This PDF file contains the front matter associated with SPIE Proceedings Volume 12709, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Although the optical character recognition (OCR) is a mature technology in theory and application, some obstacles still exist in studying scene text recognition (STR). The STR that the electricity nameplate text recognition belongs to always is limited by various factors such as the crappy quality of image, the instability of environments, and the differentiation of fonts. To correctly identify the information on the electricity nameplate, we propose a novel architecture which consists of two pipeline networks, an improved text recognition network based on Transformer OCR and an enhanced spelling error correction network based on Soft-Masked BERT. The former ensures that the glyph knowledge is left, and the latter is to preserve the expressions of diversity. To validate the effectiveness of our method, we evaluated it on a self-annotated electricity nameplate corpus and reported the across-the-board performance gains compared to competing prior models. We further discuss the ablation results for dissecting the gains obtained above.

In view of the complex substation environment and the low efficiency and accompanying safety hazards of the traditional inspection method, technical innovation of the traditional inspection method can be carried out by the image recognition technology based on deep learning. This paper presents a deep learning-based recognition technique for transformer and converter image components. First, image calibration is performed on the transformer converter components, then simple features are combined into complex features using deep learning and feature extraction is completed using VGG 16. Finally, part recognition is performed by Faster-R-CNN framework. The proposed transformer equipment image recognition technique in the paper is tested and compared with the accuracy of previous recognition of different number of images. The results show that the proposed deep learning-based transformer/converter equipment image recognition technique can complete the recognition of transformer and converter components, and the recognition accuracy will improve with the increase of recognized images, which can realize the self-learning of image recognition.

Under cloudy weather, PV power will fluctuate dramatic due to the sporty cloud shading. However, most existing PV power prediction models do not utilize the cloud map sufficiently. In this paper, we propose a novel method to improve the prediction accuracy through combining cloud map features and hybrid neural network. Firstly, we extract the static and dynamic features of the ground-based cloud map using image processing techniques. Secondly, we establish an PV power prediction model based on the ensemble empirical mode decomposition-bi-long short-term memory (EEMD-BiLSTM) with numerical weather information and historical power data. Finally, in order to further improve the power prediction accuracy of cloudy weather, we build an error correction model using the cloud map features, based on the light gradient boosting machine (LightGBM) decision tree. Experimental results on the data from a PV plant in western China shows that our method can effectively improve the accuracy of PV power prediction under cloudy weather.

This paper introduced a calibration mechanism that can calibrate the index angle. The control system is designed based on the mechanical body design of the mechanism, which realizes the integration and portability of the calibration device. In the experiment of measuring the displacement angle, the leveler is used to adjust the levelness of the loading mechanism. A bidirectional long-term and short-term memory network (BLSTM) based on deep learning is introduced for artificial intelligence calibration, which processes and analyzes the experimental data. After experimental verification, the online calibration accuracy of the angle sensor has been greatly improved.

Submarine cables can provide energy and communications for offshore platforms, offshore wind farms, and other offshore projects, and are known as the "Deep Sea Lifeline." It is of great significance to ensure the safety and reliability of deep-sea exploration and development to evaluate the health status of submarine cables, give early warning of their possible faults, and carry out targeted maintenance work accordingly. Therefore, this paper proposes a state assessment method for submarine cables based on long short-term memory (LSTM) network. Firstly, a submarine cable state evaluation index system considering dynamic and static data is constructed. Then, the prediction model of marine cable online monitoring data based on the LSTM network is established, and the ultra-short-term prediction value is obtained. Secondly, combined with the inspection and test data, the improved analytic hierarchy process (AHP) and entropy weight method were used to calculate the subjective and objective weights, and then the improved triangular trapezoid membership function and fuzzy statistical test method were used to finally get the multi-period state assessment results of submarine cables. Finally, the marine cable of an offshore oil and gas project is taken as an example to test the feasibility and effectiveness of the method.

Aiming at the problems for electricity policy researchers to obtain policies related to specific entities, a deep learning model based on Bert-BiLSTM-CRF is proposed to address the problem by recognizing related entities. The model encodes the power policy text with characters through BERT, extracts features from the characters vectors through the BiLSTM network, and generates label sequences through CRF decoding. The experiment reveals that the BERT-BiLSTM-CRF model outperforms other standard models for the recognition of named entities in electricity texts, with the highest F1 score of 88.47%.

Dynamic load identification of simply supported rectangular plate was studied in this paper. As an inverse problem, ill-posed problem is prone to occur in identification process. In order to reduce ill-conditioned degree, Tikhonov regularization method was used for load identification, and regularization parameters were selected by ordinary cross validation, generalized cross validation and L curve method, respectively. The influence of response’s signal-to-noise-ratio on load identification accuracy is analyzed from a statistical point of view to provide some suggestions for engineering practice.

Aiming at the problems that the tension of underwater towing cable is not easy to monitor and the calculation time of the time domain coupling analysis process of towing cable is long, a tension early warning system of underwater towing cable based on long short-term memory (LSTM) neural network is designed. This paper considers the influence of single time variable and multiple variables (six degrees of freedom of tugboat) respectively. The tension time series of the key nodes of the towing cable are predicted and analyzed, and the predicted results are compared with the tension data obtained based on OrcaFlex numerical simulation to verify the accuracy of the prediction. Meanwhile, the predicted results are compared with the back propagation (BP) model and the kernel extreme learning machine (KEML) model. The experimental results show that the prediction accuracy based on LSTM neural network has obvious advantages, the maximum determination coefficient R² is 0.8787, and the minimum root mean square error is 4.9082, which can achieve the effect of tension prediction, and has certain significance for the early warning of underwater towing cable tension.

Aspect Based Sentiment Analysis (ABSA) aims to determine the sentiment polarity of a review towards a given aspect terms. One main challenge of this task is to correlate relevant opinion words with aspect terms, especially for the aspect terms consisting multi words. To this end, attention mechanism plays an important role in previous models due to its capability of modeling the relationship between the context and aspect words, and solving the defect of parallelization which can't be realized in conventional neural network like recurrent neural network. However, the attention mechanism may introduce irrelevant noise and reduce models’ performance. To alleviate the above problem, graph neural network is proposed which utilize the syntactic information for ABSA task. Following this line, we propose a semantic enhancement graph attention network model named SEN-GAT, which innovatively introduce the aggregation and aspect-aware attention mechanism. The experimental results on standard datasets show that our proposed model can substantially improve ABSA tasks performance.

Facial expression recognition has always been a complex problem in computer vision. In recent years, with the rapid development of deep learning, some methods based on convolutional neural networks have greatly improved the accuracy of expression recognition. To further improve the accuracy of facial expression recognition, this paper proposes a convolutional neural network facial expression recognition model embedded with attention mechanism. The convolutional layer is used to extract the shallow features of facial expressions, the SE attention module is embedded into the network model to realize the extraction of salient regional features of facial expressions. To effectively prevent gradient explosion and gradient disappearance, and speed up the training and convergence speed of the network, a BN layer is added after each convolutional layer. The convolutional layer is used instead of the fully connected layer, and 1×1 convolution is used to achieve classification, which reduces the number of parameters, avoids the overfitting problem caused by a large amount of calculation, and improves the generalization ability of the model. The experimental results show that the facial expression recognition accuracy of the proposed method on the FER2013 and CK+ public datasets has been improved to a certain extent, which verifies the method's effectiveness.

In view of the lack of interpretability and data dependence of RNN-based deep knowledge tracking model, a deep knowledge tracing model integrating attention mechanism is proposed. First, learn the embedded representation of students' historical interaction, and then learn specific weights based on the topic's attention mechanism to identify and strengthen the impact of students' historical interaction on the knowledge state at each moment to different degrees; comparative experiments will be carried out on two data sets, and the best performance will be achieved in the ASSISTments2012 dataset, and the problem of long sequence dependence will be alleviated to some extent. This model can capture students' knowledge state more accurately and predict students' future performance more efficiently.

Lettuce height is one of the important phenotypic parameters of lettuce. Currently, evaluating plant height information based on RGB images is the most convenient and fastest method. However, these tasks often require complex camera calibration operations or reference objects with known heights, which are somewhat limiting. Therefore, we utilize a deep neural network to solve this problem to directly obtain the height of lettuce from RGB images without camera calibration and reference objects. This paper integrates four high-performance image recognition networks MobileNetV1, Densenet-121, ResNext-50, and EfficientNet-B3 using average fusion and linear fusion methods to predict lettuce plant height. Experiments show that we have achieved good results, the accuracy of the four basic models has been improved, and the average absolute error has been reduced to varying degrees. The smallest reduction is Efficient-Net, which has decreased 0.01 mm, the largest drop is the ResNext-50, which is 0.15 mm lower. After the model is averagely fused, the average absolute error is further reduced to 1.22 mm, and the maximum absolute error is reduced to 6.48 mm, which is not only better than linear. The fusion results are better than the prediction results of a single model.

Primary liver cancer is one of the most lethal cancers worldwide. Polyploid giant cancer cells seen on pathological examination are associated with prognosis of it. However, its quantification requires doctors to interpret region by region on pixel-level images, which is time-intensive and subjective, not been applied clinically yet. For the first time, we proposed the PGCCI index based on the quantity of polyploid giant cancer cells and the size of the tumor region, developed a deep learning based PGCCdet framework for detection and calculation, and proved the validity of PGCCI by survival analysis. In our framework, U-Net and Resnet are used for foreground segmentation and tissue classification respectively. For polyploid giant cancer cells detection, we proposed YOLOCN, a detection network combining multi-scale feature fusion and multi-dimensional attention mechanism, to achieve accurate quantification. For survival analysis, Kaplan-Meier method is implemented. Experiments show that on the self-owned HCCP dataset, the F1 score of detection is 0.814 and the map@0.5 is 0.871, better than other methods. Survival analysis on the TCGA dataset with 323 cases shows that the median survival time of the low PGCCI group was 84.4 months, while the high PGCCI group was 38.3 months. The Pvalue was smaller than 0.0001, indicating a significant survival difference. The results show our framework can accurately detect polyploid giant cancer cells and the PGCCI can effectively predict survival. This framework can change the way of diagnosis of liver cancer with great application potential in the clinical diagnosis and treatment in the future.

Posture changes in pig behaviors are a helpful way of detecting early signs of precision livestock farming. In commercial settings, computer vision-based approaches have been widely used to obtain individual pig health and welfare information such as body condition score, live weight, and activity behaviors. For this, precisely estimating pig posture is a prerequisite, which is an important step in obtaining real-time pig information. In this paper, a cross-stage stacked hourglass network is conducted to solve pig posture estimation in a real feedlot environment. In addition, a cross-stage connection method that increases the flow of information to mitigate possible information loss is proposed, which obtains 2.1% improvement of PCKh than that of the original network. We trained and tested the proposed approach on a challenging pig dataset, which includes not only real pig images but also images from different websites. The results indicate that it is possible to recognize pig posture from images without any manual interference automatically, and it has great potential for application in the early detection of health and welfare challenges of commercial pigs.

In recent years, with more and more high-rises developing, the area of public buildings in our country is growing continuously. Each commercial space needs to decorate independently. The demand for the wall spraying robots is becoming more and more large, and the requirement is becoming more and more high. The spraying method proposed in this paper is based on the normal vector of the wall. The spray gun adjusts its attitude according to the normal vector of the wall, making the axis of the spray gun coincide with the normal vector of the wall, which can achieve better spraying effect. A sliding mode control algorithm based on adaptive adjustment is proposed, which combines adaptive motion control with sliding mode variable structure control to achieve the optimal trajectory tracking effect.

To address the problem, existing graph neural network-based text classification models can only effectively learn pairwise binary relationships between words while ignoring the multivariate higher-order relationships between phrases and their inadequate representation of semantic information and local features in the text context. This paper proposes a text classification model for hypergraph attention networks with multi-feature fusion. Firstly, this paper learns multivariate higher-order relations between words by introducing a hypergraph structure instead of the original graph structure. Second, this paper constructs sequential hypergraphs, syntactic hypergraphs, and semantic hypergraphs based on textual information to enrich the textual representation of the graph neural network, thus compensating for the inadequate information representation of the graph neural network. A dual graph attention neural network is then used to learn the embedding representation of word nodes in the hypergraph and the embedding representation of relational hyperedges, respectively. At the same time, an attention-based text pooling module is used to extract discriminative and critical word nodes to help the graph neural network capture the deep local feature information of the text so that the model can better represent the text information. Finally, an adaptive fusion method fuses the three different types of text features to generate a final text feature representation for more effective text classification. The accuracy of this paper's approach on the publicly available datasets Ohsumed, R8, R56, MR, and 20NG reached 71.22%, 98.42%, 96.55%, 79.82%, and 88.32%, respectively, and the experimental results all outperformed the compared baseline models.

In this paper, a detection model of power stealing behavior based on improved fuzzy c-means clustering was proposed, which was suitable for the situation without a large number of known power stealing user samples. The model includes factor analysis, local outlier calculation based on improved fuzzy c-means clustering, model evaluation and parameter adjustment with ROC curve, and the selection of the best detection threshold. Through factor analysis, the dimension specification of user’s power consumption characteristics was carried out to improve the efficiency of model detection. The paper improved the FCM clustering algorithm with genetic simulated annealing algorithm. Then, the paper detected the user's power consumption characteristics, compared with the existing mature algorithm, the result indicates that the model has a high detection accuracy for electricity theft. The detection model can output the power consumption behavior outlier and the power stealing probability order of all the tested users. The output of the model of detection can detect the power stealing users with high precision, and the efficiency of anti-stealing work can be improve.

