Abstract:Along with the rapid development of internet services, the distributed microservice-based application has gradually replaced the traditional application as one of the main forms of Internet applications. Distributed microservice-based applications boast scalability, high fault tolerance, and great availability, but they are often challenged by cumbersome installation, complicated deployment, and difficult maintenance. Kubernetes, as the most popular container-based cluster management system, is affected by coarse grains, inaccurate fault location, and other weaknesses. To address the above issues, this study proposes a fault detection method based on trace similarity matching: First, use injecting proxy to forward request traffic to collect tracking information about microservices. Then, collect the state information during normal operation of the system and record the performance of the system after the failure occurs by injecting known faults. Finally, take string edit distance as the standard for the execution tracking models of unknown and known faults. The edit distance serves as a standard to measure the similarity, and the possible cause of failure is identified. Experimental results show that the method can accurately describe the processing and execution tracking information of the request and find the cause of system failure with microservices as the granularity.

Abstract:As urban development has a great impact on land use, it is an important driving force for land use change. Landsat remote sensing images of Haitan Island in 1984, 1990, 1996, 2003, 2010, and 2017 were taken as data sources for defining land use categories by random forest to study the impact of the development of Pingtan Comprehensive Experimental Zone on land use. Land use changes, urban expansion and its influence on other land types in 5 periods were quantitatively analyzed. Results show that under the premise of feature variables optimization, the classification accuracy of each period is higher than 85%, meeting the requirements of subsequent analysis. Besides, the area of farmland in each period is the largest, followed by that of woodland (including forest and shrubland), and bare land has the smallest area. The overall farmland showed a decreasing trend, with the largest reduction in 1984–1990 (8.07%). Construction land expanded in all six periods, with the maximum expansion from 2010 to 2017 (11.82%). Moreover, urban expansion in each period was dominated by edge expansion, followed by leapfrog expansion and then infilling expansion.The development of cities on the island has a major driving effect on the changes of other land types, and the main source and destination of urban expansion are farmland. The main transformation modes of urban land include farmland, woodland, grassland, and wetland.

Abstract:The batch production process is an important part of process industry. Due to unique flexibility and high efficiency, it has been widely applied to food, chemical, pharmaceutical, plastic processing and other industries. The optimization and control of batch processes are extensively studied regarding their nonlinearity and repeatability. In this paper, the optimization and control strategies applied to batch processes in recent 30 years are summarized. Moreover, the difficulties in optimization and control are analyzed, considering nature of batch processes, and further possible development is estimated.

Abstract:Oxytocin is the first choice for labor induction, induced abortion, and prenatal fetal monitoring. Improper dose control of oxytocin during labor can increase the risk of adverse pregnancy outcomes. However, clinical oxytocin infusion mainly depends on the manual adjustment of medical staff, leading to subjective judgment errors in doses and high human cost. In addition, the existing oxytocin injection system lacks effective intelligent control means. Therefore, this study proposes to design an intelligent program for oxytocin dose control. It can extract the features of uterine contraction signals of a fetal heart monitor, and combined with fetal heart rate, electronic medical records, nursing records and other data, a prediction model of oxytocin doses was designed based on BOA-LightGBM. The experimental results show that LightGBM optimized by Bayesian is feasible to control oxytocin doses in real time compared with the traditional model. Therefore, this study can provide decision support for obstetric medical staff to adjust oxytocin doses during labor. It plays a positive role in reducing labor costs and enabling accurate drug delivery.

Abstract:Macro malware is widely used in advanced persistent threats. Macro obfuscation is low-cost and flexible, rendering traditional rule-based anti-malware systems insufficient. A gradient-boosting-decision-tree-based approach to detecting obfuscated macro malware is proposed. The approach performs large-scale feature engineering guided by the expertise of malware specialists, with fine-grained modeling for obfuscated macro malware carried out on top of lexical analysis, and massive samples are used to train the model. Experimental results show that the approach is able to precisely detect real-world obfuscated macro malware found in the network of enterprise customers, as well as those variants generated by mainstream obfuscation tools; 10-fold cross validation is carried out for a total of 4000 000 macro programs, giving a precision of 99.41% and a recall of 97.34%, which outperforms existing works.

