Abstract:In the field of MOFs research, searching for novel MOFs is still a complicated problem. After MOFs are processed by “material genetic encoding”, the Genetic Algorithm (GA) can be used to rapidly explore novel MOFs, but their performance depends on the setting of individual fitness functions, and the effective evaluation of the novel MOFs also contributes to the effectiveness of this method. As one of the representative methods of machine learning, the Artificial Neural Network (ANN) can uncover the non-linear constitutive relationships. In this paper, the neural network is introduced to predict the adsorption capacity for CH₄ gas by the novel MOFs generated by GA, thereby facilitating the search for novel MOFs by GA. The experimental results show that the neural network can thoroughly evaluate the novel MOFs materials, demonstrating the feasibility of combining the neural network and GA for the search and screening of the novel MOFs.

Abstract:In today’s general trend for ubiquitous computing and software definition, formal methods have gradually become an important way to guide the definition of software requirements, analyze software design schemes, and verify the correctness of software products, which penetrates the entire life cycle of software engineering. Event-B, as a “correct by construction” method, supports the application of formal methods in software engineering. This paper classifies and expounds on the existing formal methods in software engineering based on Event-B, which are mainly divided into Event-B control structure, object-oriented Event-B, reusable Event-B, as well as real-time Event-B models. It also summarizes the support from various Event-B models for the whole life cycle of software development and provides references for the formal methods in software engineering.

Abstract:Amid the vigorous development of the knowledge graph, relation extraction, as a key part of information extraction, has attracted increasing attention from researchers. In general, relation extraction can be divided into template-based extraction and machine learning-based extraction. Later, with the extensive application of the extraction methods based on deep learning, the performance of relation extraction has been greatly improved. In this study, the time sequence method is employed to summarize the extraction methods of binary relations in a restricted domain. This study first briefly introduces the concept, data set, and evaluation indicators of relation extraction. Then it systematically sorts out the related extraction methods and highlights the current research on the relation extraction methods based on deep learning. Finally, it analyzes the future research direction and application of relation extraction.

Abstract:Since the proposal of Fully Convolutional Network (FCN), applying deep learning to image semantic segmentation has attracted extensive attention from researchers in the field of computer vision and machine learning, becoming a research hotspot of artificial intelligence. The core idea of FCN is to build a fully convolutional network that accepts the input of arbitrary sizes and produces the output of the same sizes through efficient inference and learning. FCN provides a new idea for image semantic segmentation, but it also has many shortcomings, such as low feature resolution and the objects at multiple scales. As research progresses, the convolutional neural network has been gradually optimized and expanded in the field of image segmentation. In addition, the mainstream segmentation frameworks based on FCN have emerged one after another. Image semantic segmentation plays an increasingly important role in scene understanding, which is widely applied to the self-driving technique, the UAV field, detection and analysis of medical images, and other tasks. Therefore, image semantic segmentation is worth further study to better serve practical applications.

Abstract:In an era of big data, application needs to integrate various data management tools to fulfill its business requirements. However, these tools differ in their APIs and need adapters to communicate with each other. Therefore, how to integrate data management tools quickly becomes an important research problem in academia and industry. This study introduces a language translating based framework for heterogeneous data unified access, which is called the Bi-Adapter Framework for Data Unified Access (BAF4DUA). In BAF4DUA, two types of adapters are located at the data provider endpoint and data consumer endpoint to translate the query semantics and data model. The query language and data model are decoupled by the two adapters, thereby enabling many-to-many and plug-and-play data access modes between data providers and data consumers. In addition, the decoupled framework can also promote the flexibility and scalability of the application system.

Abstract:In order to predict customer churn with large sample data, a customer churn prediction model based on spectral regression was put forward from the perspective of feature expression, which took advantage of the spectral regression to reduce the dimension of feature. On the basis of the original customer features, a distinguishing feature space of low dimension is established by using the manifold dimension reduction based on spectral regression, and then we used the support vector machine to realize the binary classification of customer churn prediction. The model was evaluated on two different data sets of network customers and traditional telecom customers, and compared with different classifiers, different feature reduction or selection methods, the experiment results verify that the model is effective.