The built-in battery of smart electric energy meter is used to realize the functions of information storage, special event recording, clock normal operation and so on. It is the key component of the electric energy meter. However, in the field operation, a high proportion of smart electric energy meters occurred battery under-voltage warning, and these meters are in the service life required by the technical specifications. Research on this problem is carried out. The purpose is to find out the cause of battery under-voltage fault, and put forward improvement measures to improve the service life of smart electric energy meter. The hourly temperature in the measuring box at different installation positions is measured. It indicates that the electric energy meter in operation can be in high temperature for a long time. The capacity prediction model of the built-in battery of smart electric energy meter based on Arrhheninus formula and BP neural network was established. It was found that the self-discharge rate of the Li/SOCl2 battery would increase exponentially when the temperature increased.

As vehicles become more electronic and automated, traditional mechanical vehicles evolve into intelligent networked vehicles with electronic systems. The multi-module electronic systems form complex in-vehicle networks. Gradually, cars are opening up more and more externally accessible interfaces from the otherwise closed networks. These interfaces are connected by the car's internal buses to the control units of other nodes by direct or indirect means, and attackers can penetrate these interfaces directly into the vehicle's critical control systems by hacking into them. Automotive buses play a crucial role in in-vehicle networks, but they are poorly designed with regard to the information security of the network in mind. We introduce the current network security and the main bus technologies for connected cars, analyze the security threats that exist in bus technology, and on this basis study the security strategy for bus technology.

As an important branch of artificial intelligence, deep learning network has made remarkable achievements in image recognition, target analysis and other fields in recent years. At the same time, HRRP has been widely used in target recognition, especially in ship target recognition, because its data is easy to obtain and process and contains more target information. Under the above background, this paper proposes a target recognition method based on DBN network, and analyzes and verifies the performance of the network with the measured data of ten types of military and civilian ship targets. Through the experiments on SVM and DBN in low bandwidth, it is found that DBN model can maintain a high recognition rate even when the amount of target information is reduced, and has a certain robustness. When the attitude angle changes, the recognition performance of DBN model is relatively stable, which overcomes the problem of attitude angle sensitivity to a certain extent, and has a good application prospect.

Aiming at the problem of low mapping efficiency of Simultaneous Localization And Mapping (SLAM) algorithm for single robot, a real-time fusion scheme for multi-robot raster maps based on improved map_meiging package is designed, and the PROSAC algorithm with improved RANSAC algorithm is used to eliminate mismatched feature points and improve the efficiency of map fusion. Firstly, the two single robots construct local maps based on the Gmapping algorithm, and then extract the feature points of the raster map after grayscale processing, match and complete the map fusion after purification, which does not need to predict the initial pose of the mobile robot in advance. Finally, test experiments are carried out in the Gazebo simulation environment to verify the effectiveness, real-time and robustness of the method.

Tidal level is very important in transportation, surveying and mapping, fishery and military fields. An automatic tidal level recognition algorithm based on water gauge images was studied to solve the problem that the water gauge data in ocean observation stations cannot be observed and compared with tidal level automatically. The numbers in front of the water line on the water gauge image were extracted by image segmentation technology and recognized by convolutional neural network. Finally, the automatic measurement of the tidal level of the water gauge is realized with 91.1% recognition success rate which promotes the unmanned and intelligent process of the ocean observation stations.

To address the problems that a single Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network cannot determine the importance of each word in the text when extracting features from the text and the high time complexity of the network, this paper proposes a text classification method based on a multi-channel This paper proposes a text classification model (BWCNN-GRU-LSTM-ATT) based on multi-channel CNN, Gated Recurrent Unit Network (GRU), LSTM and Attention mechanism (ATT) to solve the problem of inadequacy of a single method, so as to achieve accurate classification of civil aviation passengers' emotions. Firstly, we use word embedding to vectorize the input text, extract the local features of the text through multi-channel CNN, and then integrate the full text semantics; make full use of the advantages of GRU and LSTM, and design a GRU-LSTM tandem network model to increase the effective extraction of contextual semantic information; then increase the weight of keywords by adding the attention mechanism to fully extract the features; finally, we adopt the Bayesian optimization algorithm is adopted to optimize the model parameters and construct the optimized sentiment classification of civil aviation passengers. The results show that the classification accuracy reaches 99.2% and achieves better classification results.

The surface of the near-Earth planets, such as Mars, is covered with fine-grained regolith and other objects, such as rocks, craters, and sand dunes. In addition, the gravity of Mars is weaker than that of Earth, making rover operations on Mars more challengeable, such as higher slip, wheel sinkage, and lower tractive force, etc. In this paper, we use the AlexNet to establish a model for identifying the types of the rover rut on Mars. The model is used to predict trafficability based on the images of the rover rut which comprise the wheel-soil interaction. The types of the rut are evaluated by the level of slip ratio. According to the value of the slip ratio, the rut is classified as three types: small, middle, and large range. The dataset used in this work includes 203 samples, which was split into the training and the validation sets consisting of 143 and 60 data samples, respectively. The experimental results show that the AlexNet based on the transfer learning can accurately predicts the type of the rut. The training and validation accuracies obtained by the proposed method are 94.44% and 91.46%, respectively. The proposed technique can be used to improve the ability of environment perception, risk assessment, and trafficability estimation. This work provides the knowledge and strategy in developing future rovers for deep space planetary explorations.

Skeleton-based human action recognition has emerged as a vibrant area of research in recent years. In action recognition based on skeleton, the method based on graph convolution network is through modeling space to explore physical dependence between body joints, has obtained the remarkable performance, however, most GCN method uses the stratification to aggregate a wide range of field information, which makes the joint characteristics in long distance transmission is weakened. In this paper, dense graph convolutional networks are proposed to overcome these shortcomings. To enhance the informative skeleton features and create a compact representation at an early fusion stage, an MMF is developed. Then, our proposed method utilizes a dense graph convolution operation to improve the local context information of each joint. Finally, feature channels and action classes are specific, In our model, it is crucial to assign weights to important features and focus on the relevant information. Therefore, to improve the identification of features in our model, we introduce a channel joint attention module. A large number of experiments on three large datasets (NTU-RGB+D, NTU-RGB+D120 and Kinetics) have demonstrated the performance of the proposed MMDGCN for skeleton-based action recognition.

Classroom action recognition is a hot research in educational teaching environment combined with artificial intelligence technology, through which action recognition can quantify students' concentration in classroom. In this study, we propose an advanced 3D and 2D branching fusion action recognition model for recognizing students' actions in the classroom environment. The experimental results show that the classroom action recognition model has high accuracy and speed, and can accurately and quickly identify specific high-frequency classroom actions, which can then be effectively used for students' classroom concentration analysis. Accurate identification of students' classroom actions and quantification of their classroom attention can not only provide reference for teachers of relevant courses, but also provide data support for teaching quality assessment at school level, ultimately helping to improve the quality of classroom teaching.

Tactile sensors integrated into the tip surface of smart prosthetic fingers are extremely important for grasping objects in patients with physical disabilities, as the tactile perception system can recognize characteristics such as roughness, material, softness and hardness. These depend on the sensor having texture recognition function. Therefore, this paper is designed to prepare a tactile finger with different shapes of bionic fingerprints as well as a tactile finger without fingerprint structure to carry out texture recognition experimental research to determine the fingerprint structure with the best accuracy of texture recognition. The results show that the tactile finger without fingerprint structure cannot recognize the surface texture of the object, while the tactile finger with fingerprint structure can recognize the texture spacing with a percentage error of 10% or less; the square bionic fingerprint tactile finger can recognize the texture spacing no matter sliding in x direction or sliding in y direction, and the percentage error between the actual texture spacing value The percentage error is only 0.69%, which is much lower than other shapes of bionic fingerprint, and the output voltage signal has higher vibration amplitude, so the square bionic fingerprint can significantly enhance the frictional vibration signal generated during the sliding process compared with other shapes of bionic fingerprint structure, and improve the texture recognition performance of the tactile finger.

The constrained multi-stage machine flow shop (FS) scheduling model with segmental and quadratic fuzzy numbers such as processing time, job weight and breakdown time is studied in this paper. The model becomes clear with the closed interval approximation method of piecewise quadratic fuzzy numbers, which introduces a non-crossover optimization sequence into the regarded solution and minimizes the makespan in fuzzy conditions. Illustrated by a numerical example, the method presented in this paper can provide the decision maker with an applicable solution that is relevant to the real problem and minimizes the makespan.

With the increasing shortage of energy in today's society, the rapid development of the multi-energy complementary microgrid system came into being, at the same time, the industry is changing day by day, and gradually formed a multi-operator participation regional integrated energy system. Due to the profit-seeking nature of capital and the urgent needs of users, society urgently needs to establish a multi-agent interaction mechanism and corresponding optimization methods for regional integrated energy systems. This article takes the energy hub as the starting point, beginning from the significant basics of the regional integrated energy system-multi-energy complementary microgrid, firstly proposed by the energy hub operator (EHO), the user cluster with photovoltaic (PV) and the energy storage operator (ESO) composed of multi-energy complementary microgrid system operation interaction mechanism, and then proposed a multi-agent multi-subject complementary microgrid system based on energy hub game model, using particle swarm optimization algorithm to solve. Finally, the energy hub is simulated, a specific case is proposed, and the proposed optimization model is used to solve, at the same time, the proposed method is analyzed in detail. Begin the abstract two lines below author names and addresses.

With the rapid development of the Internet era, the demand for data traffic has exploded, and this dramatic growth has brought higher requirements for spectrum resources. Unlicensed long term evolution (LTE-U) technology is considered one of promising solutions to address the problem. However, the unlicensed band is rich resource, which is mainly occupied by wireless fidelity (WiFi). The biggest challenge for the LTE-U technology is how to achieve friendly coexistentence with WiFi. Many Researchers in academia and industry have proposed various solutions to deal with this problem. In this paper, we propose a classification method based on self-attention mechanism algorithm in order to distinguish unlicensed LTE and WiFi to further realize the friendly coexistence of the two signals. The proposed algorithm can enhance the feature extraction capability of the model by fusing the features extracted by the convolutional neural network through the self-attentive mechanism. Simulations show that our proposed algorithm has higher accuracy and stability compared to other algorithms at the same signal-to-noise ratio. When the signal-to-noise ratio is higher than 0 dB, the recognition accuracy can reach above 90%. Especially when SNR id -5 dB, the accuracy of self-attention network is 12% higher than CNN’s.

As the E-Booster is widely used, its stability and safety are receiving more and more attention. However, the nonlinear characteristics causes interference in describing the system characteristics. In this paper, a nonlinear least squares (NLS) method is proposed to extract the hydraulic characteristics, and then the control is corrected by predicting idle stroke to improve the performance. First, the reasons that lead to the deterioration of the braking performance after the vehicle passes over a washboard surface are analyzed. The open-loop system gain is changed due to the increase in clearance, and it is demonstrated that the system performance becomes worse or even unstable with the help of Nyquist plots and describing functions. A piecewise linear function with three parameters is then proposed to characterize the dynamic system. From a complete period of increase and decrease in pressure, the parameter values can be obtained by NLS. Taking the three parameters as input, the support vector machine (SVM) and neural network (NN) methods are used to predict the current idle stroke. Finally, the simulation results show that NN has higher classification accuracy and better generalization capability. The nonlinear compensation to the controller can significantly improve the performance of the system.

Effective image semantic segmentation is an important task. In recent years, deep convolutional neural networks have played a major role in improving the effectiveness of semantic segmentation. However, previous works capture local contexts by multi-scale feature fusion which ignore the global dependencies. Moreover, as the layers deepen, the loss of feature is inevitable. Multi-Scale Dual Attention Refine Network (MSDARNet) is proposed. Specifically, we firstly append two types of attention modules on top of dilated FCN, which model the global semantic interdependencies in spatial and channel dimensions respectively. Secondly, multi-scale feature fusion is added to extract local dependencies in the above attention module. We conduct extensive experiments to validate the effectiveness of our network on three challenging semantic segmentation datasets and obtain great improvements. In particular, our MSDARNet attains mIoU accurary of 50.3% on NYUDv2, 71.1% on Person-Part.

Unsupervised person re-identification (Re-ID) focuses on studying the issues of person Re-ID in the open world where there are no available annotated labels for pedestrian images. The original transformer only learns global features. It ignores local features and fails to consider high-level local information on images. Therefore, we introduce a transformer-based multi-branch unsupervised person Re-ID framework, named GLTReID, with learning more robust features of pedestrian images. Specifically, we split the original transformer into one global branch and two local branches, with learning global and local features in parallel. The global branch learns global features of pedestrian images, and two local branches learn local features of different parts of pedestrian images. Numerous experiments are conducted on several person Re-ID datasets, including CUHK03, Market-1501 and MSMT17. Numerous experiments denote that our method is better than most unsupervised learning methods and unsupervised domain adaptation methods in person Re-ID.

Recently, neural networks have been used for handwritten Chinese character recognition (HCCR). While we have seen very high recognition accuracy for constrained Chinese handwriting, it remains a challenging task to recognize unconstrained Chinese handwriting. This paper proposes a new network architecture based on convolutional neural networks (CNNs), called RugbyNet. A special characteristic of the network is the rugby-like structure which is capable of reducing the number of multiply-accumulate operations and the required storage. And the convolutional blocks in the rugby-like structure are guided and designed by the speed, a direct indicator of computational complexity, to keep the model efficient while reducing the cost of memory access. Besides, We further proposed three data augmentation methods, extrusion, expansion, and distortion, to generate synthetic samples from the fixed sample set. These synthetic samples can then be used for training. Experimental results show that the recognition accuracy of the proposed RugbyNet on the ICDAR-2013 offline HCCR competition dataset reaches 97.77%, and enables a relative 4.70% error reduction compared to the current single-network method MCANet trained only on handwritten dataset.

