This PDF file contains the front matter associated with SPIE Proceedings Volume 12587, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

To optimize the multi-channel interactive experience of the household intelligent clothes hangers, the Kano model is used to qualitatively analyze 15 elements from the two demand levels of "practicality and ease of use", and the SII-DDI matrix analysis diagram is made. The importance ranking of 18 design requirements in two dimensions of "function and interaction" is constructed by the QFD method. The results show that: (1) "convenient storage" and "knowing process" are the requirements of household intelligent clothes hanger users "Remote control", "aseptic", "quick drying" and "ensuring night safety" are expected requirements. (2) The absolute importance of "client remote control", "human sense recognition" and "visual display operation" are 447, 428 and 428 respectively, which are three interaction modes with high user satisfaction. (3) The design points of "folding storage" and "intelligently matching the drying temperature to ensure the comfort of special fabrics" are the most important for the product function upgrading of the existing household intelligent clothes hanger Market.

Because the memory resources and computing resources of embedded devices are very limited, it is very difficult to deploy deep learning models on embedded devices. Therefore, how to efficiently and conveniently extract more feature maps from the limited feature maps in convolutional neural network has become a mainstream research direction. In the era of epidemic normalization, in order to deploy a mask face detection system more suitable for the actual scene in densely populated places such as communities, shopping malls, airports, railway stations and so on. This paper proposes the Mix-YOLOv5s model, which mainly combines the YOLOv5s model with the lightweight Ghost module, the CBAM attention module, Bidirectional Feature Pyramid Network(BiFPN), Mish activation function and Alpha-IoU loss function. They are used to promote the object detection ability of the model. We make quantitative and qualitative comparison between Mix-YOLOv5s and YOLOv5s. Compared with the YOLOv5s model, this model is able to improve the comprehensive performance by a small amount on the basis of greatly reducing Parameters and GFLOPs. Therefore, Mix-YOLOv5s model has great significance and advantages in the design of mobile terminal deep learning model, which can correctly judge whether the subjects wear masks, and has strong research value and broad application prospects.

Globally, solar power technology has become one of the most important sources of electricity for cities or households. And more and more households are choosing to use small, intelligent solar power systems from utility companies as a supplementary energy source for their homes. The energy consumption data stored by the smart system can reflect the user's household activities. The aim of this paper is to re-analyse the energy consumption data provided by Red-back for households in 2011, using big data techniques, to determine which energy information needs to be protected in the smart system by predicting household daily energy consumption using deep learning and machine learning methods, combined with weather data to predict home occupancy.

Service load prediction is a critical basis of cloud-edge autonomous collaborations which mainly considers the rapid response of tasks and load balancing of multiple terminals. Traditional load forecasting is usually in the form of point estimation with a relatively high variance. Frequent changes in point estimation may lead to scheduling errors and waste of resources, thus is not suitable for application scenarios of cloud-edge collaborations. To solve these problems, this paper proposed a service load interval prediction method for cloud-edge collaborations based on type identification and SVMs. The main function of the proposed method is to provide the upper and lower bounds of load forecasting suitable for cloud side collaborations with stronger adaptability to load changes. It mainly includes four steps: service load type identification, load history data interval construction, parameter optimization of SVM, and load interval prediction. This paper takes three types of cloud-edge collaborative tasks in robot target tracking (visual location, target analysis, route planning) as examples, and carried out a large number of experiments to verify the effectiveness of this method. The result shows that it outperforms traditional methods in normalized interval proportion of average width and comprehensive width coverage to a great extent.

Aiming at the multi-scale, multi-directional and deformation of targets in remote sensing images, a Fast R-CNN algorithm with deformable convolution is proposed in this paper, which can significantly improve the detection accuracy of ship targets and shorten the processing time.According to the characteristics of large aspect ratio of ships, the size and ratio of anchor frame are adjusted by the K-means clustering method , so that the improved algorithm is more suitable for detecting approximately vertical or horizontal objects, and can obtain high-precision horizontal region detection of ship targets.

Micro air-launched decoy (MALD) is electronic weapon aiming to interfere with enemy air defense systems. MALD uses signal enhancement subsystems and active radar jammers as its loads. This paper discusses the basic situation of MALD in detail, and analyzes the technical advantages and development trends.

Federated Learning (FL) protects users’ privacy by only uploading the training result instead of gathering all the private data. However, achieving the desired model performance often requires a large number of iterations of parameter transfer between client and central server. Currently, selecting a fixed number of clients to participate in training can slightly reduce the communication overhead during model training, but ignore the impact on model training accuracy. In this paper, we propose an adaptive chosen client number K scheme, which can give a better tradeoff between accuracy and cost. Firstly, through experiments, we find that increasing extracted clients’ number K can reduce iterations’ number T, but after K increases to a certain extent (𝐾₁), T will no longer reduce significantly. Similarly, increasing K can further improve the accuracy of model training, but K is large enough (𝐾₂ ≥ 𝐾₁), the accuracy will also no more be improved remarkably. Thus, [𝐾₁,𝐾₂] is the optimal range. Secondly, we conduct experiments on different datasets with different number of clients, and find that the optimal client’s number growth rate q’ for different conditions is 0.02. According to the experimental results, we set the initial K to be 𝐾₁ for the optimal T, when the model update magnitude in two adjacent iterations is less than a threshold, the number of clients participating in training will increase by q’ to speed up the convergence until K reaches K₂, otherwise it will remain unchanged. Finally, we use our algorithm to improve present FL algorithms. Through experiments, we demonstrate that our algorithm is suitable for existing differential private FL algorithms.

The adaptability of an electromagnetic radiation measurement system in the explosive field environment of energetic materials is studied, the installation method of the instrument is demonstrated, and the measurement point layout is studied. The effective measurement area of the explosion field is determined from the radiation source field area, the antenna measurement field area, the radius of the explosion fireball, the sensitivity of the data acquisition instrument, etc. The number of layouts at the same distance is determined and the dislocation arrangement is adopted to achieve omnidirectional coverage of the measurement range, finally a complete electromagnetic radiation measurement system has been established.

This paper introduces a design and implementation method of an access control management system based on remote control, which consists of the identity collection device and the database server access controller and realizes access control through remote reading and background data identification. When the target person wants to open the door, the door lock control device sends the acquired personnel data to the database server. The database server compares the data with the time information stored in the database in advance by the person, and quickly determines whether the door needs to be opened. If required, the access controller sends a signal to the electric control lock to open the door and the person can enter the house. Otherwise, the person cannot enter the house. This design integrates the human-computer interaction technology, data transmission technology, and communication technology to implement automatic management of intelligent devices through remote time control. Therefore, this design can efficiently identify and authenticate personal identity. In addition, this design features high reliability, low cost, small volume, complete functionality, and strong scalability on information collection. The time-based access control system fully realizes the automatic management of personnel who attempt to pass the access control system based on the identity information of the personnel stored on the database server and the time node to be reserved and by using the computer as the background processing tool.

To reduce the loss of the jujube production and develop more effective pest control measures, this paper studies the prediction model of jujube pest epidemic based on the Internet of Things technology. The data of jujube growth was collected based on the Internet of Things technology, and multi-source information was obtained by combining sensor technology and network communication technology. The characteristic bands of the jujube image were extracted. The useless bands were removed, and the overall physical information of the image was retained. The prediction model of the jujube pest and disease epidemic situation was established, and the prediction factors with poor stability were processed by BP neural network structure to realize the prediction. Through experimental demonstration and analysis, compared with the actual occurrence degree of diseases and insect pests, the prediction curve of the occurrence degree of diseases and insect pests of jujube by this model is consistent with the changing trend of the actual occurrence degree curve. Compared with the traditional model, the prediction error of the proposed model is smaller, which proves that the proposed model is more effective. In this paper, the model can obtain effective information on the basis of multi-source data, and get more accurate prediction results, which provides a scientific basis for decision-making for the formulation of pest and disease control plan to ensure the high and stable yield of jujube.

