Abstract:So far, most researches on LoRa technology are about single-application oriented IoT, low utilization of configurable parameters leaves room for further optimization of network performance. In order to adapt to the growing transmission requirements of heterogeneous multi-type services, it is increasing essential to optimize the performance of the LoRa network. To address the above issue, a dynamic parameters adaptive configuration strategy based on simulated annealing genetic algorithm is proposed, which can improve the number of end devices and data throughput supported by single gateway LoRa network while limiting energy consumption. The simulation results based on LoRaSim reveal that the proposed method outperforms ADR by 25.6%. By simulating the single gateway LoRa network of nearly over 1000 devices, the experimental results show that when packet generation rate 1/100 s, dynamic parameters adaptive configuration strategy proposed in this study can guarantee PDR above 90%. This method can adapt to the data transmission needs of multi-heterogeneous applications and effectively improve the data throughput while ensuring the PDR of each applications.

Abstract:Detecting malicious URL is important for defending against cyber attacks. In view of the problem that supervised learning requires a large number of labeled samples, this study uses a semi-supervised learning method to train malicious URL detection models, which reduces the cost overhead of labeling data. We propose an improved algorithm based on the traditional co-training. Two kinds of classifiers are trained by using expert knowledge and Doc2Vec pre-processed data, and the data with the same prediction result and the high confidence of the two classifiers are screened and used for classifiers learning after being pseudo-labeled. The experimental results show that the proposed method can train two different types of classifiers with detection precision of 99.42% and 95.23% with only 0.67% of labeled data, which is similar to supervised learning performance and performs better than self-training and co-training.

Abstract:To solve the blank of current research on the prediction of density limit disruption of EAST, 972 density limit disruptive pulses selected as data sets from the EAST’s 2014 to 2019 discharge. 13 diagnostic signals were chosen as features. Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) was used as models and the disruption risk was used as output to build the predictors. The experimental results show that for density limit disruptive pulses, under different alarming times, the successful prediction rate of LSTM (around 95%) is higher than that of MLP (85%), and for non-disruptive pulses, the false prediction rate is around 8% for both MLP and LSTM. The performance of LSTM has great improvement than MLP, shows the feasibility of building EAST density limit disruption system with neural networks and improving the response performance of disruption avoidance and mitigation system.

Abstract:Polarimetric Synthetic Aperture Radar (PolSAR) is a type of microwave imaging radar that avoids the influence of weather, light and clouds, and it has the capability of all-day and all-weather imaging. Therefore, PolSAR images have become one of the main data sources for land classification based on remote sensing image. From the perspective of technical methods, this paper discusses the methods and applications of land classification based on PolSAR image in recent years. It introduces the technical methods and experimental effects, and analyzes the development trend of land classification based on PolSAR image.

Abstract:With the rapid development of modern technology, the data center has become the IT infrastructure of the information society, storing and managing a large amount of key data. At present, the management of data centers mostly relies on experienced professional operation and maintenance personnel to use computers to automatically monitor equipment room equipment indicators and make multiple inspections of equipment, which is time-consuming and tedious. Deep learning and artificial intelligence technologies are currently attracting more and more attention and have achieved many successful applications in the Internet and industrial fields. This study designs a Gated Recurrent Unit (GRU) based deep learning framework to automatically diagnose equipment failures in cloud data center equipment rooms and combines timing information to predict future states based on past equipment operating status information. Series data are split into fixed time windows as input to the bidirectional GRU layer which makes the network learn the time dependency relationship in data points. Besides, we add an attention layer and embedding layer after the output of GRU unit, to help the neural network learning more efficient features for prediction task and further dimension reduction. At last, multi-layer perception is used to classify the data. Experimental results based on real data sets show that proposed neural network framework based on GRU can accurately detect cloud data center faults compared with LSTM, SVM and KNN.

Abstract:Big data industry has risen to the national strategy. The establishment of big data laboratory and experimental curriculum system is necessary for training big data technical personnel. This paper combs the knowledge system of big data, analyzes the training objectives and career orientation of the major of “data science and big data technology” and the major of “big data technology and application”, and clarifies the key knowledge that big data students should master and the professional skills that need to be cultivated, introduces the mainstream big data ecosystem, selects the most general big data architecture, proposes different plans to build big data laboratory in single machine environment, single machine virtualization environment, shared big data cluster environment and cloud computing environment, and designs the big data experiment curriculum system and experiment projects.

