This PDF file contains the front matter associated with SPIE Proceedings Volume 12712, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The U-K approach provides a new, general solution for the explicit motion equations of complex constrained mechanics systems. These equations are suitable for constraint forms such as holonomic or nonholonomic, ideal or non-ideal, scleronomic or rheonomic. In the resolution process, the constraint forces and accelerations can be achieved without introducing Lagrange multipliers. At present, this approach has been applied to the trajectory tracking of robotic arms, structural dynamics control, and the motion of swarm robotic systems. The process of resolving the U-K equations involves two key problems. (1) The introduction of generalized inverse matrix brings great difficulties to the actual solution. (2) The mass matrix in the representation is positive definite, but the mass matrix described by the redundant minimum number of generalized coordinates may be singular, which causes the equations can not be used. For the U-K method and the above problems, we firstly uses the Lagrange method to get the dynamic equation of the constrained multibody system. The constraint equations of Lagrange multipliers and the method of additional constraint forces are imposed to express the transformation process of the unconstrained system to constrained system. Then, three different methods are respectively introduced to replace the generalized inverse representation in the original U-K equations. In the situation of singular mass matrix, the Udwadia-Phohomsiri form of generalized inverse representation and the explicit dynamics represented by null space form are introduced. Then, by adding auxiliary systems, we get two representations that respond to both the U-K equations and the Udwadia-Phohomsiri form.

The images of apple leaf disease under natural conditions has a complex background and small diseased spots, which makes it difficult to detect in real time. To address this, a new method was proposed for apple leaf spots segmentation based on a DeepLabV3+ network. We used Xception composed of depthwise separable convolutional layers as a feature extraction network. Depthwise separable convolutional layers can effectively reduce the amount of model parameters and improve the model training efficiency. The introduction of the CBAM module allows the model to better focus on the spot area, to reduce the influence of complex background. This study densely connected the ASPP modules, expanded the model receptive field, and strengthened the model feature extraction ability. The experimental results showed that the MIoU of the model for five apple leaf diseases under natural conditions can reach 90.37%, which was increased by 7.53% compared to the original DeepLabV3+. The MIoU of the current model exceeded FCN-8s, SegNet and PSPNet , the MIoU by 20.42%, 15.91% and 13.83%, respectively.

Failure correlation is a necessary condition for source trace and violations determination in accountability of cloud services. This paper surveys the failure correlation. Most studies were involved in the correlation between services violation and failure factors only. However, an understanding of the correlation between the factors, the integrality of the factors is ambiguous and even few involve the correlation patterns of specific violations. Above all, the taxonomy of the failure factors, the methods of structure representation and analysis, some problems of recent research and suggestions are proposed in our paper.

The semantic segmentation model Deeplabv3+ model in deep learning has a great deal of parameters and spend plenty of training time, and the accuracy of small objects and edge segmentation is low. To solve this problem, an improved semantic segmentation algorithm for DeepLabv3+ network is proposed. Firstly, MobileNetv2 is used instead of Xception to reduce model complexity and speed up model training; secondly, the attention mechanism is introduced to improve segmentation accuracy; thirdly, a Strip Pooling branch is connected in parallel to the Atrous Spatial Pyramid Pooling module (ASPP) to improve the characteristics of the model extraction ability; finally, the Swish activation function is used to replace the Relu function. Experimental results show that compared with the original model, the proposed algorithm can still maintain a high accuracy when the number of parameters is greatly reduced.

Application high availability (HA) has always been the goal pursued by various core businesses. At the practical level, multiple business systems are usually deployed based on two or more physically separated data centers, which are connected by BGP or dedicated lines to ensure data transmission efficiency and latency; and specifically for each data center, applications deployed through the industry's leading cloud-native technology development can provide a certain degree of single-cluster HA. This paper does not involve physical topology, like link layer and other hardware facility level HA implementation scheme, but mainly focuses on the software level HA, under the premise of hardware level HA, in order to achieve the HA effect of minute level RTO and second level RPO. The most important thing is to solve the synchronization efficiency problem. In this paper, we design a data synchronization algorithm with dependencies, and elaborate how to combine the algorithm to achieve automatic and efficient synchronization of workloads of distributed applications developed in cloud-native way and data of stateful applications in cloud-native scenario, and at the same time, it is convenient to perform switchover rehearsal between main and standby clusters in normal times, and when failure actually occurs. It can quickly switch the flow between clusters to achieve distributed applications when a failure actually occurs, effectively solving the problem of achieving HA of cloud-native applications across multiple data center clusters.

As one of the fundamental functions of fulfilling the smart library, the research regarding the work efficiency enhancement of library novelty search service has attracted significant attention from both academic and industry communities. This paper jointly takes novelty service applicants (NSAs) novelty search service requests and novelty service providers (NSPs) assignments into consideration. The novelty search service efficiency maximization optimization problem is formulated. To enhance the novelty search service efficiency, a high-efficiency solution is proposed inspired by matching theory. The results show that the proposed solution can remarkably enhance service efficiency while decreasing the service time cost in comparison with the current approach.

A graph convolutional neural network is a special type of neural network that GCN can use to extract features from graphs and use these features for classification or regression. The current lightweight graph convolutional neural network in the recommendation field only considers the case of fixed neighborhood weights, with the increase in the number of training layers of the network recommendation effect will rapidly decline resulting in the problem of oversmoothing and overfitting. In this paper, in order to solve the above problem, this paper will introduce a DE-LGCA personalized recommendation algorithm that uses a random wandering strategy for data enhancement based on a lightweight graph convolutional neural, while introducing an attention mechanism to increase the node neighborhood weights, and show how to use this algorithm to achieve personalized recommendations. Experiments were conducted on three publicly available datasets, and three performance metrics, Precision, Recall, and Normalized Discounted Cumulative Gain (NDCG), were used to evaluate the performance of DE-LGCA, which not only outperformed the comparison method in terms of performance metrics, but also improved the efficiency of the algorithm.

