Abstract:As a method of automatically detecting application vulnerabilities, fuzzing often serves for various software and computer systems. The quality of the seed file is very important to the fuzzing test. Therefore, this study proposes a method for generating fuzzing seed input based on machine learning. It relies on sample input and machine learning-based technology to learn the rules and grammar of sample input, which are then used to generate new seed input. We also propose a sampling method, considerably improving the coverage of the new seed input.

Abstract:An earth system model is the important software for researching climate changes and modeling the earth system. Chinese Academy of Sciences-Earth System Model (CAS-ESM) is the high-performance computing software for earth system simulation developed by Institute of Atmospheric Physics (IAP), CAS. Currently, IAP has released CAS-ESM 2.0, while simulation performance is always one of the critical factors restricting its development. To evaluate and analyze the performance of CAS-ESM 2.0, we install CAS-ESM 2.0 to two high-performance computing platforms, “Yuan” and “Earth System Numerical Simulation Device”, for coupled numerical simulation. The experiment results show that CAS-ESM 2.0 has different performances affected by these two platforms; the atmosphere model has the highest proportion of running time, exceeding the sum of other component models; some component models perform poorly in scalability. Further analysis reveals that the atmosphere model is mainly restrained by ineffective communication. Future research and development of CAS-ESM 2.0 will focus on cross-platform optimization, atmosphere model upgrading, parallel algorithm improvement, and scalability of component models.

Abstract:Image enhancement algorithm mainly process the captured images to enhance the overall effect or local details, increasing the overall and partial contrast while suppressing unwanted details. As a result, the quality of the images is improved, conforming to the visual perception of the human eye. Firstly, according to the basic principles of image enhancement algorithms, this study analyzes those based on histogram equalization, wavelet transform, partial differential equations, fractional-order differential equations, the Retinex theory and deep learning, and their improved algorithms. Then, the qualitative and quantitative comparisons between image enhancement algorithms are carried out with regard to visual effect, contrast, and information entropy to indentify the advantages and disadvantages of them. Finally, the future development trend of image enhancement algorithms is briefly predicted.

Abstract:As the China’s communication market has been saturated over time, the competition among telecom operators is becoming increasingly fierce. Churn prediction of customers has turned into one of the most concerns for telecom operators. This study proposes a method based on multi-model fusion to create a churn prediction model of customers. First, through bootstrap sampling and positive-negative sample balancing, multiple training datasets are obtained from the original training data. Then, base models are trained by these datasets with ensemble learning and deep learning algorithms. Finally, the base models are merged into a high-level model. The experimental results prove that the fusion model performs better than all base models in the test datasets, with a practical value for production.

Abstract:Full-color 3D printing can quickly reproduce color prototypes, which has begun to be widely used in many fields such as medical treatment, industry, and food. These fields have increasingly higher requirements for the color reproduction accuracy of 3D printed products, and how to improve the color reproduction quality of full-color 3D printed products has become one of the research hotspots. This paper comprehensively discusses the color reproduction methods of full-color 3D printing. After the classification of full-color 3D printing methods, it focuses on the analysis of the core technologies of color conversion and printout models for color reproduction of full-color 3D printing and discusses the advantages and disadvantages of those methods. On this basis, it predicts the trend of full-color 3D printing.

Abstract:In an open academic environment, academic exchanges and scientific research cooperation are pivotal to academic innovation and development, and helping researchers find suitable academic groups contributes greatly to their inspiration. Most of the existing approaches to community detection of researchers focus on the correlation between research results and ignore that between their own academic activities. As such, this study uses Open Research and Contributor ID (ORCID) data to build a network of academic information by analyzing researchers’ academic activities. The cross-layer edge clustering coefficient is improved on the basis of similarity between nodes at all levels, and then a model detecting the researcher community based on the weighted cross-layer edge clustering coefficient is proposed. The model extracts the direct correlations between researchers by constructing multiple meta-paths and stratifies the network according to different attribute relationships. Inter-node similarity is calculated with weighted cross-layer edge clustering coefficients. Then the network is transformed into a homogeneous network which is combined with the Louvain algorithm for community detection. Experiments are carried out in both artificial and real networks, and the results are evaluated according to the actual situation of the community, improving the division while avoiding the uncertainty of parameters.