Aiming at the problem that the effect of two-typical wing forms on aerodynamic characteristics of mid-cylindrical airship, the aerodynamic characteristics of airship with X-shaped tail and inverted Y-shaped tail are studied respectively. Based on the method of Computational Fluid Dynamics(CFD), the aerodynamic characteristics of airships with the same hull shape parameters are simulated, at an angle of attack ranging from -30 to 30 degrees and angle of sideslip ranging from 0 to 20 degree. The result shows that the mid-cylindrical airship with inverted Y-shaped tail has a certain advantage in reducing hull resistance, compared with the airship with X-shaped tail. The mid-cylindrical airship with X-shaped tail has better ability to provide lift force and maintain hull stability, compared with the airship with inverted Y-shaped tail. The simulation results provide an effective reference for aerodynamic shape design and flight control strategy of mid-cylindrical airship.

In this paper, a novel policy network update approach based on Proximal Policy Optimization (PPO), Advantageous Update Policy Proximal Policy Optimization (AUP-PPO), is proposed to alleviate the problem of over-fitting caused by the use of shared layers for policy and value functions. Extended from the previous sample-efficient reinforcement learning method PPO that uses separate networks to learn policy and value functions to make them decouple optimization, AUP-PPO uses the value function to calculate the advantage and updates the policy with the loss between the current and target advantage function as a penalty term instead of the value function. Evaluated by multiple benchmark control tasks in Open-AI gym, AUP-PPO exhibits better generalization to the environment and achieves faster convergence and better robustness compared with the original PPO.

In this paper, a model-based anti-noise neural network controller for redundant robot motion control is proposed for motion control of redundant robots with uncertain kinematic parameters. The main challenge of this problem is the coexistence of parameter uncertainty, redundancy resolution, and system physical constraints. Therefore, a new model - driven neural network controller is proposed in this paper. A class of nodes are introduced to deal with the kinematic parameter uncertainty of the system. On this basis, the selection of the initial value of the hyperparameter of the neural network is deeply analyzed, and this processing has a positive effect on accelerating the convergence of the tracking error. The proposed controller has the advantages of simple structure, small computation and simple implementation. The simulation of Kinova Jaco2 manipulator verifies the effectiveness of the proposed algorithm.

Radio Frequency (RF) perception of humidity is the current research hotspot of grain condition monitoring for "three temperature and three humidity". In order to analyze the optimal detection parameters on humidity of storage environment, the Wireless Insite test platform is built to analyze the effects of antenna height, carrier frequency and grain and with different moisture content on Channel State Information (CSI) such as received power, path loss and time delay. The mathematical relations between the humidity and path loss, the humidity and time delay are deduced respectively. In simulation experiments, the frequency bands of 2.4GHz, 5GHz, 28GHz, 45GHz, 60GHz are selected. Study the changes of path loss and delay spread with humidity in the storage environment at different frequencies and antenna height. When the height of the transceiver antenna is set to 1m, the 28GHz, 45GHz and 60GHz millimeter wave signals are more sensitive to humidity changes, thereby which can provide more higher perception resolution of detecting humidity in the storage environment, compared with the 2.4GHz and 5GHz frequency bands. Besides, when space humidity is constant and the moisture content of wheat is changed, the arrival time of each ray in the channel at 45GHz is longer than that at 60GHz. The above work can provide reference for the application of RF sensing technology in the humidity detection scenario.

Since the conventional integration methods cannot distinguish the useful data when there is much heterogeneous data with the same semantic, the data integration time is long and the utilization rate of power resources is low. Based on this, an XML-based heterogeneous database integration method for electric power enterprises is designed. Firstly, the same semantic objects of heterogeneous databases are identified to reflect the integration conflicts of the same semantics of heterogeneous data in power enterprises, and then the integration framework of heterogeneous data exchange in power enterprises is constructed by using XML, and finally the similarity value of numerical attribute data in power enterprises is refined. Thus, the efficient integration of heterogeneous databases in power enterprise is realized. The experimental results show that the proposed method has a relatively high utilization rate of heterogeneous data resources in power enterprises, all above 98%, and its integration time is short, less than 4s, which is better than the comparison method. The proposed method can help power enterprises to quickly manage complex business data and improve their automation and informatization levels.

For the rail traffic, feasibility of 5th Generation (5G) mobile communication massive Multiple Input Multiple Output (MIMO) used in the tunnel is studying. In order to effectively set up base station, antenna and intelligent reflecting surface (IRS), the direction of arrival (DOA) must be captured. Aiming at the problem of poor performance of two-dimensional MUltiple SIgnal Classification (2D-MUSIC) algorithm in the environment of low signal-to-noise ratio (SNR), small snapshots and small incident angle interval signals, an improved MUSIC algorithm with uniform rectangular array (URA) based on reconstructive subspace is proposed. By reconstructing subspace, a new spatial spectrum is obtained in terms of subspace eigenvectors. Then, the DOAs are obtained by searching the maximum of the new spatial spectrum. High estimation accuracy and angular resolution are often required in practical applications of rail traffic 5G system, and there exist tunnel scenarios that we need the improved MUSIC algorithm has the ability to conduct two-dimensional DOA estimation. Simulations and the measured data of straight tunnel scenario are used to verify the effectiveness of the proposed algorithm and its higher searching precision in complex signal environments such as low SNR, strong-to-weak proximity, and coherent interference.

With the increase in energy demand, carbon emissions, environmental pollution, climate change and other issues have become increasingly prominent, China has accelerated the construction of new energy sources. Especially in the field of wind power generation, it is the most potential type of large-scale development of non-hydroelectric renewable energy. Due to the volatility, intermittency and low energy density of wind power, the power of wind power also fluctuates. However, with the early digital transformation of my country's energy industry, a large number of meteorological environments and equipment measurement points have been accumulated in wind power production sites. Power generation related data, using artificial intelligence, deep learning and other technologies can effectively predict the power generation of the station with high precision. A model algorithm of multi-loop gradient boosting decision tree is used in this paper, considering the stationarity test of time series and wind power fluctuation attribute, the accuracy of wind power prediction is effectively improved. Help the power dispatching department to pre-arrange dispatching plans according to wind power changes. Ensure the smooth and safe operation of the power grid.

As a security defense technology to protect the network from attack, network intrusion detection system plays a very important role in the field of computer system and network security. Aiming at the multi classification problem of unbalanced data in network intrusion detection, machine learning has been widely used in intrusion detection, which is more intelligent and accurate than traditional methods. The existing multi classification methods of network intrusion detection are improved, and an intrusion detection model using smote and ensemble learning is proposed. The model is mainly divided into two parts: smote oversampling and stacking classifier. The NSL-KDD dataset is used to test the Stacked Ensemble model in this paper. Compared with the other five basic learner models, the Stacked Ensemble has a higher detection rate. Stacked Ensemble has significant advantages in solving the multi classification problem of unbalanced network intrusion detection data. It is a practical and feasible intrusion detection method.

The channel estimation of IR-UWB ultra wideband wireless communication system realized by using compressed sensing theory. Firstly, sparse signal, observation matrix and reconstruction algorithm of compressed sensing theory were discussed. Secondly, discussed the composition of IR-UWB wireless communication system. The IEEE802.15.SG3a channel model was adopted for UWB multipath channel. According to the matrix calculation method of cyclic convolution, the compressed sensing model for channel estimation of IR-UWB system was derived, and GOMP algorithm was used to reconstruct the channel parameters of IR-UWB system. With the help of Matlab software, the simulation results showed that GOMP algorithm can reconstruct the channel parameters of IR-UWB system well.