In order to avoid the excessive pesticide content in the ecological environment caused by the use of a large number of pesticides, it will cause serious harm to people, livestock, water, atmosphere, soil and other environmental elements. Even if it is a small amount of residues, long-term contact will do harm to human health. In this paper, the application of data fusion algorithm in the detection of pesticide residues in food is proposed. A data fusion method suitable for repeated sampling of single sensor is applied to the processing of pesticide residue detection data. The data are grouped by statistical theory, while the support probability is fuzzy, and the support interval is determined by subsection discussion, which makes the fused data more stable than the previous results obtained by absolute support probability and support interval, and further ensures that the data fusion has better anti-interference.

Aiming at the problems of traditional particle swarm optimization (PSO) algorithm such as slow convergence speed and easy to fall into local optimum, this paper proposes to use the core idea of simulated annealing algorithm to improve the particle swarm optimization algorithm. First, the Metropolis criterion is used to select the probability of inertia weight, and for each particle in each iteration, the probability of inertia weight is selected to balance the search ability of particles; Secondly, by analyzing the information exchange of the optimal solution between each iteration, the mutation particles are constructed, and the position information of the next generation of global learning particles is determined through probability selection, which can effectively prevent the particle population from falling into the local optimal region. The simulation results show that the PSO algorithm based on simulated annealing rules has higher convergence accuracy and speed compared with other test algorithms, and can effectively improve the ability of PSO to find the optimal solution.

Aiming at the disadvantages of the standard seagull algorithm (SOA), such as discrete distribution of initial population position, low solution accuracy and slow convergence speed, a seagull optimization algorithm (L-SOA) which based on good point set mapping and integrating Levy flight and adaptive walking strategies is proposed. Uses good point set mapping to produce a better initial solution, select Levy flight and adaptive walk strategies for optimization by random probability to enhances the ability to jump out of local extreme and improves the convergence performance. Finally, compared with other four swarm intelligence algorithms on eight standard test functions, the results show that L-SOA algorithm has higher precision, faster convergence speed and more stable robustness.

With the introduction of Cloud-native thinking and container technology, the industry trend is to transform network functions into cloud-native implementation with container bearing, that is, CNF (Cloud-native Network Function). Existing CNF research focuses on both CNF collaboration with CISM systems and CNF performance enhancements. This paper focuses on the implementation of CNF itself, and proposes a general CNF framework that can be assembled on demand to provide different network services with good scalability.

JSON is a data format to transform and exchange of different data in various applications, especially in Web applications. JSON format data contains not only the data itself, but also the semantic information behind the data. Like relational data and XML data, functional dependencies and keys are very important semantic information in JSON format data. This paper studies the semantic constraints such as functional dependencies and keys of JSON format data. The concepts of functional dependency and key are proposed based on JSON tree model for JSON format data. We discussed the implied meaning behind JSON functional dependencies and keys. This research is of great significance for efficient management of JSON data, and has important reference significance for other relevant research, such as JSON semantics.

The classification of human activities plays an important role in sports, medical care and other fields. This paper is based on the measurement data of the nine-axis sensorfirstly, the time domain characteristics of sensor data are extracted, and the dimension of the data is reduced by principal component analysis. By comparison, a K- nearest neighbor algorithm classification model is designed and optimized for human activity classification.

In the application of SIMPLE algorithm, the generation and solving of linear equations with sparse coefficient matrix appear. Gauss-Seidel and ILU iteration are commonly used to solve the problem. Combined with the common characteristics of LU-SGS time advance method and ILU iterative method, the numerical iterative solutions of sparse matrix linear equations are studied. An iterative LU-SGS algorithm for solving general sparse matrix is proposed. Using C++ as programming language, a special equation solving library is constructed by using three-point data storage structure of multi-dimensional column vector and sparse matrix dot product algorithm. Combined with SIMPLE algorithm based on pressure, the transient NS equation in static grid is simulated numerically. Compared with Gauss-Seidel iterative method, this algorithm has the advantages of accurate calculation and fast convergence, so it has certain application value and reference value.

With the development of cloud manufacturing, there is an influx of cloud manufacturing resources on the Internet, which makes it a valuable area of research on how to effectively update the cloud pool of manufacturing resources and improve resource management efficiency in cloud manufacturing. To this end, recent research has focused on addressing the dynamic updating of cloud manufacturing resources to meet the time-sensitive needs of cloud manufacturing. However, due to the increasing diversity of manufacturing resources, the traditional manufacturing resource register only contains static information about manufacturing resources, which can easily have an impact on the dynamic matching and recall of manufacturing resources. So it is very challenging to achieve an update of manufacturing resources, especially video resources. So we propose a dynamic tracking method for spatial features of cloud manufacturing called CMFK (Kalman-based method for dynamic tracking of cloud manufacturing resources). CMFK consists of two main components, firstly the identification, description, and matching of manufacturing resource feature points through computer vision. Secondly, the update and registration of dynamic information are completed. The experimental results demonstrate the feasibility and effectiveness of our approach.

This article focuses on the implementation of equipment control in the Internet of Things monitoring system due to the change of environmental parameters. When the abnormal environmental parameters are proposed based on the edge calculation of the intelligent pipe corridor based on the multi-sensor data detection, the calculation algorithm of the detection process and timing correlation is designed, and the corresponding execution equipment is controlled by the algorithm, so as to achieve the purpose of correcting the environmental parameters in the corridor. A good control effect is obtained.

Deep neural networks are widely used at computer vision while the hidden layers form a black-box and make the networks uncertain. Quantifying the uncertainty in neural networks is a big challenge. Although the Bayesian neural networks can construct model for black-box problems by distributing learning weight, these networks are computationally expensive compared to standard neural networks. In this paper, we propose a novel encoder-decoder semantic segmentation network to reduce the uncertainty of the network. In addition, we innovatively introduce information entropy into semantic segmentation, which intuitively reflects the level of uncertainty in and out of the domain. The result shows the network has a good segmentation effect (79.4% mIOU in CamVid), and makes the network robust by increasing the difference between in-domain and out-of-domain information recognition.