New retail is now a new retail model based on the new technology of the Internet era, online and offline and logistics integration. With the change of traditional retail model and user consumption habits and thus the core of the consumer, the new retail to provide quality services to consumers deeply into all walks of life. Due to the special nature of tobacco products, the traditional retail brand promotion channels are relatively limited, and the method of on-site promotion by brand managers with the help of retailer stores no longer meets the requirements. Under the new retail model, tobacco offline retail terminals can use big data analysis and recommendation technology to intelligently recommend cigarettes and personalized cigarette brands with similar styles and tastes to users based on the characteristics of new retailing, to guide customer demand and improve user experience; thus targeting marketing, reducing marketing costs and improving revenue. We propose the HLKAR model, an end-to-end recommendation model that integrates knowledge graph and attention mechanism with hybrid link prediction, to realize personalized cigarette recommendation for users in offline tobacco retail terminals.

In the conversation system based on background knowledge, the core step is knowledge selection, but the existing model often has the following problems: the dialogue jumps too much before and after, and the answer is not answered by the topic; the dialogue is repeated, and the same sentence is constantly generated as a reply; the conversation contains too little information. Therefore, in response to these problems, in this paper, we propose a new model based on the research of previous researchers, which can combine the theme of the dialogue and the difference in knowledge selection. Experimental results show that our model has been improved in terms of BLEU and ACC values, and the generated content fully contains background knowledge information, and there will be no large jump without repeating.

The form of war has changed from platform operation to system operation, and joint operation has gradually become the trend of development. The operational network is the core of joint operations. At present, the research in this field is mostly focused on the performance evaluation of traditional combat networks. Military operations put forward higher requirements for the intelligence of combat networks. The research on intelligent combat network modeling is of great significance for improving the combat capability of systems under modern conditions. In this regard, we have changed the topology of the traditional combat network, established an intelligent combat network model, and proposed an intelligent combat mode of composite laying of the original combat network under the cooperation of the polymerization nodes, and setting up virtual links to assist in combat. The polymerization node makes the battle command power sink, the network organization mode changes from centralized to distributed, and makes the network have a rapid response mechanism of marginal response; virtual links enhance the survivability of the network, and two types of elements enhance the intelligence level of the combat network. It lays the foundation for the research of intelligent warfare and provides support for the development of intelligent warfare theory.

To address the problems that the traditional two-stream network has limitations in effectively fusing spatial-temporal information and the extracted temporal information hard to capture complex motion patterns, which resulting in low recognition accuracy, an improved two-stream convolutional neural network method for human motion recognition is proposed based on multi-attention mechanism. Firstly, in the temporal network, the C3D network is used to replace the original two-dimensional network to solve the problem that the temporal information cannot be extracted effectively; secondly, in the spatial network, multi-scale convolutional Transformer encoder is coded based on the contextual relative position, features are integrated adaptively at different scales under the action of the adaptive scale attention mechanism. Experimental findings conducted on the UCF101 dataset indicate the approach described in this paper performs better in terms of the number of parameters and recognition accuracy.

In order to improve the anthropomorphic nature of intelligent speech products, the academic research on speech emotion recognition is getting hotter and hotter. Currently, the speech emotion recognition system mainly consists of two steps: speech feature extraction and speech feature classification. In order to improve the accuracy of speech emotion recognition, the Mel Frequency Cepstrum Coefficient (MFCC) of speech signal, which has a good effect on the feature capability in the field of speech at this stage, is chosen as the input of the deep learning network, and the ResNet-BiGRU network based on the attention mechanism is used to extract the MFCC information is extracted using ResNet-BiGRU network based on the attention mechanism. The experimental results show that the introduction of attention mechanism in the model can effectively focus on useful information and reduce the interference of redundant information. The accuracy rate on the Chinese sentiment corpus CASIA reached 84.83%.

Soil environmental remediation technology has been an important tool to solve environmental pollution, and the technology mainly focuses on the effective degradation of polycyclic aromatic hydrocarbons (PAHs) organic substances. And this kind of substances will volatilize toxic substances during the whole degradation process, which brings some difficulty to the research of environmental remediation process, but only effective understanding of the degradation process can propose more effective degradation and analysis methods, so an effective research method is needed to complete the real-time monitoring of the whole degradation process. In this paper, we propose a digital twin-based visualization early warning and analysis system, which only needs to collect images of the experimental process to produce a three-dimensional visualization twin platform, on which sensors can be monitored to obtain data, and the data can be analyzed and stored in real time. With this system, effective monitoring and analysis without contact can be achieved, bringing a new effective monitoring and analysis method to environmental remediation technology. The digital twin platform system adopts the RGB threshold limiting function of openCV to identify sub-buildings, save the topology and location information of sub-buildings' 2D pixel coordinate arrays, reorganize them with 1D arrays containing height, and then form a 3D digital twin base map with almost zero cost and real-time change through software, and then embed the sensor locations and transmitted data to finally form a digital twin platform that can display data in real time and monitor and warn in three dimensions. The study of this system continues to improve soil remediation solutions by analyzing the results, giving early warnings by comparing the collected data and thresholds, and also proving that interventions can be made by mapping virtual to real in areas with high levels of pollution.

Voice controlled multi-color lamp, combining components such as speech recognition chip, single chip microcomputer, LED lights and relay, applies the voice control technique to the lighting circuit in family oriented setting. In this paper, based on STC11L08XE single chip microcomputer and LD3320 speech recognition chip, a voice controlled multi-color lamp is designed and manufactured, and its function is verified This lamp enables the users to give speech commands to change the lightening modes. It not only can illuminate, but also enriches the users’ life by bringing the fun and convenience of technology.

Since the outbreak of COVID-19, the widespread use of face masks has prompted significant research efforts in face mask recognition. Based on the previous work, it is crucial to develop a face mask recognition system swiftly using offthe-shelf techniques, while it is equally important to ensure that the system's performance and efficiency meet the requirements of real-life scenarios. In this paper, we propose a two-fold face mask recognition method, by first detecting faces in an image using YuNet and then recognizing whether the faces are wearing a mask or not using MobileNet. After validation on a self-built dataset, compared to similar related works, our method can effectively improve the face detection rate, even in situations of occlusion and blurring, and the detection and recognition efficiency has been significantly improved too.

In the motor vehicle inspection scenario, accurate identification and positioning of vehicle nameplates is an important prerequisite for automated robotic inspection operations. In view of the current problems that vehicle nameplates are mainly checked manually and the level of intelligence is low, a target recognition and localisation method based on binocular vision combined with improved SURF algorithm is proposed. Firstly, feature points are extracted based on the SURF algorithm, and target recognition is achieved by combining the nearest neighbour ratio method with the RANSAC algorithm based on the gradient constraint of the parallax; then the edge contour of the target in the template image is obtained based on the Canny edge detection method, and the target shape centre on the template image is obtained based on the contour, and the left and right scene map target shape centres are obtained using the single-response matrix mapping; finally, target localisation is achieved by combining the principle of binocular vision 3D reconstruction. The experimental results show that the method has good recognition efficiency and accuracy, and can successfully identify and locate the target, which has certain application value.

Aiming at the problem that features of communication emitter data without label information are different to extract and the classification acquisition is not high, this paper cites the contrastive learning theory, constructs a residual network with two parameters sharing as the backbone network, conducts contrastive learning on the rectangular integral bispectral features of signal augmented samples, and further extracts the feature presentation with more differentiation. To this end, the feature separability between samples from different emitters is enhanced. Then, the new features extracted are used for contrastive learning at the cluster level to complete the tasks of classification and identification. Compared with other unsupervised learning algorithms, the proposed method achieves a better identification accuracy of about 78% through experiences on the dataset of measured ultra-short wave communication stations.

With the rapid development of power devices, the problem of thermal reliability of devices becomes more and more prominent, so how to accurately obtain the thermal resistance of devices is particularly critical. The traditional structure function method is the most effective method to analyze the thermal characteristics of power semiconductor devices at present, but the existing deconvolution algorithm has some problems, such as noise, uncertain number of iterations and long time. This paper presents a new structure function method based on ANN. Build model through ANN, and the sample set measured by Thermal resistance tester is used to train the ANN. Finally, the experimental results of discretization prove that discretization can be used as the output variable of the structural function method of artificial neural networks.

Deep neural networks have made significant improvements in pixel-level semantic segmentation. However, the existing semantic segmentation algorithm still faces the problem of weighing between the accuracy and calculation cost of the segmentation. In response to this issue, this article proposes a hybrid network structure (HAM), looking for a balance point in the calculation accuracy and calculation speed. In this method, we construct a dual-attention module. The role of this module is to guide high -level characteristics through underlying characteristics to obtain more context information. Among them, the shape flow branch retains low-level space details, and semantic flow branches capture senior context information. These two branches are fused to strengthen information dissemination between different levels, thereby achieving higher segmentation accuracy. Experiments on the dataset show that this method achieves higher accuracy and speed under relatively small parameters. Compared with other real-time semantic segmentation methods, our network has achieved good compromise between parameters, speed, and accuracy.

An embedded vision ranging system is designed and developed to solve the problem of workpiece positioning. Firstly, this paper has a theoretical analysis of distance measuring based on binocular vision. Then proposing an improved stereo matching algorithm CSGM based on SGM. Also having an algorithm performance testing for CSGM on its accuracy of stereo matching and time consumption based on the Middlebury datasets by using the evaluation toolkit MiddleEval3. After the stereo calibration of the binocular camera and rectifying the stereo image pair, a distance-measuring experiment was carried out, showing the embedded vision ranging system has a millimeter precision for positioning workpieces.

Aiming at the defects of the existing part recognition methods in the industrial production process when identifying small and dense parts, and the problem that the number of model parameters and large volume are difficult to deploy on the mobile platform, an improved YOLOv5 lightweight part target recognition model is proposed. On the basis of the YOLOv5 model, the Ghost module is first used to replace the backbone network to reduce the number of parameters of the model. Secondly, a SPP_F feature pyramid structure is constructed to speed up the training speed of the model and strengthen the ability of the model to extract information. Remove the detection layer for large target scales, thereby speeding up model training and detection; Finally, the GIOU function in the YOLOv5 model is replaced with the Alpha-SIOU function to improve the positioning accuracy of the bounding box. Experimental results show that the average accuracy (mAP) of the improved model reaches 99.5%, and the volume only accounts for 3.2 MB, which is 35% faster than the original YOLOv5 model, reducing hardware costs and meeting the needs of deployment in low-computing equipment.

Text style transfer is mainly to modify the text style to suit various application scenarios without changing the semantic meaning of the text, which is a great significant issue in natural language processing. To expedite research progress, this survey conducts a systematic review of existing literature. Drawing from a vast body of research, this survey first extracts the essential connotations of style information in text of varying granularity across different tasks, and then provide a clear definition of text style transfer, summarize the main challenges at present and comprehensively codify and discuss the current datasets used for evaluation, as well as their indicators. Moreover, this survey compares the mechanisms, advantages, and shortcomings of unsupervised classical methods across word, sentence, and paragraph levels. Finally, the future directions of fighting are also given, hoping to facilitate more comprehensive solutions.

The operation status of the power distribution network is mainly reflected by the line loss and line loss ratio, which are reduced to facilitate the efficient use of electrical energy. The user electricity theft behaviors are harmful and are considered to be important aspects affecting the line loss rate. The identification of electricity stealing users is conducive to reducing the losses of power companies. In this paper, the user electricity consumption behavior evaluation index system is constructed by combining the causes of the user’s electricity consumption, line loss, and abnormal. The index system can dynamically reflect the power consumption behavior of users and help the fine management of power system. An anti-theft system based on Levenberg Marquardt (LM) neural network model is proposed to identify the user stealing electricity behavior. The 4-month electricity consumption data of 138 users of State Grid Shanghai Electric Power Company is taken as a research object. The classical decision tree model Classification and Regression Tree (CART) is compared, and the results showed that the accuracy of LM neural network model is as high as 97.2%, and it has more advantages in terms of training speed and computational accuracy. This study is beneficial for power enterprises to accurately and quickly find suspected users of electricity theft, , so as to reduce the loss of power energy and the economic loss of electric power enterprises, and promote the further improvement and development of the electric power system

As a major part of clean energy, wind energy plays an important role in the process of world energy transformation. As offshore wind power has the advantages of high wind density and large power generation capacity, it has gradually become an important approach for countries to use wind energy. However, wind power output is random and unstable. While its construction scale is expanding, it also has unstable factors on the grid, including the impact of offshore wind power access on the grid voltage stability. On the basis of considering the static voltage stability, this paper uses the static power flow method to analyze the impact of offshore wind farm access on the grid voltage, and determines the output limit of offshore wind power and the static voltage stability limit of the grid according to the power flow convergence conditions, and uses the actual data of Zhejiang Province power grid for simulation.