Entity Linking is the task of mapping entity mentions in text to real entity in a given knowledge base. One of the main obstacles to this task is the variations of name and the ambiguity of the entity. Prior studies have explored various methods for modeling local semantic relatedness between each entity-mention pair and global topical mention consistency to address these issues. Nevertheless, most of these existing methods are relatively complex so that they are impractical in some situation, e.g. on-line search-engine. This paper proposes a simple approach that combines the merits of the local and global paradigm, which aims at simplifying the entity linking process and achieving considerable performance. This proposed approach benefits from a semantic relatedness measure for local modeling and a collective score for global coherence guarantees. Our approach is evaluated on different types of datasets and the results show that, despite its simplicity, our model can achieve comparable performance compared to other similar metrics.

The key to constructing an event knowledge graph is to acquire event knowledge. At present, the method of event extraction from text is not accurate enough, while the event information obtained through encyclopedia knowledge bases has the advantages of high accuracy, good structure and rich multimedia resources. However, acquiring event entities from a single encyclopedia knowledge base has the problem of missing information, so the fusion technology for multisource encyclopedia knowledge bases is needed urgently, in which entity alignment is the core technology. Aiming at the shortcomings of current alignment methods focusing on static entity in event entity alignment, we propose an event entity alignment method based on event elements, which calculates entity similarity according to multiple event elements. In contrast to the algorithm based on latent Dirichlet allocation (LDA) model and the method based on representation learning using BERT, the proposed method provides a significant performance improvement in event entity alignment. Especially, the method optimizes the threshold setting so that it enhances the ability to identify the presence of aligned entities.

As the most popular cryptocurrency now, Bitcoin's transaction data is easy to obtain, so de-anonymizing Bitcoin becomes possible. This paper constructs a data set of Bitcoin addresses including 5 categories, analyzes and extracts the transaction features of Bitcoin addresses in more detail based on related work, and proposes two new features of fourth-order transaction moments and sample distribution. New features improve the performance of Bitcoin address classification. The accuracy of the LightGBM model was 0.94 and the F1 score was 0.91. This method can identify unknown types of Bitcoin addresses, which improves the ability of relevant agencies to investigate Bitcoin illegal activities.

To improve the accuracy of grain yield prediction, a grain yield prediction model based on wavelet transform and long short-term memory (LSTM) is proposed. Firstly, the original data is decomposed by wavelet transform algorithm to obtain a series of sub-sequences of different scales, and then LSTM prediction models are built for the sub-sequences, finally wavelet reconstruction is used to obtain the predicted yield and analyze the model performance. The article uses China's 1999-2018 grain yield as experimental data. The experiment shows that the method proposed in this article has excellent performance in both short-term and medium-term predictions compared to the existing methods.

As the quality controller of the marketing business of power grid enterprises, marketing audit has realized the management support of the information system to the core business. With the continuous expansion of the business scope and the continuous advancement of the data depth, on the basis of the gradual improvement and optimization of the marketing audit management system, it is necessary to carry out digital intelligence transformation according to its own business characteristics. RPA, as a software automation tool, can replace the original labor-consuming and labor-intensive repetitive low value-added work, and plays an important role in improving the business digitization level. Based on the application characteristics of RPA technology, this paper combs the characteristics of the marketing audit business. By constructing typical business application scenarios, refining the RPA simulation operation process of the business scenarios, this paper presents the intelligent transformation process of the audit business management, and provides experience for the intelligent application of robot process automation in other professional fields of marketing.

Objective To explore the prediction of hepatitis C incidence by stacked LSTM model. Methods Aiming at the incidence trend and the number of cases of hepatitis C in China from 2007 to 2017, the ARIMA, NNAR, SVR and stacked LSTM were used to train them. The model was used to predict the incidence of hepatitis C in the second half and the last quarter of 2017, and compared with the actual values. The prediction effects of the four models were compared and analyzed using the Root Mean Square Error (RMSE) and the Mean Absolute Percentage Error (MAPE). Results The SVR model performs not well. However, ARIMA model, NNAR model and stacked LSTM model can identify the incidence trend of hepatitis C in China from 2007 to 2017, and the RMSE values of the ARIMA model and the NNAR model are larger, and these two are relatively similar. On the contrary, the RMSE value of the stacked LSTM model is smaller. On the whole, compared with ARIMA model and NNAR model, it decreases by at least 20 percentage. The predicted MAPE value of the stacked LSTM model is less than 1%, meanwhile it is lower than the value of ARIMA or NNAR models. Conclusion The stacked LSTM model has the best predictive effect on the incidence of hepatitis C.

With the new operational styles such as mosaic warfare and multi-domain warfare of the US military moving from concept to application, the winning elements and operational modes of future maritime operations have been profoundly changed. Firstly, the military capabilities of the US military to break through anti-access/area denial are analyzed, and then the characteristics and effects of the implementation of multi-domain warfare under the mosaic warfare concept in breaking through anti-access/area denial are extracted from the military capabilities, operational styles, and practical applications. On this basis, countermeasure suggestions are proposed to deal with the implementation of multi-domain warfare under the mosaic warfare concept. The practical significance of building unmanned maritime combat clusters is explained from the perspective of military needs, and the force design, equipment development, and combat style of unmanned maritime combat clusters are described in detail. A typical combat scenario is used to show the combat process of unmanned maritime combat clusters. Finally, the key problem model of intelligent command and control decision of maritime unmanned cluster is established for this typical combat scenario, and the corresponding algorithmic framework is constructed for the input and output of the typical combat scenario, which provides research ideas for further empirical research.

Adversarial training proves to be the most effective measure to classify adversarial perturbation, which is imperceptible but can drastically alter the output of the classifier. We review various theories behind the relationship between generalization gap and adversarial robustness and then raise the question: is it the input near the decision boundary that provides guidance for the classifier to learn the ideal decision boundary and therefore yield a more desired outcome? We provide quantitative confirmation that the expected required sample size correlates favorably with sample distance and further investigate the relationship between the robust classification error and the expected distance from the decision boundary to samples. Experimental results reveal that applying the data near the decision boundary as training sets can significantly promote adversarial generalization, which keeps consistence with the main conjectures presented in this work.

Knowledge distillation (KD) transfers knowledge of a teacher model to improve the performance of a student model which is usually equipped with a lower capacity. The standard KD framework, however, neglects that the DNNs exhibit a wide range of class-wise accuracy and the performance of some classes is even decreased after distillation. Observing the above phenomena, we propose a novel Class-Wise Knowledge Distillation method to find the hard classes with a simple yet effective technique and then make the students take more effort to learn these hard classes. In the experiments on image classification tasks using CIFAR-100 dataset, we demonstrate that the proposed method outperforms the other KD methods and achieves excellent performance enhancement on various networks.

Although the mimic system can effectively defend against known or unknown vulnerabilities / backdoor attacks, some encryption protocols such as SSH will produce different encryption results on different executors, even with the same processor, the same operating system, the same encryption protocol and the same plaintext, which leads to difficulty in output arbitration. To solve this problem, this paper proposes an encryption source normalization method, which can make different executors generate same ciphertext by normalizing the source of the random number and synchronizing the length of output data, so that the output of heterogeneous executers can be successfully arbitrated by the scheduler. This method is verified by experiments using SSH protocol. Test results show that this method can effectively solve the encryption problem of mimic system.

In order to obtain the scale information of ship targets effectively, we proposed an improved Faster R-CNN algorithm which integrated multi-scale region proposal and pooling of ROI, visual attention mechanism and rotation region regression and suppression, and ship targets can be positioned by the rotate quadrangle bounding boxes to obtain the scale information of them. Our improved model is based on the standard Faster R-CNN and is maintained through end-to-end training.

