Abstract:Considering various factors affecting automobile sales, the penalty coefficient and kernel function parameters of support vector machine are optimized by cross validation and grid search, and a prediction model suitable for automobile sales is established. The simulation results show that the forecasting effect of the improved support vector machine optimized automobile sales forecasting model is better than that of the current model adopted by a company. The model has higher forecasting accuracy and greater credibility, and can provide more accurate sales forecasting reference for enterprise decision-making level.

Abstract:To solve the problem of low efficiency caused by random mutation, a more effective mutation strategy is proposed in this study. The proposed approach synthesizes different kinds of information to help the Fuzzer mutate seed file, i.e., the CFG of program, the characteristics of input seed file, the information of abnormal input detection, and the branch courage of the Fuzzer. Based on this strategy, we design a new Fuzzer which continuously monitors the execution path of each seed file used as input of target program. Meanwhile, as most path constraints depend on only a few bytes in the input, periodical byte-level taint tracking will be necessary in the whole fuzzing process. After all this, we can infer a one-to-many mapping relation between the new execution path and key bytes in seed files, which can highlight the target start-end tuples of the seed file with more possibility to explore new branches in the target program to mutate. The result shows our design can improve the branch coverage of target program and the efficient of Fuzzing.

Abstract:In view of the incompleteness of the clinical diagnosis method of liver fibrosis and the incompleteness of the feature extraction of traditional machine learning methods, by the deep transfer learning method, this study uses the pre-trained ResNet-18 and VGGNet-11 models for the diagnosis of liver fibrosis. Different degrees of transfer training were performed using the rat liver fibrosis nuclear magnetic resonance image dataset provided by Southern Medical University. The two models were trained using six network migration configurations on the MRI image datasets collected by four different parameters. The experimental results show that the use of T1RHO-FA parameters to acquire nuclear magnetic resonance images and the use of VGGNet-11 model can improve the accuracy of liver fibrosis staging diagnosis. At the same time, compared with the ResNet-18 model, the deep model migration learning method can stably improve the accuracy and training speed of the VGGNet-11 model for liver fibrosis staging diagnosis.

Abstract:This paper presents a texture synthesis and 3D reconstruction algorithm based on a single face image combined with standard skin color. Firstly, ASM algorithm is used to extract face feature points, and color conversion is realized by editing propagation based on Local Linear Embedding algorithm, so that the tone of face image is consistent with the standard skin color of 3D face model. Then, the facial features are fused with the standard skin color map, and the eyebrow occlusion is considered. The eyebrows are restored by using the facial symmetry or eyebrow template. Especially for semi-occluded eyebrows, the method of combining Li model with corner detection is used to reconstruct the eyebrow contour and get the final face texture. Finally, the face texture map is mapped to the three-dimensional face model by texture mapping, and a better personalized three-dimensional face reconstruction effect is obtained. Experiments show that the proposed algorithm can be applied to face images captured under different complex background and illumination conditions. It has a faster processing speed and can be applied to real-time face reconstruction products.

Abstract:Infrared images have low contrast and low signal-to-noise ratio, which is always a huge challenge for the segmentation of infrared photovoltaic panel images. In order to solve the problem that the traditional Fuzzy C-Means (FCM) clustering algorithm is susceptible to the uncertainty of the initial clustering center and does not consider the spatial information, a clustering algorithm based on FCM is proposed. The algorithm uses the histogram, meanwhile, the characteristics of the graph determine the initial clustering center, and based on the traditional FCM and Fuzzy Kernel C-Means (KFCM) algorithm, the traditional FCM is improved by the relationship between the spatial information among pixels and the neighboring pixels. The objective function is clustered to derive a new objective function. The experimental results show that the proposed algorithm has significantly lower over-segmentation and mis-segmentation rate than the Otsu algorithm, the adaptive k-means algorithm, and KFCM algorithm. The effect is very close to the manual segmentation map.

