results show that compared to the traditional AdaBoost sensitive fault diagnosis test group and the traditional parallel fault diagnosis test group, the maximum absolute difference obtained by the artificial intelligence fault diagnosis test group designed in this article is relatively small, indicating that the application fault diagnosis error of this method has been greatly controlled, with better results, and has practical application value.

According to the motion characteristics of the six-axis manipulator, the D-H method is used to create the coordinate system of each joint, obtain its parameters, and derive the forward and inverse kinematics equations. The simulation model of the manipulator is created by using the MATALB Robotics toolbox, and its forward and inverse motion and trajectory planning are simulated. Through simulation, under the joint space planning, the relationship diagram of each joint parameter with time during the movement of the manipulator model is obtained. Under the Cartesian space planning, the linear motion trajectory of the end of the manipulator is obtained. By comparing the theoretical data and simulation results, the correctness of the manipulator model and the accuracy of the forward and inverse motion formulas are verified.

At present, most of the spectral detection methods with plant diseases are based on machine learning, we present a novel approach to spectral analysis using a spectral feature pseudo-graph-based residual network, referred to as SFPGRN, for data in the near-infrared band. The proposed method involves constructing a residual network model using a characteristic surface obtained via natural neighborhood interpolation based on preprocessed near-infrared spectral reflection signal and first-order differential spectral index. The input layer is a pseudo spectral map that is generated through surface projection of the characteristic surface. The generated spectral pseudo image contains a diverse set of spectral features that augment the representation capability of the pseudo-image while retaining the spectral band reflectance information. We evaluated the proposed method on the 2015 spectral dataset of apple leaf diseases and insect pests in China and found that it outperforms traditional machine learning algorithms, achieving a classification accuracy of 93.21%.

In order to solve the problem of mission success degradation caused by the obstructed communication of UAVs in the communication uncertain environment, we propose a semantic communication reinforcement learning for UAVs intelligent following method for maintaining UAV communication in the scenario with high interferences to keep following the leader UAV. This method employs a semantic communication model and a leader behavior learning model as the follower intelligence so that the follower can resist communication interference without deploying additional communication anti-jamming devices. The follower using this method continuously follows the leader to improve the probability to complete the task in a communication-restricted environment. In experimental studies, the proposed method effectively improves the task success rate in the event of communication interruptions, which significantly outperforms benchmark methods in complex environments.

Pathfinding algorithm has a decisive impact on unmanned aerial vehicle flight. In practical applications, aerial vehicle is often required to respond to emergencies quickly and plan a new path in time. Therefore, we designed a pathfinding optimization algorithm, combined with the object detection algorithm based on deep learning network, to identify objects that suddenly appear on the original path, timely avoid and re-plan the path. Experiments show that the proposed method can re-plan the new safe path in a short time which satisfies the requirements in response time and path planning.

Semantic segmentation technology is an important research direction in the field of computer vision, with the continuous improvement of computer GPU and other hardware arithmetic power, as well as the more clear and accurate remote sensing images. In recent years, deep learning technology has developed rapidly in the field of semantic segmentation, and the level of semantic segmentation technology has improved significantly. Based on unsupervised, automated, end-to-end deep learning networks have gradually become the object of research in scene object segmentation, pre-person background segmentation, face recognition and other related fields. In this paper, we conduct research on deep learning semantic segmentation algorithms for water body extraction, focusing on specific analysis of typical problems such as long freezing period of lakes in cold and arid regions, low precipitation, complex seasonal changes of lakes, and fragile water bodies. The cold and arid zone lakes distributed in Inner Mongolia region of China are selected as the research objects, and the water body extraction experiments are conducted on the remote sensing images of Hulun Lake and Wuoliangsuhai Lake. Propose a SER34AUnet model, the model was improved based on the Unet network model of Encode-Decode architecture, and finally achieved good results in the water extraction experiments. The main work of this paper is as follows: (1) Establishing image segmentation dataset: remote sensing images of several lakes in cold and arid regions in different seasons, different satellites and various remote sensing images with specific representative features typical of cold and arid regions are used to enhance the generalization ability of the dataset. (2) Improving the Unet model: Firstly, the Resnet34 network, which is more capable of extracting image features, is used in the Encode part. Secondly, the SE (Queeze and-Excitation) module based on the channel attention mechanism is added to the BasicBlock module inside the Resnet34 network to obtain the SEResnet34 network. Finally, the Dual Attention Network (DANet) is added outside the SEResnet34 network to enable the model to adaptively enhance the perception of water body parts. In the Decode section, upsampling is performed by bilinear interpolation and fused with the feature extraction layer layer by layer to finally obtain the segmented image for water body extraction. (3) SER34AUnet model training: Firstly, the established dataset is divided into training and test sets in the ratio of 8:2. Next, the FCN_Resnet50, DeeplabV3_Resnet50, Unet semantic segmentation network model and the improved model in this paper are trained and tested. Finally, the visual and numerical analyses of the two lake water segmentation results and the three evaluation indexes of the test set, respectively, demonstrate the good applicability of the method proposed in this paper for lake water extraction in cold and arid regions.

As people pay more and more attention to privacy protection, federated learning with privacy protection ability comes into being, and gradually becomes one of the research hotspots in the field of machine learning. As a distributed machine learning framework, federated learning can prevent data from being directly leaked, but some private information can still be derived from the gradient. Therefore, some cryptographic schemes are used to solve this privacy leakage. In this paper, Paillier algorithm in the semi-homomorphic encryption scheme is introduced to encrypt the gradient uploaded by clients in federated learning, which is aggregated by the server and sent back to each client for model update. Relevant experiments are also designed to verify the efficiency of Paillier algorithm through the change of the computing cost with the increase of the number of customers.

The representation learning of knowledge graph aims to represent the semantic information of entities and relations as dense low-dimensional real-valued vectors and map them to the same low-dimensional space. Existing methods often focus on the single-modal information of the text and ignore the information of the image modality, resulting in the ineffective use of entity feature information in the image. And there is entity-related descriptive information in most knowledge graphs, which are not well used in current multimodal knowledge representation learning methods. In this regard, a multimodal knowledge representation learning method combining description information is proposed. This method combines multimodal (image, text) data to construct a knowledge representation learning model and combines its corresponding brief description information to improve the representation effect of multimodal data. Experimental results show that the method performs well on triple classification and link prediction tasks on the constructed WI-D dataset.

Hardness recognition of objects can improve the grasping ability of robots. Controlled robots can obtain useful information in complex environments and have wide applications in industrial fields and special material measurement. Existing hardness recognition networks usually only consider tactile or visual information, for many fields where it is not possible to distinguish between areas that cannot be viewed directly or collected pressure information data and cannot be widely used. To address such issues, this paper proposes a multi-time scale convolutional neural network based on visual tactile fusion network (MTSCNN-VTF), which consists of two parts: Multi Time Scale Resnet Network (MTSRN) and Separable Multi Time Scale Resnet Network (SMTSRN) and extracts spatiotemporal feature vectors of data through a two-part network, The fused feature vector MTSCNN-VTF network is used to identify the hardness level of an object. The ablation study conducted on the STAG dataset in this article has verified the rationality of our method design. The comprehensive comparison with other advanced methods proves the effectiveness of the MTSCNN-VTF model.

This paper takes the autonomous landing of four-rotor UAV as the specific research object. Aiming at the problem of identifying the feature points of landing signs and improving the accuracy of autonomous landing, an improved multi-objective soft subspace clustering algorithm is proposed. The anti-redundant mutation operator and forward comparison operation are designed, it is proposed to improve the population diversity and convergence speed of the algorithm. The iNSGA-II is used as the base algorithm, and the repair operator and local search operator are designed to reflect the clustering characteristics. An autonomous landing control model based on visual recognition is established. A large number of position feature points are clustered and analyzed to decouple the highly nonlinear mapping relationship in the feature space, and then the redundant information between the dimensions is stripped to achieve the maximum stability of the sampling points, the importance sequence of all features and the accurate classification.