Abstract:Federated learning is an emerging privacy-preserving machine learning paradigm widely applied to the Industrial Internet of Things (IIoT), where multiple clients (e.g. IoT devices) train models locally to formulate a global model under the coordination of a central server. Blockchain has been recently leveraged in IIoT federated learning to maintain data integrity and provide incentives to attract sufficient client data and computation resources for training. However, there is a lack of systematic architecture design for blockchain-based federated learning systems to support methodical development in IIoT. Also, the current solutions do not consider the incentive mechanism design and blockchain scalability. Therefore, in this study, we present a platform architecture of blockchain-based federated learning systems in IIoT, where each client hosts a server for local model training and manages a full blockchain node. For verifiable integrity of client data in a scalable way, each client server periodically creates a Merkle tree in which each leaf node represents a client data record and stores the tree root on a blockchain. To encourage clients to participate in federated learning, an on-chain incentive mechanism is designed based on the size of client data used in local model training to accurately and timely calculate each client’s contribution. A prototype of the proposed architecture is implemented with our industry partner and evaluated in terms of feasibility, accuracy and performance. The results show that the approach ensures data integrity and has satisfactory prediction accuracy, and performance.

Abstract:For the unified intelligent management of university seats and independent setting of seat rules in different regions, a set of university seat management system based on film pressure sensors and C/S architecture is developed, which is in line with the Android MVP design mode. The requirements for university seat management are analyzed, and the system’s functional modules, data flow, and system architecture are designed and analyzed in detail according to the principles of user-friendliness and friendly interface. The system includes the C/S three-tier architecture and relies on the Android MVP model for layered development to achieve data mapping, access and persistence. Through the design and debugging of a user APP and a detection terminal, the reliable detection of seat usage and comprehensive information-based management of data are realized, and the seat management and use efficiency are improved, which is of positive significance to university seat management and information-based campus management.

Abstract:At present, most of the pointer recognition methods are based on the traditional image processing technology, and the extraction process is complicated with many steps. To effectively solve the problems of difficult pointer axis extraction and poor reading recognition accuracy of a pointer instrument, this study introduces a method of pointer instrument recognition based on deep learning. First, the Faster R-CNN algorithm is used to detect the instrument disk, and then the method based on deep learning is adopted to detect the pointer. According to the position information of the target frame, the pointer image is obtained by clipping. The final reading of the instrument is identified by binarization, thinning, Hough transform, and the least square fitting line. Compared with the traditional image processing directly on the image of the panel target frame or the original image, this method greatly reduces the interference in the process of locating the line where the pointer axis is located. The experimental results show that the average accuracy of pointer detection based on deep learning proposed in this study is up to 96.55%. It has high accuracy and stability for pointer detection of the pointer instrument under a complex background.

Abstract:An interactive virtual roaming system integrates many elements, such as hardware, interactive technology, and content, and its designers consider these elements to meet users’ requirements. Because the current virtual roaming system is subject to the limitations on the application scope and cost, its development environment, methods and technology are not conducive to promotion. First, this study analyzes and compares the language development tools and their functions in the applications of an interactive virtual roaming system. It then proposes a feasible design and development process for the interactive applications of the virtual roaming system with regard to an actual project. On this basis, application software and hardware platforms are designed for the interactive virtual roaming system (namely the combination of hardware and interactive technology), and the construction and optimization process of the interactive digital campus virtual roaming system is illustrated with specific cases . The platform has many advantages, such as favorable portability, flexible combination, high standardization, and low cost, and the design carefully considers users’ personal preferences and designers’ self-evaluation.

Abstract:The domestic Linux operating system running third-party GUI application software needs to solve the software dependency library problem. The official dependent software cannot meet the configuration of the dependent library environment, resulting in a large amount of third-party GUI application software that cannot be installed and used in the domestic operating system. A solution is proposed to package third-party GUI application software and its operating environment into independent application software with container technology, so that the third-party GUI application software can run on a domestic operating system. With the open source distributed rendering system, Equalizer, as the target object, the docker container technology is used to package the dependent libraries required for its compilation environment and running environment into a mirror. In the X11 service in the container, the Equalizer in the container parses the X11 file in the operating system and displays the graphical interface on the host screen. This study uses the existing docker technology to create an independent image and configures the container to share the graphical interface service and graphics card driver of Linux system with the host system and finally realizes the normal use of the Equalizer program in the domestic operating system. Experimental results show that the scheme is feasible and can be extended to other GUI application software.