With the development of the takeaway industry, the demand for space to deposit takeaway food increases rapidly. Currently, an increasing number of office buildings and communities are equipped with food delivery cabinets. However, the conventionally inconsiderate site selection has incurred problems of poor experience for takeaway consumers. In this case, this paper applied the center-of-gravity method to build models for consumers' complaint rate and distance to the food delivery cabinet. The models are solved using particle swarm optimization (PSO) while they are subject to nonconvex functions due to the large volume of data. Compared with conventional site planning, this approach evades local optimal solution, and PSO starts to converge at a favorable rate from the early phase (the 40th iteration). Therefore, the PSO algorithm is efficient in the site planning practice of food delivery cabinets and generates accurate results, offering a brand-new perspective for it.

In recent years, sudden events such as public health events have occurred frequently, which has seriously hindered China's economic development. Reasonable scheduling of emergency supplies can effectively reduce the harm caused by emergencies, and it is of great significance to ensure the production and life of people in disaster areas. In order to ensure that emergency materials can be quickly and efficiently distributed to the affected areas after major emergencies, the vehicle scheduling optimization problem in emergency logistics is studied, that is, to quickly screen the best driving route in a complex road network system to deliver emergency materials to the demand location in the shortest time. In this paper, a model with hard time window is established for emergency logistics vehicle routing optimization problem, and the improved saving algorithm in the shortest path algorithm is selected to solve the emergency material dispatching example in Wuhan area. The experimental results show that the algorithm in this paper has a good effect in solving the problem of emergency supplies scheduling. The algorithm has a good application prospect in reducing the cost of emergency supplies scheduling, shortening the travel time of emergency logistics vehicles and improving the satisfaction of emergency supplies demand points.

In order to improve the positioning accuracy of the centroid algorithm, this paper proposes the RSSI centroid node positioning algorithm based on chimpanzee optimization (SRSSI-CHOA). The algorithm first optimizes the received signal strength value by velocity constant filtering, and then calculates the unknown node coordinates using the distance-weighted centroid algorithm, and then constructs the fitness function based on the distance information and the obtained position information; Finally, the chimpanzee algorithm is used to find the optimal solution position and optimize the node coordinates. The simulation results show that the proposed algorithm is 30.78%, 24.42% and 8.3% lower than the RSSI - centroid, Gaussian model RSSI centroid average and RSSI-PSO algorithm.

In this paper, a new point cloud feature extraction algorithm is proposed to address the problems of fixed point cloud feature extraction method, large number of point clouds and low computational efficiency of the current open source laser SLAM. The algorithm follows the original LOAM of calculating point cloud curvature and the screening mechanism of anomalous points, on which a new mechanism of selecting edge feature points and plane feature points is designed to reduce the number of point clouds transmitted to the back end; Finally, this paper transplants it in the open source LIOSAM scheme and completes the experimental validation of the algorithm in different datasets. The evaluation results show that the localization accuracy of this paper's algorithm is better than that of the LIO-SAM algorithm under the same conditions, and it also reduces the computing resources required.

An important component of intelligent driving technology is the recognition of traffic signs based on convolutional neural networks (CNN). How to design a traffic sign recognition system with high accuracy and good real-time performance is crucial for the safe driving of vehicles. For the current traffic sign detection algorithm, there are high network complexity, large amount of computation, and high difficulty in edge deployment. This paper proposes a deep neural network compression strategy, which skillfully uses model lightweight, pruning, knowledge distillation and quantification methods. The lightweight full connection layer is used to accelerate reasoning, and the knowledge distillation technology is innovatively used to assist the pruned student network to recover the lost accuracy. The teacher network is used to help pruning restore the original accuracy better, improve the generalization ability, and avoid that the small network cannot work after excessive pruning, so as to achieve a higher pruning rate. This experiment shows that knowledge distillation can assist pruning recovery in a more accurate manner than ordinary pruning. On the traffic sign GTSRB dataset, the mainstream network models VGGNet and AlexNet are used for training and testing. The models are compared before and after compression. Based on the results, the model can be compressed to 0.08% and will have a 97.32% accuracy.

Based on the fluid-solid coupling theory of unsaturated soil, MIDAS GTS NX was used to simulate and analyze the effects of anti-slide pile, anchor bolt, anchor bolt + anti-slide pile reinforcement on the plastic zone, overall displacement and safety factor of slope under heavy rainfall conditions, and the reinforcement effects of different supporting measures were discussed according to the numerical analysis results. The results show that the slope is easy to produce landslide phenomenon without reinforcement under the condition of continuous rainstorm. The anti-slide pile + prestressed anchor reinforcement can improve the shear strength and stability of the slope soil, and the overall displacement of the slope decreases obviously.

Epilepsy is a chronic disease caused by sudden abnormal discharges of neurons in the brain, which leads to transient brain dysfunction. Tens of millions of people in the world suffer from epilepsy. Predicting the coming seizures will help to save the lives of patients. In this paper, a mixed convolutional neural network (CNN) seizure prediction algorithm based on attention mechanism is proposed. We first processed the original EEG signals as 3D features, and then designed a mixed CNN to automatically extract spatial-temporal features and completed the classification task. In addition, we applied the attention mechanism to the proposed mixed CNN model, which can not only obtain the importance between different channels, but also further learn more abstract features. We used the CHB-MIT scalp EEG dataset to evaluate our seizure prediction algorithm. We also compared our seizure prediction algorithm with the previous deep learning algorithm, and the experimental results have showed that our proposed seizure prediction algorithm is effective.

It is a very important method in nondestructive testing to detect defects using infrared or electroluminescent (EL) images of solar cell modules. Traditionally, this work is completed by skilled technicians, which is not only time-consuming but also vulnerable to subjective factors. The surface defect detection method of solar cells based on machine learning has become one of the main research directions for its high efficiency and convenience. Therefore, this paper proposes a defect detection method based on machine learning for solar cell EL image defect detection. First, the EL image is preprocessed to provide a clearer image with more obvious defects for subsequent feature extraction; Then, the gray, contrast, brightness and edge gradient features of the preprocessed image are extracted, and the edge gradient information of the defect is extracted using the improved LBP algorithm. Combined with the gray value, contrast and brightness information of the image, the training and testing data set of the classifier model is formed. Finally, the optimized decision tree algorithm - C4.5-G algorithm is used to classify the surface defects of the battery chip. While improving the performance of the algorithm, the convergence speed is accelerated and the over fitting phenomenon is avoided. The experimental results show that the method is feasible, effective and high detection rate in the defect detection of solar cells.