Abstract:Recognition of human actions in videos is an important research field in computer vision in recent years. However, existing methods have insufficient representation of video and cannot focus on significant areas within the image. We propose a deep convolutional neural network based on visual attention, which can effectively add a weight to the video representation features, pay attention to the beneficial regions in the features, and achieve more accurate behavior recognition. We conducted experiments on HMDB51 and our own Oilfield-7 dataset to verify the validity of the model proposed for human actions on the oilfield. The experimental results show that the proposed method has certain advantages compared with the two-stream architectures which have achieved excellent performance.

Abstract:The forest is the mainstay of terrestrial ecosystems. So the protection and restoration of forest ecosystems are the significant parts of ecological civilization construction. In order to solve the problem that forest ecological models are deployed in local machines and developed repeatedly, we manage to design and implement a web processing system for the study on forest ecology. According to the interface specification of spatial data sharing and interoperability, which is formulated by Open Geospatial Consortium (OGC), and by combining with the technology of WebService, the system manages to realize the multi-model integration for the study on forest ecology and displays the calculation results, which facilitates the communication of researchers and the public. It could reduce the work of developing repeatedly, and increase the sharing of data and service.

Abstract:The paper analyzes the role of code dependency visualization in intelligent software development. On the basis of summarizing the characteristics of code dependency and the general process of information visualization system, the design of code dependency visualization system is proposed. The system uses force-directed node-link graph and hierarchical edge bundles as visualization types. Based on in-depth analyses of the layout characteristics of the two visualization types, the interaction design of filtering children nodes is created for force-directed node-link graph, and the interaction design of aggregating leaf nodes is created for hierarchical edge bundles. After implementing the code dependency visualization prototype system dpViz using a variety of software technologies, the system is tested in enterprise software development. The experimental results show that the visualization prototype system can effectively improve the efficiency of code analysis.

Abstract:Aiming at the real-time environment state of vehicle battery voltage and working temperature, remote monitoring communication system platform is designed and built by using embedded micro-processing system combined with 4G communication technology and Wireless Local Area Network (WLAN). The client terminal of the system collects the data and realizes the short-range and long-distance data communication. The server of the system uses the PC of Linux system and the mobile phone of Android system to receive the data. In the course of the experiment, the whole system can monitor the voltage of battery pack and the operating temperature environment accurately and remotely, and the error of the data shows that the voltage is less than 1%. The whole system runs stably, the performance is reliable, which demonstrate the convenience for the management personnel to the spot understanding.

Abstract:With the rapid development of information technology and network, the information resources of the Internet have increased dramatically. The problem of information overload has promoted the development of customized recommendation technology. Collaborative filtering algorithm is widely used in various fields of e-commerce by establishing links between users and information. In this study, it is proposed that the user's personalized information data can be obtained by using the WeChat mini program, and through the collaborative filtering algorithm, an intelligent assistant designed for users can recommend customized life service information for users. In this paper, the design method of intelligent assistance is introduced, and the realization of system functions and customized recommendation function introduces elaborated in detail.

Abstract:In order to improve the efficiency of retrieval and query of meteorological automatic station data, Redis database technology based on memory key-value structure is adopted. By building a Redis database cluster, the data are cached in memory and the master-slave replication is realized. A data storage structure model suitable for the characteristics of meteorological automatic station data is proposed, which makes high frequency access possible. Meteorological automatic station data can be read directly from memory, effectively reducing the time of data query response. In this study, the hourly automatic weather station observation data transfer and reporting time characteristics are analyzed. Based on Redis, the hourly automatic weather station data are stored and designed from business database, index database, and time database respectively. Finally, through experimental comparison, it is concluded that the hourly automatic weather station observation data are cached by Redis, and the retrieval efficiency of proposed method is higher than that of CIMISS. This method can also be applied to other information retrieval and invocation business scenarios, and has a good application value.

Abstract:This article introduces functions and system operation of satellite resource management system. The Party Central Committee pays much attention to the remote education for members of the Chinese Communist Party. In this article, the basic service of the system is supported by satellite and domestic independent controlled information platform. Multi-language satellite terminal resource management is built using Symfony software to manage the content and learning process of satellite terminal resource, which is meaningful to the development of remote education for Party members in Xinjiang.