With the increasing awareness of privacy protection in recent years, various encryption techniques are gradually applied to network traffic, which makes encrypted traffic classification an indispensable part of network management. Recent studies show that the approaches based on deep learning are compelling for the traffic classification task. However, most of them take the encrypted payload as input, which not only requires high computational overhead to make classification, but also limits the performance improvement due to the unavailability of the plaintext. In this paper, we treat the encrypted traffic as sequences and solve the classification task from the perspective of sequence modeling, which only depends on several sequence fields obtained from the traffic header. We properly design a lightweight model and name it TGA by its structure, which consists of a temporal convolutional network (TCN), a gated recurrent unit (GRU) and the attention mechanism. TGA first extracts short-term features from sequences by applying the TCN, and then captures the long-term dependencies by exploiting the GRU, and finally focuses on valuable features through dynamic assignment of attention weights. Through these three steps, TGA is expected to obtain the most effective but lightest temporal features. Experimental results on the public dataset demonstrate that TGA shows superiority in terms of classification accuracy and time efficiency, while the number of parameters is reduced to at most 30% of the state-of-the-art models.

This paper mainly discusses processing the video stream obtained from the network camera by using OpenCV image processing function to detect the face. Face recognition classifier uses OpenCV comes with the face classifier that trains by HAAR characteristics , which draw out of the face part with a blue box. Secondly, we identify the meaning of gesture. The recognition of gesture uses the classifier which uses Haar characteristics to extract the target feature to make the classifier. And also the use of pattern matching method become the second choice. Finally we use JNA to call the Windows API, to make the Windows platform browser switch to the web page. Also we use JavaFX to create the GUI, to take pictures, set the value of HSV, and display the meaning of gestures. Key words-OpenCV; Gesture Recognition; Pattern Matching; Face Recognition

Humans can easily learn to recognize every object in life, every landscape, and describe the things around them in detail from the process of growing up, but computers cannot. How to make computers learn to describe things in pictures has become the research direction of many scholars. If this technology is mature, it will bring great boon to people with visual impairments. They can understand the things around them and the beautiful earth through hearing. Robots recognize objects and understand their surroundings. With the development of artificial intelligence, the power of convolutional neural networks is more and more comparable to that of the human brain. In recent years, many scholars have proposed different methods to seek better solutions to this problem, including generative adversarial networks. Based on the classic structure of Encoder-Decoder, this paper first compares the code implementation and results of ResNet101 as an Encoder on the COCO dataset, and then proposes a new solution that integrates YOLOv5 and LSTM, aiming to improve the model inference speed and inference accuracy.

With the outbreak of covid-19 in 2020, timely and effective diagnosis and treatment of each covid-19 patient is particularly important. This paper combines the advantages of deep learning in image recognition, takes RESNET as the basic network framework, and carries out the experiment of improving the residual structure on this basis. It is tested on the open source new coronal chest radiograph data set, and the accuracy rate is 82.3%. Through a series of experiments, the training model has the advantages of good generalization, high accuracy and fast convergence. This paper proves the feasibility of the improved residual neural network in the diagnosis of covid-19.

Vehicular networks improve traffic safety and efficiency by wireless communications among vehicles and infrastructures. However, security has always been a challenge to vehicular networks, which may cause severe harm to the intelligent transportation systems. Sybil attack is considered as a serious security threat to vehicular networks since the attacker can disseminate false messages with multiple forged identities. In this paper, we designed a behavior feature and multi-scenario-based Sybil attack detection method, named BMSAD in vehicular networks. In BMSAD, we propose solutions for two scenarios: normal traffic and traffic congestion. A node is allowed to verify the authenticity of another node by estimating their geographic distance based on received signal strength, and compare them to its claimed localizations. Then we design long- and short-term pattern to analyze the similarity of vehicles’ behavior features and trajectory for Sybil nodes detection, which act as the second line of defense in normal traffic scenario. At last, the Sybil nodes of the traffic congestion scenario are detected by logical inference based on kinematics theory. And experiment results demonstrate that the proposed scheme achieves high detection rate and low false positive rate.

As the most important module in recommendation systems, click-through rate prediction has attracted the attention of industry and academia. Due to the powerful learning ability of deep learning, it is widely used in click-through rate prediction. Behavior sequences based on user is an important direction of click-through rate prediction. Although some results have been made in related directions, existing methods still have some problems, such as the inability to learn feature weights better, the presence of noise in user behavior sequences, not fully mining the hidden information in features, etc. In this paper, we propose a method for related problems, named DISFMN, which can dynamically learn the importance of features as well as filter out the noise in user behavior sequences. The method also combines high-order and low-order feature interactions to uncover more valuable information in features. Comparative experiments are conducted on different datasets and the experimental results showed the effectiveness of the proposed method.

Aiming at the problems of single monitoring and management mode, poor real-time performance, low transparency, and difficulty in operation and maintenance of the current data room, a digital twin machine room dynamic environment monitoring system based on the Drools inference engine was constructed. In the virtual scene, the Drools rule engine is used to build an expert system for fault analysis and prediction in the data room, which improves the interactivity of the dynamic loop system in the data room, greatly improves the accuracy and timeliness of fault diagnosis, and has great application value.

As a cornerstone of cloud-native systems, Kubernetes uses YAML, a data description language, to configure resources. However, YAML does not meet the configuration requirements of complex scenarios and has three major problems. First, YAML has no type checking mechanism and therefore data type mismatches cannot be detected during compilation. Second, YAML does not have the ability to reuse data descriptions and there are duplicate code for largescale data. Third, YAML lacks a type merging algorithm that meets the needs of multi-team development in enterprises. This paper implements an experimental cloud resource provisioning language, Hermias, which is based on the functional programming language OCaml. Hermias is used to describe the resource configuration of cloud servers, which solves the above three problems in YAML. The novelty of this work is to propose a type merging algorithm that computes records with common labels by union types and subtyping.

The essence of enjoying music is that people can track music beat anytime and be brought into the scene expressed by music. Music beat tracking is a common task in Music Information Retrieve (MIR). While numerous studies have been done in this field, most works focus on the offline beat tracking. However, tracking music beat in real time is a challenging task for computers. In the past few years, people attach more importance to this field. Researchers care about the music beat without taking music style or context into consideration. In this paper, we propose a method for tracking music beats in real time in conjunction with music genre. Specifically, the proposed model is based on a widely-used framework of Hidden Markov Model (HMM). By recognizing the genre of input music, we narrow the range of beats per minute (BPM), which significantly reduces the number of hidden states in HMM. Consequently, the inference time of beat tracking decreases. We experimentally verify the model on the open-source Ballroom dataset, and its accuracy remains at a competitive level while having a much shorter inference time.

Aiming at the problems of fuzzy boundary and poor segmentation effect of SwinUNet in melanoma image segmentation, an improved SwinUNet network segmentation method was proposed. Firstly, Dice loss function is used to alleviate the background and regional imbalance. Secondly, each decoder layer is made to fuse the smaller scale from the encoder, the same scale feature map and the larger scale feature map from the decoder, so that the fine-grained semantics and coarse-grained semantics at the full scale can be captured . Finally, the size of the sliding window is increased, the receptive field of the model is enlarged, and the Dice coefficient is used to evaluate the segmentation results. The average Dice values of the original SwinUNet and the three improved models were 0.8311, 0.8689, 0.8719 and 0.8661, respectively. The experimental results show that the improved model proposed in this paper can effectively improve the accuracy of the original model, which is extremely important for the early diagnosis and treatment of melanoma.

In this paper, we propose a lightweight machine learning (ML) framework based on unsupervised spectral domain discretization for hyperspectral image (HSI) classification. Firstly, the high-dimensional HSI data is mapped into a discretized image by unsupervised learning method, and then the histogram statistics of discrete features are performed to align feature vectors. Finally, supervised ML method is used for classification, thus achieving a lightweight ML method of high-dimensional HSIs. Practical applications and comparative studies on three publicly available HSI datasets show that the framework approaches and surpasses deep learning models in classification accuracy while significantly compressing computational time consumption. The performance of six unsupervised clustering methods in HSI spectral domain discretization is compared in the study. Among them, K-means and GMM are superior in terms of classification accuracy. And SOM provides high classification accuracy while its discretization results are better interpretable due to better maintenance of topology during discretization.