In order to solve the problem that the detection accuracy of YOLOv4-tiny algorithm decreases due to the lightweight and small weight of the model, a YOLOv4-tiny technology method based on attention fusion mechanism is proposed to detect pedestrians. In this method, CBAM is added to the effective feature layer and upper sampling layer extracted by FPN feature pyramid network based on the original YOLOv4-tiny algorithm to screen advanced semantic features. The 26×26 detection layer of the prediction feature layer was fused with the second layer of CSP structure with CBAM, and a 52×52 detection head was added to expand the coverage of pedestrian detection. Use IOU_NMS non-maximum suppression to remove some redundant candidate boxes and speed up the regression of prediction boxes. The Python script was used to extract the Person-class data image from PASCAL VOC2012 data set. The experimental results show that this method has improved the accuracy, recall rate and accuracy in the data verification set, achieving the purpose of accurately detecting pedestrians.

Deep learning has led to remarkable success in many research fields,such as machine translation,speech recognition. Combined with the characteristics of the Tibetan and the current development of Tibetan, with the help of transformer, conformer and other frameworks, we studies how changes in modeling units and encoder types affect Tibetan speech recognition. The modeling unit in this paper generates subwords through bpe algorithm or automatically divides Tibetan characters, and the encoder types are transformer and conformer. Finally, we verify it on the open-source Tibetan language dataset xbmu-amdo31. The experimental results show that the Tibetan characters as the modeling unit and conformer as the encoder has the best performance. The CER of conformer encoder whose modeling unit is character is 11.89%, which is 44.51% lower than the transformer model. The WER of conformer encoder whose modeling unit is subword is 17.40%, which is relatively reduced by 32.27%.

In recent years, the technology of camouflage semantic segmentation using machine learning and computer vision has been widely concerned as one of the emerging research hotspots in the field of artificial intelligence. There are several methods have been developed for camouflage object segmentation, which still can be improved to solve the existing problems such as incomplete segmentation, edge dispersion, and too many parameters. This paper proposes a lightweight camouflage segmentation model called ConvNeXt Orientation Aggregation net (COA-net). In this paper, the camouflage segmentation model PFnet is selected as the basic network, and the feature extraction ability is enhanced by using the lightweight backbone network ConvNeXt. Futhermore, a plug-and-play mixed attention mechanism, called CDSEnet, is proposed, which enhances the ability to identify effective features and interference information. The experimental results show that the proposed method can achieve better segmentation performance compared with the original model. The experimental results show the proposed method can achieve the improvement of the index structure measure Sα, the adaptive measure Eφ, the weighted measure Fβ and the average absolute error M by 3.3%, 2.7%, 6.3% and 20.1% respectively and the subjective quality is better as well compared with the original method.

In order to improve the classification accuracy of weeds associated with lily in real field condition, this paper constructs a dual attention module from both channel attention and spatial attention, and the dual attention module is embedded into the simplified Inception network. It not only reduces the scale of the model but also improves the detection accuracy. The experimental results show that the training parameters of the model decreased by nearly 90%, and the classification accuracy increased from 94.26% to 98.32%. Furthermore, before embedding, 3.10%, 1.32%, 9.79% and 4.15% of S. wightianus, S. vulgaris, S. viridis and E. pilosa Were mistakenly detected as lily leaves and after embedding, only 1.03% of S. wightianus and 1.81% of E. pilosa were mistakenly detected as lily leaves. This study can provide reference for the next step of pesticide variable spraying research.

In order to improve the target detection performance of the detection model pair in the target tracking scenario, this paper abandons the use of the CSPDarknet53 backbone feature extraction network in the original algorithm and selects the lightweight network MobileNetV3-large as the new backbone feature extraction network. And the channel attention SE module in the original inverted residual structure of the network is changed to the integrated attention module based on the channel dimension and spatial dimension. According to the training results on the dataset CrowdHuman, compared with the original Yolov5 model, the improved detection algorithm improves the detection accuracy by 1.5% and the detection speed by 10.2 frames per second, which is more in line with the real-time requirements of the automatic driving scene.

In English teaching, grammar teaching is still an indispensable part. Teachers pay more attention to the correction of written form than oral form in grammar teaching so that students can see their mistakes intuitively and clearly. Whether from the perspective of teachers or students, the correction of grammatical errors and good learning results are inseparable. At present, the public grammar error correction corpus is too small and the quality is uneven, which makes the parameters of the grammar error correction model can not be fully trained. The performance of the model is also a bottleneck. Graph neural network-based model for grammar error detection is studied. The errors in the text data are detected. Neural network modeling is adopted as the basic structure of the model. In addition to predicting the correct label of each word in the sentence, an auxiliary task of predicting the context word of the word is introduced to further improve the detection performance of the model. Furthermore, the graph neural network with a gating mechanism is adopted to model the dependency syntax tree of the statement, which provides important information features for error detection and effectively improves the performance of model checking. Finally, good results in English grammar error detection and test data sets are achieved. As one of the core technologies in online education and other fields, the research of grammar error correction has great research and application value.

Millimeter wave radar, compared with cameras and other detection equipment, can better protect privacy. At the same time, millimeter wave cost is low, and is not affected by rain, fog and other weather, so it is used more and more widely. In this paper, millimeter wave radar AWR1642 was used to collect 8 kinds of daily movements at home. Based on the obtained radar data spectrum, the millimeter-wave radar action database is constructed. Finally, the residual network is used to recognize and classify the spectral graph, with an accuracy of 99.79%. Abstract: In this paper, millimeter wave radar AWR1642 was used to collect 8 kinds of daily movements at home. Based on the obtained radar data spectrum, the millimeter-wave radar action database is constructed. Finally, the residual network is used to recognize and classify the spectral graph, with an accuracy of 99.79%.