Aiming at the problem of low reliability of ship model recognition from remote sensing images, we propose a ship model recognition method based on multi-view learning. Firstly, the multi-view feature data is constructed by different feature operators. Then the single feature view data is used to train SVM classifier respectively while the multi-view data is fused and used to train classifier by CPM-Nets. Finally, we fuse the results of single feature view classifier and multi-feature view classifier by classifier aggregation to enhance the accuracy rate. Experiments show that the proposed method can improve the accuracy of ship type recognition.

Machine learning algorithms advance the classification of cultural relics in recent studies. A part of the studies adopts image processing methods as a solution. We believe that the blurring of the image limits this approach. On the other hand, we notice the studies conducted from the perspective of chemical composition, which avoids the problem of blurring images and shows interpretability. In this paper, we follow the direction of classifying by chemical composition and aim to provide our method with better interpretability. Hence, we propose a strategy that combined both supervised and weakly supervised methods as the solution. It visualizes the decision tree of classifying the main classes of cultural relics at first and extracts the chemical composition of the subclasses with the pseudo-labeling according to limited subclass labels next. Besides interpretability, it also performs accurately in subclasses classification by using the radial basis function network and random forest. Our results indicate that our weakly supervised method achieves the closest reduction to a finer classification.

In the method of sleep staging task, the artificial sleep staging process is complicated, while the traditional machine learning only realizes feature classification and relies on manual feature extraction. This paper presents a deep learning method which can automatically extract and classify EEG features. At the same time, in view of the slow improvement of staging accuracy and high model complexity of existing studies, and the high cost of data due to the dependence on annotated data sets, this paper constructs the end-to-end self-supervised sleep staging model for these problems and continuously improves it. An automatic sleep staging model based on Inception structure and BILSTM was proposed. The convolutional neural network of Inception structure extracts the corresponding features from different time scales, combines the application of attention mechanism and residual structure to carry out the EEG signal characterization learning, and uses BILSTM to carry out the sequence information feature learning. Finally, a self-supervised learning method based on comparison is used to learn the potential feature representation of EEG data from a large number of unlabeled data. In the experiment, the sleep-EDF dataset was used, and the results showed that self-supervision improved the overall classification accuracy of the model. Meanwhile, the classification results of various categories were equalized, and the overall performance of the Sleep staging model was improved.

IoT data mining is an important link in IoT technology. Traditional data mining can no longer meet the requirements of IoT technology, and the emergence of cloud computing, with its powerful computing power and ability to store large-scale data, has provided a new direction for IoT data mining. In such a context, a cloud based IoT data mining technology has emerged, which has attracted wide attention worldwide. In this paper, we study the K-means algorithm in data mining to solve the problems of K-means algorithm such as unstable clustering and inability to handle large-scale IoT data. This paper mainly. The experiments show that the clustering effect of ALCDK-means algorithm is significantly improved, and the accuracy rate is greatly improved compared with K-means algorithm. Finally, the shortcomings of ALCDK-means algorithm, such as the inability to effectively handle large-scale data sets, are addressed by parallelizing the algorithm based on the Hadoop platform to solve the problem that traditional data mining cannot handle large-scale data. The experimental results show that the parallelization of ALCDK-means algorithm solves the problems of traditional data mining such as long time consuming and inefficient to a certain extent, which in turn is beneficial for large-scale data mining.

High Performance Computing (HPC) system has excellent computing power. The router is at the heart of a high-performance interconnection network, which is an important building block of HPC systems. Router architecture is classified into three types: crossbar, hierarchical, and network structures. The crossbar is the fairest, and the most basic component of the router, but it is also the most difficult to extend. Hierarchical structure using the hierarchical scheduling algorithm reduces arbitration complexity. According to the idea of network design, multiple crossbars are connected in a specific topology using integrated circuit technology implemented on a single circuit chip. This paper analyses the characteristics of the three structures and related studies, which have guiding significance for the subsequent high-radix router design. This work also demonstrates the performance of various router architectures in terms of channel throughput and latency.

In order to promote the work of "reduction and control" in public security traffic management, and to clarify the temporal and spatial evolution characteristics of traffic fatalities, this paper is based on the data of road traffic fatalities from 2019 to 2021, with the help of the AutoNavi map open platform to complete the spatial location of accidents. Using the improved density clustering algorithm, the clustering analysis was carried out on the key illegal behavior accidents, using geographic information analysis technology to analyze the spatial distribution characteristics and temporal evolution characteristics of fatal traffic accidents caused by three types of key violations. In the regional distribution of fatal accidents of drunk driving and unlicensed driving, the situation is mainly stable and improving, and the fatal accidents of failing to give way according to regulations are mostly deteriorating; in addition, the construction of the new city in District VII and the construction of the sub-center of the city in District VIII have been carried out steadily, the traffic demand has become increasingly strong, and many types of illegal activities and accidents have shown a worsening trend, and XI and IV districts are mainly stabilizing and improving. The results of the study are helpful for the traffic police department to implement refined management of regional traffic accidents and make corresponding preventive measures.

In order to improve the performance of traditional SOM model, an adaptive SOM neural network algorithm is proposed. By introducing the error square sum solution formula, the clustering error of the model sub-data set is calculated by using the method of cutting the data set, and the change trend of the clustering error is quickly obtained. Then the learning rate of the SOM network is adaptively adjusted according to this, and the convergence speed is accelerated on the basis of ensuring that the network is fully trained, thereby improving the performance of the SOM network.

This paper mainly uses the Python language and the pyecharts library to mine the earthquake data that occurred around the world in 2020, and finally displays it with pie charts, bar charts and other graphics. It lays a foundation for better use and processing of earthquake big data in the future, and also demonstrates the application of big data in other industries.

Complete meteorological data is essential for meteorological research. However, due to sensor failure or occlusion, data loss always occurs. In order to deal with this problem, geoscience often uses Kriging and other traditional interpolation methods. Nevertheless, because the traditional method only considers the existing data, it cannot provide accurate reconstruction results when the spatial missing rate is large. Inspired by image inpainting algorithms, we propose a new deep learning method that combines other variables related to missing meteorological variables to provide more effective information for missing regions and proves the good effect of the method through a series of experiments. To illustrate that our method can maintain good performance under different missing rates, we use the reanalysis data of the European Centre for Medium-Range Weather Forecasts to create a data set with three different missing rates of 30%, 50%, and 70% through the artificial mask. Under different missing rates, the average RMSE of the method based on deep learning decreased by 77% and 31% respectively compared with Kriging and DINEOF. In addition, by combining relevant variables, the accuracy of the deep learning model can be further improved by about 16%.

The Z-FFR (Z-Pinch driven Fusion Fission hybrid Reactor) contains two physical processes, nuclear fusion and nuclear fission, and has a complex structure itself. Using artificial intelligence and big data technology to construct the digital Z-FFR, a decision decomposition method for writing the source code of the digital Z-FFR (Z-Pinch driven fusion fission hybrid reactor) by an artificial intelligence programmer is also proposed, including the following steps: using big data to build a normalized description of the digital Z-FFR; based on the normalized description of the digital Z- FFR, a dimensional decomposition method is used to split the digital Z-FFR modeling, digital Z-FFR simulation and digital Z-FFR writing structure to obtain the decision splitting set of digital Z-FFR. The decision selection method is determined according to the digital Z-FFR decision splitting set and the digital Z-FFR source code writing is completed. The decision splitting method for writing digital Z-FFR source code with artificial intelligence and big data proposed in this paper decomposes the writing logic of digital Z-FFR and uses different artificial intelligence decision methods to complete the writing of digital Z-FFR source code according to different writing logics, which overcomes the disadvantages of long development cycle, repetitive development workload and high learning cost of various existing simulation systems.