Abstract:Fault localization is one of the most expensive activities in software debugging.The Mutation-Based Fault Localization (MBFL) assumes that the mutants killed by most of the failed test cases can provide a good indication about the location of a fault. Previous studies showed MBFL could achieve desired results in a Single Fault Localization Scenario (SFL-Scenario), but its performance in a Multiple Fault Localization Scenario (MFL-Scenario) has not been thoroughly evaluated. Recently, Higher Order Mutants (HOMs) have been proposed to model complex faults that are hard to kill, but whether HOMs can improve the performance of MBFL is still unknown. In this study, we investigate the impact of First Order Mutants (FOMs) and HOMs on MBFL in an MFL-Scenario. Moreover, we divide HOMs into three groups, i.e., accurate, partially accurate, and inaccurate HOMs, considering the mutation location in the program, to find which type of HOMs is more efficient in fault localization. Based on the empirical results on five real-world projects, we find that in an MFL-Scenario, HOMs can behave better than FOMs. The influence of the types of HOMs on the effectiveness of MBFL cannot be ignored. In particular, accurate HOMs can contribute more than inaccurate ones. Therefore, researchers should propose effective methods to generate this type of HOMs for future MBFL studies.

Abstract:Amid the widespread application of face recognition technology for authentication, various attack methods against face recognition systems have emerged over time. An identity authentication system based on lip reading recognition is proposed to cope with such security issues. It requires users to read the verification code during face recognition for authentication. The system not only checks the face, but recognizes the content of the speech through the lip reading recognition technology to compare it with the verification code. The user’s identity can be authenticated only when both the two links are passed. Finally, an identity authentication system based on lip reading recognition is designed, mainly including front-end, gateway, and back-end.

Abstract:The study of reservoir storage capacity can provide a valuable guide to the rational protection and utilization of water resources. With regard to the current situation and shortcomings of the estimation of reservoir storage, the remotely measured water levels of a reservoir are extracted by a Radar Altimeter (RA), and the surface water area is collected with a Multi-Spectral Instrument (MSI) and a Synthetic Aperture Radar (SAR). Then a reservoir storage estimation model based on data assimilation of multi-source remote sensing is established. Finally, a reservoir storage estimation system based on multi-source remote sensing data is developed, which can obtain remotely measured water levels and water area and estimate water storage capacity without contact. Experimental results prove that this system has stability and versatility. It is expected to be applied to flood control, drought resistance, reservoir dispatching and other fields to provide a technical support for the scientific and efficient management of reservoirs

Abstract:An automatic fare collection system for temporary parking lots based on plate number recognition is proposed to resolve parking difficulty and enhance the operation efficiency of the parking lots in densely populated areas. First, cameras at the entrance will take snapshots of vehicles. Subsequently, the system will identify, adjust and split license plate characters with image processing techniques, and then recognize the specific characters through a convolutional neural network. After that, in-out time, duration, and fare will be recorded and calculated in the MySQL database. Finally, an operation dashboard is designed to visualize and manage the system. The whole system is simulated by OpenCV and C++, and results prove a recognition rate of 97% or above. The device is portable with highly automatic operation, greatly facilitating the management of temporary parking lots.

Abstract:Deep learning has shown visible advantages in the artificial intelligence-based image classification. It usually costs plenty of time on manual information annotation for preparing image datasets. Then this study proposes an online collaborative system for image dataset annotation based on microservice architecture to improve the efficiency of annotating datasets and thus to accelerate the generation and iteration of deep learning models and applications. More users can join for image annotation after the heavy annotation task is divided into smaller ones. Besides, the system performance has been improved by introducing an object storage system and microservice architecture, and the integration efficiency of the system in the development progress has been enhanced by continuous integration and deployment.