To improve the accuracy of the power supply enterprise service network public opinion crisis early warning, the fuzzy reasoning theory is introduced to carry out the design research of the power supply enterprise service network public opinion early warning method. Based on public opinion topic intensity, development heat and public attitude, the power supply enterprise service network public opinion early warning index system is constructed. Combined with fuzzy reasoning theory, the index membership degree and early warning level membership degree are calculated. Through the learning method, the public opinion early warning level judgment rule is learned, and the public opinion early warning level judgment and early warning display are completed. The experiment proves that the new public opinion early warning method can accurately judge the degree of public opinion crisis, and give a reasonable and intuitive early warning display result.

The project is based on Window10+Python3.6 environment, and uses Python libraries such as OpenCV, Sklearn and PyQt5 to construct a relatively complete gesture recognition and translation system, which can recognize all kinds of static gesture signals in life, and translate them into Chinese or Arabic numerals through image processing.Due to the limitation of the training amount of Support Vector Machines (SVM), this design is only used to recognize gesture actions 1-10, and the interface is designed using PyQt5 for real-time display of the results of gesture recognition translation.The paper focuses on the noise elimination, contour extraction, RGB colorspace and YCrCb feasibility comparison of still images of 1-10 gesture features extracted by computer camera, GUI page design, Fourier operator extraction of gestures, training SVM model, and debugging of the system. The system is also able to be integrated on different development boards and can be embedded in device carriers to meet multi-scene adaptability.

Given the excellent globality and parallelism, Transformer has been widely applied to image tasks. Visual Transformers demand modeling the spatial correlations among visual tokens. However, those existing methods either only emphasize the relative position between two tokens, or only concern on their contexts. Intuitively, a rational attention distribution should hinge on both. To this end, this paper proposes Reference Aware Attention (RAA). RAA decomposes inner-tokens dependency into three intuitive factors, in which reference bias is introduced to model how a reference token attends to a region. Experimental results suggest that RAA can effectively promote the performances of visual Transformers on various medical image diagnosis tasks.

More and more complex and diverse network security problems bring great challenges to the analysis of abnormal network behavior. In order to detect the abnormal connection behavior of the network more accurately, this paper first uses the knowledge graph technology to extract the graph feature parameters that can reflect the node and the overall situation of the network, and then proposes a two-stage unsupervised anomaly analysis method for the abnormal changes of the feature parameters. In the first stage, the anomaly analysis of the whole network graph features is carried out based on clustering technology, so the rough positioning is carried out. In the second stage, the abnormal trend analysis is performed on the graph features of important nodes to determine the category of abnormal connection behavior. On this basis, the time series prediction method is used to predict the node graph features, so as to provide early warning for network security. The experimental results show that the method can effectively extract the network abnormal behavior and predict the development trend of the network in the future, and provide a good support for the understanding of network security situation.

Nowadays, due to the increasing need for cloud storage and cloud computing services, digital images need to be kept confidential in order to secure the privacy of individuals' information. Since the currently available codomain reversible data masking techniques are based on gray images, this paper introduces a data masking approach using color image channel correlation to achieve codomain reversibility. In the encryption stage, the security and homomorphism of encryption are achieved through XOR encryption and block scrambling. The data hiding processed by making full use of the texture information data of the reference channel to predict the pixel data more accurately, and an adaptive prediction error expansion technique is used to embed the encrypted information into the R, G, and B channels. The experimental results prove that the method has better security and better performance than the existing algorithms.

It's hard for the classic TextRank algorithm to differentiate the degree of association between candidate keyword nodes. Furthermore, it readily ignores the long-distance syntactic relations and topic semantic information between words while extracting keywords from a document. For the purpose of solving this problem, we propose an improved TextRank algorithm utilizing lexical, grammatical, and semantic features to find objective keywords from Chinese academic text. Firstly, we construct the word graph of candidate keywords after text preprocessing. Secondly, we integrate multidimensional features of candidate words into the primary calculation of the transition probability matrix. In this regard, our approach mines the full text to extract a collection of grammatical and morphological features (such as part-of-speech, word position, long-distance dependencies, and distinguished BERT dynamic semantic information). By introducing the dependency syntax of long sentences, the algorithm's ability to identify low-frequency topic keywords is obviously promotional. In addition, the external semantic information is designed to be imported through the word embedding model. A merged feature-based matrix is then employed to calculate the influence of all candidate keyword nodes with the iterative formula of PageRank. Namely, we attain a set of satisfactory keywords by ranking candidate nodes according to their comprehensive influence scores and selecting the ultimate top N keywords. This paper utilizes public data sets to verify the effectiveness of the proposed algorithm. Our approach achieves comparable f-scores with a 5.5% improvement (4 keywords) over the classic. The experimental results demonstrate that our approach can expand the degree of association differentiation between nodes better by mining synthetic long text features. Besides, the results also show that the proposed algorithm is more promising and its extraction effect is more robust than previously studied ensemble methods.

Most existing video denoising methods based on the PatchMatch algorithm and optical flow estimation often lead to artifacts blurring and poor denoising effect on scale-varying data. To tackle these issues, we propose a multi-scale feature fusion network based on different pyramid blocks and adaptive spatial-channel attention, which enables to effectively extract multi-scale feature information from noisy video data. Furthermore, we develop a spatial-temporal alignment module with deformable convolution to align the implicit features and reduce blurring artifacts. The results show that the proposed method outperforms the state-of-the-art algorithms in visual and objective quality metrics on the public datasets DAVIS and Set8.

Reducing network power consumption plays an important role in the research of wireless sensor networks. This paper focuses on energy efficiency in environmental data collection scenarios. The data collected in this scenario are usually redundant due to spatial and temporal correlation. Consequently, selecting some nodes for dormancy can reduce power consumption. It is the key issue that how to select dormant nodes, and existing research mainly focuses on uniform clustering and optimal routing algorithms. However, the algorithms cannot guide the selection of dormant nodes because of their less consideration of attribute characteristics. Therefore, this paper proposes a node dormancy strategy for temporal-spatial correlated nodes in wireless sensor networks. The temporal-spatial correlation of the data is firstly verified; then the attributes combined with the location information are provided for FCM clustering; after, dormant node selection and head node selection are performed according to the clustering. Experiments on real temperature datasets demonstrate that using this paper's strategy, data accuracy can still be maintained at more than 95% of what no dormant node perform when 50% of nodes are dormant and around 90% when 80% of nodes are dormant. The improvement even reaches at most 80% against the traditional strategy with the same percentage of dormancy.

An improved RAKE (Rapid Automatic Keyword Extraction) algorithm is proposed to solve the problems of phrase conglutination and inability to obtain professional words in the process of extracting keywords from Chinese professional texts. Through the TTF-IDF (Total Term Frequency-Inverse Document Frequency) method, professional field stop words are extracted and added to the general stop word dictionary for phrase segmentation. Professional domain entity words are introduced into the general word segmentation dictionary, and appropriate weight is given to them in the degree calculation to ensure that professional entity words get higher scores and are effectively extracted as keywords, because in professional field texts, professional entity words contain more core information. The experiments show that this algorithm is better than the basic RAKE and other algorithms in keyword extraction for Chinese professional field texts.

With the rapid development of artificial intelligence technology, emotion recognition has been applied in all aspects of life, using eye movement tracking technology for emotion recognition has become an important branch of emotion computing. In order to explore the relationship between eye movement signals and learners' emotional states in the online video learning environment, we used machine learning and convolutional neural network methods to recognize eye movement signals, and classify learners' emotional states into two categories, positive and negative. The study of eye movement data under different time windows mainly includes four stages: data collection, data preprocessing, classifier modeling, training and testing. In this paper, a Eye-movement Feature Extraction Classification Network(EFECN) based on convolutional neural network is proposed for small sample data and the classification of emotion state based on eye movement. The eye movement data were transformed into images through cross-modal conversion as input of multiple different deep convolutional neural networks, and the emotional states were classified in positive and negative directions. The accuracy was used as the evaluation index to evaluate and compare the different models. The accuracy of the eye movement emotion recognition model reached 72% in the SVM model and 91.62% in the EFECN model. Experimental results show that the convolutional neural network based on deep learning has a significant improvement in recognition accuracy compared with traditional machine learning methods.