Abstract:The Internet of Things (IoT) is widely used in all walks of life. The central idea of IoT is to parameterize the real world and combine with sensing devices, network communication technique, data technologies, and other means to realize the direct usability of the IoT data for users. Based on the Shancheng intelligent gas data management system, this study proposes a hybrid scheme combining the message middleware Kafka and the NoSQL database MongoDB for performance bottleneck of the application service platform in the IoT platform architecture when encounters large scale device access scenarios. The scheme is implemented based on the application background of gas companies and equipment manufacturers, the application service platform concurrency performance and data persistence efficiency are improved.

Abstract:In this study, an improved A* algorithm is proposed for the traditional A-star algorithm, which has many redundant path points and long-term one-way search. The proposed algorithm uses a bidirectional preprocessing structure to reduce the number of redundant nodes, and further optimizes the evaluation function to improve the traversal speed by normalizing the processing and adding node marker information. Simulation software is used to simulate the improved A* algorithm and compare with other classical path planning algorithms. The simulation results show that the improved A* algorithm can complete the global path planning with lower search node number and search duration than the traditional A* algorithm.

Abstract:In order to improve the accuracy of human body image segmentation under complex background, a new human body image segmentation algorithm is proposed. This algorithm solves the problem of specifying the number of pixel blocks in the super-pixel block segmentation for the simple linear iterative algorithm (SLIC). By referring to the CV energy model, it is constructed by minimizing the image into multiple regions for horizontal set iterative segmentation. The adaptive super-pixel block is made such that each super-pixel block after the segmentation fits a single color block in the image. Then combined with the human body average template, the human body standard posture area of interest is marked on the picture, which improves the anti-interference ability of the algorithm against the complex background. Finally, each super-pixel block is clustered as a node by k-means clustering algorithm to realize standard human body image segmentation. The experiment is carried out by collecting multiple sets of pictures in different environments. The results show that the proposed algorithm improves the segmentation accuracy of the human body's standard posture and ensures strong anti-interference ability for complex backgrounds with rich chroma.

Abstract:At present, the resolution of the image captured by the camera is much higher than that of the screen. A large number of moires are appeared after the image is scaled. Aiming at the phenomena of moire interference, large memory, and short focal length distortion, this study puts forward a kind of image reconstruction algorithm for the first time. First, the adaptive high-frequency linear filtering algorithm is used to separate the high frequency signal from the network. Then the high-frequency information of the image is extracted by the fast Greedy algorithm and Bernoulli random walk. According to the prior knowledge of the structure of the liquid crystal point, a complete liquid crystal layout model is constructed and mathematically fitted to separate the information of the screen and the reticulation of the liquid crystal point. Finally, the reconstruction of the image matrix space and the interpolation and restoration of the color channel are carried out. The reconstruction algorithm is consistent with the original image in matrix space. The histogram similarity of the experimental sample and the original image is more than 80%, which is much higher than that of the screen image and the original image 38.7%. The algorithm retains the image information, eliminates the high-frequency mesh noise, restores the original matrix space, and makes the image more in line with the actual needs.

Abstract:The research is carried out for the purpose of emotion recognition, and the signal feature method of wavelet filtering transformation combined with autoregressive model extraction is proposed on the basis of Principal Component Analysis (PCA). Besides, sentiment classification is realized on the basis of gradient promotion classification tree. The focus of feature extraction is laid on the changes of Electro Encephalo Gram (EEG) signals and the changes of wavelet components as features of EEG signals. The multimodal standard database DEAP proposed by Koelstra et al. to analyze human emotional state is adopted to extract eight positive and negative emotions to represent 14 channels of EEG data in each brain region. The results suggest that the average accuracy of the algorithm for 8 kinds of emotions in pairwise classification is 95.76%, and the highest accuracy is 98.75%, making it possible to help emotional recognition.

Abstract:The application of bipartite graph theory in collaborative filtering recommendation based on substance diffusion theory of complex networks has attracted more and more attention from scholars. Existing algorithms mainly consider the positive rating when calculating neighbor users, ignoring the negative rating of users. In order to improve the accuracy of recommendation algorithm, a collaborative filtering recommendation algorithm based on improved bipartite graph and user reliability is proposed. The algorithm quantifies both positive ratings and negative ratings into the weight of the path, which controls the user's energy distribution, and takes users' reliability into account when predicting the rating, therefore, the accuracy of recommendation result is significantly improved. A series of comparative experiments are carried out on MovieLens and Eachmove datasets. The experimental results show that the improved algorithm has lower mean absolute error.