Under the renewable energy power consumption guarantee mechanism, the power purchase and sale decisions of electricity retailers should not only pursue the maximization of benefits, but also assume the responsibility of renewable energy consumption. In this paper, we propose a model for power purchase and sale decisions of electricity retailers that takes into account the responsibility of renewable energy consumption: based on medium and long-term transactions and spot transactions, we introduce renewable energy power purchase agreements (PPA) and green certificate transactions; coordinate the TOU price and energy storage leasing strategy; the conditional value at risk is used as the risk evaluation index, and establish a combined purchase and sale decision model with the comprehensive trade-off of expected benefits and risk losses as the decision objective. The example analysis shows that enlarging the peak valley price ratio and reasonably developing energy storage leasing can improve the expected income and reduce the risk loss of electricity retailers, which verifies the effectiveness of the model.

The frequency regulation problem of power system with high proportion of new energy under marketization is prominent. The comprehensive frequency modulation (FM) performance index can reflect the FM capability of the FM service provider. In this paper, on the basis of the existing FM rules, the Analytic Hierarchy Process(AHP) is used to optimize the weights of the sub-index of the comprehensive FM performance index in the frequency control ancillary service (FCAS) market, and the winning results before and after the optimization, the clearing and settlement price is compared with the comprehensive interests of the market. It is concluded that the uniform weights value can not reflect the FM performance of the units well. After the weights value optimization, the comprehensive FM performance index can better reflect the FM performance of the units and get better clearing result.

As the development of electric vehicles is receiving more and more attention from more and more countries, it has become an important research idea to make use of the huge regulation potential of electric vehicles to participate in the frequency regulation auxiliary service market. Therefore, this paper proposes an optimization strategy based on distributionally robust chance constraint for electric vehicles to participate in the day-ahead frequency regulation auxiliary market. First, a virtual energy storage model of electric vehicles is established to evaluate their frequency regulation potential; then, a day-ahead optimization strategy for electric vehicles to participate in the frequency regulation auxiliary market is established based on distributionally robust chance constraint considering the uncertainty of regulation signals; finally, the effectiveness and feasibility of the proposed strategy are verified by arithmetic examples.

An urgent problem in the process of substation anomaly detection is the scantiness of abnormal substation equipment samples. Conventional anomaly detection methods which use object-based detection are difficult to distinguish every type of equipment and unknown anomalies, leading to poor generalization. A substation anomaly detection method based on lightweight Siamese network is proposed to solve these problems, the distance between the image to be measured and the image of normal will be calculated in the method for detecting anomalies, so our method is able to deal with unknown samples, and data enhancement methods such as rotation, mirroring, and symmetry are used to expand the number of negative samples. The improved Siamese network adds ECA attention module to decrease the interference of background environment, ECA module performs channel feature learning on the compressed feature map, multiplies the original input feature map and the channel feature maps, and finally outputs the feature map with channel attention. The full connection layer calculates the distance of feature vector instead of the conventional Euclidean distance or cosine distance, expanding the difference value between positive and negative samples, and clarifying the threshold selection. This method has a good universality, it can meet the detection of various background environment and has high accuracy of detection work.

Applying optimization techniques to hull shape and mooring configuration designs has become an effective way to reduce the life-cycle cost of a floating offshore wind turbine (FOWT). This paper aims to investigate the influence of met-ocean data integrity on the FOWT dynamic performance optimization. Two optimization loops with different levels of met-ocean integrity are set up. In the first loop, the FOWT dynamic performance is evaluated by the long-term response indicators, e.g. the lifetime accumulative fatigue damage at tower. The long-term responses are predicted based on an integrated 25- year met-ocean database which includes 73000 3-hour wind and wave data. In the second loop, the dynamic performance is evaluated by the short-term responses, which are predicted based on the simplified met-ocean data. It includes 41 representative load cases that are selected from the 25-year database. Both loops employ the NSGA-II multi-objective optimization algorithm to find out the Pareto designs that have the best trade-off between dynamic performance and manufacturing cost. By comparing the two optimization results, it is found that the Pareto designs of the first loop tend to have a smaller tower base fatigue than the second. The first loop has smaller mooring fatigue in the high manufacturing cost region, while an opposite trend is seen in the low cost region. These findings demonstrate the importance of metocean data integrity to FOWT optimization.

Flexible electronic technology has laid the foundation for complex human-computer interaction system, and has attracted great attention in the field of human motion detection and software robot research and development. Graphene has high electron mobility, high thermal conductivity, excellent mechanical properties, and is an ideal material for preparing flexible conductive networks. In this study, a simple preparation process of fractal flexible sensor based on graphene ink is proposed. The fractal structure is printed with PDMS graphene ink using Ecoflex 00-30 as the substrate by using direct ink writing method. After curing, it has high tensile and piezoresistive sensing properties. The structure thickness is 2 mm, which is convenient for assembly and adhesion. The sensor has been tested under different bending angles, showing excellent sensitivity and wide response range, and can be used for attitude recognition. In addition, repeatability and good mechanical properties make it have potential applications in wearable devices, electronic skin and soft robots.

Oral infusion system is an automatic injection system that can replace manual injection for precise injection and solve the problem that traditional manual injection is easy to cause patients' abnormal pain. Based on the analysis of the functions of the oral infusion system, this paper determines the design idea of using PMSM as the driving unit. On this basis, a model reference adaptive composite control algorithm based on fuzzy adaptive law is proposed to realize the speed sensorless control of the system. The simulation results show that the speed response of the motor is very fast, the speed stability with load is better than the traditional control algorithm, and is equivalent to that with speed sensor. Finally, the dynamic response of the infusion system is simulated, and it is concluded that the system can achieve high-precision and reliable control of the infusion rate to ensure that the drug can enter the patient's body with uniform speed, accurate dosage and safety, and meet the clinical needs.

MPSOC chip is the heterogeneous multiprocessing platform chip with the most extensive application and the highest stability in the global market. It has important applications in artificial intelligence, autonomous driving, and other fields. The MPSOC has a flexible power domain inside, and multiple power rails make its dedicated power module design very important. As long as the accuracy of the power supply is guaranteed, the safety of the chip can be guaranteed. A new power module is designed through the combination of DCDC, LDO, and PMIC to make the power supply more flexible and reliable and to provide more security for the application of MPSOC.

Reliability and accuracy of electric energy metering are the basis to guarantee fairness and justice of electric power market. In China, there are high power users equipped with self-supplied power supply, and the users generally adopt the operation mode of double-bus loop to ensure safety. However, when the power generation changes within a certain range of the load consumption, two metering points of double bus will have a situation of reverse power flow. In the conventional trading mode, the positive and negative electric energy of the metering point cannot accurately correspond to the electricity price of difference in electricity price between the on-grid price and the off-grid price. This paper analyses the electric charge transaction of the metering point under the condition of double-bus loop operation, and designs a calculation method of electric charge based on digital sampling. Through experimental verification and engineering application, it is proved that the calculation method of electric charge based on digital sampling can effectively solve the problem of inaccurate calculation of electric charge, and has certain engineering value.

The safe and stable operation of substation equipment is becoming more and more important, especially in recent years, the construction of smart grids has correspondingly increased its requirements. Because of the different locations and severity of substation equipment failure, infrared thermal imaging technology can quantitatively determine the hidden danger or the degree of failure. The practice has proved that this is an effective means of substation equipment fault diagnosis. In the working environment of power equipment, there will be a situation that the equipment will be covered, which will lead to the absence of targets in infrared imaging. In view of this situation, this paper randomly chooses some normal samples, constructs different missing levels and puts them into the test set for experiments. It also compares the classification effects of various algorithms in different missing levels. The results indicate that the classification accuracy of this paper is significantly lower than that of normal conditions, and it is still superior to its similar algorithms under different sample missing levels. The results also show that the method is robust and the Zernike moment invariants have certain adaptability when the target region is destroyed. It is of great significance to enhance the operation safety and economic benefits of substation equipment, reduce the maintenance cost of equipment and economic losses, and ensure the reliable operation of substation equipment.

In the field of radiation calibration of micro-optical loads, the stability of the light source is a technical problem that researchers must solve, the stability of the light source determines the accuracy of the radiation calibration results, at the same time, the stability of the light source is determined by the stability of the driving power supply, so based on the driving requirements of the radiation calibration light source, a tungsten halogen lamp driving power supply with an output current stability better than one thousandth of a percent is designed as a tungsten halogen lamp light source The thesis first analyses the advantages and disadvantages of switching power supplies and linear power supplies, and selects the linear power supply as the power supply structure according to the requirements, and analyses the working principle of the linear power supply, and then designs the corresponding circuit structure based on the working principle; finally, the experimental platform is built for power supply experiments, and after testing, the experimental results show that the output current stability of the power supply reaches After testing, the experimental results show that the output current stability of the power supply reaches three ten thousandths, and the stability of the light source reaches one ten thousandth when driving a 5W tungsten halogen lamp, which can meet the requirements of the tungsten halogen lamp driving power supply index.

Accurate Photovoltaic power (PV) power forecasting is the basis and key to grid dispatch management. With machine learning algorithms and the latest swarm intelligence algorithms being proposed, a reasonable combination of the two will produce good prediction results. This paper addresses the problem of optimal selection of hyperparameters for the XGBoost algorithm in the PV power prediction problem. This paper establishes an XGBoost long-term PV power prediction model based on the optimization of ISMA algorithm, firstly, the dataset is pre-processed and the training set and test set are divided, then the data is trained, the model with the best practical optimization results is predicted, and finally the predicted results of the model are compared with the classical models. The experimental results show that the optimized XGBoost PV power prediction model based on ISMA algorithm can achieve better PV prediction results with MAE of 0.154, MSE of 0.256 and R² of 0.964, which are better than other models.

Irony is the use of counterwords, irony and other techniques to reflect the essential characteristics of people and things, and is a way to express emotions. Irony detection is an important research content in the field of sentiment analysis, and it is also one of the hot issues in the research of artificial intelligence in natural language processing. In this paper, a sentimental ironic detection model based on multi-feature fusion is proposed, which takes text features, image features and image attribute features as inputs, and image attribute features are represented by adjective-noun pairs abbreviated as ANPs. In this model, the neural network is first used to extract features from the three inputs, and then the text features and image features are fused through the shared fusion network, and the contrast fusion network fuses the text features and image attribute features, so as to express the contrast and difference of graphic and text information. Finally, satirical classification is carried out through the classification layer. Experimental results show that the model achieves good performance on the public multimodal irony detection dataset.

DC distribution network has a good consumption effect on new energy, and DC fault detection is one of the key technical problems for the large-scale application of DC distribution network. Aiming at the problems of difficult threshold setting and long detection time in existing fault detection schemes, a fault detection scheme for DC distribution network based on generalized regression neural network is proposed. Firstly, the transient voltage change rate of positive and negative lines is used as the fault start criterion. Then the disturbance of DC distribution network is eliminated by low voltage protection. Secondly, the Spearman correlation coefficient is used to analyze the correlation of the transient voltage of the positive and negative pole lines to distinguish symmetric faults and asymmetric faults. For symmetric faults combined with the transient voltage change rate of positive and negative pole lines, bipolar short circuit faults and AC faults can be distinguished. For asymmetric faults, GRNN is trained offline with simulation data, and the fault poles are distinguished according to the online output results of GRNN. The fault detection scheme is verified on the MATLAB / simulink platform to build a simulation model of the DC distribution network at both ends, which can quickly and accurately identify the fault.

Inductive program synthesis is the process by which a computer automatically constructs a program that satisfies all the examples based on a given set of input and output examples. To address the problem of incomplete description of the program statute by examples, we proposes a two-stage program synthesis method and proposes an MA-FRENet model based on a multi-headed attention mechanism, which calculates an attention score based on the similarity of the example state to the state of the example executing a program that satisfies some of the examples by introducing a multi-headed attention mechanism, thus enhancing the discrimination of representative example features with complete and accurate constraint descriptions. Compared with the baseline synthesis accuracy, the experimental results show that the method is effective.

Point cloud data has been widely used in many fields, and processing them using deep learning methods has become a popular research topic. However, the irregularity and unordered data structure of point cloud makes it necessary to design neural networks that are different from those used in image and natural language processing to accommodate its characteristics. In this paper, we propose a novel network called the Point Distance Mask Attention Network. This network utilizes the position relationship between the centroid and the generated neighborhood to weight the features of each point, making the network more focused on the features close to the centroid. Additionally, we use a masking operation to ensure the permutation invariance of the network when using the ball query method to find the neighborhood. Furthermore, we propose a residual-like connection in the network architecture, which achieves better results without changing the network feature extractor and depth. We evaluated the network model on the ModelNet40 dataset and achieved an accuracy of 93.7%. Our experiments show that our network achieves better results than the original baseline network by 1.1%, with almost the same parameters and inference time.

At present, the power system Al model has been initially applied, but the model lacks standard system verification. In order to improve the quality of the power system simulation Al model, this paper builds a set of Al model verification platform based on the performance test method of cloud computing. This platform is based on the B/S structure, and uses the powerful features of cloud computing to achieve the test and evaluation of Al models, improving efficiency and accuracy.

Based on the analysis of magnetoresistive sensor vehicle detection principle, and the characteristics of output signal, an improved adaptive threshold finite-state machine vehicle detection algorithm is proposed. This algorithm realizes the online real-time update of the reference value and threshold, meanwhile solve the problems such as: the baseline drift caused by environment; the accuracy reducing caused by the fixed threshold; the indistinguishableness of leaving vehicle caused by the intermediate-state machine. The actual road experiment shows this algorithm can be used in the magnetic vehicle detection system, it's accuracy could be over 97%.