Brain computer interface constructs the direct connection between human brain and external devices, which is becoming a promising method in reconstructing human’s motor abilities who suffered from body disability. Among several kinds of BCI technologies, Motor Imagery based Brain-computer Interface (MI-BCI) has attracted more and more attentions since it’s a more intuitive method. In the procedure of MI-BCI, the signal recognition methods play a significant role. Therefore, this paper would search into the classification techniques utilized in the processing procedure in MI-BCI systems, including machine learning techniques, naïve bayes classifier (NB), support vector machines (SVM) and linear discriminant analysis (LDA). For deep learning techniques, sparse autoencoder (SAE), convolutional neural network (CNN), recurrent neural network (RNN) was introduced. Then the paper would compare them in terms of accuracy, classification speed and data requirement. This paper would give an overview on the commonly seen classification method used in MI-BCI, and also present researchers who are selecting classification methods the most suitable choice.

Brain-computer interface (BCI) is a link between the human brain and a computer or other peripheral devices for communication and control. The most frequently utilized BCI paradigms at the time are motor imagination (MI) BCI. In the procedure of MI-BCI, one of the most important roles is the feature extraction of EEG signals. This article examines various feature extraction approaches in four distinct domains: time, frequency, time-frequency, and spatial. Various approaches are introduced in each domain, including the ERD/ERS computation, the FFT method, the Wavelet Transform (WT), the Discrete Wavelet Transform (DWT), Common Spatial Patterns (CSP), and Sub-band Common Spatial Patterns (SBCSP). This paper also compares the advantages and disadvantages of different methods in practical application, which can provide reference for future research.

As the most common captcha, text captcha can prevent others from maliciously using computer programs to log in or attack, and is an important safeguard in Internet authentication. In recent years, with the development of the Internet, the field of artificial intelligence has also developed at a high speed, and convolutional neural networks are widely used in various fields. In this context, for the common problem of character-based captcha recognition, this paper investigates captcha recognition based on a deep learning neural network framework used by the TensorFlow framework with modifications based on the VGG16 convolutional neural network. The 4-digit captcha randomly composed of 64 characters is then converted into an image, and after operations such as image processing and encoding of the captcha, a large number of training sets are generated and the recognition of the captcha is done by the convolutional neural network. Finally, the design GUI interface is deployed to mobile devices with a final accuracy rate of 85% on the test set.

Feature extraction is an important research topic in the field of image processing.In autonomous driving, it is of great importance to extract the feature information of the picture obtained by the vehicle camera for the agent to better understand the environment information. In order to improve the quality of feature extraction, this paper combines meta-learning and deep learning-based feature extraction methods, and proposes a Meta-VAE-WGAN-GP (MVWP) feature extraction algorithm, and applies it to automatic driving. Firstly, aiming at the problem of parameter centralization in Wasserstein generative adversarial network (WGAN) and the problem of gradient explosion and gradient disappearance caused by improper manual parameter adjustment, a generative adversarial network based on gradient penalty and Wasserstein distance (WGAN-GP) was proposed, and it was combined with VAE. The VAE-WGAN-GP model is formed. Secondly, aiming at the problem that the feature extraction model needs to be trained from scratch every time it is faced with a new task, and the training time is too long, the MVWP model is formed by combining meta-learning with VAE-WGAN-GP (VWP) mentioned above. Finally, the experimental results show that compared with VAE, VAE-WGAN and VWP, the training speed of MVWP model is increased by about 6 times, the reconstruction loss is reduced by 55.9%, 37.8% and 20.2%, respectively, and the reconstructed images are clearer.

Currently, more and more video data and terminal devices accessing video resources are available to users. Video platforms such as Tiktok and Youtube are gradually rising, and the user scale and video resources are increasing day by day, which brings an urgent practical demand for video-text data cross-modal retrieval. This paper proposes a video-text cross-modal retrieval algorithm based on multiple encoding. By encoding the global features, serial features and local features of video and text, the encoded features are mapped to the common embedding space for training, loss function calculation and optimization. Through experimental verification on MASR-VTT data set and comparison with existing methods, the overall performance R@sum increased by 9.22% and 2.86% respectively, which proved the superiority of this method.

To address the problem of low detection accuracy of YOLOv5s target detection algorithm in foggy traffic environment, an improved YOLOv5s-based pedestrian detection algorithm for foggy skies is proposed. The algorithm uses image defogging techniques to preprocess the data, expands the sample size by manually generating the Foggy Cityscapes-Person dataset through a new fog simulation pipeline algorithm, and enhances the network's ability to sense small targets under foggy skies by adjusting the loss function and the training method to improve the detection accuracy of pedestrians under foggy skies, resulting in an increase of the mAP value from 64.97% to The mAP value increases from 64.97% to 81.29%. The experimental results show that the YOLOv5s-ACE network model proposed in this paper effectively reduces the missing detection rate and false detection rate, and the model can quickly and accurately detect pedestrian targets in foggy sky scenes.

An improved algorithm based on fast extended random tree (RRT) was proposed to solve the 3d route planning problem of uav in complex environment. Firstly, the planning space modeling is carried out according to the threat factors of flight route. Secondly, in view of the large randomness of THE RRT algorithm, the heuristic distance function is introduced as the basis for the selection of nodes to be expanded, so as to increase the probability of nodes near the target being selected as nodes to be expanded, and improve the way of generating new nodes in the random tree to accelerate the convergence speed of the algorithm. Then, UAV dynamics constraints were incorporated into the new node to meet the flight path requirements. Finally, b-spline curve was used to optimize the smoothness of the initial route curvature discontinuity problem. Simulation results show that the improved algorithm has certain advantages in planning speed and route length, and can effectively solve the problem of UAV 3D route planning.

In this paper, we propose a labeling algorithm to deal with the shortest path problem with additional constraints take the forms of node Y must be included in or excluded from the shortest path if node X is passed through. We use a three dimensional label for each node in the network and define domination condition to delete unpromising paths so that the algorithm can quickly find the optimal path satisfying all the constraints.

Aiming at the problems of traditional Rapidly Exploring Random Tree (RRT) algorithm in route planning, such as slow speed, poor route quality and low flightability, a route planning algorithm based on integrated improvement of RRT was proposed. Firstly, in the selection of nodes to be expanded, the minimum sum of the distance between nodes and the target and the random sampling point is taken as the selection basis instead of the original method of determining nodes only according to random sampling points, so as to increase the probability of nodes near the target in the random tree being selected as nodes to be expanded. Secondly, in the process of node expansion, the reachable region of the next waypoint was determined according to the UAV dynamic constraints, and then multiple alternative nodes were randomly generated in this region. Then the route cost function is designed and the comprehensive generation value of the route formed by the alternative nodes is taken as the judgment criterion for node addition. Finally, B-spline curve smoothing is carried out to further improve the route quality. The simulation results show that the improved algorithm has obvious advantages in improving the planning speed and air route quality, and the obtained air route satisfies the UAV dynamic constraints and has high flightability.

The health bracelet composed of multiple sensors has large data acquisition, low data accuracy and poor fault tolerance. Therefore, the market application of health bracelet is limited. To solve the above problems, a multi-sensor data fusion method based on BP neural network is proposed. The simulation results show that the BP neural network model for multi-sensor data fusion processing, greatly improving the data accuracy, operation speed and robustness of multi-sensor.