Abstract:An online mobile water quality monitoring system, which is based on wireless sensor networks, is developed to solve the existing problems in the static off-line monitoring methods, such as large sampling errors, low monitoring frequency, dispersed monitoring data and lack of real-time feedback on continuous dynamic changes of water quality. The system is composed of an underwater monitor, buoy nodes, and visual software. The exterior design of the underwater monitor mimics tuna appearance, and the internal is equipped with water quality sensors (regarding temperature, pH, turbidity, conductivity, etc.) to acquire the parameters of water quality. Buoy nodes, depending on solar power, are responsible for receiving, processing, relaying, and forwarding the data from multiple underwater monitors. The underwater monitors communicate with each other and the buoy nodes through a 170 MHz wireless radio frequency module or underwater sound, while the buoy nodes communicate with shore stations by ZigBee. In addition, visual software includes PC and mobile terminals. This system overcomes the limitations of traditional monitoring based on manual sampling, achieving online monitoring of water quality at any point.

Abstract:With the increase in the number of street lamps in Chinese cities, the proportion of the power cost of street lamps in the government’s financial expenditure is also rising rapidly, while the traditional street lamp system cannot effectively reduce energy consumption due to poor stability and flexibility. In this study, an intelligent lighting management platform based on micro service architecture is designed and implemented by separating specific functions to form multiple micro-services. At present, the system has been developed and tested, which has been put into use. The practical applications show that the platform can reduce energy consumption substantially, which should be widely promoted.

Abstract:At present, doctors mainly use multiple two-dimensional images to diagnose the bone deformities of lower limbs. It is difficult to comprehensively observe the deformed parts from multiple angles, which is detrimental to detailed planning before surgery. Therefore, the Web development technologies, such as HTML5, JS, and CSS, are employed to independently develop an online auxiliary diagnosis simulation system for lower limb bone deformities in this study. Various required functions are implemented, such as online 3D modeling of lower limb bones and artificial joints based on DICOM images, 3D model browsing in an all-round way, dynamical display and comparison between DICOM images, auxiliary diagnosis tools (including measurement tools, gray scale transformation, detail viewing, and precise positioning), and patient information management. The system integrates open source architectures such as VTK, Cornerstone, and Three. It has the advantages of strong practicability, online sharing, and easy scalability, providing an efficient and powerful online auxiliary platform for the diagnosis of lower limb bone deformities and surgical planning.

Abstract:Building an online service platform complies with the requirement of “The Education Informatization 2.0 Action Plan” for the smart campus construction in colleges and universities. In this study, combined with the actual cases of online service platforms, the integrated modeling, elements, and steps that meet the needs of multi-campus management are proposed from the view of integrated management of multi-campus services. The component system based on the workflow-as-a-service is developed and a unified service user interface is presented. At last, the effectiveness of the modeling method and design is illustrated through the practical example which is applied in a practical scenario.

Abstract:To meet the needs of meteorological archive management and service, Anhui meteorological archives business system is developed and put into operation. This paper describes the basic design idea, function structure, basic platform architecture, and information flow of the system under the background of smart archives construction. The business system consists of five sub-systems for data collection, file arrangement, file storage, file utilization and file identification, respectively. The main key technologies include meteorological archive standardization system, meteorological archive metadata, knowledge graphs, and Internet of Things (IoT) technology. The design and implementation of the system also provides a reference for the development of industry archives business systems.

Abstract:Wax deposition in oil wells seriously affects the normal production of oil wells during the development and production of oilfields. This phenomenon will block oil flow channels and reduce oil production during the production of oil wells. Wax deposition prediction in oil wells and advance maintenance of oil well equipment are pivotal to higher production capacity, lower maintenance cost and more intelligent management. To solve the problem of serious imbalance between the normal data and wax deposit data of oil wells, this study introduces two processing methods of non-equilibrium samples, ADASYN and ENN, which deal with the non-paraffin and paraffin samples separately. Finally, the random forest algorithm is used to integrate the training data set, and the intelligent AERF model is constructed to predict the wax deposition in oil wells. The experimental results show that the AERF model proposed in this study has a better prediction effect in the wax deposition data set of oil wells, greatly improving the prediction accuracy.