Abstract:For the limitations of the Cellular Automata (CA) model and the original Artificial Fish Swarm Algorithm (AFSA) in describing the conventional evacuation behavior of the comprehensive transportation hub personnel, a kind of pedestrian evacuation model based on the CA-Improved AFSA (CA-IAFSA) is proposed with considering the difference of walking speed and the difference of view between individuals. As the queuing mechanism and the export (entrance) selection behavior, the guiding behavior, and the memory function are added to the original AFSA. The top layer adopts the IAFSA for mobile location updating, and the bottom layer uses the CA model to solve moving position conflicts. Experiments show that the model can truly reflect the evacuation process of people transferring vehicles in an integrated transportation hub. Under the same environment, compared with the original AFSA, the proposed model realizes the orderly movement of individuals according to guidance, avoiding falling into local optimum. Compared with the CA model, it is better in terms of reflecting the individual's herd, obstacle avoidance, and export (entrance) selection behavior, thus effectively reduces the time complexity.

Abstract:Short text classification is one of the hotspots of research in natural language processing. A new model of text representation is proposed in this study (N-of-DOC), and in order to solve the problem of sparse representation in Chinese, the Word2Vec distributed representation is used, finally, it is applied to the improved Convolution Neural Network (CNN) model to extract the high level features from the filter layer, the classification model is obtained by connecting the Softmax classifier after the pooling layer. In the experiment, the traditional text representation model and the improved text representation model are used as the input of the original data, respectively. It acts on the model of traditional machine learning (KNN, SVM, logistic regression, naive Bayes) and the improved CNN model. The results show that the proposed method can not only solve the dimension disaster and sparse problem of Chinese text vectors, but also improve the classification accuracy by 4.23% compared with traditional methods.

Abstract:Data center energy optimization problem is an important research direction in cloud computing field. However, in the real world, relevant research needs to bear high research costs and long experimental period. Therefore, simulation technology has been widely used in this field. In order to improve the accuracy and reliability of the data center energy-aware simulation experiment, this study analyzes the built-in energy consumption model of the simulation platform and the energy consumption evaluation methods proposed by other scholars, and puts forward the energy consumption evaluation method, which considers the impact of CPU utilization on memory energy consumption, based on CPU and memory usage rate. Also, a multivariate nonlinear model is used for regression analysis. The experiment proves that the energy consumption evaluation method proposed in this study can be applied to the simulation platform and has high prediction accuracy, which effectively improves the accuracy of the energy consumption evaluation of the cloud computing simulation platform.

Abstract:Aiming at the insufficiency of the classical Douglas-Peucker data compression algorithm, such as low recursion efficiency and uncertain threshold selection, the study proposes an improved feature point extraction method. The algorithm calculates the frequency of data points by histogram, according to the distance from the data point to the baseline and the turning angle between adjacent data points, considers the "isolation" and frequency of data points, and entropy value method is used to obtain the ultimate evaluation value, the curve data is compressed automatically with the given data compression ratio. The simulation experiments are carried out on MATLAB, and using the self-developed control system platform, incremental self-learning for the flow characteristics of inlet metering valve, the corresponding experiments are carried out on pump test bench and diesel engine test rig. Experimental results show that this algorithm can effectively compress data to meet the data compression requirements of measurement system and control system.

Abstract:For practical application in the design of fractal graphics demanded in clothing design, a method of automatically extracting core basic pattern and generating quadrangle continuous patterns is put forward based on fractal graphics content. The method first uses the Canny operator for edge detection, analyzes the region of the main pattern, and then extracts and analyzes the texture features using the gray level co-occurrence matrix. On this basis, the optimal stitching method is selected to splicing the extracted patterns to generate a square continuous pattern. This method produces a square continuous pattern with beautiful appearance, low computational complexity, strong universality, and sound practicability.