Abstract:Amid the continuous innovation in medical technology, medical staff’s demand for relevant information systems is increasing. As required by the clinical practices of doctors and nurses, an information management system suitable for the catheterization laboratory is developed and connected to the original systems of HIS and PACS in the hospital. This system builds on the Client/Server (C/S) model, which is equipped with the C# programming language and the SQL server database. In light of clinical needs, it serves the information management system for the Digital Subtraction Angiography (DSA) catheterization laboratory, with the functions including registration before operation; interventional therapy, medication and consumables recording, nursing evaluation, coronary angiography, and image acquisition during operation; report template setting, operation logging, operation recording and complex query after operation.

Abstract:In response to the lack of domestic open-source role-customization system plug-ins in the field of computer game development, a Unity3D-based role-customization system plug-in, Ark Avatar Customization (AAC), is designed and implemented. AAC supports to change game characters’ dress-up and adjust their bone parameters, appearances and colors in a visual way, enabling role-customization required by developers and players. This study takes Unity3D as a platform for software development to fulfill customization functions and optimize system performance with the improved texture merge algorithm, editor extension, and prefab optimization. Performance comparison reveals the time and space cost of AAC are lower than that of existing techniques, proving the practicability of the designed role-customization system plug-in.

Abstract:With regard to the limitations of traditional chemistry teaching and the deficiencies of existing virtual chemistry classrooms, a virtual chemistry intelligent classroom system based on Unity3D has been designed and implemented. A virtual scene is constructed and imported into the Unity3D engine, and virtual simulation experiments are performed through the C# script and a particle system. For authority allocation in multi-user virtual scenes, the intelligent auxiliary module with authority strategy is designed to assign different ranges of authority to students and provide prompt functions such as text and video. Besides, the system functions are split through the Spring Cloud-based framework, and the system is optimized by Redis. This system is suitable for chemistry teaching in the middle school. It implements highly-simulated virtual experiments with authority control to solve authority problems in virtual scenarios and improve learning efficiency, with high availability, etc.

Abstract:As an important part of financial informatization, the database faces the challenges of continuous business growth and high availability and scalability, while the traditional single-point architecture of databases, represented by MySQL, Oracle, etc., fails to meet the current requirements of financial services in terms of availability, scalability and storage capacity. Distributed databases are designed to address the challenges faced by single-site databases and provide the more flexible architecture, ensuring stable system operation. To this end, this study, subject to actual financial service requirements, researches and implements a distributed database equipped with a distributed SQL engine, which is capable of distributed transaction support and hybrid transactional/analytical processing. The system is designed with full component redundancy. In addition, high availability of the storage layer and strong consistency of the data are ensured by the Raft-like enhanced consistency algorithm, while high availability of the scheduling layer is guaranteed by the Zookeeper-based cluster scheduling scheme.

Abstract:A remote-control mobile video surveillance system is designed to expand the video monitoring range and enhance flexibility of a single camera. The system is composed of four modules. The smart car based on the Arduino system is equipped with a camera to receive user instructions for collecting videos. The embedded Linux system makes real-time acquisition of video data feasible through the V4L2 interface. Meanwhile, it sends the data to the forwarding server through the network and forwards the control commands from users to the smart car. The server ransmits the video to the client while the user control instructions to the Linux system. Additionally, Android-based mobile terminal presents monitoring videos and provides a user control interface. Compared with the existing system, the new system enables monitoring without blind spots by a single camera.

Abstract:The safety of automobile tires is crucial to passengers’ travel security. Abnormal tire wear is easy to cause a blowout, and serious wear threatens the life of passengers. Automated tire wear tests are necessary, since tire treads are mainly worn when cars run on the road. This research makes the automatic collection, transmission, and processing of tire images feasible, with the help of a Visual Studio 2017 development platform, C++ programming, and the OpenCV API interface of a computer vision library combined with the processing order of independent programming. The system can accurately extract tread images and determine the degree of tread wear through the characteristic values of gray level co-occurrence matrices of tread images, enabling the automatic detection of tread wear for automobile tires.