Abstract:Road-side sensing is indispensable for a cooperative vehicle infrastructure system, through which vehicles could have sensing ability beyond the visual range by receiving road information via V2X communication. For the optimal sensing results in reality, RSU configuration needs to vary according to scenarios, which is both time consuming and labor intensive. Meanwhile, recognition of traffic participants based on machine learning is crucial to road-side sensing, requiring a huge amount of labeled data, and it is proven to be an inefficient way to label manually. However, these two problems can be solved by building a simulation system of road-side sensing. Experiment I shows the vehicle occlusion on extreme occasions by adjusting the height and orientation of lidar in the simulation system, which provides a recommended height for installment in reality. Experiment II proves the virtual data derived from the simulation system can be complementary to real data by mutual verification.

Abstract:At the moment, the detection technology of malicious code based on deep learning is a research hotspot in the field of malicious code detection. However, most researches focus on how to improve the algorithm to enhance the detection accuracy of malicious code, but ignore the lack of sample tags in the data set of malicious code, failling to train high-quality models. In this study, the problem of detecting isolated islands of data samples and data trustworthiness of malicious code is solved by Blockchain technology, and code features are extracted with the Markov graph algorithm. The training fusion block chain based on distributed deep learning has the advantages of decentralization, traceability and non-tampering, and the contributors of different computing power adopt synchronous training to update model parameters. The feasibility and great potential of this method are verified by simulation experiments and theoretical analysis.

Abstract:Microservices have been fused into the design framework of Internet applications over time. It is necessary to evaluate the application value of microservices fairly and transparently on a regular basis to improve the value of microservices, promoting developers to upgrade low-cost microservices. Therefore, a microservice evaluation system based on blockchain technology is proposed. It records the evaluation data of each dimension of microservices with the distributed accounting and consensus algorithm of Hyperledger Fabric blockchain technology, protecting the non-tampering and traceability of evaluation data. Besides, this study combines the analytic hierarchy process and the entropy method to build a comprehensive evaluation model of microservices and then calculates the comprehensive score of them. Experimental results demonstrate that the system can trace the source of the evaluation results of microservices, producing more reasonable results than a single evaluation model. It can provide effective data support for intelligent management of microservices.

Abstract:In recent years, Visual Question Answering (VQA) based on the fusion of image visual features and question text features has attracted wide attention from researchers. Most of the existing models enable fine-grained interaction and matching by the attention mechanism and intensive iterative operations according to the similarity of image regions and question word pairs, thereby ignoring the autocorrelation information of image regions and question words. This paper introduces a model based on a symmetrical attention mechanism. It can effectively reduce the overall semantic deviation by analyzing the semantic association between images and questions, improving the accuracy of answer prediction. Experiments are conducted on the VQA2.0 data set, and results prove that the proposed model based on the symmetric attention mechanism has evident advantages over the baseline model.

Abstract:Amid the continuous progress in technology, artificial intelligence technologies have been widely applied to the social life. This study develops a system that can identify abnormal behaviors in videos with predictive values. Firstly, we employ a Two-Stream Inflated3D (Two-Stream-I3D) convolutional neural network to extract features from the video. Secondly, we rely on Python to transform the features into those that can be recognized by a deep learning network. Finally, we perform GRNN training for abnormal probability regression. Experimental results show that the system can achieve the average accuracy of nearly 74% for abnormal behavior recognition during the detection of nearly 50 cases.

Abstract:On the basis of new IT infrastructure, such as cloud computing environment and big data platform, information application has obtained flexible and reliable underlying services. At the same time, through new service forms, such as Internet of Things application and mobile service, the business service ability is enhanced. However, these new techniques make information operation and maintenance face a lot of problems, such as frequent changes of operation and maintenance objects, constant adjustment of object relationships, diversity of operation and maintenance data formats. According to the current operation and maintenance status of power grid enterprises, we present a design of operation and maintenance data storage based on a graph model, which improves the ability of processing dynamic and unstructured operation and maintenance data. The experimental results show that the model has wider universality and stable data reading and writing performance. Therefore, the proposed scheme in this study effectively solves the business difficulties in the information operation and maintenance system of power grid enterprises in the new form of operation and maintenance.