The application of welding robots has effectively improved labor productivity, improved the labor intensity of workers, and has good stability. Visual sensing technology can be empowered for intelligent welding automation, which can help welding workstations to complete welding tasks more accurately. It can detect the shape, size, position and welding quality of the welding seam. This paper describes the structure of the intelligent welding robot workstation system, the carrying mode of visual sensing, and the working principle of various visual sensing technologies in welding seam positioning, welding seam tracking, welding quality detection, etc. Detailed descriptions of the current mainstream line structure light method, point structure light scanning method, multi-line structure light method and surface structure light method are given respectively. These technologies have their own characteristics, and need to be matched with the application scenarios in actual application. Welding vision technology can help enterprises better control welding quality, improve production efficiency, reduce production costs, and provide more possibilities for the development of welding industry.

In this paper, a method for predictive analysis is proposed based on BP neural network optimized by genetic algorithm and BP neural network optimized based on particle swarm algorithm. By optimizing the initial weight and threshold of the neural network by the algorithm, showing the error between the predicted value and the real value, and comparing the mean squared error of the three, the MSE of the GA-BP prediction model is reduced by 56.53% compared with the BP neural network prediction model. The MSE of the PSO-BP network predictive model was reduced by 50.00%. The results show that this method can accurately predict the thickness of the blasting protective water belt, which provides a new and reliable method for the study of controlling blasting flying objects.

Aiming at the problems of vague information acquisition, insufficient accuracy, and poor robustness of a single sensor, this paper investigates the information fusion algorithm of multiple sensors. Firstly, the motion of the moving body is modeled using staged CV, CA, and CT models, and the complete motion of a single maneuvering target is simulated based on 2D planes or 3D space. Secondly adaptive implementation of Kalman's nonlinear filter estimation in maneuvering targets under a single sensor based on the Interactive Multi-Model (IMM) algorithm. The multi-sensor IMM results are then estimated and fused based on the Simple Convex Combination (SCC) fusion algorithm. Finally, the single-sensing and multi-sensing fusion performances are simulated and evaluated by experiments to prove the effectiveness of the designed track fusion algorithm under multi-sensing.

As dispersed power sources become more common and power markets open, power system operation becomes increasingly difficult. One is source-grid uncoordination. When several dispersed power sources are linked to the grid, their uncontrollability and unpredictability may cause an imbalanced power system, impacting grid security and stability. Hence, power system optimization research focuses on how to coordinate scattered power sources and power grids for synergistic source-grid optimization. This work addresses the imbalanced power issue generated by dispersed power sources linked to the grid and proposes a source-grid co-optimization strategy based on the two-layer hybrid optimization theory to ensure safe, dependable, and cheap power system operation. We will apply a hybrid particle swarm algorithm and genetic algorithm double-layer nested method to solve a two-layer optimization model with the integrated cost of distribution network on top and distributed power supply on bottom. With imbalanced power, cooperative source network optimization may increase power system operating efficiency and stability.

In recent years, machine learning has made considerable progress in various fields. The training of deep learning algorithm model requires a large amount of data, and a large amount of data annotation is very costly. In this case, the use of simulation data can provide great convenience. Using the three-dimensional reconstruction algorithm to construct a three-dimensional model of the real scene can speed up the construction of simulation data. The model has problems such as model damage and model fragmentation. The repair model takes a long time and the repair effect is poor. To solve these problems, we designed and developed a simulation training data generation system. The system uses Blender as a development platform, introduces a three-dimensional reconstruction algorithm to initialize the model, transforms the model into a point cloud for repair operation, and proposes a point cloud region adjustment method and a point cloud fast repair method. Finally, the refined model is used to generate a data set with true values. The experimental results show that the system we designed can improve the efficiency and accuracy of data generation and efficiently generate simulation training data for machine learning.

In order to fully understand the carbon emission of grid users and improve the accuracy of carbon emission forecasting results, this paper proposes a carbon emission forecasting method for grid users based on the convolutional neural network (CNN) and long-short term memory network (LSTM). The mapping relationship between energy consumption data and carbon emissions is explored by taking the historical energy consumption data of grid users as samples. Then the high-dimensional mapping relationship of carbon emission variables is extracted based on the convolutional layer and pooling layer of the CNN network to construct a high-dimensional time series characteristic vector, which is input to the LSTM network. A carbon emission prediction model is established based on CNN-LSTM by training LSTM network parameters. Through the actual data verification, this research finds that compared with a single CNN, LSTM and ISPO-BP, the MAPE and RMSE of the CNN-LSTM model are reduced to 8.48% and 0.0526, while the R2 is improved to 97.33%, which indicates that the model constructed in this paper has a significant advantage in the accuracy and generalization ability of carbon emission prediction for grid users.

Based on the application requirements of physical feature recognition of integrated circuits, this study will use YOLOv5, are presentative of the YOLO series algorithms, to improve and optimize them from the theoretical and experimental perspectives, in order to obtain a faster and more accurate detection model in precision-type target detection tasks. This experiment will detail the basic architecture of the YOLOv5 model, and make up for the shortcomings of the YOLOv5 series model in terms of detection accuracy and break through the model bottleneck by appropriately replacing core components, introducing spatial attention mechanism, channel attention mechanism, and improving the loss function and nonlinear function.

Multi-sink wireless sensor networks are widely used because of their prominent advantages in extending network life and balancing energy. However, the current research on cascading failures of wireless sensor networks mainly focuses on the impacts of network topologies or routing schemes, and rarely involves the impacts of multi-sink placement. Therefore, this paper proposes a realistic multi-sink-oriented cascading model, in which a multi-oriented betweenness is designed to describe the load distribution of sensor nodes and wireless links considering sink nodes, and the overload dynamics of nodes and links are considered at the same time. On this basis, several topological metrics are considered to quantify the placement properties of sink nodes, and then how they affect the cascading robustness under node attacks and link attacks is investigated. Experimental results show that node attacks are more likely to trigger cascade failures. More importantly, no matter what kind of attack the network suffers, if sink nodes are placed as hubs, and these nodes are decentralized, the networks can better resist cascading failures.

In the management of power grid infrastructure construction projects, the management of construction sites plays a very important role. At present, the company's construction site management business lacks supporting management tools and software, and it is difficult to achieve timely adjustment of the current construction with traditional evaluation methods. To this end, this paper studies the sensing information collection and virtualization construction of smart grid infrastructure scenarios. This paper establishes an objective and intelligent evaluation index system, pure evaluation standards and evaluation methods, and introduces and evaluates the impact on construction, which can improve the efficiency of construction and has very important practicability. Experiments show that the BiLSTM+CRF model using the word embedding vector as input can effectively capture long-distance semantic dependencies, which can not only improve the accuracy of named entity recognition, but also greatly reduce the training time of the neural network model. It shows better performance in mining tasks.

With the continuous acceleration of development in recent years, energy has become an important strategic reserve, and energy-saving technology and environmental protection concepts have become important research directions in development. In automatic transmission, low efficiency, high energy consumption, easy to generate electric sparks are important problems at present. This project is a wireless direct drive circulation transport device, which uses non-contact power transmission to supply power to the mover in real time. The dynamic coil structure, contact (sliding touch) power supply, easy to produce electric sparks and powder and other defects, and the dynamic magnetic structure needs to lay a large number of drive controllers and position sensors for stator segmental switching power supply control along the way, resulting in complex control systems and high cost defects.

The electric power industry is the key field of carbon emission reduction, and promoting the carbon peak of the electric power industry in advance is of overall significance to the realization of the “double carbon” goal across the country. This paper proposes a power system optimal allocation method that couples the power CO₂ emissions with the traditional power system operation simulation and aims at minimizing the total CO₂ emissions, and takes the power system of Shandong province as an example for empirical analysis. Taking the current power status data of Shandong province as the input parameter, a power system production simulation model that minimizes the total carbon emissions while achieving the goal of reaching the carbon peak is constructed, and the carbon emission trajectory and the optimal allocation results of various power sources and flexible resources are obtained; At the same time, the Mann-Kendall trend test method is used to verify and analyze the carbon emission trajectory, which proves that the carbon emission trajectory has a significant downward trend, which also verifies the correctness and effectiveness of the method proposed in this paper.

The traditional transmission line disaster monitoring equipment has some problems, such as relatively single function, high environmental requirements for the deployment of monitoring equipment, and the accuracy of monitoring data needs to be improved. In this paper, the high-precision positioning ability of Beidou is applied to the real-time monitoring of the wire position of transmission lines, and the on-line monitoring equipment of wire position based on the integrated application of Beidou is developed. Reducing the volume of traditional equipment, reducing the weight of equipment, effective integration of different monitoring functions, easy installation and high reliability can greatly improve the efficiency and accuracy of the operation and maintenance of overhead lines.

Substation personnel behavior detection is mainly based on the two-dimensional image taken by the camera for discrimination. The substation has the characteristics of complex environment and multiple types of personnel operations, which leads to the low accuracy of two-dimensional image of personnel behavior discrimination. To solve this problem, this paper studies a human skeleton point detection model based on mask initialization. First, the operator image in the substation is processed by using the joint detection and segmentation network Mask R-CNN to obtain the boundary frame and mask of the human target, and then the target area is detected from bottom to top. Experiments show that this method not only reduces the impact of complex environment and job types on the detection accuracy, but also greatly reduces the invalid convolution operations, thus improving the detection speed.

At present, the detection of pavement diseases mainly focuses on improving the algorithm to improve the accuracy of pavement disease detection, but there are few studies on the measurement of pavement disease size. This paper mainly proposes a new measurement method, which combines the YOLO algorithm with RealSense D415 to measure the depth, and after binarization, dilation and erosion and flood fill of the pothole image, the external rectangle is obtained, and the length and width are calculated by the rectangle pixel difference. The experimental results show that the measurement effect of the pothole size is very good, which can meet the requirements of engineering applications. Among them, the average relative error for length measurement is only 3.58%, followed by width is 7.86%, and the average relative error of depth is slightly larger at 9.38%.

Accurate estimation of the state of health (SOH) of Li-ion batteries is an important guarantee to ensure safe and reliable operation of battery systems, and is also a key indicator for battery management systems. In this paper, a method for SOH estimation of lithium-ion batteries based on double-stage attention-based gated recurrent unit (DAGRU) is proposed. Firstly, depth wise separable convolution (DSC) was used to extract the health features (HFs) of battery charging voltage. Secondly, the model uses double-stage attention, which enables more information on input features to be obtained at both the temporal and spatial scales. Finally, NASA battery dataset was used to verify the proposed method. Experimental results show that the proposed method can accurately estimate the SOH of lithium-ion batteries.

Genetic Programming (GP) algorithm has been successfully applied to classification tasks for many years, with a long history and remarkable effects. It is worth noting that GP tends to evolve biased classifiers when it is used for class imbalanced datasets. Biased classifiers are often unreliable and will seriously affect the classification performance. In addition, when GP is used to train the model to adapt to the high-dimensional imbalanced datasets, it is easy to be affected by the majority class and ignore the existence of the minority class. Therefore, classification tasks based on GP in high-dimensional and imbalanced datasets need to be further studied. In this paper, we propose a GP approach for imbalance classification tasks with the aim of improving the classification performance and minimizing the time consumption, named HGPWOA. A multi-objectivization fitness function is adopted to solve the problem, supplemented by Whale Optimization Algorithm (WOA) for feature selection, and then feature augmentation strategy is adopted to improve the classification efficiency. The method was tested on five imbalanced datasets. The experimental results show that this method is superior to the traditional classification algorithms, which provides a new way to study the class imbalance problems.

Aiming at the problem of measuring the height of surrounding obstacles using the monocular fisheye camera of AVM system in parking scenarios, a method of measuring height through inverse perspective transformation and multiple calibration is proposed and verified by simulation experiments. First, the distortion correction of the fish-eye camera and the calculation of the homography transformation matrix are performed to complete the inverse perspective transformation, and the transformation relationship between the original camera image and the vertical plane image is obtained. Secondly, the YOLO-v4 target detection algorithm is used to identify obstacles and return the pixel height of obstacles. Then, according to the calculation formula of monocular camera height measurement and the relationship between pixel distance and object distance established by calibration, the method of measuring obstacle height is determined. Finally, the effectiveness of the method is verified by simulation experiments.

In recent years, the novel coronavirus has been raging around the world, which has brought a huge impact on the lives and health of people around the world and social and economic development. Therefore, the research on pneumonia detection methods is of great significance. At present, the pneumonia detection method based on deep learning has attracted the attention of scholars at home and abroad because of its excellent effect, low cost, and fast detection speed; however, there are still many problems in this technical field, such as insufficient detection accuracy, slow response speed, detection Poor reliability, no end-to-end implementation of segmentation and detection; this paper proposes a binary tree cycle attention convolutional neural network (BRA- CNN), the network can identify two discriminative regions at the same time at each layer, forming a binary tree structure, which greatly reduces the disadvantages of the complete failure of the rear network due to the recognition error of the front network, and designed a dedicated loss function for the network, make the network training process go smoothly.

A recognition method based on the enhanced Transformer model is proposed to solve the task of human abnormal action recognition in surveillance videos. Video Swin Transformer (VST) is used to extract video features, and the 3D Adaptive Spatial Pyramid Pooling (3DASPP) module is used to enhance video features. Human body detection is performed on the key frame in the video through the target detection algorithm, and the video features corresponding to the target are extracted. Finally, the human body action category in the video sequence is identified, and whether there is an abnormality is judged. The mean Average Precision (mAP) is used as the evaluation metric. Experimental results show that the proposed algorithm can effectively recognize abnormal human actions in videos, providing strong technical support for intelligent surveillance, intelligent security, and other related fields.