Abstract:An indoor positioning system based on inertial navigation is designed to track indoor pedestrian movements. In this system, the accumulated error in inertial navigation is reduced by the improved zero velocity update, increasing the positioning accuracy of the positioning system to less than 1% mileage. The real-time positioning information obtained by the system is uploaded to the cloud through 4G communication, and the tracks of pedestrian movements are simultaneously drawn and displayed at the user interface. The test is performed in a square-shaped building with stairwells on the opposite sides for fire fighting. The testers equipped with the system enter this building from the first floor, go through a long corridor and then arrive at the stairs in the north. Then they climb the stairs to the third floor and come to the smoke-proof stairwell in the south after a long corridor. From this stairwell, they go down stairs to the first floor and return to the starting position. At this point, the whole process is completed in a closed loop, involving the movements of testers such as walking, trotting, and going up and down stairs. Multiple tests reveal the real-time positioning error is less than 1%, proving the superior real-time autonomous navigation of the indoor positioning system without any auxiliary system.

Abstract:An intelligent operation and maintenance mode of the distribution network based on the Internet platform is proposed to obtain all the data of the operation status of the distribution network as well as the possible anomalies and faults of the power equipment, thus improving economic benefits. An intelligent operation and maintenance platform of distribution network is built after the data from other professional systems is integrated into the Internet platform. The normalized spectral clustering algorithm is used to analyze the historical normal and abnormal data of multi-dimensional state variables, obtain the shape coefficient and contour coefficient of historical data curves, and extract the fault characteristics of multi-dimensional state variables. Besides, the submodules are designed on the basis of the submodule of knowledge discovery and a decision-maker to analyze the health of power equipment in the distribution network. According to the association rule mining, the importance index of power equipment is determined to assess the risk of decision-making for operation and maintenance, managing the decision-making for intelligent operation and maintenance of the distribution network. The experimental results show the model collects the real-time data to timely identify the power equipment with possible anomalies and faults, improving economic benefits of the distribution network.

Abstract:The traditional JDBC framework is featured by low code reusability, high coupling, and difficult transplantation, with connection object failures caused by frequent network faults. In this study, we propose a novel JDBC framework combined with several design patterns and the database reconnection mechanism. Through the DAO pattern, the proposed method provides the persistence logic interface, the decouple business logic and the persistence logic to the business logic layer. The concrete implementation of DAO is encapsulated through templates, strategies and factory patterns, improving uniformity of persistence codes while reducing code redundancy. This novel framework is applied to the performance evaluation system in a university. The results demonstrate that the improved JDBC framework decouples the data persistence layer from the business logic layer and improves the reuse rate and development efficiency of the codes in the data persistence layer, enhancing the robustness of the system.

Abstract:Amid the progress in science and technology, the application scenarios of intelligent devices in a power system have been increasing, especially in power automation as well as data acquisition and communication. Many platforms rely on an embedded Linux operating system. An Energy Remote Terminal Unit (ERTU) is taken as an example for analysis. Communication parameters should be preset into the device in the form of configuration files, since ERTU needs to communicate with electric energy meters and supports data calling of different master station systems during operation. If the device fails, the configuration files may require reconfiguration in the subsequent recovery process, increasing the workload and affecting the consistency of parameters. If the automatic backup and restore of configuration files are realized simply and quickly with the script program and external memory in a Linux system, then the workload of field engineering operation and maintenance can be reduced, improving the management level of the whole engineering application.

Abstract:To deal with the insufficient competence for real-time detection and generalization of the existing methods for fatigue driving detection, this study proposes a detection method of eye closure features, which integrates the Convolutional Neural Network (CNN) and Support Vector Machine (SVM). The CNN is employed to extract facial feature points and locate the eye Region Of Interest (ROI). Then the Histogram of Oriented Gradient (HOG) of the ROI serves as the feature classified by SVM to determine whether there exists the eye closure feature of fatigue driving in the original image. There into, graying and histogram equalization contribute to weakening the impact of illumination variation. The proposed method is implemented on both the PC platform and the ARM embedded platform, which is verified with regard to examinees subject to different levels of illumination. Experimental results prove that the method reaches an accuracy of above 94% for detecting eye closure features, with strong generalization and satisfied real-time reaction.