Abstract:On the basis of the application of enterprise-oriented human resource management systems, we design a system development platform for enterprise management. Specifically, the platform supports sustainable development and the developed system has good scalability, expansibility, and portability. First, the technical solution of the proposed platform is emphatically described, and the advanced nature of the platform is analyzed in detail. Then, in combination with the actual business needs of the enterprises, the human resource management system built on this platform is described and the attendance calculation rules and implementation examples in attendance management are emphatically elaborated. Finally, through the smooth operation of the system, it is verified that the system development platform can develop projects in a modular, efficient, and rapid manner and adapt to complex business and demand changes, having broad promotion and application value.

Abstract:In order to ensure the safe and smooth operation of new energy vehicles, accelerate their development and promotion, and alleviate the environmental crisis, we adopt a “normalized voltage difference curve” to analyze the safety features of power battery packs. Furthermore, we formulate the control strategies of online early warning about faults in the charging process of electric vehicles on the short-term and mid-long-term scales. Thus, the basic performance of single batteries is analyzed and the trend in battery voltage can be accurately and clearly characterized. In this study, we mainly investigate the early warning control strategies of mid-long-term charging affected by the temperature rise in battery charging. Then, an objective function is established and the genetic algorithm is used for optimal control. Finally, the data about the charging status provided by the charging-pile monitoring platform verify the proposed online early warning of the charging data based on the battery model.

Abstract:To address the problems of blurred image boundaries, unclear image texture, and poor visual effect after inpainting, we propose a generative adversarial inpainting model that combines edge detection with self-attention mechanism in this study. Through this model, the contour information of the images can be extracted by edge detection, avoiding the problem of blurred boundaries after inpainting. Since the self-attention mechanism can capture the global information of images and generate precise details, a texture inpainting network incorporating the self-attention mechanism is designed. The proposed model is composed of an edge complement network and a texture inpainting network. First, the designed edge complement network completes the edges of a damaged image to obtain an edge complement image. Secondly, the texture of the missing region is accurately inpainted by the texture inpainting network combining the complemented edge image. Finally, the model proposed in this study is trained and tested on the CelebA and Place2 image datasets. The experimental results show that compared with the existing image inpainting methods, the model can greatly improve the accuracy of image inpainting and generate vivid images.

Abstract:In order to study the impact of the new satellite hall on the flight connection of transit passengers, analyze the transfer tension of transit passengers, and improve the utilization efficiency of airport resources, we investigate the distribution of boarding gates. On the premise of minimizing the number of boarding gates, considering the transfer tension of transit passengers, we establish a 0-1 integer programming model for aircraft-gate distribution. In order to improve the search ability of traditional heuristic algorithms, we propose an improved simulated annealing algorithm based on beam search by combining the neighborhood construction idea of variable neighborhood search and comprehensively employing the advantages of beam search and simulated annealing algorithm. Furthermore, the algorithm is solved by Java language. The results show that compared with the tabu search algorithm, variable neighborhood search algorithm, and ant colony algorithm, the proposed algorithm has better optimization effect.

Abstract:For the curriculum evaluation in colleges and universities, data-driven teaching management and decision-making issues are investigated in this study. First, the index system of curriculum evaluation from a school determines the data structure of multi-dimensional evaluation data covering students, teachers, peer experts, and teaching supervisors. After clean and conversion of the collected questionnaire data, a data set for data mining is constructed. Then, considering misleading suppression, we apply the improved Apriori association rule mining algorithm based on varying interest degrees to extracting the association rules between the evaluation indices. Finally, a comparison of the discovered relational patterns with the results using the traditional Apriori algorithm shows that the improved Apriori method used in this study can increase the efficiency and accuracy of knowledge discovery and has a prominent guiding role in curriculum construction.