Abstract:The traditional microgrid electricity transaction mode has such problems as centralization and data transparency, and because of the features like decentralized, tamperability, traceability and untrusted characteristics of the Blockchain is widely used in electricity transaction. Current research focuses on consensus mechanism, trading mechanisms and privacy protection, and most of them require traders to make constant bidding every cycle. This paper presents a consortium Blockchain without bidding electricity trading mechanism based on the consensus of proof of authority (PoA), electricity consumers do not need to bid the pre-deposit of electricity, can automatically trade electricity based on smart contracts before the electricity bill runs out, the settlement of funds using energy tokens. Finally, the comparison and analysis of the other 4 existing Blockchain electricity trading schemes, and the simulation of 5 producers and 5 consumers of electricity transactions, analysis of the buyer’s and and seller’s earnings. The results show that the scheme benefits both buyers and sellers at the same time, with lower economic costs and high consensus efficiency.

Abstract:Ecosystem changes have an important impact on our production, life and health. The ecological cover data contains important characteristics of ecosystem changes. In order to use ecological cover data to study the spatial division and time-series changes of ecosystems, we establish the regional ecological cover change data model based on ecology data transfer matrix. Then the ecological change visual analysis system ECOVIS is designed based on the dimensionality reduction algorithm, in which the improved Sankey chart is used to realize the visualization of the ecological cover change data, the Scatter plots are designed for interactive cluster analysis, and map-based heat maps are used to display the spatial distribution of selected cluster data. The system is used to analyze the data of ecological cover in China. We analyze the time sequence of forest and town changes in local areas, and to cluster and divide the whole ecological space. The analysis results show that the method has good spatial clustering and time sequence comparison functions for the data of ecological system, which can improve the analysis of ecological cover data effectively.

Abstract:The research on the health management of weapon systems is a hot topic in many countries around the world. Based on the working characteristics of the new-type artillery control system, this work studies the implementation framework, key technologies, and system development of the data-driven health management system for the artillery control system. The multi-task program design and implementation based on VxWorks health management system based on embedded operating system was expounded. Focusing on data storage and real-time data visualization, the embedded database design and Tilcon-based interface engineering are designed and implemented. After actual engineering verification, the health management system has a reasonable task division, high system real-time reliability, and significantly improved the supportability of the artillery control system.

Abstract:The grid wiring diagram adopts the entire grid model of the dispatching and control cloud as the basis for graphic display and interaction. Aiming at the single display style of grid graphic and the deficiencies of visualization of power grid operational status of traditional power grid dispatching and control system, this study proposes a holographic display technology of power grid wiring diagram based on dispatching and control cloud, which is able to show the operation status of the power grid in different temporal, and achieve the historical scene inversion, real-time status monitoring and future operation mode analysis of the power grid, and then provide users with holographic display of various business scenarios based on graphics. This technology has been deployed and formally launched in the Area III of provincial and above power dispatching and control centers in multiple regions, meets the needs of displaying various business scenarios on the graphics, and improves the entire process control ability of the power dispatching and control centers.

Abstract:China Telecom NetCare service is a unified cloud network integrated monitoring platform. The system monitors client’s network device, cloud applications, virtual resources, VPN and Internet. It is developed by China Telecom to provide the monitoring management and analysis services of network layer, transmission layer and application layer for enterprises. It serves the government and enterprise customers of China Telecom, and has very strict requirements for availability of the system. The high availability of a system is influenced by hardware, network, operating system, database, middleware, application itself and other aspects. This study focuses on the application layer’s high availability of the NetCare system, and calculates the availability of the system with actual measurement and reasonable estimation of data. Finally, we come to a conclusion according to the paper: when the remote disaster recovery system is excluded or included, the whole system’s availability is 99.9978% and 99.45% respectively.