Convolutional Neural Network (CNN) has always been a hot topic in deep learning. With the increasing demand for network models in daily production, the optimization of convolution calculation process is very important. This paper starts with the process of back propagation in convolutional neural network, introduces the derivation of convolutional neural network back propagation and the conversion process of im2col, uses implicit to convert the calculation of convolution on the domestic acceleration platform, and optimizes the convolution back propagation operator through a variety of general matrix multiplication optimization strategies. The final performance reaches more than 70% of the performance of NVIDIA operator, which meets the expectation of the experiment under the performance bottleneck of the platform.

The progress of scientific development requires the use of high-performance computers for large-scale simulations, resulting in a significant communication overhead and thereby constraining the computer's performance. To address this issue, optimizing the topological mapping of application processes to computing nodes is essential for enhancing the communication performance of high-performance computers. However, this topic has not been extensively explored in the literature. In order to reduce the communication overhead of high-performance applications, this study formulates the optimization of topological mapping from application processes to computing nodes as a quadratic allocation problem. The proposed method collects communication features to assess the communication intimacy between processes and considers the communication relationship between application processes and network topology. To overcome the limitations of traditional genetic algorithms, this study introduces elite learning and adaptive selection into the mutation operator. In this algorithm, individuals undergoing mutation learn from fragments of the best individuals in the current population. Additionally, three functions are selected to control the probability of selecting the elite learning mutation during the mutation process, thereby enhancing the algorithm's efficiency and accuracy. The results of the experiments demonstrate that the suggested methodology yields a noteworthy enhancement in communication performance compared to the widely adopted round-robin approach in NPB test suites. Furthermore, the enhanced genetic algorithm displays superior optimization efficiency in comparison to conventional genetic algorithms and other heuristic approaches.

With the dense development of power network and the progress of flexible transmission technology, the problem of short-circuit current overshoots are beginning to emerge. In this paper, the development trend and limitation measures of short-circuit current in domestic and foreign power grids are analyzed, and the short-circuit current levels in different development stages are summarized. Furthermore, the factors that affect the short-circuit current level are analyzed, and a calculation method of short-circuit current with flexible DC is proposed. Finally, new types of short-circuit current limiting devices that limit the level of short-circuit current are introduced.

With the improvement of comfort requirements in noise environment, the research on vibration noise has developed from noise reduction to sound quality optimization. Under the premise of combining subjective and objective evaluation, the method of sound quality evaluation is constantly improved. This paper carries out the test design of subjective evaluation of sound quality, collects and screens sound samples, and conducts the subjective evaluation test of sound quality using the grade-based scorecard method. Subsequently, establish subjective and objective evaluation model of sound quality, and put forward Discrete Firefly Algorithm and Back Propagation neural network method (DFA-BP) sound quality evaluation model based on Back Propagation neural network method (BP) sound quality evaluation model. Compared with BP sound quality evaluation model, the relative error of prediction accuracy is within 10%, which provides a new method for further improving the subjective and objective evaluation model of sound quality.

This paper introduces a robust decentralized regulation approach of linear interconnected time-varying systems. If the optimal optimization problems of time-varying systems are considered, the solution of time-varying Riccati differential equation should be involved and pre-calculated. The weighting parameters of the optimal performance index in this paper are all time varying. The optimal solution relies on the initial conditions of the solutions of Riccati differential equations, and these initial conditions are derived by the backward Euler’s method. Then the obtained optimal control gain is modified to accommodate the interactions among the subsystems. The overall system is mathematically proved to be exponentially stable with a prescribed degree of stability. Computer simulations of a time-varying example composed of three subsystems are conducted to demonstrate the feasibility of this approach.

In the field of electricity safety inspection, with the increase of the scale and complexity of the business process, the business process has dynamic changes. But the data model is not adapted to the dynamic changes in the business process. Therefore, in the business process of concurrent access to the database, it is easy to lead to data access conflict, reduce database concurrent access efficiency, and lead to increased business development costs. At present, there are many research results in business processes and database optimization, but there are still some mature and perfect optimization of the theory of databases, which can consider the optimization of business processes and business process data models. The data model optimization algorithm proposed in this article can be used to reduce database access conflicts of concurrent business processes, so that the data model can adapt to the dynamic changes in the business process, thereby improving the efficiency of concurrent access to the database.

Traditional video keyframe retrieval algorithms have certain limitations in terms of retrieval accuracy and user experience, such as the inability to accurately capture the semantic information of the video and weak correlation between keyframes. To address these issues, this paper proposes a video keyframe retrieval algorithm based on CLIP technology. The algorithm extracts video keyframes using a clustering method based on SIFT features and uses a pre-trained CLIP model to extract both visual and semantic features of the keyframes, achieving more accurate retrieval. Additionally, the algorithm uses a bidirectional LSTM model to model inter-frame contexts, enhancing the correlation between keyframes and improving retrieval accuracy. Experimental results show that the proposed algorithm outperforms traditional algorithms in terms of retrieval accuracy and user experience.

The existing social network matching algorithms have problems in processing text attribute information, as they cannot handle polysemy issues of word meanings well and cannot effectively extract deep semantic information from the text. To address the aforementioned issues, this paper introduces pretrained models to extract information from attribute text and proposes a social network alignment algorithm called NABP (Network Alignment based on Pre-trained Language Models). NABP first preprocesses the attribute text data to extract the text feature vectors; then, it uses the DeepWalk algorithm to perform graph representation learning on the two networks to extract their structural features. Next, the attribute feature vectors, and structural feature vectors are concatenated to form the final feature vector. Then, a projection function is learned using supervised information to further improve the algorithm's performance. To verify the effectiveness of NABP algorithm, this study conducted experiments on two real social network datasets. The results show that the NABP algorithm performs better than existing algorithms in both the Accuracy and Precision@k metrics, demonstrating the effectiveness of the algorithm and pre-trained language models.

Non-motor vehicle detection is an important part of the road information collection task in autonomous driving systems. To address the problems of small target omission and dense target misdetection of existing non-motor vehicle detection models in complex traffic scenarios, A non-motor vehicle detection algorithm based on improved YOLOv5s is proposed in the paper. It mainly includes adding an efficient channel attention mechanism to the feature extraction network of YOLOv5s, using a weighted bidirectional feature pyramid network for feature fusion, and improving the loss function and non-maximum suppression algorithm. The experimental results show that the precision, recall and mean average precision of the improved algorithm for non-motor vehicle target detection in complex traffic scenes reach 0.935, 0.883 and 0.949, respectively. Compared with the YOLOv5s algorithm, the precision is increased by 3.4 percentage points, the recall is increased by 2.6 percentage points, and the mean average precision is increased by 3.8 percentage points. Compared with other mainstream target detection algorithms, the improved algorithm in this paper has fewer parameters, less computation and higher detection accuracy. The weight of the improved algorithm is only 13.8MB, and the number of frames per second can reach 90. While maintaining lightweight, the improved algorithm shows good performance in terms of speed and accuracy.