Abstract:Underwater acoustic target recognition is to classify the targets through collecting the signals of the underwater acoustic targets and has very important and extensive applications in the fields of ocean exploration and monitoring technology. Due to the complexity of the marine environment, the diversity of target ship engines, and the background noise, underwater acoustic target recognition is difficult. Traditional feature extraction methods cannot extract effective feature representations to fully represent the targets. In order to solve this problem, we propose an underwater acoustic target recognition algorithm based on the improved bag of visual words. Specifically, this algorithm adopts the bag of visual words to extract high-dimensional features in a spectro gram and then adjusts the weights of the obtained feature vectors using the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm commonly used in the field of natural language processing. Furthermore, the vectors are input to a Multi Layer Perceptron (MLP) to classify and recognize the underwater acoustic targets. The experimental results show that the recognition algorithm proposed in this study achieves an accuracy of 92.53%, which is a significant improvement in comparison with the best Deep Boltzmann Machine (DBM) algorithm at present.

Abstract:To address the optimization of tool combination in the NC machining of complex pockets, we propose a tool optimization method for complex pocket parts based on the genetic algorithm in this study. First, a feasible tool set for pocket machining is constructed by an offset loop algorithm. Secondly, with machining efficiency and tool cost as the optimization objective, an optimization model of tool combination for complex pocket parts is established to decrease the difference between each feature of the pocket parts and the overall optimal tool combination. Furthermore, a directed graph and an improved genetic algorithm are developed to solve the optimization model of tool combination. Finally, the feasibility of the proposed method is verified by testing the optimization of NC machined tool combination for a complex pocket part.

Abstract:Collaborative filtering algorithm widely used in the recommendation systems has the problems of sparseness and cold start. For this reason, a recommendation model based on deep neural networks and dynamic collaborative filtering is proposed in this study. The model combines a pre-trained BERT model with bidirectional GRU to extract hidden feature vectors from users and commodity reviews. Furthermore, coupled CNN is used to construct the score prediction matrix and the temporal changes in user interests are incorporated through dynamic collaborative filtering. Finally, the experiments on an Amazon’s data set show that the proposed model increases the accuracy of commodity score prediction.

Abstract:When a genetic algorithm is used to solve job shop scheduling, in order to obtain the optimal solution and increase the convergence speed of the algorithm, we propose an improved genetic algorithm in this study. The goal of the algorithm is to minimize the maximum completion time. First, the population size is doubled during the initialization to increase the diversity of the population and a new fitness function is adopted to make chromosome distinguishing easier in the iteration. Then, chromosomes are selected via roulette. Furthermore, crossover is completed by Precedence Operation Crossover (POX) and mutation by Reciprocal Exchange Mutation (REM). Finally, the optimization ability and convergence speed of the proposed algorithm are improved by adjusting the crossover and mutation probability with self-regulation. The simulation results show that the improved genetic algorithm has faster convergence, stronger optimization ability, and better optimal solution than the traditional one and thus it is more suitable for the processing and production in job shops.

Abstract:Nowadays, researchers are paying special attention to the classification algorithms related to facial expression, and improving the accuracy of classification is of practical value to frontier fields such as artificial intelligence. The classic methods for image classification are linear discriminant analysis and sparse representation. This study proposes an improved collaborative representation algorithm, aiming at the high computational complexity of image classification, feature utilization, and classification accuracy. First, the block weighted local binary patterns are applied to the texture feature vector of each sub-block. Then, principal component analysis is used to avoid the curse of dimensionality and also increase the running speed of the proposed algorithm. Finally, a collaborative-competitive representation algorithm is adopted to obtain the final classification results. In conclusion, the combination of feature extraction with collaborative representation algorithms has a good classification effect.