We propose a real-time infrared pedestrian detection method, which achieves higher performance than the state-of-theart YOLO series in infrared pedestrian detection. Our method is based on the YOLOv6-n with some new technologies. In particular, Mixed Downsampling Module (MDM) and Cross-Stage Connection Structure (CSCS) are designed to reduce the inference latency of the network. In addition, Simplified Spatial Attention Module (SimSAM) is proposed to improve detection performance with negligible overheads. We validate our method through extensive experiments on FLIR detection datasets. Experimental results show that our method can achieve 84.86% AP with inference latency of 0.54 ms on an NVIDIA 1080Ti GPU.

Based on capsule network and virtual adversarial training, a multi-label text classification method is proposed for the multi-label text classification problem to solve the problems of insufficient feature extraction and weak model robustness. This method used BERT model to obtain the word vector representation of the sample and used the dynamic routing mechanism of capsule network to extract text features. At the same time, the method added disturbance construction adversarial samples into the input samples, and updated the model parameters by virtual adversarial training method, finally, improved robustness and classification effect of the model. Experiments on AAPD and RCV1-V2 datasets show that the micro-F1 value is better than those of other benchmark models. Experimental results show that the proposed method can be effectively applied to multi-label text classification tasks.

In order to solve the problem of analysis and warning of distribution network planned maintenance outage range, a method of analysis and early warning method of power outage range for planned maintenance of distribution network based on logic coding is proposed. Firstly, the outage range analysis model is established through the planned outage range information matrix and the planned outage range associated equipment information matrix. Secondly, the real-time outage comprehensive alarm logic information of the distribution transformer is described through the distribution transformer power failure alarm information logic matrix and the real-time logical coding information matrix of planned outage range associated equipment. Then, the expected standard outage comprehensive alarm logic information of the distribution transformer is described through the standard logical coding information matrix of planned outage range associated equipment. Finally, the difference between the actual outage range and the planned outage range in the maintenance process is described by the planned maintenance outage range alarm information matrix and the planned maintenance outage range difference value. This method can timely alarm the abnormal power outage range, guide relevant personnel to eliminate defects in time, and avoid the occurrence of dangerous events, which has good practical significance.

Due to the complex physicochemical properties of lithium-ion batteries, it is difficult to identify the internal changes that cause battery degradation and failure. Ultrasonic testing, as a non-destructive characterization method, has the advantages of high sensitivity, low cost, convenient use, and fast speed, and has great potential for application in battery characterization. Nowadays, the application of ultrasonic waves for state characterization of lithium-ion batteries has achieved initial success, but research on building capacity degradation models for lithium-ion batteries using ultrasonic testing is still limited. In this paper, a capacity degradation model for lithium-ion batteries is proposed based on ultrasonic non-destructive testing technology. The established capacity degradation model is used for extrapolation to reach the failure threshold, providing theoretical support for data-driven lithium-ion battery life prediction and health management.

To reduce the transmission latency of the grid differential protection scenario, a low-latency data transmission technique is proposed in this paper. A multi-path low latency routing optimization strategy based on task assignment is considered to find the optimal and sub-optimal transmission paths in the current network environment with the transmission latency as the weight, and then jointly assign the number of tasks to be transmitted on both paths according to their link bandwidth, node queue length, and other network resources to achieve multi-path task transmission and minimize the task transmission latency. The simulation results show that the proposed multi-path task transmission has lower transmission delay and higher network resource utilization than single-path task transmission and evenly split task transmission.

RGBT tracking receives increasing interests due to its flexible application in all-day and all-weather environments. However, the training of deep RGBT trackers usually relies on large-scale aligned RGBT pairs, which usually require high human labor and time cost. Considering the strong commonality and specificity of multi-modal data, we propose a novel two-stage learning framework to capture modality-shared and modality-specific features using large-scale unpaired RGBT data, and thus achieve state-of-the-art performance in RGBT tracking. In specific, in the first stage, we aim to learn the modality-shared representations and thus design a generic transformer network which only requires mixed modal data for training. In the second stage, we aim to learn the modality specific representations and achieve adaptive back-propagation using unpaired data. To achieve these goals, we design a modality transformer network, in which two modality encoders are used to capture modality-specific features and a modality-adaptive attention module is designed to enforce the interchange of information between different modalities in a separate-gather way. Since the two training stages do not rely on paired or aligned multi-modal data, the power of unpaired multi-modal data is unveiled in the training of deep RGBT tracker. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our method against state-ofthe-art RGBT trackers.

With the further deepening and upgrading of digitization and informatization of industrial enterprises, the industrial network needs higher flexibility and reliability. As a comprehensive, unified, and standard new industrial Ethernet technology, Time-sensitive network technology makes industrial network technology and industrial ecology more open and dynamic. Time-sensitive Software-defined network introduces the concept of software-defined network based on time-sensitive network, which makes the whole network more real-time and flexible. With the emergence of dynamic applications, time-sensitive networks also need to be updated in real-time. The old update mechanism cannot ensure that the new and old flow is separate from one another, resulting in network congestion. In this paper, a new scheduling update scheme is proposed, which eliminates the conflict between the old and new traffic by separating the old scheduling scheme from the new. The experimental results on two different network topologies show that the update time of the proposed scheduling scheme is shortened by 5.68% compared with the traditional scheduling scheme. The network resource utilization of the two network topologies decreased by 3.3% and 10.02%, respectively.

The application of combat network in modern warfare makes combat operations have overall effectiveness. However, it is often impossible for commanders to observe the complete network topology of the enemy. The link prediction of the combat network can help the commander to reasonably implement the distributed network operation, which has important military value. This paper proposes an Adaptive and Interpretable Probabilistic Spreading (AIProbS) method for combat network link prediction. It controls the resource propagation in the ProbS framework through the network representation method of determinable dimensions and interpretable elements, thus simulating the information transmission in the real battlefield environment. Through experiments, the effectiveness of the method we proposed are proved.

Stance detection estimates the stance of a user towards a specific target through the text that the user posts. As an important step in public-opinion mining, stance detection has great potential for wide application and provides necessary data for decision-making by governments and enterprises. On social media, people exchange ideas by replies, in which the texts of replies are closely related to the context. Existing works in this regard focus mainly on the method of aggregating the context, but overlook the fact that the theme of texts alters with the progress of the response. Therefore, we propose a stance detection model that can effectively learn the thematic information, where contrastive learning is used to enhance the learning effect of thematic information. Experiments were performed on CreateDebate and Kialo datasets, and our proposed model achieved an accuracy of 68.0% and 69.6% on the two datasets, respectively. Experiments have proved that our method can improve the performance of stance detection.

Convolutional neural networks have shown amazing recovery in image super-resolution in recent years. Despite the great success of CNN-based methods, deploying these model methods to devices with low computational power is difficult due to the large number of parameters to which the super-resolution methods of recent years have been applied. To address this problem, this paper proposes a multi-scale progressive feedback pixel attention network. Specifically, a multi-scale progressive feedback attention block is designed to complete the attention structure using a multi-scale feedback module, which references the architectural approach of the self-attentive mechanism in the design of the multi-scale module by removing its secondary complexity and adopting The multi-scale module is designed to accept all vectors as input and to learn in an incremental manner. Depending on the depth of the network, the MPFAB is able to retain information at different depths and can adequately fuse information from networks of different depths to mirror each other. Experiments show that the proposed MPFAN has better reconstruction results than existing methods with a smaller number of parameters.

In order to solve the problem of measuring the size of the truck carriage and positioning the vehicle, a lidar-based non-contact dimensional measurement device for truck carriage is designed. The measuring device is composed of a vertical two-dimensional laser radar, a four-degree-of-freedom manipulator, a sliding rail and an industrial computer. The point cloud information of the length and width of the carriage is obtained by the laser radar. The linear features are extracted by the probabilistic Hough transform algorithm, and then the outline segments of the carriage are extracted according to the prior knowledge, so as to measure the length, width and height of the carriage and position the vehicle. The measurement result of the carriage model shows that the device is characterized by small measurement error, high measurement efficiency and simple structure and so on.

Digital signage is a visual management tool. Through the visual interface, the data is organized and presented with various charts, and various information and problems are displayed intuitively. In order to grasp the current situation and necessary information in time and guide the rapid development of relevant measures. In order to improve the management level of distribution network fault and power outage business, a digital dashboard for distribution network fault and power outage information analysis based on multi-purpose digital technology is proposed. This paper discusses the design ideas of digital signage, constructs the technical route, and analyzes the application effect of digital signage. Combined with the actual situation of distribution network fault and power outage business, digital signage can effectively reduce the probability of distribution network fault and power outage events, improve power supply reliability, and meet business management requirements.

When a nuclear power plant is in normal operation, the generated electric energy is divided into two circuits. One is boosted and transmitted to the State Grid and the other (about 5%) is used for the electrical equipment in the nuclear power plant after voltage reduction. As an important part on the user side of smart grid, the Energy Management System (EMS) is of great significance for reducing energy consumption. Meanwhile, under the unique environment of high humidity and high salinity of nuclear power plants, Heating, Ventilation, and Air Conditioning (HV AC) accounts for most of the energy consumption except nuclear power unit. An EMS optimization decision problem is usually formulated in multiple timescales. In order to make the electric dispatching for electrical equipment (represented by HV AC) except nuclear power unit and improve environment comfort, a two-timescale optimization decision model is built in this paper. According to the characteristic of nuclear power plant, the state relationship of two-timescale is in mean value type, and the consistency of two-timescale is guaranteed. With the actual data of a nuclear power plant in China, the results of case study demonstrate the validity of the proposed model and it can satisfy the environment comfort with energy limited.

Nuclear energy is one of the most economical, environmentally friendly, and efficient energy sources. However, the development and utilization of nuclear energy also have significant risks, the most typical of which is the occurrence of nuclear accidents. The radiation release caused by nuclear leakage will seriously affect the personal safety of surrounding residents. Vehicle evacuation is one of the effective methods of large-scale evacuation. A vehicle scheduling optimization model based on the dual constraints of tin1e and exposure dose is proposed in this paper for the uncertainties of road resistance, dose field variation, and personnel distribution in real scenarios. The road resistance formula is introduced to modify the vehicle travel time, and the dose field is meshed to realize a real-time update of the dose rate. A genetic algorithm solves the model to obtain vehicle quantity allocation and route arrangement. The model reasonably arranges vehicle schedules and routes to assist in emergency decision-making. With the actual data of the Fuqing Nuclear Power Plant in China, the case demonstrates that the evacuation time and exposure dose are reduced by 44.17% and 40.22%, respectively, compared with the traditional solution. This research improves the safety performance of nuclear power and promotes the sound development of nuclear power.

With the development of new power system, traditional and new problems in distribution network are becoming increasingly prominent and facing unprecedented challenges. Soft open point (SOP) can provide key technology and equipment support for the implementation of intelligent distribution. In order to solve the needs of high proportion distributed generation access, power quality improvement, self-healing control and operation optimization in the intelligent distribution network, this paper studies the coordinated control technology of distribution network based on SOP, and finally, the digital real-time simulator RTDS is used to verify the relevant strategies.

Edge computing node has large data flow and large resource occupation, which is not conducive to data transmission and processing, and is not conducive to the application of edge computing in the distribution of IoT. This paper proposes an optimization algorithm for edge computing nodes based on the deep Q-learning network theory. The node deployment problem is modeled according to the goal of minimum energy consumption and minimum traffic of the network to solve the optimal strategy. The performance of the algorithm is verified by setting different network sizes and different requests. The test shows that the method used can solve the problem of excessive resource consumption on the edge side.

As a major discipline in the field of fluid mechanics, flow control is commonly used in academic research and engineering applications. However, because the flow problem is more complex as a nonlinear problem, it has been very difficult to carry out closed-loop control in its research process. Machine learning has developed rapidly in recent years, where genetic programming provides an optimization method for solving flow control problems. This paper reviews the successful cases of combining genetic planning with flow control in the experimental field by showing the application methods of genetic planning and flow control, and analyzes the advantages and disadvantages of genetic planning in flow control applications. The article describes the overall overview of the application of genetic programming in flow control, but because the application of genetic programming in flow control is still in its infancy, through this paper we also hope to promote the application of genetic programming in the field of flow control engineering.

Drainage emergency vehicles are used as drainage equipment for rapid emergency response, and the performance of its pumping system plays a vital role in emergency rescue. This paper analyzes the design and working conditions of the pumping hydraulic system, and selects the load sensitive system as the hydraulic power system. Using Amesim simulation software to build a simulation model of the drainage emergency vehicle pumping system to verify the working characteristics of the load sensing system, and obtain the change curves of the pump outlet flow and pump outlet pressure. The results show that when the output flow is stable, the pressure difference between the pump outlet pressure and the load is constant as the set pressure and the load is constant as the set pressure; the output power of the load sensitive pump always matches the power required by the load.

Redirected Walking (RDW) technique allows users to walk indefinitely in a limited physical space while keeping the feeling of real walking. Developers use RDW controllers to manage RDW techniques based on physical and virtual environment information. However, traditional RDW controllers still suffer from many problems. For example, generalized controllers are less optimized, and scripted controllers are difficult to handle unexpected movements. Based on reinforcement learning, we present a novel RDW controller that allows the user to explore complex and large virtual environments while minimizing the number of collisions with obstacles in the physical environments. Our RDW controller directly prescribes the translation, rotation, and curvature gains by analyzing real-time information of the physical environment. The simulation-based experiments show that our controller significantly reduces the number of resets caused by collisions between user and obstacles of physical spaces compared to steer-to-center (S2C) and current state-of-the-art controllers using reinforcement learning.