As the basic component of the power system, the prediction of power transmission line icing plays an important role in the power security protection. As a key factor affecting transmission line icing, cooling is widely introduced into the prediction model of power transmission line icing. Accurate understanding of the nonlinear characteristics between cooling and power transmission line icing has become the key to improving the level of transmission line icing prediction. At present, there is a lack of systematic and comprehensive analysis of the relationship between regional cooling and power transmission line icing in China. In this paper, a nonlinear analysis method of cooling and power transmission line icing is proposed, and a nonlinear analysis system is designed. In the specific analysis, the meteorological transmission line icing is first separated from the original transmission line icing, and then the maximum information coefficient is used to measure the linear or nonlinear relationship between different regions, different times, different cooling and meteorological transmission line icing. Finally, the efficiency and practicability of the proposed system are proved through the transmission line icing prediction experiment. The analysis results of the system can play a basic guiding role for the relevant research of transmission line icing prediction.

In order to solve the problem that the existing helmet detection algorithms have low accuracy in detecting dense small targets in a complex environment, helmet wearing detection algorithm based on improved YOLOv5 is proposed. Firstly, the original CSP structure is replaced by a new redesigned DB_CSP module to improve the ability of backbone network feature extraction. Secondly, the coordinate attention mechanism is added in the backbone feature network layer to make the model pay more attention to small targets. Then, Bidirectional Feature Pyramid Network (BiFPN) is used to replace PANet to enhance the ability of feature fusion. Finally, a special detection layer for small targets is added to form four scale detection. The experimental results show that the mAP of this model is 92.28%, which is 4.9% higher than that of YOLOv5s model. The average detection time is 0.0276 seconds, which improves the detection accuracy while maintaining the model speed.

With the rise of new technologies, people hope to have a better experience and feeling in their daily life, and the technology and application of various industries are developing towards the direction of digitalization, networking and intelligence. The digital reform of colleges and universities has become an important means of the informatization development of colleges and universities. Its significance lies in improving the teaching quality, strengthening the teaching management norms, optimizing the construction of teaching resources, etc. In this paper, a class behavior recognition algorithm based on two-stream convolutional neural network is studied. The purpose of the research is to analyze and evaluate the teaching quality of the classroom through the identification and statistics of students' behaviors, so as to improve the teaching methods according to the feedback results, and finally achieve the purpose of improving the teaching effect.

This paper implements Timing-Go, a distributed timed task scheduling system based on a loosely coupled architecture, which solves the problems of poor manageability, poor parallel processing capability, poor task scheduling capability and poor availability of existing timed task scheduling schemes. The system implements the automatic task registration mechanism through Java annotation technology, designs the task visualization management platform, and improves the efficiency of task management; proposes a timing task scheduling trigger scheme based on Quartz core elements, realizes the decoupling of scheduling process and execution process, and improves the performance of timing task scheduling; extends the functions of Quartz scheduling framework, introduces the workflow hierarchical processing mechanism of DAG model and the large task parallel processing mechanism based on consistency Hash algorithm, and improves the execution efficiency of timing tasks. Meanwhile, this paper also proposes a distributed synchronization control mechanism and a series of compensation mechanisms to guarantee the high availability of system operation and task execution. The Timing-Go system implemented in this paper has good performance and scalability, and its effectiveness has been verified by testing.

UAV inspection of overhead transmission lines is a common means for power grid companies to conduct defect detection, but in bad weather, the performance of the target detection algorithm deployed on the onboard equipment often decreases significantly. Based on the classical atmospheric light scattering model, the atmospheric light estimation is first done more accurately by separating the sky area. In terms of transmission estimation, the synthetic transmission is estimated from comprehensive information loss and image texture information weighting, and finally the color bias is corrected to realize the target image fog processing. The experimental results show that the proposed algorithm can significantly improve the clarity of transmission line images in foggy environment.

In this paper, we propose a reconfigurable framework optimized for resource-constrained platforms to accelerate CNNs using the high concurrency and data-proximate characteristics of edge computing devices. The framework is designed from three aspects: control flow, data flow, and storage flow. To address the impact of memory access cost on network computation efficiency, we introduce a parallel ping-pong data scheduling framework to compensate for it. Experimental results show that the system can support convolutional kernel operations of any size under structural constraints, and the computational efficiency is improved 2.28 times compared to CNN acceleration systems based on traditional data scheduling schemes.

With the increasing scale of wind power, the impact of wind fluctuations on the power network must be considered. Considering the randomness of wind have a serious impact on power grid, we use Gaussian mixture function method to approximate the wind power output error probability density distribution. Specifically, we use Davidson-Boding Index (DBI) to determine the optimal number of cluster center. Then, the fastest gradient descent method is employed to solve the weights of Gaussian function. Finally, the simulation results show that the Gaussian mixture function can effectively approximate the wind power output error probability density distribution.

In order to solve the problem of semantic segmentation difficulties of city street scene pictures due to uneven color and strong light changes, this paper proposes a semantic segmentation algorithm of city street scene pictures based on DeepLabV3+ architecture with dark channel a priori theory. The algorithm firstly inverts the city streetscape image and then performs color balancing through the dark channel a priori module, then extracts contextual information hierarchically using a multi-channel parallel network, fuses low-level features with high-level features hierarchically to obtain the optimized feature map, secondly completes the fusion of multi-scale features through the spatial pyramid structure, and finally fuses the obtained feature map with the decoder twice to generate the final prediction results. The experimental results show that the algorithm in this paper outperforms the original DeeplabV3+ model in terms of subjective perception and objective indicators in urban scenes, and provides a new solution and technical idea for semantic segmentation of urban scenes.

The classified collection of urban waste by waste collection vehicles involves one-way and two-way roads. At the same time, the capacity of collection vehicles is limited. Therefore, the waste sorting and collection problem is modeled as a mixed capacitated arc routing problem (MCARP), hybrid refers to the presence of edge and arc elements, which approximates a real road network and is solved using the Epsilon constraint method. Relative to the existing literature, the model also considers the classified collection of different types of waste and the fairness of the workload of collection workers, and a Bi-objective MCARP model is developed and solved using the Epsilon constraint algorithm method. Finally, a set of cases of different scale is designed to be solved in numerical experiments to determine the solution scale of the model.

As a gradient-guided search method, differentiable architecture search greatly reduces computational costs and improves search speed compared with traditional reinforcement learning methods and evolutionary methods that search for network structures in discrete spaces. However, when the number of search epochs is too large, the searched architecture will contain a lot of skip connections, resulting in a sharp decline in network performance. Aiming at this phenomenon, this paper designs a phased search process. With the deepening of the search stage, different early stopping rules are designed, so that the units located in different positions of the network can present different structures, effectively solving the performance crash problem caused by skip connections. At the same time, the design of edge normalization is introduced on the connection between nodes, and the difference between the weight parameters of different operations is enlarged by improving the loss function, which effectively improves the stability of the architecture search. The experimental results show that this method trades off a small amount of parameters and search time in exchange for an improvement in accuracy, and the verification accuracy on the cifar10 and cifar100 datasets has increased by 0.15% and 1.36%, respectively.

Support Vector Machine (SVM) is an excellent Machine Learning (ML) method and Kernel Function (KF) in SVM can effectively accelerate the calculation and affect the results in ML. Therefore, the selection of KF has become a research hotspot. In small-sample medical detection learning tasks, there are often imbalanced data categories and data overlaps. To alleviate these situations, we adopt feature selection in the data preprocessing stage and optimize the support vector machine learner with fusion kernel function. In this paper, SVM based on Fusion KF(FKF-SVM) is used for Parkinson's Disease (PD) detection to improve the performance of traditional SVM. The improved KF used in the paper includes Analysis of Variance (ANOVA) kernel, triangular kernel, log kernel and wavelet kernel. The proposed KF is evaluated by using two popular and publicly available data sets from University of California Irvine (UCI) repository with different parameters like accuracy, recall, F1score, Area Under the Curve (AUC) in classification and Mean Square Error (MSE), Mean Absolute Error (MAE), R2score, Explained Variance Score (EVS) in regression. According to the experimental results, the triangular kernel is the winner in the classification that achieves an accuracy of 93%, recall of 96%, F1score of 96% and AUC of 97%, which is increased by ten percentage points at least compared to the traditional KF. In the regression, the ANOVA kernel is a superior KF which obtains an MAE of 2.3, R2score of 0.81 that is 200% lower than the traditional KF at least.