Abstract:In this study, regarding a sequence submarine image for underwater vehicle navigation, a ribbon matching method based on image information and prior knowledge is proposed to increase the number and accuracy of matching points of submarine images. First of all, the deep submarine image is pre-processed with proper light compensation and linear enhancement. Then with navigation data as a guide, adjacent image ribbon areas are calculated on the basis of the physical position offset of images. The global calculated feature points are replaced by the calculated features in the ribbon area to increase the number of image registration points and reduce the mismatching rate. In experiments on deep sea images with repeated and few textures, compared with registration rely solely on image information, this method improves the number and accuracy of corresponding points in image registration.

Abstract:Designing a group secret key scheme to secure communication between nodes represents a huge challenge to wireless networks. To address this issue, we propose a group key extraction scheme based on three nodes. In this scheme, we first select a trusted authorization system to group the user nodes in the network. Then we extract the short key with the wireless channel characteristics of the physical layer and exchange the information signed by Schnorr. Finally, we make each user node authenticate its neighboring nodes. If the authentication is passed, a group key is established for the nodes in the network. The simulation results reveal a positive correlation between the group key rate and the signal-to-noise ratio, and the group key rate does not decrease with the increase in the number of nodes. It proves this scheme is applicable to wireless networks.

Abstract:Traditional topic models rely largely on word co-occurrence patterns to generate text topics. The data sparseness of short texts due to insufficient context has restrained traditional topic models from achieving good results with regard to short texts. On this basis, this study proposes a short text topic model based on semantic enhancement. The algorithm integrates the Dirichlet Multinomial Mixture (DMM) model with a word embedding model. It obtains the vector representation of words by training global word embedding and local word embedding and calculates the semantic correlation between word vectors with cosine similarity. Besides, it enhances the semantic meaning of words by calculating the weight of topic-related words. Experiments demonstrate the proposed model is more accurate in consistence of topic representation and improves the classification accuracy of the model in regard to short texts.

Abstract:At present, most benign applications in the Android market adopt a shelling method to protect themselves from being decompiled so that the detection of malicious applications can only rely on the permissions from AndroidMnifest.xml. However, the machine-learning-based classification algorithm based on permission features has a poor detection effect because of a small difference between malicious applications and benign applications. If a more fine-grained Application Program Interface (API) is taken as a feature, a serious imbalance in the number of positive and negative samples will be caused due to application shelling. In response to the above problems, with a large number of malicious applications as training samples and some benign applications as the point of novelty, we use the one-class SVM algorithm to establish a detection model for malicious applications. Compared with two-class supervised learning, this method can effectively distinguish malicious applications from benign applications, which has practical significance.

Abstract:Aiming at the pickup of mobile phones in narrow locations such as sewers and jointing, this study proposes a method of servo mechanical arms for grabbing based on machine vision. Firstly, the camera on the eye-in-hand mechanical arm is calibrated, followed by image preprocessing and target detection. In the pose detection, we propose an algorithm to solve the pose of the mobile phones based on the two-dimensional coordinate system. It turns out that the best pose angle is only related to the difference in the pixel coordinates of the clamping point, and the size of the pose angle determines the rotation angle of the gripper. Then, the pose detection is simulated by Matlab, including target detection with SURF invariant feature points and pose calculation. Finally, the right arm of a two-armed Rethink robot grasps a mobile phone for verification. The results show that within the allowable error range, the proposed algorithm accurately guides the servo mechanical arms to grasp mobile phones.

Abstract:In the inspection of a wind turbine in a drone, abnormal illumination may seriously affect the quality of the captured images. As a result, the blades of the wind turbine have abnormal brightness in the images, and small defects such as cracks on the blades cannot be found effectively, which affects the stable operation of the wind turbine. For this reason, this study proposes an illumination analysis method for the inspection of wind turbines. Before the inspection, the illumination situation is prejudged according to the planned trajectory and solar orientation. During the inspection, the illumination of the key parts is analyzed according to the segmentation results of the blades and tower barrel. In this process, we introduce a method of illumination analysis about the whole image based on weighted mean. In conclusion, the proposed method can predict the abnormal illumination in the whole inspection process and provide a basis for efficient and accurate inspections.