With the rapid growth of computer communication technologies such as "Cloud Compute", "Big Data", "IOT", "Mobile Internet" and "AI", the different types of the message middleware have been widely used in multiple heterogeneous systems. In order to solve the message communication requirements between heterogeneous systems of "Dispatching and Control Systems", this paper proposes the message gateway technology of heterogeneous systems and completes the presidential architecture design of this technology. To achieve the reliable and stable message interaction between heterogeneous systems, some technologies such as message subscription and publishing technology, message persistence and cache technology, elastic expansion technology of message gateway, message communication fault location technology are proposed. These technologies significantly improve the ability of various types of data interaction between heterogeneous systems.

For two types of common faults in automotive fuel cell reactors, namely flooding and drying, a PEMFC fault diagnosis approach combining support vector machine (SVM) and adaptive particle swarm optimization (APSO) is proposed. In this study, principal component analysis (PCA) is used to reduce the dimensionality of the original data samples, extract the two key variables characterizing the faults, establish the APSO-SVM diagnostic model, and compare and contrast it with RNN, CNN, and BP diagnostic methods to demonstrate the effectiveness and superiority of the algorithm. To compare and evaluate the diagnostic model's generalizability, several sample set sizes are used. The simulation outcomes demonstrate that the fault diagnosis method of optimized SVM using APSO has excellent performance in terms of diagnostic accuracy, model training speed, and generalization ability, providing a method reference for further research on automobile fuel cell fault detection.

Barcodes have played an essential role in our daily life. Localizing, or detecting them in real scenes in a fast and robust way has many practical applications. Recently, some deep learning-based methods have shown great potential in object detection. However, because barcodes are placed at any angle, vertical bounding boxes cannot sufficiently capture accurate orientation and scale information. In this paper, we propose a barcode detector that performs dense prediction to accurately locate the position of pixels belonging to the barcode region. For better detection performance, we design a spatial attention module to integrate global information adaptively, which can be easily plugged into the prediction backbone. Meanwhile, we employ the knowledge distillation training strategy to train a small student network with the help of a heavy teacher network. Extensive experiment results demonstrate that our method can perform real-time speed on CPU environments and locate barcodes in images with complex scenes.

Mashup is an information system development project based on Web services. It can build new business processes by mixing data, resources and APIs from different sites or applications, and meet new user needs. However, in recent years, the number of Web API is increasing quickly, and it is difficult for developers and users to find suitable Mashups, which influences the development and functional efficiency of information systems. Therefore, designing an accurate and efficient Web service recommendation method to improve the functional characteristics of information systems has become a research hotspot. At present, many algorithms, like matrix factorization and collaborative filtering, often have serious cold start and sparse problems. In order to alleviate the above problems and capture as much information as possible, we proposed a multi-modal knowledge graph enhancement recommendation method (MKE), which combines structured information and unstructured information of Mashup and Web API. Firstly, the feature information is mined from the text description of API and Mashup to assist the construction of the service recommendation system. Then, the knowledge graph is constructed by using the related attributes and relationships of Mashup and API, as well as the entity and relationship of Mashup and API, and the text features. The low-dimensional embedded Trans model provides Top-K recommendation results for the specified Mashup description. Finally, experiments are carried out on the latest ProgrammableWeb dataset, and the experimental results show that the recommendation performance of this method is better than that of RippleNet, NCF, CKE, KGCN.

Scene text detection aims to detect text regions in the complex background, and deep learning-based methods have been the mainstream. To obtain robust performance, deep learning-based methods are data-hungry, and one efficient technique to obtain sufficient data is data augmentation. However, current data augmentation for scene text detection is 1) changing the whole image, which ignores the instance-level diversity, or 2) generating synthetic data using generative models, which requires extra training data. In this paper, we propose a self-compositional data augmentation (SDA) for scene text detection. Our SDA generates new data by changing the original text regions of one image with four types of variations: translation scaling, rotation, and curving, and putting the changed text regions back into random places of the same image. In specific, our SDA is an instance-level augmentation, which could be combined with image-level augmentation; SDA requires no extra training data, which could be easily adopted in different methods. We conducted extensive experiments with three state-of-the-art scene text detection methods on two public datasets. Using our SDA improves all methods on all datasets, and the improvements demonstrate the effectiveness and generality of our SDA.

In relation to influence from the domestic and international environments, the cost of raw materials is frequently subject to multiple seasonalities that are more difficult to predict^[1]. To predict the price of raw materials, it is critical to employ a reliable and scientific model. The autoregressive moving average (ARIMA) model, the long short-term memory (LSTM) model, and the combined LSTM-ARIMA model were used to forecast the spot price trend of the steel in the short term by using data on a steel plate spot price as the research object.The actual data were empirically analyzed using Python software, and the forecasting results of the ARIMA model, the forecasting results of the LSTM-ARIMA model, and the combined LSTM-ARIMA model The comparison results indicate that the LRST-ARIMA model has the highest prediction accuracy, followed by the ARIMA model, whereas the LSTM-ARIMA model has the highest prediction stability. The results show that the use of combined LSTM-ARIMA models to forecast changes in raw material prices not only promotes the application of time series models and deep learning techniques in enterprise forecasting but also effectively helps enterprises make decisions in the raw material procurement business.

The current mainstream object detection algorithms are designed based on natural distribution images, which results in low detection accuracy when detecting remote sensing data images with a large number of densely packed small targets. Traditional deep detection networks mostly use 3x3 convolution kernels, which makes the receptive field of the network relatively fixed, and is not conducive to detecting remote sensing targets with large scale differences and dense distribution. Inspired by the receptive field (RF) structure in the human visual system, we propose to design feature extraction modules with different receptive fields in the deep neural network, which are used to collect the spatial information of the feature maps. This method takes the multi-branch convolution module in Inception V4 as a reference idea, and uses dilated convolution to enhance the receptive field and improve the extraction ability for densely packed targets. The operation of having different receptive fields for the network provides a more refined resolution ability for the model. To verify its effectiveness, we conducted experimental comparisons on multiple datasets based on YOLOX as the benchmark model, and the experimental results confirmed the effectiveness of our multi-receptive-field approach.