Abstract:The collaborative filtering algorithm has been widely used in the recommendation system. It has significant effects in implementing the new recommendation function, but there are still problems such as sparse data, poor scalability, cold start, etc. New design ideas and technical methods are needed for optimization. In recent years, deep learning has achieved outstanding results in the fields of image processing, target recognition, and natural language processing. Combining the deep neural network model with the recommendation algorithm has brought a new opportunity for the construction of a new recommendation system. In this study, a new hybrid neural network model is proposed, which consists of stack denoising autoencoder and deep neural network. It learns the potential feature vectors of users and projects and the interaction behavior model between users and projects, effectively solves data sparseness, and thus improves the quality of system recommendations. The recommended algorithm model is tested by the MovieLens film scoring data set. The experimental results are compared with traditional recommendation algorithms such as SVD, PMF, and classical autoencoder model algorithms, the recommendation quality is significantly improved.

Abstract:Real-time communication mainly transmits real-time audio and video with low latency and high bandwidth consumption. In the scenario of large number of users, the single server solution cannot meet the overall needs, so it is necessary to build a distributed cluster to provide services, and how to properly distribute these requests to different servers, balance the load of servers in the cluster is particularly important. This paper first analyzes the real-time communication flow of single server. Then the common load balancing algorithms are analyzed. In order to meet the consistency requirements of forwarding from the same group of clients to the same server, an adaptive load balancing algorithm based on consistency Hash algorithm and genetic algorithm is proposed and verified.

Abstract:Aiming at the characteristics that many cluster intelligent algorithms are easy to fall into local optimum and have slow convergence rate, a new algorithm (PK algorithm) with less parameter settings and strong global search ability is proposed. The comparison of 10 benchmark functions with particle swarm optimization algorithm verifies the effectiveness of the algorithm, because the average and minimum values of the PK algorithm under 30 trials are better than the particle swarm optimization algorithm. Then using the PK algorithm to optimize the BP neural network, and 11 test data sets were classified. The experimental results show that the BP neural network based on PK algorithm has better performance than the original algorithm on 11 test sets, and the performance is superior to BP neural network based on genetic algorithm on most test sets. Thus, we conclude that the BP neural network based on PK algorithm can effectively improve the classification accuracy and enhance the robustness.

Abstract:In the semi-supervised learning process, the performance of the classifier is often degraded and unstable due to the random selection of unlabeled samples. At the same time, the performance of the traditional semi-supervised learning algorithm is not sufficient for the classification problem of high-dimensional data containing only a small number of labeled samples. In order to solve these problems, this study proposes a safe semi-supervised learning algorithm S3LSE, which combines stochastic subspace technology with ensemble technology from the perspective of exploring data sample space and feature space. Firstly, S3LSE decomposes the high-dimensional data set into B feature subsets using random subspace technique, and optimizes each feature subset according to the implicit information among the samples to form B optimal feature subsets. Then, each optimal feature subset is sampled to form G sample subsets, and a safe sample marking method is used in each sample subset. The learning algorithm generates G classifiers and integrates G classifiers, and then integrates B classifiers generated by B optimal feature subsets to realize the classification of high-dimensional data. Finally, a high dimensional data set is used to simulate semi-supervised learning and the experiment result shows that the algorithm has better performance.

Abstract:Nowadays, with the rapid development of science and technology, a number of new technologies have emerged. New scientific fields such as data mining and machine learning have been deeply studied. Many intelligent algorithms have emerged and applied to different fields. This paper constructs a combined model based on BP (Back Propagation) neural network and SVR (Support Vector Regression). Based on the agricultural product price data, the example verification analysis shows that compared with the single prediction model, the BP-SVR-BP prediction model has greatly improved the prediction accuracy. The fitting effect is closer to the real data curve, which can objectively and truly reflect the law of agricultural product price changes.

Abstract:To address the corpus collection, the corpus collection system based on semantic relevancy focused crawler is implemented. Word vector trained by Wikipedia and HowNet are used for calculating page information semantic relevancy with descriptive information according to topical keywords, and the URL structural information is used for calculating the topical relevancy. Experimental results show that this system has better effect on party-construction corpus collection with high precision of average accurate rate 94.87%, while the average accurate rate for web pages is 64.20%.