Abstract:In order to improve the level of information construction in colleges and universities, it is necessary to acquire the information needs of teachers and students deeply and comprehensively. WeChat at Work, as a mature information development platform, has built a good application ecology, with consistency and easy-to-use user experience. Through the research on the application and development mode of WeChat at Work, and combined with the characteristics of information-based demand management in colleges and universities, an easy-to-use demand management system is developed. This paper describes the system development architecture design, module composition, implementation process and implementation results, and tries to use Spring Boot to optimize the rapid development of WeChat at Work. The system has been developed and put into use, teachers and students can easily use WeChat at Work to submit all kinds of information needs, the information administrators can efficiently carry out demand management, has achieved sound application effect.

Abstract:To solve problems like automatic testing line production fault is not easy to find, the found fault is difficult to solve, and other issues, this study introduces a kind of online fault diagnosis method based on improved Hierarchical Signed Directed Graph (HSDG). This method firstly sets up models and hierarchical clusters, gets automatic testing line’s hierarchical SDG model, then outputs SDG model samples combined with the Dynamic Kernel Partial Least Squares-Support Vector Regression (DKPLS-SVR) model to calculate all nodes symbols. Finally, the paper presents the complete fault diagnosis process based on “offline analysis” and “online diagnosis”. The results show that, this method is robust which can improve the diagnostic speed and resolution, the fault diagnosis results have certain ability of explanation and guiding significance in the production.

Abstract:Composite materials are composed of at least two materials with different properties. Based on this feature, high-throughput material calculation and multi-scale simulation methods and concepts are particularly suitable for the theoretical design of composite material formulations. To this end, we have developed a high-throughput calculation and screening interface application software based on Materials Studio that supports composite material formulation design. At present, the software mainly supports two calculation modules, Amorphous Cell and Force Plus in Materials Studio. Through the call of the module interface, various formulas of high-throughput composite materials are realized, and high-throughput automatic process calculation and screening based on molecular dynamics. Compared with directly using Materials Studio software, this software has the characteristics of “one-click, automatic process, high-throughput screening” and so on. At present, the software has realized micro-scale automatic process screening based on the molecular dynamics module Forceite Plus. In the next step, we will develop a Mesocite module interface using dissipative particle dynamics method on the mesoscopic scale, realize cross-scale calculation simulation and high-throughput automatic process screening, and carry out formula design of epoxy resin matrix composite. Users must have Material Studio copyright to use this software.

Abstract:Aspect level sentiment analysis is a more fine-grained sub task of sentiment analysis tasks, the purpose of which is to predict sentiment tendencies of a certain aspect. At present, most aspect level sentiment analysis tasks use neural networks to extract semantic information of sentences, and then predict emotional polarity. Based on this, this study proposes a semantic representation method based on sentence structure information, that is, the fusion of sentence structure information in the part of speech sequence of the statement. In this work, two Bi-LSTM are used to extract the semantic feature and the structural feature of the statement, and the semantic representation based on sentence structure is constructed. Then, the given aspect level vectorization is embedded into the semantic representation based on the sentence structure, and then sent to the Softmax layer for sentiment classification. Experiments show that the semantic representation method based on the information of sentence structure is more effective.

Abstract:Suicide is a serious public health problem in today’s society. It is of great social significance to conduct in-depth research on suicide prevention. This work studies the suicide risk assessment method based on Microblog text. According to Microblog text features, in order to solve the bottleneck problem of the current neural network single structure in the prediction accuracy improvement, this study proposes a hybrid architecture neural network model nC-BiLSTM and applies it to the Microblog text suicide risk identification. The model extracts local feature information by using multiple convolutional layers of different convolution kernels, and extracts contextual semantic feature information of sentences by using Bidirectional Long Short-Term Memory (BiLSTM) network layer. The experimental results show that the recognition accuracy, recall rate, and F value of the nC-BiLSTM model are better than other models. The results of this study can be applied to the early intervention of suicide prevention.

Abstract:More details may be lost and considerations of the surrounding environment of the road are inadequate when extract the road from GF-2 remote sensing satellite which based on the deep neural network. Aiming at these problems and based on the existing researches results, this study proposes an improvement proposal which using the full convolutional neural network to extract road from remote sensing images. The scheme innovatively researches the algorithm principle of the full convolutional neural network and outputs the pre-graded GF-2 images in a certain size. Then, the output images and the corresponding labels are input into the improved full convolutional neural network. At last, a road extraction image with higher recognition accuracy is obtained by combining residual unit and increasing the number of network layers. Experiments show that the effect on road extraction of GF-2 satellite images is improved in the same sample, the integrity and accuracy of the road are also improved.