To solve the problem of video repair, we propose a new optical flow guidance solution that uses parallel structured convolution and attention networks to jointly infer video missing regions. In the network, a parallel structural model based on convolution and attention networks guided by optical flow is used to extract feature information to integrate the spatial and temporal context of video frames. This method integrates information between the target frame and the reference frame. To enhance feature learning capabilities using convolution and attention mechanisms, the feature fusion module fuses local and global features in an interactive manner, maximizing the retention of local and global representations. Our model produces visually satisfactory and time consistent results, while demonstrating on two benchmark datasets that our method outperforms the most advanced methods in terms of quantity and user research.

Maneuvering target tracking technology is widely used in military scenarios. To solve the hysteresis problem of maneuvering target motion model transition in interactive multiple-model technology, an adaptive interactive multiple-model algorithm with time-varying Markov transition matrix is proposed. The proposed method constructs time-varying Markov transition matrix based on the differential between the motion model probabilities at adjacent time steps. Meanwhile, it is combined with particle filter to get a better estimation accuracy for maneuvering target tracking. The proposed algorithm is verified in experimental simulation compared with traditional interactive multiple-model algorithm, and it is shown that average tracking error decreases by 18.5% and the probability of matched motion model of maneuvering target can increase rapidly when the target maneuvers.

This paper proposes a ship detection method based on contour extraction networks and minimum bounding box generation algorithms. The proposed method utilizes contour extraction networks to extract the contour of the ship target and generate a set of key points. Then, the minimum bounding box generation algorithm is used to transform these key points into the rotational detection box of the target, which can more accurately describe the shape and orientation of ships, thus improving the accuracy and reliability of detection. The experimental results show that the proposed method can effectively detect ships of various orientations and postures and has better performance than traditional rectangular box detection methods. The paper compares rotated bounding box detection methods based on traditional machine learning and deep learning and concludes that rotated bounding box detection methods based on deep learning can better solve the problem of ship detection in complex situations, but require a large amount of data for training and may require a higher computational cost. Finally, the paper introduces the Minimum Bounding Rectangle Algorithm used in the instance segmentation process.

The transformer has revolutionized Natural Language Processing and its exceptional performance has extended its application to computer vision. However, as more complex tasks arise, the demand for models to process longer sequences has increased. To address this, efficient transformers with reduced time and memory complexities have been developed. Linear transformer is an outstanding one of them with linear time complexity with input length. Causal product is the core computation of causal linear attention in linear transformer, which are commonly used to perform autoregressive tasks. However, in order to preserve the linearity, the causal products of queries are computed sequentially in Linear Transformer. We propose an innovative CUDA kernel function which allows for parallel computation of the query attention instead of sequential. Furthermore, we optimized the GPU computations of linear attention in Linear Transformer by double pipeline that decreases the time of data transmission. Our experiments demonstrate that our method achieves a 9x speedup compared to the original causal product function when implemented on a CPU.

Data link network is the foundation and key to support the efficient aggregation and sharing of various resources in the battlefield. How to flexibly organize and use the limited network resources according to the task requirements in the future strong countermeasure communication environment is one of the problems faced by the data link. In this paper, a flat network structure is designed according to the flexible and adaptive task requirements of unmanned cluster and the characteristics of invulnerable and reconfigurable network. The key to the operation and maintenance of the flat network structure is the topology management technology. Task oriented management of the flat network topology can effectively avoid network degradation and improve network robustness and survivability while meeting the task requirements.

When ships navigate in the island area, they often face problems such as high danger factor, difficult navigation and complicated route planning. Based on the above problems, we propose to adopt the improved adaptive genetic algorithm after rasterizing the map to select a route that considers safety, efficiency and smooth road strength for drivers' navigation. The environment of ship island area navigation is simulated in the raster map, and a reasonable path is planned by applying the improved adaptive genetic algorithm. The results show that the path is shorter, smoother and with fewer iterations than the traditional genetic algorithm. With the continuous development of automatic control technology, it can provide theoretical support for the future unmanned ship navigation in the island area.

To improve capitation accuracy by residual network and multiscale image training, there are two methods to detect caphead characteristics. Previously based on pre-described head characteristics, and methods based on the statistical training of the models, the latter is very robust, using a neural network to train, a better effect can be obtained. Based on the presence of overlapping heads in dense multi-human images and head features with varying near and far scales, this requires more special treatment for the training sample of character avatar. For problems that cannot be detected in real time, using an alternative method of YOLO V2 detection can remedy the defects and help obtain the relevant application model.

As for current network intrusion detection technology, it has become critical to improve the problems of low detection efficiency and low accuracy caused by feature redundancy. To improve this problem, we propose a multi-strategy Improved Tuna Swarm Optimization (ITSO) algorithm and apply it to feature selection for network intrusion detection. First, an elite backward learning strategy is used to initialize the population and improve the quality of the solution in the initial stage to lay the foundation for the global optimum. Then, an improved sine cosine strategy is added to further optimize and balance the ability of local exploitation and global search before the tuna swarm starts feeding. Finally, the Levy flight-Cauchy variation strategy was added to the spiral foraging and parabolic foraging behavior of tuna schools to increase the diversity of group locations and improve the global search capability of the algorithm in network intrusion detection feature selection. The ITSO algorithm was tested against the original TSO algorithm and other classical algorithms in six benchmark functions. In terms of accuracy, it improved by 2.95% on the UNSW-NB15 dataset and 4.1% on the NSL-KDD dataset compared to the original TSO algorithm. Intrusion detection time is also greatly reduced. This effectively improves the performance of network intrusion detection.

With the advancement of deep learning and the growth of AI, a new concept for classifying music score images has been introduced. This paper presents a classification method for guitar tabs and numbered musical notation based on the ResNet50 network. The ResNet50 residual network model is used to extract features from music score images, and then the classification is performed based on the recognition probability. Results of the experiments indicate that the recognition accuracy of this model can reach 99.154%, suggesting that the ResNet50 network-based classification method holds good research potential in music score image classification.

Gas Chromatography-Ion Mobility Spectrometry (GC-IMS) was applied to identify rice varieties and adulteration in order to address the inefficiencies caused by the dependence on intricate biochemical procedures for the identification of rice species and adulteration. The original migration profile data for the identification of rice varieties and adulteration were provided by testing the volatile flavor substances of five kinds of rice. The original data was compressed through Principal Component Analysis (PCA), which was then used as the classifier's input. An improved Multi-classification Maximum-interval Twin Support Vector Machine (MMTSVM) was proposed to construct a multi-classifier with increased classification efficiency and accuracy compared to the conventional support vector machine. The experimental results on the testing set demonstrated that the accuracy of rice variety identification was close to 94.00%, which was obviously better than that of conventional support vector machines, decision trees, and other models. At the same time, the experimental results also showed that the accuracy of the GC-IMS method in identifying adulterated rice reached 93.3%, outperforming the traditional chromatographic or spectral analysis methods.