Abstract:Inertial Measurement Unit (IMU) is widely used in the acquisition and control of human motion information due to its small size, low costs, high accuracy, and strong timeliness. However, it still has obvious limitations in the time-series feature extraction and the data about gait environment during gait recognition. Aiming at the complexity and poor applicability of lower-limb gait recognition based on feature extraction, this study proposes a new method of human gait recognition based on Tsfresh-RF feature extraction. Firstly, an algorithm of human gait recognition based on Tsfresh time-series feature extraction and Random Forest (RF) is constructed by a human gait data set acquired by IMU. Secondly, experiments including nine gaits are carried out by this algorithm on different sensor positions, such as climbing, walking, and turning. Finally, the average classification accuracy of the proposed method reaches 91.0%, which is significantly higher than that of traditional Support Vector Machine (SVM) and Naive Bayes (NB) methods. In addition, the proposed algorithm is robust, which will provide a favorable basis for subsequent control of lower-limb exoskeleton robots.

Abstract:The existing Point Of Interest (POI) recommendation method operates based on the POI access frequency of an individual user and the access habits of his/her partakers in the location-based social network, with the geographical location of the POI as one of the recommendation conditions. However, most of the POI recommendations only take the geographical location of POI as the preferential reference, rather than the access cost of the users. Therefore, some of the POI candidates generated based on similar methods may meet the user preference but have poor accessibility. To solve the above problems, this study proposes a POI recommendation method based on LSTM and distance optimization. This method supplements the interaction matrix between a user and POI according to the user’s social network and then decomposes the matrix into the hidden vectors of the POI. Finally, according to the user’s POI access record, the temporal relationship between the hidden vectors is established, and the sequence is learned in a recursion-like model to infer the possible POI sequence accessed by the user in the future. In addition, experiments on the Gowalla and Yelp data sets demonstrate that in the limited data dimension, the proposed method has slightly higher recommendation accuracy than other representative models and can generate POI sequence easily accessed by current users.

Abstract:Aiming at the diverse and complex factors affecting expressway traffic flow, this study proposes a Bi-LSTM prediction model based on multiple factors. Firstly, the original data are cleaned up and analyzed with respect to their correlation to improve the research accuracy and reduce the data dimension. Secondly, a multi-factor time series matrix for traffic flow is constructed based on the time sliding window and the proposed model is trained and optimized with MAE and RMSE as the evaluation indicators. This model considers high-correlation influencing factors such as weather conditions, holidays, and toll, as well as changes in the preorder and postorder of traffic flow. With the expressway toll data in Shaanxi Province as the object, the results show that the proposed model is more applicable and accurate than GRU and LSTM in the short-term prediction of expressway traffic flow.

Abstract:This study establishes the relationship between image restoration and foggy days based on the atmospheric scattering model and restores the images according to the dark channel theory. For large-area dense fog and sky that do not conform to the prior assumption of dark primary color, the reason behind image distortion is analyzed, and the transmittance is corrected by introducing a tolerance parameter. As a result, the fog removal of the sky is not too difficult and the restoration of images with the sky is improved. This study sets an intensity threshold to avoid difficult defogging due to over high atmospheric light intensity. In addition, it adjusts the hue distribution of the restored image by an automatic color scale algorithm for a more natural image. Furthermore, the validation process of the image defogging algorithm is designed to obtain adjustable parameters. The subjective and objective performance verification and analysis of the three algorithms prove that the improved algorithm is better than the first two algorithms in defogging.

Abstract:This study proposes a neural network named Hierarchical Bilinear-YOLOv3 for human behavior detection due to a large disparity in the same behavior and high resemblance between different behaviors in human behavior detection, as well as problems such as visual angle, occlusion, and incapability of continuous real-time monitoring. YOLOv3 is first designed for prediction on three scales, and certain layers in its feature pyramid networks are used as inputs for Hierarchical Bilinear to capture local feature relationships between layers in the feature maps and predict the results on three scales. The integrated results of both YOLOv3 and Hierarchical Bilinear show that the improved network only adds a few parameters compared to the original one. It improves the detection accuracy of the original algorithm without lowering the detection efficiency and thus is superior to the current behavior detection algorithms.