Abstract:Traditional time series forecasting models perform poorly in forcasting the time series data with long-term and short-term time relevance, nonlinearity and non-stationarity. To improve the accuracy and efficiency of the time series forecasting model, this study proposed a time series forecasting model (Attention Temporal Convolutional Neural Network, A-TCNN) combining temporal convolution, residual structure, and the attention mechanism. The model considers the efficiency of temporal convolution to extract temporal features, the superiority of residual structure to accelerate model convergence, and the strengthening effect of the attention mechanism on the parameters. Firstly, the long-term and short-term features are extracted from the data through multiple residual temporal convolutional layers; secondly, the weight of the parameters that have a greater impact on the output is strengthened through the attention layers; finally, the output result is obtained through a fully connected layer. On the dataset of actual hospital finance, a variety of multi-step prediction strategies are compared with those in conventional networks. The experimental results show that this model has higher prediction accuracy and efficiency compared with conventional models.

Abstract:In the optimization stage, Kubernetes determines the score of a node only according to its utilization of CPU and memory. This can only guarantee the resource utilization of a single node but fails to achieve the load balancing of cluster resources. In response to this problem, a genetic algorithm-based Kubernetes resource scheduling algorithm is proposed. In the algorithm, two evaluation indicators, i.e., network bandwidth and disk IO, are added and assigned with different weights. In addition, a check dictionary is introduced to check and repair the individuals that do not meet the configuration in the new population generated by the genetic algorithm. Experimental results show that compared with the Kubernetes default resource scheduling strategy, this algorithm takes into account the resource utilization of all nodes in the cluster and performs better in ensuring cluster load balancing.

Abstract:Small-sample problems are common challenges for training models. Because small sample data with insufficient information fails to represent the whole dataset, the data-driven models will have lower accuracy. This study proposes a Generative Adversarial Network (GAN) algorithm based on meta-learning for small sample data. It aims to train a generative adversarial network on various data generation tasks and find the optimal initialization parameters of the model. Consequently, new data generation tasks can be tackled with fewer training samples. The algorithm is applied to a water-cooled maglev unit for data generation. Experiments show that the algorithm can find the optimal initialization parameters under the condition of insufficient samples, which reduces the requirement for the dataset size. The failure classification experiment of mixed data verifies that the generated data is authentic, which is helpful for failure diagnosis and analysis.

Abstract:A pipeline vibration perception algorithm based on computer vision is proposed to solve pipeline damage or working fluid leakage at connecting flanges and valves due to a failure of timely warning induced by abnormal pipeline vibration in the production area of power plants. First, a convolutional neural network is used to estimate the optical flow information of the pipeline to be measured. Then, the information is analyzed to detect whether the pipeline vibrates or not. Finally, a vibration measurement module is employed to measure the vibration frequency and amplitude of the vibrating pipeline target in the monitor display for the perception of pipeline vibration. The experiments on the vibrating pipeline data taken by the original camera of a power plant show that the speed of the proposed method is about 4 f/s, and the measurement error of vibration frequency is less than 0.08. This method provides new ideas for computer vision technology to accomplish real-time pipeline vibration detection and measurement tasks without changing the original hardware devices of the power plants.

Abstract:The auxiliary information in recommendation systems can provide real help for recommendation, while the traditional collaborative filtering algorithm has a low utilization rate of the auxiliary information and high data sparsity in the calculation of user similarity, which leads to low recommendation accuracy. In response to this problem, this study proposes an improved collaborative filtering algorithm that integrates user preferences and multi-interactive neural networks (NIAP-CF). Firstly, the information about the item attribute preferences of users is collected according to the rating matrix and the item attribute feature matrix. Then, the SBM method is used to calculate the similarity of item attribute preferences between users to improve the calculation formula for user similarity. In the process of score prediction, we build a multi-interactive neural network prediction model integrating user and item attribute preference. Dynamic trade-off parameters are introduced to integrate the predicted scores calculated by user similarity and by the model. Experimental verification based on the MovieLens data set shows that the improved algorithm can improve the recommendation accuracy and reduce the MAE and RMSE of score prediction.