Action recognition is a cutting-edge research direction in the field of computer vision, widely used in video surveillance, video retrieval, human-computer interaction, and other fields. However, existing action recognition algorithms improve prediction accuracy at the cost of having too many model parameters and high computational complexity, which limits their practical application. In order to address this issue and improve recognition accuracy with higher computational efficiency, we propose a dense residual 3D depth-separable convolutional network model based on channel attention mechanism. By extending the depth separable convolution to 3D, the model can maintain high accuracy while greatly reducing the number of model parameters and operational complexity. The dense residual structure is introduced to enhance the training ability of the network, improve feature extraction, and further improve network performance by learning the correlation between channels through the channel attention mechanism. Compared with other action recognition algorithms on two datasets, NTU RGB-D and Fall Dataset Lei2, the proposed action recognition algorithm has better robustness under lightweight condition, with 415.9 MFLOPs of computation, 205.4k of parameters, and 90.5% of CS, 95.5% of CV and 95.97% of accuracy, respectively. The proposed algorithm has a high recognition accuracy and greatly reduces the calculation cost, which can provide better performance for practical applications.

Aiming at the problems of missed detection, false detection and rough edge recognition in the detection of water body changes in high-resolution remote sensing images of the same area before and after, an improved network model is proposed to better improve the extraction accuracy of water body change areas. Combining deep learning with machine vision we introduce the transformer block into the network. We design a Siamese change detection network with stronger contextual change semantic feature extraction capability SCTNet (Siam-CNN-transformer-Network), transformer block directly connects information at any location through a multi-head intention mechanism efficiently compute long sequence information mining long-distance dependencies. As the front-end of the network, the CNN model uses convolution to focus on the correlation between two-bit local data, ensuring the correlation between adjacent pixels to enhance the local modeling ability of the model and reduce the model's dependence on the amount of data. The proposed model performs well on the water body change dataset made from the WFV image data of GF-1 satellite. The experimental accuracy rate reaches 97.4%, F_1-score and mIoU 97.9% and 90.9% respectively. Compared with other change detection models, all indicators have been significantly improved, the situation of false detection and missed detection has been improved, and the effect of water edge recognition has been enhanced.

Federated learning is a new privacy protection framework for machine learning. The central server aggregates multiple participants to decentralized optimized parameters, then distributes the generated model to the client, and finally converges the global model. The model obtained by performance approaching centralized data training is trained under the condition that the data is not leave local. However, many studies have shown that this centralized federation system is vulnerable to confidentiality attacks by "honest but curious" attackers, using the gradient parameter information transmitted during federation model training to carry out reconstruction attacks or inference attacks, obtain the privacy data of participants or deduce some member information, which poses a severe challenge to the privacy protection of federated learning. In this paper, a hybrid defense strategy based on confusion self-encoder combined with localized differential privacy is proposed. On the one hand, the labels of the participants' local data are confused through the self-encoder network, so as to cut off the relationship between the gradient information and the original data. On the other hand, the localized differential privacy mechanism is used to disturb the transmitted gradient parameter information, and a model performance loss constraint mechanism is designed to reduce the impact of noise addition on the model performance. Experiments show that the hybrid defense strategy proposed in this paper can effectively resist reconstruction attacks and inference attacks in the process of federated learning model training, and achieve a better balance among computing overhead, model performance and privacy security.

Magnetic Resonance Imaging (MRI) is one of the most common imaging methods in the diagnosis and treatment of brain tumors. Multimodal brain tumor MRI automatic segmentation is significant for assisting the diagnosis and treatment of brain tumor diseases. In this paper, we first introduce the background and status quo of multimodal brain tumor MRI segmentation. Then, we divide the multimodal brain tumor MRI segmentation methods in recent years into three ways according to the network architecture. The basic theory, development process, research status, advantages and disadvantages of these three network architectures are also summarized respectively. In addition, we also introduce several publicly available datasets and evaluation metrics of brain tumor segmentation. Finally, we summarize and prospect the segmentation methods of multimodal brain tumor MRI. This article mainly introduces the recent research progress of multimodal brain tumor MRI segmentation, which will provide some help for researchers and practitioners in this domain.

Aspect-level sentiment analysis aims to determine the sentiment polarity of different aspects contained in text sentences. To address the problem that current neural network models based on LSTM and attention mechanisms cannot effectively encode aspect features and sentiment features, thus leading to a less than reasonable representation of text information, this paper proposes a fused syntactic information tree SITM-Bi-LSTM model for aspect-level sentiment analysis. First, the text sequence is passed through a bidirectional LSTM neural network to obtain its hidden output representation containing contextual information. Then, the syntactic information is used in the syntactic path to focus on the influence of contextual words with different distances from the aspect on its sentiment polarity, which in turn achieves the effect of enhancing the aspectual feature representation. Finally, the Tan-Relu gating unit is constructed to selectively extract emotional features that match the given aspect information for determining the emotional polarity of the aspect. The experimental results on the Laptop and Restaurant datasets show that the accuracy and values of the SITM-Bi-LSTM model are better than those of the comparison models, which confirms the model's effectiveness.

In various transformers, positional encoding is used to compensate for the inability of the attention mechanism to capture positional information between words. Previous research on transformers' temporal modeling has utilized recursive and relative positional encoding based on the Recurrent Neural Network (RNN). Recursive positional encoding captures linear text structure but lacks parallelization, hindering speed. In contrast, relative positional encoding ignores linear text structure, leading to weaker performance in short text classification compared to recursive positional encoding. To address the issues, we propose a model, sumformer, which mainly includes two parts different from the other transformers: cumsum calculation and summer initialization. Cumsum calculation simplifies the feature extraction part of RNN by a substitution method, replacing the dynamic rate function of RNNs with static trainable position parameters, and preserves the recursive structure, which enables the model to capture the linear structure information of the text through cumsum calculation method and maintains a low time overhead compared to RNNs. In addition, the summer initialization method, which limits the highest standard deviation of the positional parameter, enables the model to pay attention to the multi-level information of the text during initialization, with the richer optimization space, thereby improving the convergence ability of the model. The experimental results show the sumformer achieves roughly a 3% improvement in performance and a 58% improvement in speed compared to existing transformers based on recursive positional encoding. It achieves better short text classification faster, and summer initialization also can improve the performance without increasing training and inference time.

With the rapid increase in the types of electrical appliances used by users, the number of electrical accidents caused by this has increased year by year. In view of this, real-time monitoring of the use and operation of electrical appliances of power users is of great significance to better protect the safety of life and property of power users. Considering the monitoring cost and the acceptability of the monitored users, this paper proposes a machine learning fusion model based on V-I trajectory characteristic curve from the perspective of non-intrusive load decomposition, which realizes the effect of remote monitoring of abnormal electrical appliances with lower cost and higher user acceptance.

Recently, Convolutional Neural Network-based Compressive Sensing (CNN-CS) have drawn attention of extensive researchers since it can achieve faster reconstruction speed than traditional ones. Instead of traditional iterative reconstruction algorithms, a CNN reconstruction network is used with the goal of extracting key features of the images. For this purpose, the network should be trained on large amount of image data. However, most images usually poses both abundant low-frequency components and sparse high-frequency components, thus it is unreasonable to extract image features through a single network as in most CNN-CS algorithms due to that they may degrade the feature extraction of each other. In order to get a better result, a two-branch CNN-CS is proposed in the paper. In the network, the image signals are decomposed into two parts according to frequency firstly, and one branch is dedicated to extract the features of low frequency components, while the other is used to learn the high frequency ones. Considering the sparsity of high frequency, Mean Absolute Error (MAE) is chosen as the training penalty loss function. As such, both low and high frequency components can be well reconstructed respectively. Experimental results show that our method is prior to all traditional CS methods and single network CNN-CS method (CSNet+). The gain of PSNR over CSNet+ can reach 0.45dB.