Abstract:Internet financial institutions have many credit businesses, and some of the newly launched businesses cannot establish an effective credit scoring model due to the lack of customer data. This work studies the application of transfer learning ideas to this problem and uses existing customer data from other businesses to help new businesses build effective credit scoring models. This study proposes a deep learning method based on the combination of Triplet-Loss and domain adaptation to re-encode existing business data, and transfers the knowledge obtained after re-encoding to the model of the new business, and finally uses XGBoost as the classifier. In view of the above problems, the model proposed in this study has improved the effect compared to traditional machine learning methods, and solved the problem to a certain extent.

Abstract:Aiming at the difficulty of small target detection in current target detection technology, a small target detection algorithm named improved Bi-directional Single Shot multibox Detector (Bi-SSD) based on Single Shot multibox Detector (SSD) is proposed. This algorithm designed a small object feature improvement module for the shallow features of SSD. In the classification and regression parts of the network, a 6-scale Bi-directional Feature Pyramid Network (BiFPN) is designed as classification and regression sub-network according to multi-scale feature fusion method and BiFPN structure. Experimental results show that Bi-SSD has better detection performance than the original SSD on PASCAL VOC and MS COCO object detection datasets. On VOC2007+2012, Bi-SSD achieves 78.47% mAP, which is an increase of 1.34% compared to the original SSD algorithm. On COCO2017, Bi-SSD achieves 26.4% mAP, which was an increase of 2.4% compared to the original SSD algorithm.

Abstract:The quality of the credit evaluation classifier can directly affect the profitability of credit financial institutions. The traditional grid search takes a lot of time for parameter optimization. Based on this, we propose an improved grid search to optimize the XGBoost (GS-XGBoost) personal credit evaluation algorithm. After using the feature selection based on random forest, the algorithm uses the improved grid search method to optimize the parameters of n_estimators and learning_rate in XGBoost to establish an evaluation model. We analyze the credit data selected from the UCI database to compare with support vector machines, random forests, logistic regression, neural networks, and unimproved XGBoost. Experimental results show that the F-score and G-mean values of the model are improved.

Abstract:In order to cope with problems related to malicious copying and illegal use of software, security authorization of software with intellectual property rights is an effective means to ensure software security. In the software authorization process, it is particularly important to use a highly secure encryption algorithm for the authorization data. This study proposes a new authorization encryption method, that is, a hybrid encryption algorithm of DPAPI encryption algorithm and RSA digital signature algorithm. This algorithm uses the DPAPI encryption algorithm to encrypt the client’s application for authorization information, while ensuring the encryption and the correctness of the software authorization at the same time, and then uses the RSA digital signature algorithm to digitally sign the server-side authorization information to ensure the unforgeability of the authorization information. The verification of the hybrid encryption algorithm shows that the algorithm has certain feasibility in the software authorization process.

Abstract:A remote detection method for vehicle collision events is proposed. This method is used to analyze the speed and acceleration signals of vehicles by using machine learning so as to monitor the vehicle running states online. The front-end equipment of a vehicle collects the speed and acceleration signals in real time, preliminarily identifies possible collision signals and sends them to the back-end server through wireless network. The back-end server accurately identifies the collision signals and estimate the damage degree of vehicles. The detection methods of collision events and collision damages are given in this article, and the experimental results show that the proposed method is effective.

Abstract:In view of the shortcomings in accuracy and complexity of community discovery algorithm based on seed node selection, a Node2Vec overlapping community discovery algorithm is proposed. First, the vector representation of each node in the network is learned by using Node2Vec algorithm to calculate the similarity between nodes. Second, the node influence function is used to calculate the node influence and find out the seed node. Then the community extension optimization is carried out based on each seed node. Finally the high quality overlapping community structure is excavated. In this study, several real networks are selected for comparative experiments, and the results show that the proposed algorithm can find high quality community structures under the premise of ensuring sound stability.