Abstract:Accurate load forecasting is one of the important tasks for power workers, and power load forecasting can be generally divided into short-term forecasting and medium- and long-term forecasting depending on the forecasting period. Compared with short-term power load forecasting, medium- and long-term forecasting is little explored by cutting-edge workers. Therefore, this study proposes an XGBoost-DNN-based algorithm that can be applied to mid-term power load forecasting. The algorithm combines the tree model with the deep neural network and introduces short-term forecasting into mid-term forecasting. According to the characteristics of the tree model, the data features are processed into high-order cross features, and in combination with the original data, the deep neural network is used to learn rich feature information. Algorithm analysis with the data of the 2017 Global Energy Forecasting Competition shows that in mid-term power load forecasting, the XGBoost-DNN model proposed by this method is more accurate than DNN and LSTM.

Abstract:Collaborative filtering algorithms are widely used in recommendationsystems. However, traditional collaborative filtering algorithms, which only use scoring information, have the defects of inaccurate similarity calculation and low personalization in actual scenarios and thus fail to meet user needs. For this reason, this study proposes an improved algorithm combined with user preferences and item attribute extension. Firstly, two improvements are made in the calculation of item similarity: Tag correlation is introduced to study the similarity between items; the extended attribute of items constructed according to the characteristics of the users who scored the item scan measure the item similarity in terms of item audience type. Secondly, considering the subjective preferences of users, a support vector machine is adopted to train the preference prediction model for each user, which can help to modify the item prediction score and improve the personalization and accuracy. Experimental results based on MovieLens dataset show that the proposed algorithm can calculate the similarity more accurately between items and get more accurate prediction scores according to users’ personalized preferences.

Abstract:The main part of drawing the knowledge map of electrical power systems is the extraction of power knowledge. In the traditional supervised-learning-based single neural network models, CNN performs well in extracting the most important local features but is not suitable for processing sequence input, and RNN is strong in tackling serialization tasks but weak in extracting important features. To solve these problems, this study puts forward a model based on GRU and PCNN. Compared with traditional models, this model combining the advantages of the GRU helped model and the PCNN model can obtain impressive results and effectively extract the knowledge of electrical power systems.

Abstract:Creating a convolutional neural network consumes substantial human resources, and much computing power is needed during training. The application of dilated convolution instead of the pooling operation in the convolutional neural network can considerably increase the receptive field and reduce the computational complexity, but the dilated convolution will bring about the loss of spatial hierarchy and information continuity. This study proposes a parallel asymmetric dilated convolution module, which can fill in the information lost by dilated convolution and be embedded in the current convolutional neural networks to replace the 3×3 convolution for network training. As a result, network convergence is accelerated and network performance is improved. The experimental results show that the proposed module can significantly improve the classification of various classical networks on CIFAR-10 and other data sets.

Abstract:The Steered Response Power-PHAse Transform (SRP-PHAT) localization algorithm has high accuracy but poor real-time performance. In response to this problem, this study introduces a localization algorithm based on Time Difference Of Arrival (TDOA) to improve real-time performance and then proposes a combination algorithm based on TDOA and Search Space Clustering (SSC), search space shrinking clustering, to optimize SRP-PHAT. First, the TDOA localization algorithm is used to estimate the range of the sound source in the direction angle and radial distance after outlier correction. Then, the search area is shrunk according to the estimated sound source range. Finally, fine-grained (5 cm) space search calculations in the shrinking area are performed by the SRP-PHAT-SSC algorithm to obtain the three-dimensional (3D) coordinates of the estimated sound source. A five-element microphone array and the virtual source method are employed to simulate the indoor sound field, and the 3D localization of the sound source is simulated by Matlab. The experimental results show that compared with the Full Grid Search (FGS) algorithm and the SSC algorithm based on SRP-PHAT, the improved algorithm has great real-time performance and robustness in 3D localization.