Generated Adversarial Network has been widely used in the field of image style transfer. Among them, ChipGAN is specifically aimed at the study of Chinese ink painting style transfer. In previous studies, it has demonstrated its excellent transfer effect at the visual level. In this paper, using CycleGAN and ChipGAN model, we adjust the parameters by controlling the variable method, quantitatively compare the image style transfer of the two models on the basis of visual effect comparison, and prove the effectiveness of ChipGAN in ink style transfer and the complexity in training.

In order to meet the requirements of lightweight power distribution data processing, raise the utilization rate of new energy resources, improve the processing efficiency of refined services, and minimize the delay. In this paper, the cloud-edge collaboration-based data processing capability reconfiguration architecture is proposed to meet the multi-service requirements of low-voltage power distribution. Secondly, the resource block matching-embedding cloud-edge collaboration data processing capacity reconfiguration algorithm is proposed, achieving embedding of resource blocks to edge servers to minimizes data transmission delay and data computing delay. Simulation results show that compared with traditional matching-based data processing reconfiguration algorithm (TMPR), the proposed algorithm can better guarantee the delay requirement of high priority service.

The queuing theory is widely applied to air defense weapon, which is illustrated by an example, to analyze the penetration process and calculate the penetration probability. The paper firstly analyses the service time statistical characteristics of air defense weapon and the maximum queuing time statistical characteristics of the target, then provides simulation experiment data to verify the theoretical analyses.

When resources are limited, active learning seeks to optimize the dataset labeling process by choosing the most valuable unlabeled samples through sampling strategy. The majority of active learning techniques in use today are intended for balanced datasets. Machine learning, however, frequently deals with unbalanced datasets in the actual world. We think that unbalanced datasets make active learning more difficult. Here, we suggest an innovative active learning strategy for unbalanced datasets called the imbalanced feature-mixing K-center method (IFMK). In order to lessen the imbalance in the unlabeled samples, it develops interpolation between the representation of unlabeled samples by learning the more pertinent generic properties of the class. We employ a regularization term to enhance the proximity and diversity of unlabeled and labeled samples during sample querying. Results from two datasets with a wide range of imbalance variables demonstrate the superiority of our strategy over more current active learning techniques. On a dataset with more imbalance, these improvements are more noticeable.

There is a serious data island problem for many power data sets at present, and centralized storage of large amounts of data will cause the privacy disclosure of the original data owner, and also face security and regulatory requirements. This patent proposes a power consumption prediction model and method based on federated learning, which includes cloud computing center, central node, edge computing node and communication network. The cloud computing center builds an improved LSTM based power consumption prediction model and initializes it. The central node publishes prediction tasks and cloud computing publishes initialization models to edge computing nodes. After receiving the prediction tasks and initialization models, edge computing nodes use local load data to perform local training. After training, the central node receives the training parameters of each edge computing node, The root mean square error and average relative error of the parameters are calculated. If the error is less than the set accuracy threshold, the new model will be sent to the edge node to complete the model training. If the error is greater than the set accuracy threshold, the local model update parameters of the edge node will continue to perform the training task and perform the iterative training task, which improves the accuracy of power consumption prediction, and isolates the data to ensure that the data will not be leaked to the outside, further meeting the needs of user privacy protection and data security.

Removing the rain stripe from the image obtained in a wet setting improves visual identification accuracy. This study focuses on both classic and deep learning models to track the research progress and trends. Various metrics including SSIM and PSNR have been designed for evaluating these models, based on some widely used open image datasets, e.g., Rain12 and Rain100L. Both synthetic datasets, which are based on additive composite model and rainstorm model, and real-world datasets are used in the literature. The improvements of several models over baselines are summarized. This paper also covers the major unresolved topics in current research. When applied to the actual world, several powerful algorithms still struggle to suppress edge artifacts, weigh definitions, and understand structural subtleties. Open issues including rain water direction, quantity, and future directions are also highlighted to encourage further research.

Several machine learning algorithms for image processing and computer vision applications have been proposed in the past decade. LBP, HAAR are some popular algorithms that are widely used in face recognition and produce excellent results. However, most of these algorithms are not suitable for real-time recognition in unconstrained environments. Recently, the most advanced deep learning technology has become the new favorite of traditional machine learning algorithms. In this paper, the obstacle recognition technology based on optimized deep learning algorithm, based on the traditional GS optimization algorithm, the gray wolf optimization algorithm is used to optimize the image features .Firstly, this paper creates an obstacle image database to provide deep learning data. Create cnn model in MATLAB to learn the features of pictures in the database and identify the categories of pictures.

The quality of transferred embryo is the key factor to determine the success of the in-vitro fertilization (IVF) procedure. Traditionally, embryologists assess the quality of embryos by experience. To some extent, it is a subjective matter. Recently, deep learning methods have been introduced to address the problem. In our study, we applied four network models (i.e. EfficientNet variant, Swin Transformer, STORK, and AlexNet) on quality assessment of the human blastocysts and cleavage stage embryos, respectively. Approximately 20,000 embryo images are used to train and test these models. Experimental results show that the Swin Transformer performs best among the four models, and its accuracy is about 99.5%.

The Special things such as the pandemic are changing people's work and life style, and there additionally has been a remarkable change in the inflow and outflow rules of bike sharing, whether it is the tidal law in time, or the gathering points in space. These dynamic changes bring new challenges to the bicycle scheduling. Based on the flow characteristics of bike sharing during the pandemic, this paper presents a solution for dynamic scheduling of bike sharing: combining K-means and K-medoids algorithm to solve isolated site scheduling problems like nucleic acid detection stations; establishing a scheduling model based on fixed cost and transportation cost between stations; and finding the best scheduling path based on the positive correlation between distance cost and the mileage of dispatching vehicles. The scheme is conducted on the actual operation data of bike sharing and distribution of nucleic acid detection station in a certain area, and the results show that the scheme is feasible and effective.