To evaluate the resource usage in the scheduler queue, assist users in selecting appropriate queues for fast calculation, and improve the throughput and utilization of the system in the high-performance computing platform, it is necessary to use the historical job data in the queue for data analysis, and then make timely and effective predictions on the number of nodes occupied by jobs. In this paper, a queue node occupancy prediction method based on improved Convolutional Neural Network (CNN) and Long Short Memory network (LSTM) is proposed. Cluster and group the historical data of high-performance clusters to obtain balanced samples. In the improved CNN-LSTM network, the weights of different channels are determined by replacing the pooling layer with the customized attention mechanism layer. Samples are selected and extracted, and L2 regularization is used to prevent over fitting training to obtain more accurate results of node resource occupation. The test results of historical operation data of a supercomputer show that the improved CNN-LSTM hybrid network model proposed in this paper is more accurate than CNN-LSTM hybrid network model, LSTM model, MLP model and random forest model.

The current communication network can complete communication, entertainment, shopping and other applications, in which mobile phones play an important role, and mobile phone communication and Internet access need to rely on base stations. This paper proposes the optimization of communication network from the perspective of network structure, analyzes the factors affecting the communication network, and optimizes the establishment and deployment of communication base stations. On the one hand, it provides customers with better communication quality and faster communication speed, and on the other hand, it improves the investment income of communication operators.

In order to solve the issue of inaccurate recognition of time series of badminton action by 2D networks and the high cost of using skeleton data for badminton action recognition, this article aims to explore the use of RVM (Robust Video Matting) to extract the silhouette of individuals in videos, combined with an improved I3D network (referred to as the SI3D network) for action classification in badminton matches. We collected a dataset of badminton match videos featuring multiple players and labeled them according to different action categories. By using the SI3D network for training, this article achieved good performance on the test set, with an accuracy of 91.5%. This study demonstrates the effectiveness of the I3D network in badminton action classification and provides a new research direction for action recognition in sports.

With the development of smart education, personalized course recommendations are getting more and more attention. Recently, some in-depth learning-based solutions have been proposed. The main principle is to make sequential recommendations based on the learner's historical learning records, rarely use the semantic information covered by the course profile and user profile to complete the recommendation task. User preferences change over time. To solve these problems, a course recommendation model (FIHA) that integrates semantic information and hierarchical attention network is proposed. next class. Specifically, TextCNN is used to perform semantic processing on the course introduction to generate vectors, and then the learner's historical access course records are used to capture the learner's long-term and short-term preferences using hierarchical attention mechanism, and the learner's dynamic preference representation is obtained. The two are fused and recommendations are made based on their scores to courses. Experiments on real datasets show that this method is superior to other mainstream advanced methods.

BERT has shown remarkable performance in several natural language processing tasks, but it fails to exhibit the same high performance in cross-lingual tasks, particularly machine translation. To address this issue, we propose a BERT-enhanced neural machine translation (BE-NMT) model that optimizes the use of the information contained in BERT by NMT. Our proposed model comprises three components: (1) A MASKING strategy to mitigate knowledge forgetting caused by fine-tuning of BERT on the NMT task; (2) Serial and parallel processing of multi-attention models for incorporating BERT into the NMT system; (3) Fusing multiple hidden layer outputs of BERT to supplement the missing linguistic information of its final hidden layer output. We conducted experiments on several translation tasks, and our proposed model notably outperforms the strong baseline by improving 1.93 BLEU points on the United Nations Parallel Corpus English→Chinese task. Additionally, our model also achieves remarkable performance on other translation tasks.

Multi-hop Knowledge Base Question Answering (KBQA) aims to find the answer entities in a Knowledge Base (KB) that are multiple hops away from the topic entities mentioned in a question. The key challenge of multi-hop KBQA is the significant semantic gap between Natural Language (NL) and the relation paths in KB. Recently, the Case-Based Reasoning (CBR) approach has been successfully applied to KBQA, which first retrieves other similar questions from the training set and then referencing their solutions (e.g., reasoning path) to derive a solution for given question. However, existing CBR-based methods are hard to provide correct solutions while the gold reasoning paths are unavailable. Moreover, they are limited to the assumption that the reasoning path required to answer a question is repeats for similar questions, thus suffer from the generalization issue. To address above problems, we propose a novel framework, which consider both the semantic matching between NL question and KB paths, as well as the matching between KB paths of similar question and given question, to alleviate the inconsistency between NL and KB. Specifically, it uses above two matching mechanisms to generate more accurate solutions for training questions and then reuses it to rank the candidate reasoning paths of given question. To cope with the generalization issue mentioned above, our method adopts different ranking algorithm according to different NL similarity. Experiments over two benchmarks demonstrate the effectiveness of our approach, which achieves state of the art on MetaQA and new state of the art results on PathQuestion-Large.

Recommendation system aims to provide effective and personalized recommendation for users and solve the problem of information overload. However, the existing recommendation system lack of effective utilization of heterogeneous data and have the problem of information loss in the process of semantic information fusion. In this paper, we propose a Multi-scale Semantic Fusion Recommendation model (MSFRec) based on heterogeneous information networks to solve the problems above. First, we use heterogeneous graph and metapath to describe the complex semantic structure in the recommendation tasks. Then we divide the neighborhood guided by metapath into multiple layers and use the multi-layer interaction to capture multi-scale semantic information. Finally, we use a two-stage relational attention to guide the multi-scale semantic fusion. Extensive experiments show that MSFRec achieves competitive results in recommendation tasks.

Network representation learning maps network nodes to a unified vector space, which has found widespread use in various complex network analysis tasks. The resulting embedding vectors can reflect the original network's features and improve the accuracy of downstream tasks. However, current multi-layer network representation learning models have limited node sequence semantics and cannot fully capture the node's local and global structural information. To address this issue, we propose a multi-layer network representation learning model that integrates structural entropy information. Our model comprehensively considers the local and global structure of a multi-layer network, using local structure entropy and network standard structure entropy to guide the intra-layer and inter-layer migration processes, and then trains the node sequence using the skip-gram model to obtain the node vector. Experimental results on five public datasets demonstrate that our proposed model's embedding vector quality is significantly improved, with AUC scores on different datasets improving by 5% to 10% in the link prediction task.

Objective action quality assessment (AQA) is a complex machine vision task because existing AQA assessment models can’t effectively fit the subjective assessment. To address this issue, we propose a novel blind action quality assessment method. By processing the video data with spatial and temporal features, the performance of the model is effectively improved. In addition, we also proposed a new loss function to better train the model, which combines the information entropy of the data. Finally, the experimental results show that on the existing datasets AQA-7 and JIGSAWS are significantly improved, reaching 0.63 and 0.57, respectively.

As an important step in the IoT workflow, IoT data aggregation transmits, filters, analyzes and synthesizes data collected by IoT devices such as sensors to help people make decisions and evaluate. However, the traditional IoT data aggregation model relies on a central data aggregator, such as a third-party cloud server, and uses traditional encryption technology to process data, which may face problems such as original data leakage and calculation errors, and it is difficult to ensure that data participants can safely participate in the aggregation. Due to its homomorphic characteristics and performance improvements in recent years, homomorphic encryption technology makes it possible to use homomorphic encryption technology to protect data privacy during data transmission and calculation. In addition, as a non-tamperable ledger technology, the blockchain can be used to store evidence, specifically to store the promises made by the commitment mechanism, and complete the verification of data consistency and correctness.

To address the problem of low accuracy of traditional scene recognition methods in indoor environment, an indoor scene recognition method combining multi-scale features and attention mechanism is proposed. The method uses Efficientnet-B3 as the backbone network, introduces channel and spatial attention modules to improve the refinement capability of the network for features, and designs a multi-scale feature fusion structure to enhance the adaptability of the network to scales, based on which the model parameters are optimized by adding a spatial pyramid, thus further improving the model calculation accuracy. The experimental analysis shows that the average accuracy of this model reaches 94.4% in nine types of indoor scenes, all of which are better than the calculation results of AlexNet, VGGNet16, GoogLeNet, ResNet34, EfficientNet and other models, providing a new way of thinking for indoor scene recognition.