Abstract:This study proposes an improved HRNet based algorithm to solve the common problems of microvascular detail loss and lesion information misjudgment in the existing retinal vascular segmentation algorithms. In the pre-processing stage, the contrast between the blood vessels and the background is improved by contrast-limited adaptive histogram equalization and adaptive Gamma correction. During coding, HRNet original convolution is replaced by deformable convolution to improve the adaptability of convolution to complex vascular morphological structures. Concerning multi-scale feature aggregation, spatial pyramid pooling and multi-scale convolution are introduced to expand the receptive field and enhance the attention to the local features of the target. Consequently, vascular artifacts and subtle information loss can be improved. Simulation on the DRIVE database shows that the accuracy, sensitivity, and specificity of the proposed algorithm are 95.79%, 80.33%, and 98.12%, respectively.

Abstract:Pipelines are welded by steel plates and their inner surfaces may have scratches, internal fractures, pits, and other problems. If the abnormality on the inner surfaces cannot be found in time, a large number of unqualified products will be produced and the enterprises will have losses. This study designs a method for detecting anomalies on the inner surfaces of steel pipes based on image saliency. First, the image information after discrete cosine transform is collected and then fused with the phase spectrum of the image to obtain the final saliency map. Finally, the detection results are mapped to the original image through the connected region detection. Experimental results show that this method has a more remarkable detection effect, higher accuracy, and better stability and practicability than its counter parts.

Abstract:In the context of big data, the informatization of China’s power industry is becoming more important, especially the analysis of power consumption data with computer technology. For the analysis of abnormal user power consumption, traditional methods are time-consuming and labor-intensive. This requires the introduction of machine learning related methods to automatically identify anomaly information. At this stage, the analysis of abnormal power consumption is mainly based on traditional anomaly detection algorithms or deep neural networks. Anomaly detection algorithms have insufficient accuracy and calculations with deep neural networks are quite slow. In response to the current shortcomings, this study adopts an anomaly detection model of user power consumption based on sampling technology and LightGBM. The detection of abnormal power consumption is regarded as a classification problem, and the popular classification model LightGBM is applied to training. The detection accuracy is improved while fast speed is maintained.

Abstract:The classical TF-IDF algorithm only considers the feature term frequency, inverse document frequency, etc. but overlooks the distribution information of feature terms between and inside categories. In this study, we calculate the weights of feature terms through the TF-IDF algorithm in the corpus with different scales and analyze the impact of category information on weights. Based on this, a new method is proposed to measure the distribution information of feature terms between and inside categories. Furthermore, an improved TF-IDF-DI algorithm based on category information is proposed by adding two new weights and discrete factors between and inside categories to the classic TF-IDF algorithm. The Naive Bayes algorithm is used to validate the classification performance of the improved algorithm. Experiments show that the algorithm is superior to the classic TF-IDF algorithm in precision, recall, and F1 values.

Abstract:Concerning the shortcoming that the classical K-means clustering algorithm is easy to fall into the local optimum, an improved genetic clustering algorithm based on Hadoop is proposed and implemented. The algorithm overcomes the above shortcoming with the globality and parallelism of the genetic algorithm. On this basis, the genetic algorithm is improved and then combined with the classical K-means algorithm. To improve the implementation efficiency, we implement the improved genetic clustering algorithm on Hadoop. The proposed method is compared with the classical clustering algorithm through experiments. The results show that the proposed method can greatly improve the clustering accuracy and efficiency.

Abstract:Traditional network embedding approaches rely heavily on random walk in a node perspective to get the local sampling sequence of networks and then maximize the co-occurrence probability between adjacent nodes to represent nodes as low-dimensional vectors. The empirical analysis of this study on a real-world network shows that random walk in node and link perspectives can respectively produce network sampling results with different node frequency distributions, resulting in various partitions of the network. To this end, this study proposes an approach to Dual Perspective Based Coupled Network Embedding (DPBCNE). DBPCNE gets the network sampling sequences by random walk in a link perspective and then combines node sequences sampled in a node perspective for coupled training. Experiments show that compared with other network embedding approaches, this approach can well preserve network structures and improve the effectiveness of network embedding for the downstream classification and prediction tasks.