At present, the air compressor unit in the fast unloading system of the semi submersible lifting and disassembling platform adopts the power frequency step operation mode, which has low automation degree and large energy consumption. This paper proposes an optimal control method for the semi submersible platform fast unloading system. Based on the fractional order PI^λD^μ controller and the particle swarm optimization theory, a fractional order PI^λD^μ controller based on the particle swarm optimization algorithm is designed to optimize the operating frequency of the variable frequency air compressor. It is of great significance to excavate the potential of energy saving through frequency conversion transformation and optimization of frequency conversion control algorithm.

With the rapid development of society and increasing improvement of economic level, a lot of energy consumption is required. However, the output of vine ranks forefront in the world in Chinese agricultural field. If the vine and straw belonging to the renewable energy are utilized, they can effectively replace the natural energy, and save a lot of nonrenewable mineral energies. Therefore, while the characteristics of substance of vine are controlled, it is very necessary to deeply study and implement the processing technology of vine crushing. In order to provide the pre-crushing processing for vine so as to make it the high-grade fuel, the existing crusher shall be improved at present. The variable speed control system of the crusher is modeled and simulated by Matlab/Simulink software based on the hydraulic speed control system as the control core, and the feasibility of the design is verified through repeated test.

This article is based on the methodology logical framework approach (LFA), developed by the USAID organization, suitable as well to guide large-scale equipment manufacturing project evaluations. The main application of the indicator system is to determine the feasibility in evaluation of a new civil aircraft project, from the view of enterprise‘s “project portfolio” management. In the mean time, LFA could also help enterprise management level to monitor the projects in process, as review techniques, to help make decisions whether to resume or end a project in time.

The calculation of gas radiation properties has important applications in many fields such as infrared target detection, remote sensing atmospheric transmittance calculation and so on. Traditionally, the line by line calculation method is used for the calculation of gas radiation properties. However, this method is inefficient for the calculation of large scale flow field radiation data, so this paper proposes a method using a pre-training agent model based on gradient boosting decision tree algorithm instead of the line by line calculation method. After experimental comparison, it is found that the new model based on gradient boosting decision tree algorithm greatly improves the computational efficiency compared with the traditional line by line calculation method on the basis of guaranteed computational accuracy, and greatly reduces the storage space occupation compared with the pre-tabulation method. It can replace the line by line model to a certain extent when dealing with the needs of extensive and complex flow field calculations and it can greatly reduce the consumption of computing time.

The Organic Rankine Cycle (ORC) system with the advantages of high thermal efficiency, environmental friendliness, simple structure and availability of components has been used to recover medium-low grade renewable energy. However, the ORC system involves many operation parameters, and the system performance is the embodiment of the coordinated matching of various parameters. In order to analyse and optimize the low temperature ORC system with the evaporation temperature from 346K to 400K, the ORC performance prediction model using BP (back propagation) neural network has been carried out. The power output of the system has been analysed by combining BP neural network model and genetic algorithm (GA). The results showed that the system power output increases with the increase of the evaporation temperature and mass flow rate, while with the decrease of the condensation temperature and pressure. The maximum power output of the system is 16.3kW, and the corresponding condensation pressure, rotational speed, condensation temperature, mass flow rate and evaporation temperature are 0.1 MPa, 3126 rpm, 290.5 K, 0.4 kg/s and 389.9 K, respectively.

The pointing accuracy of a telescope is a fundamental performance metric, and the prevalent method for correcting telescope pointing deviations involves selecting stars that are uniformly distributed across the sky, measuring the difference between the telescope's observed calibration position and the expected position from a star catalog, and then applying correction models to compensate for the deviations. However, the uniformity of traditional star selection algorithms can be suboptimal. To address this issue, this study proposes a star uniform selection algorithm based on maximizing Kozachenko-Leonenko entropy (MKLE). Specifically, the proposed algorithm employs KL entropy computed based on the k-nearest neighbor approach as a metric of uniformity, where a higher KL entropy indicates more uniform star selections. To adapt the algorithm to the cylindrical surface formed by the telescope's azimuth and elevation axes, a KD-tree search strategy is proposed. Furthermore, to enhance the uniformity of selected stars, a simulated annealing method is utilized to maximize the KL entropy and select the most uniformly distributed stars. The performance of this method was verified by calculating the KL entropy using MKLE, rectangular grid method (RGM), and first-order self-organizing selection method (SO). The results indicate that the MKLE algorithm outperforms the other methods in terms of uniformity of star selection.

To address the problem of target recognition of the tongue body during tongue image extraction. A method to improve YOLO V4 target detection is proposed to detect the tongue. The tongue region of the image is detected by the YOLO V4 target detection algorithm combining MobileNet V1 with an attention mechanism. The network combining MobileNet V1 can well reduce the size of the model, but may reduce the accuracy, which can be well compensated by imposing an attention mechanism. The experimental results show that the proposed algorithm achieves 91.98% mAP on tongue detection, which is 3.66% higher compared to YOLO V4, 17.7% higher single detection speed, and 80.7% reduction in the amount of parameters of the model, which has a better detection and localization effect, and has some guiding significance for the subsequent research on embedded devices for tongue diagnosis.

Real-time UAV monitoring is an important means of battlefield reconnaissance, and machine interpretation of UAV images has become the main form of image interpretation, so the merit of the algorithm becomes an important factor limiting UAV reconnaissance. To address the problems of insufficient graphics card arithmetic power, low detection accuracy and difficult deployment of algorithm models at the embedded end of UAVs, this paper proposes an improved lightweight target detection algorithm based on YOLOv5s, adding K-means++ algorithm and CA attention mechanism module to the original algorithm, and training the improved YOLOv5s-CA network using tank dataset, and the simulation results show that: the improved YOLOv5s-CA has an mAP value of 97.50%, an F1 value of 0.96, and an FPS value of 74.8, which can be deployed on UAVs for real-time detection.

For the existing target detection algorithm with large number of parameters and computation, which cannot be broadly applicable in the vehicle detection field, a lightweight vehicle detection network model is designed in this paper. The vehicle detection model mainly uses RFB-PANet to replace the PANet feature fusion network of YOLOv5s to obtain rich feature information to make the model more accurate, and uses GhostConv structure to optimize the traditional convolutional layer of YOLOv5s to have less parameters and computation of the model, so as to achieve the purpose of lightweights vehicle detection model. The experimental results show that proposed vehicle detection network model reduces the number of parameters by 25.0% and the computational effort by 18.75% compared with the original YOLOV5 network model. The mAp is increased by 2.1% and 1.2% on Pascal VOC dataset and vehicle dataset, respectively, thus achieving a good balance of detection accuracy and light weight

Online car-hailing is now an essential mode of transportation for people to travel. The rule engine system plays an important role in the matching mechanism of online car-hailing orders and drivers. The existing network car-hailing rule engine system is mostly implemented in the form of manually written configuration files. However, manual handwritten configuration files are prone to errors, and the efficiency of manually writing rule configuration files is low. This method needs to be faster to calculate rule matching in high-concurrency scenarios. This paper proposes a rule engine system based on the bitmap algorithm and inverted index mechanism. Different rules are divided into corresponding rule templates at the rule configuration level according to the practical dimension, which users can automatically generate rule expressions according to the templates. The system reduces the difficulty of user configuration rules, increases the efficiency of operation rules, reduces the time-consuming of rule matching, and completely releases the productivity of engineers.

In order to solve the problems of low efficiency and low safety caused by manual operation of coiled tubing machine, automatic control technology is widely used in coiled tubing operation equipment. At present, most coiled tubing equipment is still hydraulically controlled, control equipment operation by manually operating hydraulic valves. Since the concept of "intelligent coiled tubing" was put forward at the sixth International Petroleum Industry Summit Forum in 2014, it is considered as the key development direction of coiled tubing technology; therefore, in the context of global energy transformation and upgrading, the development of coiled tubing equipment generally tends to automation, informatization and intelligence, carry out the research on automatic technology of fully electronic control continuous tube equipment, lay the foundation for the digitization and intellectualization of the control system, Improving the performance and safety of equipment is an important means to solve the efficient and safe development of underground operation. After investigating a large number of documents, this paper summarizes the application of automatic control technology in coiled tubing operation equipment at home and abroad, and forms relevant conclusions and suggestions. The research can provide strong guidance for technical development and on-site engineering services.

For the existing refrigeration room design is mostly manual design, low efficiency, in the design of complex refrigeration room, it is difficult to determine the optimal path problem. For this reason, this study uses a genetic algorithm for the automatic arrangement of pipeline orientation with the basic idea of reducing system resistance and rational use of space. The algorithm models the refrigeration plant room, encodes the coordinate points and judges the merit of the route based on the path of the pipeline, the number of elbows and the compactness of the pipeline arrangement and avoidance of obstacles as evaluation indicators. The results, verified by arithmetic examples, show that the method is practical and feasible in the design of pipelines for refrigeration plant rooms. Combining the AI algorithm with the traditional industry design phase helps to change the traditional design model and improve design efficiency, as well as solving the problem of manual difficulty in determining the optimal pipeline route in complex situations. The algorithm, when studied in more depth, can assist in the path planning of a single pipeline in complex situations and has good prospects for assisting in the design of large complex refrigeration plant rooms.

In the bipolar DC network, the unbalanced loads and distributed power generations will result in unbalanced voltage on the DC bus and load side. The unbalanced voltage will lead to power losses and influence the regular operation of the power system. Voltage balancer (VB) can be introduced to adjust the unbalanced voltage in the bipolar DC network. This paper studies the regulation of unbalanced voltage in a bipolar DC network with voltage source converter (VSC), photovoltaic (PV) power generation and energy storage system (ESS) with the participation of VB. In this paper, it is proposed that the coordinated control based on VB can realize the voltage regulation of all parts in the bipolar DC network. It is pointed out that the proposed control strategy can realize adaptive adjustment of voltage droop coefficient. The MATLAB/Simulink platform is built to verify the effectiveness of the proposed coordinated control method for DC bus and load voltage regulation.

This paper proposes a novel algorithm about generating trajectory planning for quadrotor unmanned aerial vehicle (UAV) using fusion algorithm, In the path search part, this paper uses the fusion algorithm based on rapidly-exploring random tree (RRT) and jump point search (JPS), and uses the algorithm based on flight corridor in the trajectory generation stage. Firstly, we use grid map to store the environment information, which is the best mapping way for JPS. Based on JPS algorithm, we adopt RRT-JPS algorithm by combine JPS with the improved RRT algorithm, to enhance the smoothness and speed of path search. Then, by expanding the path block selected by RRT-JPS algorithm, a safe flight corridor can be generated. Finally, by adding time and energy consumption to the cost function in the trajectory optimization algorithm, optimization result can be faster under the premise of ensuring smoothness. The simulation and experimental results show the application of the method is smoother and fewer resource consumption.

Aiming at the difficulty of locating cable open circuit faults, the paper proposes a location method of the improved adaptive genetic algorithm combined with the BP neural network. The improved adaptive genetic algorithm is used to optimize the selection of the weight threshold of the BP neural network, and the predicted location results are compared with the same type of algorithms. The results show that the combined method of the improved adaptive genetic algorithm and BP neural network has a better effect on the location of cable fault distance. Finally, an example of faulty cables in a phased array radar cabinet is used to verify the feasibility of the proposed method in locating the cable open circuit fault.

Aiming at the poor positioning accuracy of the APIT positioning algorithm in wireless sensor networks, this paper proposes an APIT positioning algorithm 3DTC-APIT based on tetrahedral centroid cutting. First, the algorithm reduces the occurrence of misjudgment by setting the node counter. Second, use the centroid of the tetrahedron to cut the node location area multiple times to solve the node coordinates to be located. Finally, the node to be located whose position has been determined is added as a temporary anchor node to assist other anchor nodes in positioning. The simulation results show that compared with the comparison algorithms, the algorithm 3DTC-APIT can better improve the positioning accuracy and positioning coverage of nodes.

In view of the low photovoltaic absorption rate and hydrogen energy utilization rate in the off-grid PH coupling system at present, this paper proposes an optimal configuration (OC) method of off-grid photovoltaic-hydrogen (PH) coupling system based on model predictive control (MPC). The off-grid PH coupling system model is established with photovoltaic array, hydrogen energy storage system (HESS), battery and load as the core equipment, and the operation of the PH coupling system is regulated by model predictive control algorithm. At the same time, taking economic cost, power loss rate and light rejection rate as indicators, the optimal capacity allocation results among photovoltaic array, hydrogen storage system and storage battery can be obtained based on particle swarm optimization algorithm, which can take into account the economy and power supply reliability. Finally, in this paper verify the effectiveness of the OC method proposed in this paper.

With the continuous development of distributed power supply, its use in the power system is more and more, which makes the power system more and more vulnerable to power fluctuations and system failures, resulting in the stability of the system decline. Therefore, how to improve the stability of grid connection becomes a research hotspot. Virtual synchronous generator can simulate the moment of inertia and damping characteristics of synchronous generator, but its parameter design is complicated. To solve this problem, the mathematical model of virtual synchronous generator is established in this paper, and the parameters are optimized by using the improved hybrid leapfrog algorithm. By comparing the simulation results of grid voltage and frequency, it is proved that the improved hybrid leapfrog algorithm is more suitable for the grid stability of virtual synchronous generator.