Abstract:As a significant task in natural language processing, cross-lingual sentiment analysis is able to leverage the data and models available in rich-resource languages when solving any problem in scarce-resource settings, which has acquired widespread attention. Its core is to establish the connection between languages. In this respect, transfer learning performs better than traditional translation methods and can be enhanced by high-quality cross-lingual text vectors. Therefore, we propose an LAAE model in this study, which uses Long Short Term Memory (LSTM) and an Adversarial AutoEncoder (AAE) to generate contextual cross-lingual vectors and then applies the Bidirectional Gated Recurrent Unit (BiGRU) for subsequent sentiment classification. Specifically, the training in the source language is transferred to that in the target language for classification. The results prove that the proposed method is effective.

Abstract:Affected by the full traversal environment, the paths planned by existing methods are too long. For this reason, this study proposes a planning method for the full traversal path of omnidirectional mobile robots based on an improved ant colony algorithm, hoping to improve the path planning and obtain the optimal path. On the basis of the topological modeling diagram, a new environment model is established by angle conversion according to the position information of a mobile robot in the original coordinate system. Considering the problems in the ant colony algorithm, the decreasing coefficient is introduced into the heuristic function to update the local pheromone, and an iterative threshold is set to adjust the volatility coefficient of pheromone. Finally, the path planning process is designed to plan the full traversal path. The results show that the proposed method can shorten not only the full traversal path but also the planning time to obtain the optimal path, thereby improving the planning for the full traversal path of omnidirectional mobile robots.

Abstract:Aiming at the complex process and low recognition rate of traditional methods, this study proposes a voiceprint recognition method based on ResNet-LSTM. In this method, ResNet and LSTM are respectively used to extract the spatial and temporal features of voiceprints. Thus, the deep voiceprint features including both spatial and temporal features are obtained. The experimental results show that the equal error rate of the proposed method is 1.196%, which is 3.68% and 1.95% lower than that of the baseline methods d-vector and VGGNet, respectively, and the recognition accuracy reaches 98.8%.

Abstract:This study proposes a deep learning recommendation algorithm based on attention mechanism to solve the problem that the current recommendation algorithms based on comment texts have insufficient extraction of text features and implicit information. The comment text representations of users and items are constructed, and the context dependency of texts is extracted by bidirectional gated recurrent units for text feature representations. Moreover, the attention mechanism is introduced to obtain the interest preference of users and the attribute features of items more accurately. The two sets of hidden features of the generated user and item comment data are respectively input into the fully connected layer and then merge into the same vector space for rating prediction. As a result, the recommendation results are obtained. Experiments on two public data sets, Yelp and Amazon, show that the proposed algorithm has better recommendation performance than other algorithms.

Abstract:The scientific experiment of near-space exploration is to obtain multi-scale, multi-level, and multi-type detection data by aerostats, UAVs, and other platforms, which supports the construction of near-space cognitive systems. In view of the complexity of near-space exploration, the diversity of detection data, and the great application value, the real-time visualized display of near-space detection data is conducive to the safe and stable development of near-space exploration. During the processing of data streams, the traditional synchronous-blocking I/O mode is difficult in constructing the driver architecture of asynchronous events and thus the data processing efficiency is low. For this reason, this study introduces the real-time analysis and display method of non-blocking mode for Constant-Bit-Rate (CBR) data of near-space exploration based on NodeJS+ElectronJS+EchartsJS. This event driven method is of asynchronous programming and can achieve real-time decoding of CBR data in the scientific experiment of near-space exploration, as well as efficient and accurate visualized display of the downward transmission data.

Abstract:Click-Through Rate (CTR) prediction is a fundamental task in personalized advertising and recommendation systems. This study proposes a model named ADditional Variational AutoEncoder (ADVAE) based on an improved denoising autoencoderto improve CTR prediction and cold-start. It adds random Gaussian noise to the input data and generates new embedded features by the improved denoising autoencoder. Then, multi-level features interact to predict the users’ clicking. This method can learn the relationship between feature embedding and interactions in data sparse and cold-start situations. In addition, it has strong robustness since it focuses on the interaction in feature domains and dynamically repairs feature embedding of low-frequency data. Besides, this method can be dynamically applied to other deep learning models, with high flexibility. The results show that the proposed approach outperforms its counter parts in terms of CTR prediction and cold-start.