As an appropriate method for mining negative signals from large-scale unobserved data, negative sampling plays a key role in recommender systems. Existing negative sampling methods, however, either static or adaptive ones, are insufficient to explicitly learn the structural representation information of user-item interactions. In this study, we hypothesize that indirect interactions between users and items, which enrich the structural representation information, could be useful to explicitly yield informative and high-quality negative samples. To fully exploit the indirect user-item interactions to generate negative samples, we construct a recommender system based on improved knowledge graph policy network (IKGPolicy), which works as a valid module to explore high-quality negatives. Specifically, by improving the graph learning module of the KGPolicy model, it is explicit to capture indirect and high-order interactive information, and yield a informative negative item to train the recommender. Experiments on the Amazon-Book, Last-FM, and Yelp2018 datasets show that our proposed method presents significant improvements over both negative sampling methods and knowledge graph-based models. Further analyses provide sufficient insights in capturing and learning the structural and interactive representation information.

In response to the traditional Dijkstra algorithm, the long search time of pheromone and the occasional redundant inflection points When the final path is obtained can lead to the inefficiency of the algorithm due to its many traversed nodes. Therefore, this paper proposes the use of pheromone calculation method to improve Dijkstra's algorithm. The experimental results show that the optimized algorithm can largely reduce the inflection points generated during path planning and reduce the movement cost of the mobile robot path finding.

In the classroom, the behavior of students can directly reflect the teaching quality of teachers. Based on the deep learning model, the analysis and evaluation of classroom behavior can help promote the intelligent evaluation of teaching and improve the teaching quality. The research on the detection of students' behavior in the classroom which is based on deep convolution neural network , has made great progress, mainly focusing on the classification of classroom behavior and the accuracy of detection, but rarely involving the evaluation between learners, teaching quality and detection results. According to the actual needs of teaching quality evaluation, this paper proposes an individual evaluation method for students and a method to evaluate teaching quality in the class, based on the application of deep convolution neural network in target detection and student classroom behavior recognition. This paper uses ResNet50 pre-training model to conduct parameter learning and fine-tuning on the classroom scene behavior recognition data set. The results show that ResNet50 can optimize the parameters faster than other pre-training models, and can effectively identify students' classroom behavior. By calculating the test results based on the evaluation model, the quantitative evaluation of students’ behavior and teaching quality can be implemented.

Gesture recognition is one of the most intuitive and convenient ways of human-computer interaction. Accurate and realtime hand detection is the key to correct gesture recognition. Deep learning is a mainstream solution to visual tasks such as hand detection and recognition, but detection models based on deep learning usually require large storage resources and computing resources, which seriously hinders its actual deployment and application. Therefore, in order to deploy the hand detection model based on deep convolutional neural network to embedded devices with limited storage resources and computing resources, it is necessary to compress and accelerate the hand detection model. In this paper, YOLOv3 was used as the basic model for hand detection, and a channel pruning method based on local sparse rate attenuation training was proposed to compress and accelerate the model. Meanwhile, knowledge distillation was used instead of fine tuning to conduct distillation training on the pruned model, so as to compensate the accuracy loss caused by channel pruning. The model was trained and tested on the Oxford Hand data set. The results show that the compression method of the hand detection model based on the channel pruning and knowledge distillation training of local sparse rate attenuation can realize the compression acceleration of the model while maintaining a high detection accuracy, which provides support for the deployment of embedded devices of the hand detection model.

In this paper, a self-organizing interval type-2 fuzzy neural network based on singular value decomposition and QR decomposition with column pivoting (SVD-QR)method is proposed to solve the identification problem of nonlinear systems. The network model can realize both structure learning and parameter learning. Firstly, in the aspect of structure learning, error criterion and completeness criterion of fuzzy rules are used to judge whether the rules are growing or not, and at the same time, SVD-QR method is used to find the less active rules to delete. Secondly, a sliding-window secondorder algorithm with forgetting factor is used to optimize the parameters. Finally, the proposed model is applied to one typical nonlinear examples for identification. The experimental result shows that the proposed model can produce a relatively compact network structure and has the high identification and the forecast precision.

Wireless Sensor Networks (WSNs) communicate by broadcasting, and are usually deployed in unattended locations, making them vulnerable to attacks. Selective Forwarding (SF) attack is an internal attack and is difficult to detect because of its uncertain packet loss. In order to improve the accuracy and speed of detection SF, this paper uses Deep Belief Network (DBN) to detect compromised nodes. The scheme clusters the nodes and then trains the deep network to get the prediction value of forwarding rate. Based on the prediction error, malicious nodes can be found out. Experimental results show that the scheme can detect all malicious nodes in an ideal condition. In harsh environment, both Miss Detection Rate (MDR) and False Detection Rate (FDR) are around 5% on average. Moreover, it takes a short time to perform a detection, which can reduce the harm of the attack as soon as possible.

Continuous Integration (CI) helps developers to integrate code changes continuously and instantly. Each integration is verified by an automated build (including tests) to detect integration errors as quickly as possible. However, the build process is often time and resource-consuming as running failed builds can take hours until discovering the breakage. Therefore, it is crucial for developers to preemptively detect when the state of the code is most likely to fail at build time. In this paper, we propose a build failure prediction technique based on a deep learning time series model LogTTCN-CIBuild, which predicts the result of the next build by semantic embedding and temporal embedding of the log information and temporal information of the build logs. Extensive experiments on real-world datasets demonstrate the effectiveness of our approach.

With the continuous development of artificial intelligence, it embodies important value in more and more scenarios, such as speech recognition, recommendation systems, computer vision, etc. The deep learning techniques behind them are built on a large amount of data to learn and extract features from different data. The deep learning techniques behind them are built on a large amount of data, learning from different data and extracting features. To protect data security, differential privacy is a good mechanism which is independent of the background knowledge possessed by the attacker and shows excellent results in privacy protection. In this paper, based on differential privacy and combined with knowledge in deep learning optimizer, We propose Adam-DP to improve the learning efficiency by adapting the learning rate and noise, and the adaptive noise also reduces the impact of noise on the model accuracy to some extent. The experiments show that the method has higher accuracy and learning rate compared with the DPSGD method.