Abstract:We propose a Topology Aware Data Aggregation (TADA) method to address the problems of energy consumption and reconstruction errors in the data aggregation of a wireless sensor network. Firstly, a data stream including network initialization, data framing, and data preprocessing is constructed to form the communication process of the wireless sensor network. Secondly, a measurement matrix is built to decompose the data into multiple paths for forwarding and then the vector allocation of the whole network is carried out. Finally, a data aggregation algorithm based on a balanced minimum spanning tree is proposed. Experiments show that the proposed method is lower than other compressed sensing methods in the energy consumption of data aggregation and the error rate of data reconstruction.

Abstract:Braille is an important medium for visually impaired people to get information and learn knowledge. Because the traditional paper-based braille learning method only presents braille by tactile stimuli, it is hard for visually impaired people to master braille. The paper braille documents are too cumbersome to use, and their content cannot be refreshed. To improve the braille learning efficiency, we propose a digital braille learning method based on synchronous stimuli of vision, audition, and touch. On a braille learning machine with the functions of multi-sensorial channels, we design an information matching algorithm for text, sound, and braille dots. This method can synchronously output visual, auditory, and tactile stimuli, which provides reasonable conditions for visually impaired people to learn braille efficiently. With the short-term memory research method, the experimental results are as follows: (1) The braille learning efficiency is the highest with the combination of visual, auditory, and tactile stimuli; thus, for people with weak vision, the visual stimulus is powerful to improve their braille learning efficiency. (2) The braille learning efficiency is lower with the combination of auditory and tactile stimuli, and thus it is hard for the blind to master braille in a short period. (3) The braille learning efficiency is very low with only the auditory stimulus; thus, APP only providing auditory stimulus to learn braille is less useful.

Abstract:Concerning the real-time monitoring problem of warping dams in northwest China, this work studies the scheduling method of warping dam monitoring and early warning tasks. To avoid the delay in discovering the hidden trouble of warping dams and improve the timeliness of the warning system, this study considers the average waiting time from task unloading to edge servers and proposes a collaborative task scheduling method based on edge computing in warping dam monitoring. A task completion time model is built according to the task computation, computing power of edge servers, and other information. Then, a simulated annealing algorithm is used to optimize the unloading position of computing tasks. A task scheduling strategy is designed in which multiple edge computing servers cooperate. Experimental results show that this method can greatly reduce the calculation time of monitoring tasks and improve the timeliness of monitoring and early warning.

Abstract:Multi-energy systems provide a new solution to energy and environmental problems through the collaborative optimization of production, transmission and consumption of different forms of energy, such as electricity, heat and gas. Energy Hub (EH) is the coupling link of a multi-energy system, and its configuration scheme is crucial to the optimal operation of the multi-energy system. In this context, a scheme for optimal operation and configuration of EHs with different forms of energy, such as electricity, heat and natural gas, is proposed, taking into account the CO₂ emission. On the basis of considering CO₂ emission, a multi-objective optimization problem is raised, which is solved by Genetic Algorithm (GA) overall with the aim of maximizing social benefits and minimizing CO₂ emission. Finally, the effectiveness of the proposed method is verified through the analysis of different configuration examples. This study can provide theoretical and technical support for the construction and operation of EHs.

Abstract:In this study, according to the characteristics of specific application scenarios for the IT equipment identification of State Grid, accurate identification and positioning of the equipment is realized with improved Faster-RCNN, thereby improving the management efficiency of the grid data center. The algorithm is improved mainly in terms of the attention mechanism, the initial anchor box adjustment and the anchor box fusion. The comparison with common image algorithms shows that the improved model has the convergence speed and the accuracy increased by 30% and 1%, respectively.

Abstract:This study proposes an infrared and visible image fusion method based on latent low-rank representation and guided filtering to address the serious detail loss and the poor visual quality in the fusion. First of all, the source image is decomposed by latent low-rank representation into low-rank layers and salient layers. Then the low-rank layers are decomposed by guided filtering into basic layers and structural layers with the aim of extracting more structural information from low-rank layers. According to the characteristics of basic layers, structural layers, and salient layers, visual saliency weighting, gradient saliency weighting, and absolute maximum selection are used as fusion rules, respectively. In particular, since the initial weight is noisy and unaligned with the object boundary, it is optimized by guided filtering. Finally, the basic fusion layer, the structural fusion layer, and the salient fusion layer are overlapped to yield the fused image. The subjective and objective evaluation results of several groups of fused images are compared. The proposed method is found able to effectively extract the detail information of source images and superior to other image fusion methods in terms of visual quality and objective evaluation.