Aiming at the navigation and positioning requirements of Remote Operated Vehicle (ROV) in complex underwater environment, a data fusion algorithm for Strap Inertial Navigation System (SINS)/ Doppler Velocity Log (DVL) integrated navigation and positioning system is established by using feedback controller and Extended Kalman Filter (EKF). Then, the improved filter algorithm is compared with the common filters for navigation numerical simulation. The results show that the improved EKF filtering effect and navigation accuracy are higher. Finally, the motion control was designed based on PID controller, and three-dimensional simulation and actual pool experiments were carried out. In the three-dimensional long distance navigation simulation experiment, the navigation positioning and motion control perform well. In the experiment of manual navigation around the pool, the positioning error can be reduced to less than 10%. In the automatic control navigation experiment, the minimum position error is only 0.05 m. Experiments show that the algorithm and system design in this study have high accuracy and high stability.

Some grid economic dispatch optimization methods have the problem of high consumption costs. To improve this shortcoming, an adaptive particle swarm algorithm-based grid economic dispatch optimization method is designed. Obtaining grid operating costs, operate in maximum power point tracking mode, adopt full grid acquisition, calculate daily fixed investment costs for the grid, construct a two-level planning mathematical model, adjust battery output and design an economic dispatch optimisation method based on an adaptive particle swarm algorithm. Experimental results: The mean values of maintenance cost, fuel cost and environmental management cost of the grid economic dispatch optimisation method in the paper are: RMB 49.193, RMB 104.488 and RMB 310.045 respectively, indicating that the designed grid economic dispatch optimisation method is more effective after making full use of the adaptive particle swarm algorithm.

In order to realize the autonomous initial alignment of the vehicle-borne strapdown inertial navigation system (SINS) on the moving base, an adaptive filter based initial alignment algorithm for Doppler radar-assisted SINS is proposed. The error analysis and modeling about Doppler radar and SINS are carried out, and the alignment state variables are selected according to the established error model. The measurement equation is established by using output of Doppler radar and SINS. The initial alignment algorithm is designed based on adaptive filtering. The simulation results show that the azimuth angle alignment accuracy of the method is better than 3', and the horizontal angle alignment accuracy is better than 1' in 220 s. The initial alignment can be completed quickly and accurately in the case of large Doppler radar velocity variance.

In this paper, a simulation method of the stiffness-damping characteristics of waves is proposed through the study of the admittance control. Firstly, a traditional six-degree-of-freedom Stewart structure is used to build a wave characteristics simulator, and numerical simulation software is used to build the wave characteristics simulator dynamics model and control framework. Secondly, the admittance control algorithm is introduced into the six-degree-of-freedom positional control to realize the passive motion simulation of the sea target after being collided. The paper focuses on the analysis and simulation verification of the stiffness-damping characteristics of the target in the vertical sea level direction and horizontal sea level direction. Finally, the study proves that the wave characteristics simulator with the admittance control algorithm can simulate the motion of the target on the waves more realistically.

In supervised learning, standard algorithms are mostly designed to deal with balanced data classes, but it is inevitable to encounter imbalanced data classes in some situations. How to learn from imbalanced data is still a challenging problem. A common approach is to generate artificial data or duplicate existing classes to balance the class distribution. Many traditional algorithms can handle imbalanced classes, but they often fail to enhance enough features and tend to produce unnecessary noise points. In this paper, we propose an improved network based on generative adversarial networks (GANs) and improved K-means Smote oversampling method to oversample the data. This method replaces the noise input of GANs with the oversampled results, and then generates new oversampled data through adversarial networks. By conducting experiments on six datasets, we show that this method can effectively improve the classification results of classifiers on oversampled data.

In order to improve the efficiency of the operating system scheduler, this paper proposes an adjustable time slice process scheduling algorithm ATSPS (Adjustable Time Slice Process Scheduling) based on feature similarity clustering, which can reduce the cost of the operating system process scheduling time, reduce the process turnaround time, and improve the process execution efficiency. ATSPS forms different process clusters by clustering the characteristics of process burst time, time slice, weight, and context switching times. It uses the same specification time slice for each type of process cluster. Through queue scheduling adjustment, it accelerates the execution efficiency of computationally intensive processes, reduces the number of contexts switching times caused by the expiration of time slice, and saves scheduling time. The scheduling process of robot grabbing program in the operating system is simulated by MATLAB, and the ready queue, blocking queue and execution queue of the operating system are simulated. The robot grabbing program is simulated, including visual process, force control process and posture process. ATSPS is compared with five classical scheduling algorithms. The experimental results show that on the scheduling criteria of three CPU scheduling algorithms, namely, waiting time, turnaround time and context switching times, ATSPS algorithm reduces the time cost of process scheduling, realizes the real-time scheduling of operating system process, and provides guarantee for the realtime control of robot.

Station line loss is directly related to the operating efficiency of power grid enterprises. Traditional low-voltage station line loss management directly distinguishes abnormal line loss stations according to the index value, which is relatively extensive management. Based on the theory of station line loss, this paper proposes a method of station line loss anomaly analysis based on K-means clustering algorithm. Firstly, the LOF algorithm is used to eliminate the outliers in the platform area, and then the line losses in the platform area are clustered. Finally, the abnormal line loss platform area is identified by combining the value interval of the average line loss rate and the distance from the clustering center.

Although the traditional droop control can realize the power support of multiple converter stations at the same time, because the DC system is not sensitive to the frequency of the AC side system, it cannot provide synchronous power in the event of an AC system accident, at the expense of the mutual support of each AC system ability. Considering the frequency adjustment capability of the interconnected system, an additional frequency control strategy for the VSCMTDC system containing wind farms is proposed. This method can make the DC system respond to frequency changes on the AC side, and allow other converter stations to switch to the accident end through droop control. The stream station provides power support to reduce the frequency variation of the accident end system. Finally, the simulation software PSCAD/EMTDC is used to verify the feasibility and effectiveness of the proposed control strategy.

In recent years, flexible DC transmission has developed rapidly, and offshore wind power DC grid connection projects are using modular multilevel converters, which will become the main grid connection method for large wind farms. However, most of the research on MMC-HVDC uses two-terminal power systems, whose control strategies must be different from two-terminal active systems, and more research is needed on the application of MMC-HVDC in the field of wind power grid connection. Based on this paper the equivalent mathematical model of a wind farm is studied, the flexible DC startup process of the wind farm is analysed and the proposed MMC controlled rectifier charging strategy is validated by simulation.

In recent years, electronic commerce has emerged as one of the dominant modes of commodity trading, and online shopping platforms have become ubiquitous in the daily life. When searching for products of interest, consumers often employ image-based search to allow the platform to recommend corresponding products. To meet this demand, this paper designed a network model that employs the residual neural network ResNet101 for feature extraction of images under the deep learning framework TensorFlow. The experiments demonstrate that this network model can effectively achieve classification results for relevant images in e-commerce. Moreover, the paper conducted dataset optimization when training on ResNet101. The use of this network model enables efficient and accurate identification of submitted product images, which can be returned to users as recommended products in a manner similar to the submitted product images.

Nowadays, various servo controllers on the market are increasingly pursuing higher power density, which means that the hardware design of each module in the system should be small enough. The application of EtherCAT communication technology has become the mainstream of the development of servo system interface. Therefore, this paper designs a small-size EtherCAT slave controller applied to the servo system. By using ARM to simulate the memory function of EEPROM, sharing the clock source between the slave communication element and the servo control element, and integrating the slave communication module with the servo control module, the volume of EtherCAT slave hardware platform is reduced effectively, which provides favorable conditions for improving the power density of the servo system. After experimental verification, the final design of the small-size EtherCAT slave controller can perform standard, correct and reliable EtherCAT communication work.

The research on robustness optimization of large-scale heterogeneous combat network(HCN) is of great significance to improve the survival ability of combat system in complex battlefield and ensure that it can function even when some of its components fail. Most of the existing optimization algorithms are applied in homogeneous networks, and their application to HCNs invariably leads to information loss. In order to fill this gap, this study presents a robustness optimization algorithm, known as the HCN-MOEA, which takes into account the characteristics of the heterogeneity of HCN by introducing the functional robustness of HCN. What’s more, we design the heterogeneous mutation operator which removes the edges that are impossible in actual network to reduce the search space and improves the convergence speed. Finally, We execute optimization experiment and comparative experiment to verify the effectiveness of the optimization objective and mutation operator of HCN-MOEA, which proves that it is superior to the existing algorithms in the application of large-scale HCN. Additionally, the optimization results obtained by the experiment can also provide support for building a combat network with better combat capability.

Community detection is an important tool to analyze and understand large-scale graph networks. Traditional non-negative matrix factorization method use the complete adjacency matrix, in which the redundant information may interfere with the learning of node features and bring high computational complexity. Thus, we propose the Node2vec enhanced attributed graph matrix factorization algorithm (N2V-AGMF) which combining the graph embedding and non-negative matrix factorization. The algorithm uses the Node2vec to extract node low-dimensional features based on topological information, then combines them with the attribute matrix for joint matrix factorization. The low dimensional embedding of the adjacency matrix enriches the representation of node features, and effectively reduces the high computational complexity caused by the factorization of the high dimensional matrix. Experiments were carried out on five real world datasets to verify the effectiveness of the algorithm.

Fast pulse detection in pulse images has received widespread attention in the field of underwater computer vision, and deep learning-based fast pulse detection algorithms can clearly represent the characteristics of sonar pulses, thus gradually becoming an important research direction in this field. To address the issues of large network parameters and high computational complexity in mainstream sonar pulse detection and recognition algorithms, a lightweight convolutional network that incorporates ghost convolution is designed and applied to fast pulse detection in sonar. The ghost convolution module is designed and placed in different areas to further reduce the network parameters from the network structure. The experiment is conducted on the constructed pulse dataset, and the results show that compared with the traditional model, the sonar pulse fast detection algorithm based on this lightweight convolutional neural network can achieve higher recognition accuracy while maintaining a lower network parameter amount.

High-frequency conversion is an effective way to reduce the size and weight of power converters and to improve the efficiency and power density of converters. The LLC resonant converter has been widely used because of its ability to work at no-load and resonant tank current reflecting the lightness of the load, which is an advantage that ordinary series resonant converter and parallel resonant converter cannot match. Sliding mode variable structure control is essentially a special type of nonlinear control, whose nonlinearity is manifested in the discontinuity of the control, the difference between this control strategy and other controls is that the "structure" of the system is not fixed, but can be changed in the dynamic process according to the current state of the system (such as the deviation and its derivatives of each order, etc.) on purpose The system is forced to follow a predetermined "sliding mode" trajectory. Since the sliding mode can be designed independently of the object parameters and perturbations, this makes the variable structure control fast, insensitive to parameter changes and perturbations, does not require online system identification, and is physically simple to implement.

A suppression module is proposed based on pipeline super-resolution (SR) algorithm to address the issues of unstable restoration effect of real images and unsatisfactory details in perception level often generated in traditional image SR. This module segments the original image into several sub-images of the same size and categorizes them into three levels of simple, medium, and hard based on their restoration difficulty. It suppresses the generation of image artifacts and stabilizes the restoration of real details by applying different restoration processes to sub-images of different categories. An artifacts-map is generated based on the observed residual in the artifacts region to regulate and stabilize the model training process. Through experiments, by replacing the original model's simple classification module, more sub-images are processed through a smaller network, and the restoration effect of sub-images of different categories is improved by about 4%, and the overall image restoration effect is improved by 2%.

The CBTC system, composed of multiple subsystems, is a crucial system that ensures safe and stable train operations. In the event of malfunctions during train operation, pinpointing the source of the problem can prove to be challenging, requiring the collaboration of personnel from various subsystems. The process of troubleshooting is tedious and time-consuming, making it difficult to quickly identify the root cause of the issue. This article proposes a solution through the development of an expert knowledge conversion module based on process diagrams, coupled with fault tree analysis and the design of an intelligent fault diagnosis and tracking system. This system enables the rapid tracing of the source of the malfunction, allowing for the swift localization of the root cause and visualization of the issue. This approach will revolutionize the current process and status quo of multi-subsystem personnel collaborating in step-by-step analysis for train fault analysis, significantly reducing the advanced human and time costs required for fault analysis and tracing, and improving the utilization of maintenance resources.

Aiming at the difficulty of good control of LLC resonant converter in lithium battery chemical composition equipment, the sliding mode controller is designed by combining particle swarm algorithm. The traditional PI controller has overshoot phenomenon, and the sliding mode control parameters are difficult to select the optimal value, so combined with the particle swarm algorithm, the function of the controller output and system following error is selected as the target of sliding mode controller parameter optimization, and the appropriate parameters are selected to reduce the recovery time and dithering phenomenon. Simulation experiments show that this method has certain advantages over traditional PI controllers and sliding mode controllers.

Aiming at the problem that the existing target detection algorithm has too many network parameters to be applied to embedded or mobile devices, a lightweight YOLOv4 target detection algorithm is proposed. Firstly, the lightweight feature extraction network GhostNet and the deep separable convolution are introduced to reduce the computational load of the model; secondly, the Efficient Channel Attention (ECA) module is introduced into the feature fusion network to strengthen the learning of target features; then use the idea of deep separable convolution to improve the ASPP module and replace the SPP module. The Pascal VOC dataset is used to evaluate the optimization algorithm. The results show that the introduction of GhostNet network and deep separable convolution can reduce the network parameters by 75%, and the mAP value of the algorithm is reduced by 7.73%. After the introduction of the ECA module and the improved ASPP module, the network computing capacity has only slightly improved, and the mAP value has increased by 2.12% to 65.92%.