Online public opinion warning for emergencies can help people understand the real situation, avoid panic, timely remind people not to go to high-risk areas, and help the government to carry out epidemic work.In this paper, key technologies of network public opinion warning were studied based on improved Stacking algorithm. COVID-19, herpangina, hand, foot and mouth, varicella and several emergency outbreaks were selected as public opinion research objects, and rough set was used to screen indicators and determine the final warning indicators.Finally, the warning model was established by the 50% fold Stacking algorithm, and the training accuracy and prediction accuracy experiments were carried out.According to the empirical study, the prediction accuracy of 50% Stacking is good, and the early warning model is practical and robust.This study has strong practicability in the early warning of the online public opinion of the sudden epidemic.

In the background of changing angle and multi-dimensional posture overlapping, it is necessary to extract and recognize the emotion of multi-modal face images, so as to improve the expression ability of facial emotion. A method of emotion recognition of multi-modal face images under changing angle and multi-dimensional posture overlapping background based on convolution neural network and morphological feature parameter recognition is proposed. Constructing a face feature collection model with variable angle and multi-dimensional posture overlapping background, performing fusion filtering on the collected face images with variable angle and multi-dimensional posture overlapping background, extracting the edge contour feature quantity of the face images with variable angle and multi-dimensional posture overlapping background, and filtering and denoising the original image by using morphological convolution neural network transformation method, Multi-modal wavelet scale decomposition method is used to decompose the emotion features of multi-modal face images under the background of changing angles and multi-dimensional postures, and the detection model of pixel points and similarity features of multi-modal face images under the background of changing angles and multi-dimensional postures is constructed. Morphological convolution neural network transformation method is used to transform the emotion features of multi-modal face images, and edge corner detection and expression feature point clustering analysis are combined to realize the emotion recognition of multi-modal face images. The simulation results show that this method has good performance in feature extraction and clustering of multimodal facial image emotion recognition, good expression ability of facial emotion and high image quality.

CT examination utilizes computational functions to achieve tomography of the human body based on the basic characteristics of X-rays, thereby unavoidably producing ionizing radiation that can cause damage to the human body. So, it is not applicable to pregnant women and children; Repeated exposure to CT irradiation in a short period of time may cause leukocytosis, fatigue, dizziness, vomiting and other symptoms. In particular, pregnant women, neonates and patients with extreme weakness are more likely to develop malformation, cancers and other adverse effects after exposure to radiation. However, endoscopic examination will induce physical damage to a certain extent, leading to potential risks of inflammation, and its process will cause fear and discomfort to patients, among which children are more likely to show fear than adults. In addition, there are many practical operation problems for endoscopic examination. So, it is not an ideal method. The medical infrared thermal imaging instrument adopts the high-tech infrared detection technology, which has no radiation and does not touch the human body. When the human body is diseased, the heat balance of the diseased part will also be destroyed. The infrared thermal imaging captures this imbalance based on the infrared rays from the human body to form an infrared thermogram, which reflects the temperature characteristics of the human body and thus will not harm the human body. The instrument has now already passed the clinical verification. Infrared thermography can well reflect the presentation of sinusitis, especially performs well in distinguishing whether the inflammation is acute or chronic. And the expression on infrared thermography is better than CT. Combined with artificial intelligence imaging algorithms, it can achieve feature analysis at the level of a single pixel and provide doctors with more detailed and accurate reference data, so as to implement efficient auxiliary diagnosis. The instrument is suitable for various types of hospitals and medical institutions, and even for home medical diagnosis when it is combined with a remote auxiliary diagnosis system.

A person's state is reflected in many aspects, such as emotions and body movements. Online teaching makes it difficult for teachers to accurately understand the learning status of students due to the separation of space between teachers and students. This paper extracts images from video cameras, from which identifies the learner's emotion, head posture and fatigue, and evaluates the learner's learning state by synthesizing the three-sided information. The seven emotions were divided into three categories: negative, positive and natural. Head posture is defined by Euler angles, and fatigue is determined by blinking frequency. Hierarchical decision-making method is used in the model for information fusion. The learning state assessment method proposed in this paper integrates the performance of both internal and external aspects of psychology and behavior, and has high reliability. Real-time understanding of students' learning status can help improve the effectiveness of teaching.

With the constant updating of applications and the emergence of various encryption technologies, a large amount of new encrypted network traffic is generated every day. Therefore, it is a challenging task to achieve continual learning of encrypted traffic. Existing encrypted traffic classification techniques can only handle a fixed number of traffic classes, which is not applicable to real network environments. In this paper, we proposed a continual encrypted traffic classification method based on WGAN, called CETC. The method takes advantage of the powerful data generation capabilities of WGAN to model the data distribution of encrypted traffic. When learning from a new traffic class, the samples from the old class is generated by WGAN to train the new classifier. We use the ISCX VPN-nonVPN dataset to test the performance of CETC. Experimental results show that WGAN can generate high-quality samples of encrypted traffic and the accuracy of CETC is higher than 93%. With its efficient and continual learning capability, CETC can be applied to various encrypted traffic detection and management systems.

Big data analysis based on artificial intelligence is particularly important in the era of Internet. The data is stored in different regions in industry. Meanwhile, sending data to servers generates huge amount of communication cost for centralized training. The distributed machine learning can resolve the storage of data and decrease the cost of data communication. But different distributed machine learning frameworks are also limited with the problems of low algorithm compatibility and poor expandability. The aim of this paper is building the distributed machine learning framework based on Ps-Lite and implementing algorithms in the framework. The framework is realized with asynchronous communication and computation methods. The algorithm implementation includes gradient-aggregating algorithm (distributed Stochastic Gradient Descent) and three regression algorithms (Logistic Regression, Lasso Regression and Ridge Regression). The algorithm implementation illustrates that common algorithms fit this framework with high compatibility and strong expandability. Finally, the experiment of Logistic Regression implementation proves the performance of the framework. The computation time of unit node is saved 50% with the increase of node number. The accuracy of the training model is maintained above 70% in the framework. The convergence efficiency of Logistic Regression is 3 times higher than that of the traditional one in the multiple-node framework.

In this paper, we propose a bi-level structured classifier integrating unsupervised and supervised machine learning models, which aims to improve the model's decision-making ability on classification boundaries by dividing the sample subspace to make full use of the multivariate attribute features and spatial structure of the data. The bi-level structured classifier utilizes the unsupervised clustering algorithms for subspace partitioning of sample data in the first layer, and selects the applicable supervised models to learn on the subspace samples in the second layer. We conduct a case study on a lithology dataset from the complex carbonate reservoirs for lithology identification. The classification results indicate that the bi-level integrated classifier (98.77%) is superior to the machine learning models (XGBoost: 97.67 %). And the ability of the bi-level integrated architecture is verified in effectiveness and generalization, and effectively improves the classification performance.

Infrared small vehicle target detection plays an important role in infrared search and tracking systems applications. The target detection methods based on deep learning are developing rapidly, but the existing approaches always perform poorly for the detection of small target. In this study, we propose an improved SSD(Single Shot MultiBox Detector) to improve the detection performance of infrared small targets from three aspects. First of all, we recommend using the stride convolution layer to replace the 3~6 maximum pooling layers in the original algorithm; second, design a shallow feature layer information enhancement module, semantically fusing the feature maps of the shallow feature layer and the deep feature layer, and using a new pyramid structure to detect the target; third, introducing residual unit and use the MSRA function to initialize the weights of the neurons in each layer at the beginning of training. To evaluate the Infrared-SSD proposed in this paper, the infrared vehicle data set created by this team was used to train and test the model. Experimental results show that Infrared-SSD has higher accuracy than the original SSD algorithm. For an input of 300pixel×300pixel, Infrared-SSD got a mAP(mean Average Precision) test score of 82.02%.