Abstract:The research on the water level–water area relationship model is time-consuming and inefficient resulting from the cumbersome data downloading and processing, massive use of professional software, and frequent manual intervention in the water extraction from long-term sequence remote sensing images. We preprocess the Landsat images from 2002 to 2016 on Google Earth Engine. Then, we identify the water of East Dongting Lake by NDWI and calculate its area. On this basis, with the daily water level data from Chenglingji Hydrological Station, we identify the optimal water level corresponding to the image, and draw the water level–water area curve by the weighted average water-level method, the threshold method, and the single-point water-level method. The results are as follows: (1) The water scope of East Dongting Lake varies greatly with season, with the largest area in between July and September and the smallest area in between January and March. (2) The accuracy of calculating the optimal water level by the threshold method (R² = 0.9628) is higher than that by the weighted average water-level method (R² = 0.8322) and that by the single-point water-level method (R² = 0.9457). (3) The fifth-order polynomial model has the highest accuracy (R² = 0.9628) in fitting the water level–water area curve.

Abstract:The roadside perception algorithm is integrated with the on-board perception algorithm to achieve over-the-horizon perception. The performance of the perception algorithm based on deep learning depends on the quality of the point cloud annotation of lidar which is harder than the annotation of 2D images because it takes longer time and calls for much manpower. In addition, existing perception algorithms based mainly on the on-board lidar. In this study, we proposes a perception algorithm based on the feature clustering of roadside lidar grids. This algorithm rasterizes the point cloud of roadside lidar and extract the features, then learn the primary perception information of the grids by creating a deep learning model for clustering on this basis. We also simulate the point cloud of roadside lidar via a simulation platform, and studies the application of the hybrid data set in training perception algorithm, which is fine-tuned by the pre-training model of simulation data. Experimental results show that the proposed perception algorithm is reliable with real-time service. Besides, simulating the point cloud of roadside lidar helps with the training of this algorithm and reduces its dependence on annotation, improving its performance.

Abstract:The entity relation extraction model based on neural networks has been proven effective, but a single neural network model is unstable because it can yield various results with different inputs. Therefore, this study proposes a method to integrate multiple single models into a comprehensive one using the idea of ensemble learning. Specifically, this method integrates Bi-directional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN) into a comprehensive model through MultiLayer Perceptron (MLP), which cannot only fully take advantage of the two single models, but also make use of the self-learning ability and automatic weight allocation of MLP. This study obtains F1 of 87.7% on the SemEval 2010 Task 8 dataset, which is better than other mainstream entity relation extraction models.

Abstract:In most recognition models of Chinese named entities, language preprocessing only focuses on the vector representation of single words and characters and ignores the semantic relationship between them, hence failing to tackle polysemy. The transformer feature extraction model improves the understanding of natural language due to parallel computing and long-distance modeling, but its fully connected structure makes the computational complexity the square of the input length, which leads to poor recognition of Chinese named entities. A recognition method for Chinese named entities based on the BERT-Star-Transformer-TextCNN-CRF (BSTTC) model is proposed to solve these problems. First, the BERT model pre-trained on a large-scale corpus is used to dynamically generate the word vector sequence according to its input context. Then, the star Transformer-TextCNN model is adopted to further extract sentence features. Finally, the prediction result is received by inputting the feature vector sequence into the CRF model. The experimental results on the Chinese corpus from MSRA show that the accuracy, recall, and F1 value of this model are all higher than those of existing models. Moreover, its training time is 65% shorter than that of the BSTTC model.

Abstract:As one of the six working languages of the United Nations and a major mother tongue second only to Chinese, Spanish has complex morphological changes and grammatical rules. These result in the inability of classic term extraction methods such as C-value and thus affect the effect of Spanish text analysis. This study proposes a Spanish term extraction method to automatically construct a complete lexicon for text modeling. Given a Spanish text or corpus, the method extracts terms in three steps: preprocessing the texts, extracting candidate terms, and calculating term-hood indexes of the candidate terms based on DC-value. The set of candidate terms obtained in the first two steps can be used directly as the lexicon for text mining. Meanwhile, the term-hood indexes obtained in the third step are essential for reducing the manual workload in determining whether the candidates are really terms. According to experiments, the proposed method has a high accuracy of 80% and a recall much higher than that of classic methods, providing the effective lexicon for Spanish text mining.