Abstract:Using embedded devices to detect pedestrians at the edge can meet the basic needs of real time, security and privacy protection. The original CenterNet backbone network model usually adopts Deep Layer Aggregation (DLA), Hourglass, etc. with high complexity for multi-level features fusion, which limits the computing power of embedded devices and thereby makes the real-time detection difficult. In view of this, BiFPN and weighted feature fusion are employed for the weighted fusion of feature layers in the backbone, by which the original backbone method is improved. This strategy enhances the detection speed while ensuring the detection accuracy. Further, the Gauss kernel distribution on the HeatMap during training was modified so that the adaptability to pedestrian detection can be increased. As a result, the accuracy reduction caused by missing detection due to pedestrian occlusion is lowered. The results of the experiment on Jetson TX2 show that the Average Precision (AP) of pedestrian detection with the improved method is 0.774, and the inference time of a single image is 68 ms, which can meet the requirements of embedded devices for real-time detection.

Abstract:This study proposes a vector-based complex network modeling method on the basis of the relevant traditional method, which can model the entire complex system according to the heterogeneity of nodes. Depending on business characteristics, different business networks are modeled by virtue of the idea of hierarchical modeling, and business-driven complex system modeling algorithms are used for the networking of different networks. The network splitting algorithm is used for partial splitting of complex systems to analyze local performance. The proposed method not only reduces the complexity of modeling but also enriches the expressive ability of the model, which is of great significance to the analysis of complex systems. Finally, a smart grid is taken as an example to verify the effectiveness of the method.

Abstract:With the rapid development of cloud computing technology, more and more data-related services need to be implemented through cloud platforms. The National Microbiology Data Center has established a data platform for bioinformatics analysis based on cloud computing. To address the problems related to efficiency, security, stability, etc. that users of the platform may face when uploading files, this paper presents a transmission scheme for selecting the number of concurrent threads according to the bandwidth-delay product based on the technologies of piecewise upload and broken-point continuingly-transferring. This scheme solves the problem that all files need to be re-uploaded when accidentally interrupted and improves the utilization efficiency of bandwidth resources during multi threaded upload and the transfer efficiency of large files.

Abstract:Histopathological image analysis is the “gold standard” for cancer diagnosis, which plays an important role in the prognosis and treatment of patients. Currently, in the field of AI medical imaging, the classification of pathological images based on Convolutional Neural Network (CNN) has become a research hotspot. However, the Max/Average pooling module is widely used in traditional CNNs, which inevitably lose massive feature information in pathological images, resulting in low classification accuracy and difficult model convergence. Therefore, this study proposes a pathological image classification method based on Wavelet Decomposition Convolutional Neural Network (WDCNN). This method can make the traditional CNN model learn the frequency domain information. It introduces the multi-scale analysis of wavelet transform into a CNN model and uses wavelet decomposition to replace the traditional pooling layer, which reduces the loss of features compared with max and average pooling. In view of different characteristics of the space domain and the frequency domain, the high-frequency components after wavelet decomposition are added to the next layer through shortcut connections to make up for the detailed feature information lost in the pooling process. This paper evaluates the performance of different pooling methods and different wavelet basis functions in pathological image classification on the Camelyon16 dataset. According to the experimental results, the CNN model integrated with wavelet decomposition can improve the classification accuracy of the network.

Abstract:The convolutional network depth is crucial to accurate large-scale image recognition. In this work, we thoroughly evaluate the networks with increasing depth using the architecture with quite small (3×3) convolution filters. The prior-art configurations can be improved significantly after the depth is pushed to 16–19 weight layers. The comparison with the convolution networks of other convolution filter architectures verifies the effectiveness of the proposed network for large-scale image recognition. In addition, the network verification is conducted with some other data sets to avoid the inherent bias of training data sets. As a result, the most advanced results can be obtained from these data sets.