Abstract:In order to solve the problem that the discriminative scale space tracking (DSST) algorithm cannot track when pedestrians reappear after being completely occluded for a long time, an improved tracking algorithm (DDSST) is proposed. Under the DSST framework, pedestrian tracking is first performed. And then the high confidence index calculation strategy is introduced as the tracking accuracy credibility feedback mechanism. when the tracking is lost, the deformable part model (DPM) is used to relocate the tracking target. Finally, the accuracy of the DDSST algorithm is verified by evaluating the online Object target Tracking Benchmark (OTB) dataset and the video sequences captured in the actual environment, and compared with other tracking algorithms. Experimental analysis shows that the distance precision and success rate of the improved algorithm are improved by 4.1% and 6% compared with DSST, and the performance is better than other algorithms, and the tracking performed under conditions such as deformation, occlusion, out-of-plane rotation, motion blur, and scale transformation is more stable.

Abstract:Depth Image Based Rendering (DIBR) is the key technology of virtual view synthesis, but the generated virtual views have large areas of continuous holes. The holes repaired by traditional image repair algorithms lack semantic sense, and the existing partial convolutional neural network distort the edge of the holes area, so this study proposes a partial convolution neural network inpainting algorithm based on edge information. First, the disparity shift is used to generate the virtual view, then the virtual view is assigned and expanded to eliminate the effects of cracks and artifacts on the later holes inpainting, and the edge detector is designed to the partial convolutional neural network which make the network focus on the edge part of the pictures. Finally we use the well learned network model to inpaint large area holes in the virtual view. The experimental results show that the method presented in this paper can repair large areas of continuous holes. The repaired holes area not only has a sense of semantics, but also has finer edge details.

Abstract:Compared with two-dimensional faces, three-dimensional faces contain more feature information and can be applied to more practical application scenarios, such as face recognition, film and television entertainment, medical beauty, etc. Therefore, 3D face reconstruction technology has become a research focus in the field of computer vision. Due to real 3D face data is difficult to obtain, many deep learning-based reconstruction algorithms first use traditional reconstruction methods to construct 3D labels for a large number of 2D face images. These training data may not be accurate which will affect the reconstructive accuracy of these algorithm. To this end, this study proposes a weakly supervised learning model based on a multi-level loss function, which combines traditional 3D morphable model 3DMM and deep learning methods to directly learn 3D face feature from a large number of 2D face images without 3D labels to implement a3D face reconstruction algorithm based on a single 2D face image. In addition, in order to solve the problem that occlusion or large poses in 2D face images often affect the reconstruction of face texture, this paper uses a face parse segmentation algorithm based on the CelebAMask-HQ dataset to preprocess the images to remove the occlusion areas. The experimental results show that the quality and accuracy of 3D face reconstruction based on the proposed method have been improved greatly.

Abstract:In this study, we take a few kinds of leaf diseases of apple tree, such as Alternaria mali Roberts, as research objects, and a pathological identification method for apple tree leaf diseases based on depth-separable convolution is designed. The probability data enhancement is used to amplify the original dataset, a deep separable convolutional neural network is explored by using transductive transfer learning, and is applied to crop pathological recognition. An in-depth learning model for restricted equipment is designed to recognize and classify the apple tree leaf diseases, and the model is compressed, transformed, and transplanted to an embedded system for verification. The experimental results show that the proposed method has a good recognition effect, the recognition rate is up to 85.96% in the restricted equipment.

Abstract:Suspending unnecessary system or application processes in the background of the mobile phone while the user is sleeping can effectively reduce energy consumption, so it is of great significance to accurately determine whether the user is sleeping without compromising the user experience. Based on this problem, the coverage rate and wake rate are designed as new metrics. A sleep prediction model based on LSTM neural network is proposed, the LSTM neural network can handle time-series feature data and the evolution algorithm can optimize non derivable optimization targets. The parameters of the LSTM neural network are used as the optimization parameters of the differential evolution algorithm, and the comprehensive target of coverage and wake-up rates are used as the selection function. The selection function is re-evaluated in each iteration to use the mini-batch training. The experimental results show that compared with the traditional classification model, the prediction results obtained by training the LSTM neural network with evolutionary algorithm can achieve better coverage at low wake-up rate, with an average improvement of about 5%.