Abstract:At present, there are few studies on dust image recognition using the deep learning method, and the recognition rate of dust images is low due to the application of some traditional methods. In view of this situation, a dust identification method based on an improved residual network is proposed. The method applies ResNet-50 network to a dust data set, and the network structure is improved. Then, spatial pyramid pooling is added to solve the problem that the size of the input images is not fixed. In addition, the pyramid pooling is changed to average pooling, and the method of expanding a feature graph is applied to the backbone network, which is conducive to extract more fine-grained features, improve the performance of the model, and increase the recognition rate. In conclusion, the proposed method has high accuracy and provides an effective scheme for dust identification.

Abstract:In recent years, the field of computer vision has developed rapidly, so it is particularly important to obtain high-quality image information. Image defogging is a technique widely used to enhance the visual quality of images insevere weather conditions. The dark channel prior method achieves image defogging by estimating atmospheric light. Although it has achieved good results, there are still problems that the atmospheric light is overestimated and is not suitable for large white areas. Aiming at the existing image defogging problems, we propose a deep learning method based on the improved DehazeNet for image defogging in this study. This method introduces a depthwise separable convolutional layer inestimating the transmission map. In order to enlarge the receptive field, dilated convolutionis used in atmospheric light. The experimental results show that the improved defogging algorithm in this study can effectively restore the foggy images and improve the image quality and has an excellent defogging effect in both quantitative and qualitative evaluation compared with other comparison algorithms.

Abstract:With the development of the times, online car-hailing has gradually become an important mode of travel in today’s society. This new travel mode greatly reduces the travel costs and makes people’s lives more convenient. Online car-hailing demand forecast is an important part of the artificial intelligence transport system and has high application value. However, traditional research ignores the impact of the social attribute similarity between the destination and different regions when modeling, making the characteristics of the models incomprehensive and the forecast accuracy of the algorithms low. In response to the above problems, a Multi-Graph Spatial-Temporal Graph Convolution Neural network (MGSTGCN) is proposed to solve the forecast problem of online car-hailing demand. The network consists of spatial and temporal components. The network associated with spatial problems models the similarity of geographic information, mobile information, and social attributes through graph convolution, and the temporal problems are processed by combining the attention mechanism with the LSTM network. In the experiments, we comparatively analyze the proposed model with four mainstream network models, and the results show that this model can more effectively capture the spatial-temporal characteristics of online ride-hailing demand data and increase the forecast accuracy.

Abstract:In a recommendation system, users’ interest in items changes dynamically and is affected by various factors such as users' own historical behaviors, their friends’ historical behaviors, and even short-term hot spots. How to describe users’ temporal interests in a recommendation system and extract effective information has always been one of the challenges for the recommendation algorithms. On the basis of the Graph Neural Network (GNN) recommendation algorithm, we propose an improved graph network algorithm based on the Attention Gated Recurrent Unit (Attention-GRU) in this study. Furthermore, feature modeling is performed on the temporal interactive history of users and items, and in combination with social network, the temporal characteristics are transmitted between users and items. In addition, the proposed algorithm is verified on the Ciao and Epionions data sets and compared with other related work, proving that the model proposed in this study can effectively extract the temporal characteristics of users and items and improve the effectiveness of the recommendation systems.

Abstract:Quadrotor aircrafts often encounter problems such as low stable attitude accuracy and poor resistance to interference when performing tasks. A neuro-fuzzy PID control algorithm is proposed to adjust the fuzzy rules and membership functions of the original fuzzy PID control. The neuro-fuzzy PID control algorithm is combined with the established dynamic model of the quadrotor aircraft. The traditional PID and fuzzy PID control algorithms are used as comparison algorithms with human interference factors considered to validate the neuro-fuzzy PID controller. The Matlab/Simulink simulation experiment shows that the neuro-fuzzy PID control has better control effect on quadrotor aircrafts than comparison algorithms according to its faster response, higher stable attitude accuracy, and stronger resistance to interference.