Abstract:In the application of safety monitoring system of elevators, infrared sensor technology or traditional face detection algorithms involving Haar-like and HOG features are often used for the recognition of elevator passengers with poor effect though. With the development of deep learning in recent years, the face detection algorithm based on convolutional neural networks is more accurate than traditional face detection algorithms and has been applied in many fields. Moreover, the face detection algorithm based on multi-task cascaded convolutional neural networks is adopted to recognize elevator passengers in the safety monitoring system owing to its small model and fast operation. With the inception module introduced, the depth and width of networks at all levels are raised by the parallel operation of convolutional cores of different sizes for better extraction of network features; models are trained faster and network classification is enhanced through batch normalization. The experimental results show that the accuracy of the improved algorithm is 2% higher than that of the original one and can thus realize the highly accurate recognition of elevator passengers.

Abstract:In order to explore the application and implementation of knowledge mapping technology in intelligent agricultural production and realize the accurate query and visualization of complex and diverse agricultural production data, this study took wheat varieties as an example and collected the information of 1852 wheat varieties, 735 micro encyclopedias, and 102 349 entries by a crawler. Through knowledge mapping technology, this study designed the entities of variety knowledge graphs and their relationships, with data cleaned, extracted, and fused. A total of 258 484 entities were recognized and 328 933 relationships built. On this basis, the approach to storing wheat variety knowledge was worked out, with structured data stored in a MySQL, unstructured data in the MongoDB. Neo4j was employed to optimize knowledge query. In this way, the query about relationships between wheat varieties and entity recognition was made possible with variety data expressed accurately and visualized, proving the feasibility of knowledge mapping in visualization of information such as variety. This research can provide technical reference and theoretical support for the application of knowledge mapping in agriculture.

Abstract:Considering the real-time capacity of lithium batteries at low temperatures is hard to be estimated, the instantaneous current and voltage largely influence the change in transient battery capacity at low temperatures . The deep feedforward BP network model with the Dense fully connected layer as the main body is studied, and the influence of different added layers on the predicted and actual values of the model is analyzed; the BP network with three hidden layers [11-9-12] is used for higher accuracy; Nadam optimization algorithm and the optimization model of Log-cosh loss function that accelerate convergence with the momentum method and the adaptive learning rate method based on SGD expansion are adopted. Overfitting are reduced by regularization for better network generalization. The model is trained and tested based on the data at 0 °C under the HPPC working condition. Consequently, the prediction error of soc at different voltages and currents is about 0.04.

Abstract:Existing piecewise linear representation of time series ignores the global characteristics of time series and easily falls into local optima. To solve this, the paper studies the trend in time series and finds its fluctuation. The trends is divided into an upper layer and a lower one with their trend holding points removed. The experimental results show that the segmentation method has low time complexity and is easy to implement, and the fitting error is smaller on the premise of keeping the trend characteristics of time series.

Abstract:In order to obtain an assessment method for image quality that is consistent with the human visual perception system, this study proposed a no-reference assessment method for blur image quality by simulating the human visual perception system. The proposed method evaluates images of different blurriness by comparing the similarity of their characteristics. First, the test image is blurred by Gaussian functions to different degrees. Second, their detailed information is obtained through the retinal model. Third, singular values are decomposed to measure the intrinsic structures of images. Then, the similarities in details and singular values among the test image and its blurred images are calculated as the characteristic vectors for image blurriness, which are input into a Support Vector Regression (SVR) model for training to generate the proposed assessment method for image quality. Experimental results on benchmark databases show that the proposed method is more consistent with the subjective visual perception of human visual system than the comparison methods.

Abstract:To ensure the normal operation of the automatic knotting machine of yarn in the automatic bobbin changing system, we need to detect the yarn sucked by the pipe. Yarn is detected by image processing instead of sensors because it is thin with diverse types and colors. However, traditional image processing methods are too complex and inaccurate to identify yarn with various types, sizes, and colors. This study proposes a network of multi-scale depth separable convolution blocks modified based on Inception-Resnet-A block of Inception v4 to detect yarn in pipes. The conventional 3×3 convolution layers in the Inception-ResNet-A block is replaced with the depth separable convolution layers of the 3×3 convolution kernel, and some of the 1×1 convolution layers are removed for less parameters of convolution blocks and simpler calculation. In addition, ResNet is employed for channel fusion to prevent feature loss. According to the experimental results, this network model is remarkable in generalization and recognition.