Abstract:Human body posture recognition has far-reaching significance in the fields of human-computer interaction, games, and medical health. It is a difficult research point in this field to perform high-precision and stable recognition of various human body posture based on portable sensors. This study collects high-frequency sensor data of eight postures, and the data set is sorted out by extracting the window time-domain features of the original data. According to the characteristics of the sensor data, the human posture is divided into four stages, and the Gaussian Mixture Model (GMM) is used to fit the observation sequence of the human posture, combined with the Hidden Markov Model (HMM), then, use GMM-HMM algorithm for gesture recognition. This study compares the effects of the First Order Hidden Markov Model (1OHMM) and the Second Order Hidden Markov Model (2OHMM) under different window values. When the window value is 8, the performance of 2OHMM is optimal, and the overall recall rate reaches 95.30%, the average accuracy rate reaches 95.23%. Compared with other studies, the algorithm in this work can recognize more types of gestures, has better recognition performance, and takes less time.

Abstract:In order to excavate security threats in power grid by making full use of heterogeneous data sources in power information system, this study proposes a multi-source log comprehensive feature extraction method based on Restricted Boltzmann Machine (RBM). Firstly, the RBM neural network is used to normalize coding all kinds of log information. Then, the contrast divergence fast learning method is used to optimize the network weight, and the stochastic gradient rise method is used to maximize the logarithmic likelihood function for the training and learning of the RBM model. The data dimension reduction is realized by processing the normalized coded log information. At the same time, the comprehensive features are obtained, which can effectively solve the problems caused by the heterogeneity of log data. The big data threat early warning monitoring experimental environment was set up in the power information system, and the comprehensive feature extraction and algorithm verification of the security log were carried out. Experimental results show that the proposed method can be applied to all kinds of security analysis, such as clustering analysis, anomaly detection, etc., and it has high accuracy in extracting log features in power information system, which improves the speed and accuracy of network security situation prediction.

Abstract:In recent years, in the job shop dynamic scheduling system based on Q-learning algorithm, the state action and reward value are set subjectively by human beings, which leads to the unsatisfactory learning effect. Compared with the known optimal solution, the result deviation is larger. For this reason, based on the characteristics of job shop scheduling problem, the elements of Q-learning algorithm are redesigned, and simulation test is carried out with standard case library. The results are compared with the known optimal solution, the hybrid Gray Wolf algorithm, the discrete cuckoo algorithm and the quantum whale swarm algorithm in terms of approximation and minimum. The experimental results show that compared with the Q-learning algorithm for solving the job shop scheduling problem in China, this method is significantly improved in the approximate degree of the optimal solution, and compared with the group intelligence algorithm, in most cases, the optimization ability is significantly improved.

Abstract:Vision tracking is one of the core functions of smart robots, and widely used in automatic driving, intelligent pension and other fields. The low-cost Raspberry Pi is employed as the slave computer robot platform. The object detection and visual tracking of human hands is implemented through running the pre-trained deep learning SSD model on host computer. The SSD model is trained based on Google’s TensorFlow deep learning framework and US Indiana University’s EgoHands dataset. Both of the robot and host computer’s software is written by Python in Linux systems. Video stream and tracking control commands are exchanged between robot and host via WiFi. The practical tests show that the vision tracking function of the developed smart robot has good stability and performance.

Abstract:Stance detection tells whether the expressions of opinion holders are in favor of or against the given objects. To accurately detect stance, the information of the expressed contents must be extracted, alongside a stance match for specific objects. In this study, the Transformer structure and gating attention is applied to specific object stance detection. By effectively utilizing the tag phrase information of the posts and the matching information between posts and objects, which are a result of gating attention mechanism, it delivers a better judgment over the post’s authentic stance regarding the object. Moreover, this approach takes emotional classification as an auxiliary task to fully include emotional information into stance detection for better performance. Experimental results show that the model is superior to the latest deep learning method on the SemEval-2016 dataset.