Abstract:Aiming at the needs of multimedia information security and copyright protection, this study combines the ElGamal public key crypto system with Patchwork digital watermarking algorithms to propose a new homomorphic ciphertext domain-commutative watermarking encryption algorithm. The algorithm maps the operation of embedding the watermark in the plaintext domain to the ciphertext domain based on the multiplicative homomorphism of ElGamal, swapping the operations of encryption and embedding watermark. The watermark can be extracted in the ciphertext domain or in the plaintext domain. The experimental results show that the order of embedding watermarks and encrypting data does not affect the generation of ciphertext data containing watermarks and the extraction of watermarks from ciphertext and plaintext, which ensures the confidentiality of embedding watermarks and the security of multimedia data in distribution management. Also, the comprehensive performance of watermarking algorithms is improved.

Abstract:Smart cities depend on information and communication technology to sense and analyze the key information in the core system of urban operation. It needs to make timely and effective intelligent response to urban security threats and common cases. In order to improve the effectiveness and accuracy of identifying common urban cases, this study proposes an automatic identification algorithm for common urban violations. The improved convolution neural network extracts image features, and BP neural network is used for evaluation. On the VOC data set, the algorithm is compared with YOLO and SSD in performance. The results show that the mAP of the improved convolutional neural network can reach 76.5%, and the accuracy of identifying various types of cases is more than 72%, and the accuracy of identifying “graffiti and posted advertisement” is 83.4%. The image recognition technology of cases in a smart city developed in this study can enhance the efficiency of case processing, save human and material resources, and thus can be used to assist urban management, supervision, and administrative law enforcement.

Abstract:The category of e-commerce data is of great significance for its analysis. The classification based on human resources cannot adapt to the massive e-commerce data nowadays, and the classification based on traditional algorithm models can hardly extract valuable artificial features. In this study, the BiLSTM model integrated with an attention mechanism is introduced to classify e-commerce data. The model includes embedding layer, BiLTM layer, attention mechanism layer, and output layer. The embedding layer loads the word vector trained by Word2Vec; the BiLSTM layer captures the context of each word; the attention mechanism layer allocates weights for each word to synthesize new sample features. The experimental results show that the classification accuracy of the attention mechanism based on the inverse class frequency reaches 91.93%, which is improved compared with the BiLSTM model without the attention mechanism and other attention mechanisms introduced. This model has a good effect in the classification of e-commerce data and points out a new thinking direction for the introduction of attention mechanisms.

Abstract:Relation extraction is a core technology to construct a knowledge graph. The complexity of Chinese grammar and sentence structure as well as the limited feature extraction and poor semantic representation of the existing neural network model restrict the relation extraction of Chinese entities. A relation extraction algorithm based on a BERT pretraining model is proposed in this study. It preprocesses the corpus by extracting keywords, entity pairs and entity type and integrating them to strengthen the semantic learning ability of the BERT model, greatly reducing the loss of semantic features. Results are obtained by a Softmax classifier, which show that this model is better than the existing neural network model. In particular, the model reaches a F1-score of 97.50% on the Chinese data set.

Abstract:The IEC 60870-5-103 protocol is an information interface supporting standard applied to relay protection equipment and transmits mainly the information related to relay protection. The message is transmitted in plain text and has poor security for a lack of encryption measures and digital signature mechanism. A communication experiment environment between the master station and the DTU terminal is built to verify that there are hidden dangers in the 103 protocol of Ethernet transmission. A man-in-the-middle attack test is carried out on the system by detecting ARP spoofing. The experimental results show that the 103 protocol of Ethernet transmission faces the risk of man-in-the-middle attack. In order to improve the security of the protocol, we propose a two-way identity authentication mechanism based on an asymmetric cryptographic algorithm and rely on a symmetric encryption mechanism and digital signature technology to ensure the confidentiality and integrity of the transmitted message. Finally, the method is validated through simulation tests.