SARS-CoV-2 inhibitor plays an important role in COVID-19 preclinical drug discovery. As the existing SARS-COV-2 inhibitors showed more or less deficiencies, it is urgent to develop new SARS-COV-2 candidate inhibitors. De Novo Molecular Design plays a very important role in drug discovery. Most of the existing method use SMILES (Simplified Molecular Input Line Entry System) as the input of deep learning models. One popular way is utilizing deep learning models to automatically generate candidate drug molecules, and most of the existing models use SMILES as the input. In this study, we embed SMILES using a sub-word algorithm named BPE (Byte Pair Encoding) instead of One-Hot. First of all, the sub-word algorithm BPE learns a vocabulary of high frequency SMILES substrings from a large SMILES dataset, SMILES are then tokened according to the vocabulary learned by the BPE algorithm. Results show that the BPE algorithm can effectively learn the SMILES grammars and can help our generative model generate potential SARS-COV-2 inhibitors after transfer learning using the known 1253 SARS-COV-2 inhibitors. Generally, this paper provides an effectively method for de novo molecular design of SARS-COV-2 inhibitors.

Attention mechanism in image captioning model can help model focus on relative regions while generating caption. However, existing attention mechanisms are unable to identify important regions and important visual features in images. This problem makes models sometimes pay excessive attention to non-important regions and non-important features in the process of generating image captions, which makes model generate coarse-grained and even wrong image captions. To address this problem, we propose an “Importance Discrimination Attention” (IDA) module, which could discriminate important feature and non-important features and reduce the possibility of misleading by non-important features in the process of generating image captions. We also propose a IDA-based image captioning model IDANet, which is completely based on transformer framework. The encoder of IDANet consists of two parts, one is pretrained Vision Transformer (VIT), which is used to extract visual features in a fast way. The other is refining module which is added into encoder to model position and semantic relationships of different grids. For the decoder, we propose IDA-Decoder which has similar framework with transformer decoder. IDA-Decoder is guided by IDA to focus on crucial regions and features instead of all regions and features while generating image caption. Compared with others attention mechanism, IDA could capture semantic relevance of important regions with other regions in a fine-grained and high-efficient way. The caption generated by IDANet could accurately capture the relevance of different objects and discriminate objects that have similar size and shape. The performance on MSCOCO “Karpathy” offline test split achieves 132.0 CIDEr-D score and 40.3 BLEU-4 score.

Chemical oxygen demand value is a key indicator for water quality detection and an important link in water environmental protection. Near infrared spectroscopy is a fast, non-destructive, and green identification method applied to chemical oxygen demand detection. Near infrared spectroscopy has the characteristics of high data dimensions, severe band stacking, and difficult feature extraction. In this paper, an improved one-dimensional convolutional neural network is proposed to analyze the near infrared spectroscopy of water samples to predict their chemical oxygen demand values. Firstly, one-dimensional convolution is used to extract deep spectral features. Secondly, SoftPool is used to reduce the dimension of spectral features to improve the problem of losing some features in traditional pooling methods. Finally, the prediction results are output through full connectivity layer integration features. The proposed method is validated on a near infrared spectral dataset of chemical oxygen demand. The experimental results show that this method can accurately and quickly analyze the chemical oxygen demand value in water.

The Internet has generated a large amount of unstructured online texts that contain a large amount of argumentative content. How to automatically extract the argumentative contents from these online texts is an important issue. First, a corpus of argumentation about Chinese online current affairs reviews is built. The existing argumentation annotation scheme is also modified by adding two additional types of elements, "central conclusion" and "intermediate conclusion", and their corresponding argumentation relations to reflect the more complex structure of the long review texts. The pre-trained large models are used to identify and classify the argumentative elements in these texts. Then, the penultimate hidden state of the best model is used as the input of the pointer network for argumentative elements relationship identification. The experimental results display that the "pre-trained large model + pointer network" model improves the performance of the existing models.

Currently, remote sensing technology is developing rapidly, and many satellites and airborne sensors can provide high spatial resolution optical images, especially visible light images, which can provide the most intuitive and clear information of the earth's surface. Therefore, how to effectively process large-scale data information and obtain better detection results has become an important challenge in remote sensing image target detection research. This paper proposes an improvement to the target detection algorithm based on YOLOv5. In the input stage, dynamic anchor boxes are introduced to improve the instability of generating anchor boxes based on K-Means clustering, thus reducing model loss and convergence difficulties. In the backbone network stage, the recognition accuracy of small targets and cohesive targets with mutual occlusion is improved by introducing attention mechanisms.

Image-text matching (ITM) benefits from Large-scale Contrastive Language-Image Pre-training (CLIP) method that achieves higher accuracy. However, the CLIP method learns by contrasting global visual and textual features, which inevitably leads to the problem of mismatching in the image-text process due to the lack of inter-modal fine-grained information. Therefore, in this work, we propose a method called Turning a CLIP Model into Image-Text Matching (CIT) that focuses on combining fine-grained information between modalities to convert the CLIP model into a more efficient ITM model. The CIT method effectively improves the image-text matching accuracy of existing CLIP model and does not require additional pre-training. We demonstrate the effectiveness of our method through experiments with a range of state-of-the-art methods on two widely used datasets.

This paper proposes a Chinese word segmentation model based on RoBERTa and Conditional Random Field. The Chinese word segmentation model based on RoBERTa and CRF deals with Chinese word segmentation in the way of sequence labeling. The training of the model takes the character sequence of the sentence as the input, which is converted into number according to vocabulary and sent into RoBERTa layer. The output results of RoBERTa layer were converted into the probability distribution of word position label by linear transformation and softmax activation function. The predicted word position label sequence was obtained by Viterbi decoding operation. After experiments, the accuracy rate, precision rate, recall rate and f1 scores of RoBERTa + CRF model on SIGHAN's PKU and MSR corpora are 96.33%, 95.75%, 95.00% and 95.37% respectively (pick the highest score), which has preliminary application value and certain room for improvement.

There are many fields involved in human research on animal vocal behavior, such as biomedical science, ecology, and behavioral science. Understanding animal language and behavior remains a major focus of scientific research, and many efforts are being made in this direction. With the development of artificial intelligence technology, deep learning has gradually become a major method for humans to understand animal behavior sound. This article aims to develop an end-to-end animal behavior sound classification model using Wav2Vec2.0 (W2VCM). By combining the Wav2Vec2.0 model with the CatMeows dataset, a pre-trained model is obtained for extracting speech features and performing animal behavior sound classification tasks. Experimental results demonstrate that W2VCM achieves higher accuracy than six other deep learning models tested, both on the original and re-divided datasets. The accuracy of W2VCM reached 97.97%, 2.03% higher than the best result of other models on the original dataset, and 3.13% higher on the re-divided dataset.

In order to solve the problem of low grasping accuracy due to the influence of ambient light in the recognition of existing intelligent cooperative robots, an algorithm of classification grasping, and placement based on color information is proposed. The improved HSV (Hue, Saturation, Value) color space method is incorporated to identify the target object to reduce the impact of environmental light. Secondly, the Canny edge algorithm and the minimum circumscribed rectangle are used to identify the location of the placement area and the position and posture of the target grasping object, so as to determine the grasping center. Finally, the grasping process of the manipulator is optimized to improve the grasping accuracy. The whole algorithm is verified in the Baxter experimental platform of two-arm cooperative robot, and the experimental results show that the grasping accuracy can reach 97% and has strong robustness, which confirms that the experiment has a decisive case reference value for the research of intelligent cooperative robot and intelligent grasping of industrial manipulator.