The domestic processor DCU(Deep Computing Unit) is a GPU-like accelerator. To address the problem that DCU lacks more effective compilation and optimization tools, a DCU code generation and optimization method based on polyhedral model is proposed. First, a source-to-source compiler C2HIPC is implemented based on the polyhedral model. The input of C2HIPC is a serial C program, and the output is a HIP (Heterogeneous-compute Interface for Portability) executable on the DCU; secondly, to further improve the heterogeneous performance of the program, a kernel partitioning method based on computing unit occupancy is proposed, which completes the kernel partitioning by limiting loop merging in the polyhedral schedule calculation; at the same time, a global memory access optimization method based on data reuse is proposed, By modeling the two kinds of data reuse and global memory access costs existing in the DCU kernel, the global memory access costs in different loop orders are calculated. Using the PolyBench test set to verify the effectiveness of C2HIPC, the performance of the automatically generated HIP code after C2HIPC optimization reaches 92.4% of that of PolyBench-GPU.

The University of Melbourne has gathered data about its past clients for research opportunities that can be used to create a predictive model to estimate the probability of converting a lead to a contract (e.g., research grants, internships, etc). The model can then be used by each faculty in the University to provide business insights for the purposes of improving client engagement, profiling clients and optimizing client relationship processes. This project aims to study which supervised learning algorithms perform well in scoring leads and predicting the probability of success, then to derive further business insights from the best-performing model. A number of machine learning models were developed and the features were tested statistically for their importance. The final results showed that models like random forests and deep neural networks performed relatively well to predict the likelihood of success of each lead. Random forests are less computationally complex and more interpretable than neural networks without sacrificing too much performance, so they may be preferable in this case

Fuzzy cognitive maps (FCMs) are an efficient soft computing tool commonly used for modeling and analyzing complex systems, and have significant advantages in knowledge representation and fast reasoning. At present, many researchers have devoted themselves to solving the learning problem of FCMs. However, most of the existing learning algorithms are prone to local convergence and are not suitable for large-scale FCMs learning problems. Moreover, the learning problem of FCMs is often a complex non-convex optimization problem, which is easy to cause the final learning result of the algorithm to be only a local optimal solution. To address these limitations, this paper proposes a novel algorithm for learning FCMs by combining multimodal optimization and modified covariance matrix adaptation evolution strategy (CMA-ES), termed as MCMA-ES. First, the learning problem of FCMs is modeled as a multimodal optimization problem (MMOP), and then a multimodal optimization algorithm is proposed by combining the detect multimodal clustering (DMC) and CMA-ES, and the cubic chaos factor is combined to enhance the capability of local search. Finally the proposed algorithm is combined with the decomposition strategy to learn FCMs. Experiments are carried out on 5 benchmark datasets, and MCMA-ES is also applied to 15 groups of large-scale gene regulatory system reconstruction problems. The experimental results show that MCMA-ES has high learning generalization and accuracy.

The field of adversarial robustness has gained significant attention in recent years due to the advancements in deep learning and the increasing complexity of tasks performed by deep neural networks. Despite this progress, the underlying mechanism behind the performance of these networks in the face of adversarial attacks remains poorly understood. The existing explanations are largely based on intuition. This study provides a comprehensive explanation of various adversarial attacks by examining the multi-level interactions between input variables in deep neural networks. Utilizing the Shapley value, we believe that these attacks primarily target high-level interactions to mislead the network. The study’s effectiveness is demonstrated through application to re-identification models. Additionally, possible reasons for minor variations in results are discussed.

Knowledge graph completion (KGC) continuously fills out the knowledge graphs (KGs) by adding linkages and entities. Modern models are frequently trained using random walk sequences or negative sampling triples. These methods either neglect the neighborhood information of triples, which makes it challenging to manage long-tailed entities, or they are unable to adaptively extract the neighborhood node attributes that are most pertinent for entities, which introduces duplicate information. The joint knowledge representation model TranSage, which includes pre-trained node features as entity features, convolves the entity neighborhood features to obtain the node neighborhood information, and generates node embeddings, is proposed in this study based on the aforementioned issues. The model can enhance inductivity, improve knowledge representation performance, and address some long-tailed entities' representation issues, according to experimental findings on the KGC task. Its effectiveness confirms the efficacy of aggregated node embedding and outperforms that of the baseline model.

Domestic waste classification is of great significance for building a beautiful living environment. Waste classification is a scientific management method for effective disposal of waste. Aiming at the problem of garbage pollution faced by human beings at present, a garbage classification algorithm based on deep learning is proposed.In this paper, we first build a garbage classification dataset, identify the categories of garbage items in the picture, and then build a garbage classification model. Through the comparison of different models, we test the accuracy of different models on the garbage classification results, and find a more effective and accurate garbage classification algorithm model.The experiment shows that this algorithm analyzes the existing model, preprocesses the data image, and then uses the image enhancement method to increase the robustness of the model used. It improves the model by adjusting the learning rate, label smoothing, and data normalization processing strategies. The optimized algorithm precision increases to 95%, and makes the model have better generalization ability.

Accurate and fast identification of intersections is a key task for UGV. In this paper, we propose an end-to-end millimeter wave radar point cloud based intersection recognition network with graph based classification and radar point cloud with graph attention convolution. An urban scene millimeter wave radar point cloud intersection classification dataset is produced based on the carla simulation platform, and evaluation of the model on this dataset shows that the model can achieve intersection classification well. We investigate the performance of our model for intersection classification in different scenarios and quantitatively analyze the intersection classification accuracy to prove the model effectiveness.

The marine predators algorithm (MPA) is a swarm-based metaheuristic algorithm which present good performance, but suffers from insufficient population diversity. To enhance the population diversity and prevent the algorithm searching for a partial optimal solution, in this paper, we suggest a hybrid opposite-based learning marine predator algorithm. In HOLMPA, the performance is improved by introducing two different opposite learning strategies. The effectiveness of HOLMPA is verified by applying a group of test functions.