Abstract:In order to solve the problem of unbalanced sample distribution in a dataset in Speech Emotion Recognition (SER), this study proposes a SER method combining a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) units with data balance and an attention mechanism. This method first extracts the log-Mel spectrogram from the samples in a speech emotion dataset and devides the sample distribution into segments according to sample distribution for balance. Then, this method fine-tunes the pre-trained CNN model in the segmented Mel-spectrum dataset to learn high-level speech segments. Next, given the differences in the emotion recognition of different segments in speech, the learned segmented CNN features are input into the LSTM with an attention mechanism for learning discriminative features, and speech emotions are classified with LSTM and Softmax layers. The experimental results in the BAUM-1s and CHEAVD2.0 datasets show that the method proposed in this study has much better performance than conventional methods.

Abstract:In a multi-secret sharing scheme, a large number of public values are generated to ensure the secure and correct reconstruction of multi-secrets, and participants also need to keep a large amount of information. In order to reduce the number of public values and the information that participants should keep, this study designs a multi-secret sharing scheme based on the Chinese Remainder Theorem (CRT) and Shamir (t, n)-threshold secret sharing scheme in which shares can be used more than once. Specifically, the shares generated by polynomials are aggregated to generate public values by CRT, which reduces the number of public values. Transformed value and discrete logarithms are used to protect the shares of participants. In a multi-secret sharing scheme, multiple secrets can be shared at one time; different secrets can be shared in access structures with different thresholds; participants can verify the secrets recovered; the number of public values is fewer; each participant only needs to store one share which can be used repeatedly.

Abstract:In order to improve the accuracy of gait pattern identification for wearable sensor data, we propose a new model of deep learning-based gait pattern identification, which combines a convolutional neural network with a long short-term memory neural network in this study. This model makes full use of the local spatial features of data obtained by a convolution neural network and the excellent correlation of intrinsic feature time of data obtained by a long short-term memory neural network model. It can effectively mine the temporal-spatial gait features implied by high-dimensional, nonlinear, and random temporal gait data of wearable sensors which are closely related to gait pattern changes, improving the model’s classification performance of gait patterns. The UCI HAR data set from University of California is used to validate the proposed model. The experimental results show that the model can effectively collect temporal-spatial gait features embedded in the gait data of wearable sensors. This can reach classification accuracy of 91.45%, precision of 91.54%, and recall of 91.53%, signaling significantly better classification performance than that of traditional machine learning models, which serves as a new scheme for the accurate gait pattern identification of wearable sensor data.

Abstract:Aiming at the problems such as low efficiency, limited ability of feature extraction, and poor adaptability of traditionalclassification methods for UAV targets, this study proposes a UAV classification method that introduces attention modules to optimize deep convolutional neural networks by analyzing the characteristics of UAVs and existing classification methods. Multiple sets of comparative experiments are designed for a model structure of a convolutional neural network with three convolutional layers, three pooling layers, and two fully connected layers according to the experimental results for training to obtain the optimalclassification model for UAV targets. Then, the convolutional block attention module is introduced to strengthen and suppress feature map elements, and the batch normalization layer is introduced to accelerate convergence and improve generalization capabilities of the model. Experimental results show that after introduction of convolution block attention modules and batch normalization layers, the recognition rate of the classification model for UAV targets rises by 1.5% to 92.44%. Its advantages of high recognition rate and fast convergence over other neutral network models can basically meet the requirements of UAV target classification in actual scenes.

Abstract:To realize the fusion management of the power distribution communication network and improve the quality of service and the management of operation and maintenance, this paper studies the multi-source and multi-communication intelligent management technology of the power distribution communication network under flat management, and establishes a resource mapping model of hybrid networks based on particle swarm algorithms. The model can map WMN and PLC to the same physical network. Their advantages are brought into full play with different tasks assigned. Besides, more sub-carriers are adopted to improve service throughput. The simulation of the newly established model and its comparison with the genetic algorithm demonstrates the superiority of the proposed algorithm.