The deployments of deep learning models must be highly optimized by experts or hardware suppliers before being used in practice, and it has always been a long-term goal for the compiler community to enable compilers to automatically optimize code. However, there is no feasible solution in practice as running a program costs a considerable amount of optimization time to obtain a desired latency. Aiming at making up for the deficiency of long optimization time of TVM compiler, a novel optimized hybrid aging evolutionary algorithm is proposed to predict the running time of the code and accelerate automatic code optimization for Ansor, an auto-tuning framework for TVM. The algorithm alternately removes the worst and oldest individuals in the population during the evolution process. Unlike previous evolutionary algorithm, if an individual seeks to survive in the evolving population for a long time, it must have excellent scalability and flexibility, not just the individual's own adaptability. In this way, this algorithm not only ensures a strong search capability, but also improves the convergence speed and accuracy, significantly reducing the optimization time of tensor programs for deep learning inference. Experimental results show that the algorithm can accelerate convergence speed. For the same task, our algorithm provides 9% to 16% shorter optimization time on average while achieving similar or better optimization quality (i.e., inference time).

This paper focuses on how to ensure the availability and effectiveness of massive cloud data for industrial robots in the flexible production line, address the technical challenge in building a massive data cloud platform for industrial robots, and resolve the engineering problem of cloud based industrial robot cloud service application. To achieve this purpose, research is conducted on industrial robot hybrid cloud platform architecture, network technology, industrial robot big data system, autonomous learning cloud data processing and other technologies, which provides support for cloud service applications. It is suggested to combine knowledge atlas, digital twins, deep neural network, migration learning and other artificial intelligence technologies, which is conducive to remote monitoring and fault diagnosis cloud service applications. This has been verified and promoted in the handling, polishing, stacking, welding, assembly and other robots in 3C, mold, household appliances, automobile, furniture, electronic equipment manufacturing and other industries.

Artificial intelligence, electronics, and computers are developing faster and faster, embedded processors and machinery, intelligent robots have gained widespread attention in the market. The purpose of this paper is to design an intelligent robot navigation system based on data mining algorithm. Firstly, the navigation framework of intelligent robot based on ROS system is introduced. Then the key technologies of navigation are studied, and the path planning algorithm and self-positioning algorithm are introduced respectively. Finally, the robot navigation system is built according to the navigation framework, and the robot fixed-point navigation experiment is completed on the robot platform of this paper. In the navigation accuracy measurement experiment, A, B, C, and D are set as the coordinates of the target points, and each point is tested for navigation. The position error of the two points D in the x direction is about 0.05m, while the coordinate error in the y direction is larger, which is greater than the set 0.05m. The designed system can correctly construct the map of environmental information and can avoid obstacles and move to the set target position accurately and autonomously, which verifies the reliability and accuracy of the experimental platform and the navigation system.

There are fraudulent promotion behaviors in GitHub, which promotes Stars and Forks for specific repositories. It is harmful to the environment of the open source community, while it is not effectively detected by GitHub yet. This paper applies a heterogeneous neural network to detect repositories that are suspected of fraudulent promotion behavior. A heterogenous mini-graph neural network with attention mechanism and hyper-graph generation is proposed to detect repositories with cheating behaviors. Attention mechanism can dynamically balance the weight of semantics in heterogeneous information networks. Hyper-graph generation method can solve the problem of poor connectivity caused by many small graphs in the dataset. The experimental result shows that the model can effectively detect this kind of cheating behavior.

In the training of neural networks, the architecture is usually determined first and then the parameters are selected by an optimizer. The choice of architecture and parameters is often independent. Whenever the architecture is modified, an expensive retraining of the parameters is required. In this work, we focus on growing the architecture instead of the expensive retraining. There are two main ways to grow new neurons: splitting and adding. In this paper, we propose orthogonal initialization to mitigate the gradient vanish of the new adding neurons. We use QR decomposition to obtain orthogonal initialization. We performed detailed experiments on two datasets (CIFAR-10, CIFAR-100) and the experimental results show the efficiency of our method.

Despite the variety of cybercrimes, malicious Uniform Resource Locators (URLs) remain one of the most common threats to cybersecurity and bring huge economic losses every year. How to detect malicious URLs accurately has attracted great interests from both academia and industry. However, few focus on the multiclass malicious URL attack type detection and existing methods cannot provide robust performance due to the diversity of obfuscation strategies. In this paper, we propose a capsule-based deep neural network for malicious URL detection and classification, using character-level information from the URL string sequences. To be specific, our method transforms an input URL into character-level embedding representation firstly, then passes it into the designed convolution module to extract local features of different sizes and the local features are fed into the designed capsule module to retain the spatial hierarchical relationship of the URL string, extract accurate feature representation and output the accurate classification result finally. The experimental results on a public dataset constructed by four different classes of URLs show that compared with other baseline methods, our capsule-based method can achieve better detection and classification results, with F1-score of benign URL, malware URL, defacement URL and phishing URL at 98.94%, 95.81%, 99.63% and 94.04%, respectively. Due to the excellent performance of our capsule-based method for the detection of malicious URLs, it could be deployed in the main-stream web browsers to identify URL attack types and intercept malicious URLs effectively to protect vulnerable users against cyberattacks.

In recent years, the expansion of colleges and universities has led to a sharp rise in the number of students, and the number of poor students has also increased, which has greatly increased the difficulty and workload of financial aid for poor students. In order to improve the accuracy and efficiency of poor student identification, there is an urgent need to adopt digital and intelligent measures to assist poor student identification. In this paper, we propose a method to identify needy students using SVM and decision tree algorithm. Firstly, students' campus card consumption information is preprocessed to obtain the consumption poverty index of each student by SVM classification model. Then the decision tree algorithm is used to derive the student's family poverty index based on the student's family information. Finally, the comprehensive poverty index is calculated by weighted summation. The experimental results show that the proposed method realizes the statistics and analysis of students' consumption and family information, and it can identify poor students more accurately, which effectively improves the efficiency and accuracy of poor students' identification.

In according with nonlinear identification problem, an improved interval type-2 fuzzy c-mean clustering algorithm is proposed. A novel objective function is adapted in improved interval type-2 fuzzy c-mean clustering algorithm, which can reduce the influence of noise on clustering results. The proposed clustering algorithm is applied to T-S fuzzy model premise parameters identification and least squares is used for consequent parameters identification. The proposed identification algorithm is applied to double input single output model and actual thermal power unit main steam temperature data model, the identification results show that, the proposed algorithm has higher identification accuracy.

In the current oil fields in China, the horizontal well technology with a long horizontal interval has gradually become the core technology to develop conventional oil and gas reservoirs, and the accurate determination of the drag and torque of the drill string is the key. However, the determination of the friction coefficient is affected by many factors, and it is difficult to describe it clearly by mathematical formulas. According to the characteristics of friction factors, the method of calculating the friction coefficient of drill string is studied, and a prediction model of friction coefficient based on BP algorithm is established. Based on the predicted friction coefficient, the calculation method of drag and torque is analyzed, and a drag and torque prediction model based on BP algorithm is established. The experimental results show that the use of BP neural network can accurately predict the friction coefficient and torque, and the prediction of the friction coefficient can characterize the risk of sticking of the drill string to a certain extent, which facilitates the adjustment of drilling parameters on site to improve the safety during drilling.

With the changes of people’s consuming attitudes and the popularization of mobile payment, credit card seems increasingly indispensable in life. However, as the number of issued credit cards and credit lines is increasing, there emerges more and more fraud cases involving credit cards. Due to the rapid development of the Internet industry, the channels for capital flow have become unprecedentedly smooth, making it very difficult to prevent credit card fraud cases. If that continues, the development of banks and other financial institutions in the credit card field would be restricted, which might affect people's daily consumption and even the normal running of the society. The Bayesian Deep Learning method is used to quantify the uncertainty of credit card fraud prediction in this essay. Through experimental analysis, the accuracy of the model is over 99%. Compared with conventional classification models, this model has superior performance.