Abstract:In order to solve the problem that there are multi-source and uneven quality data in process-oriented workshop, which affect production control, the data classification and processing technology for multi-source data is proposed. Data processing is performed using a combination of data classification and sliding windows. First, a data model is established based on the characteristics of the production data, and the data is classified. It is mainly divided into three types: state data, switch data and logical data. Different types of data use different processing algorithms. As the same time, sliding windows are used to solve the difficult that different tasks have different requirements for data real-time and integrity. Finally, the data classification processing model is put into use in the actual production environment, which verifies the accuracy and real-time performance of the production data processing. The results show that by applying the processed data to production control, the control error rate is reduced to less than 1%.

Abstract:Aiming at the actual needs of network security problems faced by the power monitoring systems, research on ontology construction technology was carried out. Based on existing domain ontology automation construction technologies, a power grid monitoring domain ontology named SafeAgent was proposed from unstructured text data. Using methods such as machine learning, natural language processing and association rules to realize extraction of ontology concepts. Futhermore, this study accomplished the mining of relationships between concepts, and perfecting of the domain ontology automation construction scheme. The experimental verification shows that the method proposed in this study can complete the automatic construction of domain ontology with higher accuracy while overcoming reliance on human and expert knowledge.

Abstract:For transnational, trans-regional supply chain and the group enterprise, on the basis of Role-Based Access Control (RBAC), adding the organization and role group, three-dimensional permissions control module is proposed, which follows RBAC96 and ARBAC97 mode and controls users minimum permission set of access to web resources and business implementation. Permissions system module is developed using Java language based on MS SQL Server 2012 database and can be reusable and extended. Therefor, it can response business change rapidly.

Abstract:Aiming at the fact that there is no data processing software for multiple drilling imagers in coal mine, the overall architecture design of the software is proposed, based on the analysis of output data source of drilling imager, a trajectory display software for drilling imager compatible with multiple instruments is designed based on Visual Studio+SQLite. The software realizes multi-source data management of borehole attitude data and graphic data, 2d and 3d drawing generation and graphic operation of borehole trajectory, generation and operation of borehole histogram, etc. The application results show that this software provides friendly human-machine interface and has high application value.

Abstract:With the development of big data and artificial intelligence, it is possible to transform the manual processing of patents into automated processing. In this study, combined with the advantages of Convolutional Neural Network (CNN) to extract local features and Two-way Long and Short Term Memory neural network (BiLSTM) to serialize and extract global features, the attention mechanism is introduced in the hidden layer of BiLSTM, and a BiLSTM_ATT_CNN combination model for Chinese patent text data is proposed. The BiLSTM_ATT_CNN combined model improves the accuracy of Chinese patent text classification by designing multiple comparison experiments.

Abstract:Improving the interactivity and stereoscopic display effect of the virtual disassembly system has positive significance for the practical application of virtual simulation projects. A core drill for geological engineering is used as an object, and a set of core drill virtual disassembly system is developed using Unity3D and zSpace platforms. The system disassembly module mainly uses the six-degree-of-freedom stylus to realize interactive operations such as grasping, scaling, exploding, and resetting on the core drilling machine and other important components such as the dividing box, rotary. Through the zView plug-in package, the augmented reality display effect of the disassembly process is realized. The knowledge structure of the virtual training system is improved by adding the matching theoretical knowledge modules such as rig introduction, process flow and safety specification. The practice shows that the system has the characteristics of strong interaction and good display effect, and has achieved good results in practice teaching.

Abstract:Plastic mobile phone shell factory quality inspection. It’s time-consuming and uses traditional manual methods to identify appearance defects. The classifier uses the convolution neural network model of deep learning to train. This classifier can automatically detect the scratch defects on the appearance of mobile phone shell, which can greatly improve the work efficiency. In the experiment, a basic convolution neural network model is established firstly, the recognition baseline is obtained by training model, and then the detection accuracy is gradually improved by design modification. In order to solve the model over fitting and improve the detection accuracy in small data set training, the dropout, data augmentation method and batch normalization are used to reduce the amount of parameters, and the transfer learning method is applied. Experimental results show that the classifier model can effectively improve the accuracy and achieve a very good scratch defect recognition effect on small data sets.