Abstract:Aiming at the problems of low accuracy and over-fitting of traditional neural network model in flood forecasting process, this study takes the monthly average water level of Waizhou Hydrological Station in Ganjiang River Basin as the research object, and proposes a flood forecasting model based on regularized GRU neural network to improve the accuracy of flood forecasting. Relu function is selected as the output layer activation function of the whole neural network. To improve the generalization performance of GRU model, regularization of elastic network is introduced into GRU model, and regularizes the input weights in the network. The model is applied to the fitting and prediction of the monthly average water level at Waizhou Hydrological Station, and the experimental comparison shows that the model optimized by regularization of elastic network has a higher fitting degree, the qualified rate is increased by 9.3%, and the calculated root mean square error is small.

Abstract:Scientific and technological innovation is the main driving force of urban and regional development. Based on the evidenced data, it is particularly necessary to make a rapid and accurate quantitative evaluation of scientific and technological innovation for a specific city. Based on the evaluation index system and the visual development tools, this study designs a mobile APP for city scientific and technological innovation evaluation and applies it in the practice. The system realizes the functions of index display, index analysis and evaluation in the mobile terminal according to the index data. Moreover, combined with the geospatial information, this APP can browse and contrast the index information of different cities. This is of great significance for promoting the efficiency and accuracy of urban scientific and technological innovation.

Abstract:A method for optimizing the location of charging stations is proposed for the one-way shared electric vehicle system, with the consideration of rational capacity of charging station and the demand within the served area. Thus the problem of unbalanced vehicle inventory is solved and the customer demand is satisfied to the maximum. The method is based on the mixed integer programming model. The goal is to meet the revenue of the service, the operating cost of the station, and the depreciation cost of the vehicle. The objective function is to maximize the profit of the shared electric vehicle service provider. Finally, simulations are carried out to test the slack and performance of the method, and the model can solve large-scale problems in a reasonable time.

Abstract:Long-Term and Short-Term Memory (LSTM) network is a time recursive neural network, which is suitable for predicting events with relatively long delay in time series. In this study, a new fog proximity prediction framework based on LSTM network is constructed, which can transform meteorological observation data into time series data and model them based on time series data. In order to validate the proposed model effectively, this study transforms the surface meteorological data of 81 national stations in Anhui Province from October 2015 to June 2017 into sequence data and constructs a validation data set. Based on this data set, the future 1-4 hourly fog forecasting experiments are carried out. The experimental results show that the proposed model's TS-Scores are 61%, 55%, 36%, and 31%, respectively, which are obviously better than CNN and those of traditional machine learning algorithms such as SVM and KNN. It is an effective method for fog prediction.

Abstract:Traditional random forest classification algorithm cannot distinguish the strong and weak classifiers by using the majority voting rule, and the value of its hyperparameter needs to be adjusted and optimized. This work studies the application technology of random forest algorithm in text classification and its advantages and disadvantages, and optimizes it. On one hand, optimize the voting method, perform weighted voting by combining classification effect and prediction probability of decision tree. On the other hand, an algorithm combining random search and grid search is proposed to optimize the hyperparameters in random forest. The experimental results in python environment show that the proposed method has sound performance in text classification.

Abstract:In this study, we propose a combination approach based on group decision-making method, using random forest, neural network and GBDT as individual learners, to assess credit risk of borrowers in P2P lending. To validate the proposed method, two real-world datasets from PPDai.com and renrendai.com are examined. The results show that, compared with the individual learners, the proposed method has made a better performance.

Abstract:Object detection is widely used in surveillance systems for pedestrian detection and face recognition. It is a research hotspot of current deep learning. Supervised learning trains pedestrian detectors for specific scenes by manually annotating large datasets. However, the manual labeling method is time-consuming and laborious. In this work, the shortcomings of manual labeling of datasets for supervised learning are studied. A method of semi-automatic labeling of pedestrians is proposed. The surveillance video captured by the stationary monocular camera, using the initial foreground possibilities provided by the optical flow information, and the visual similarity across time, iteratively updates the initial foreground likelihood to segment the moving pedestrians. According to the segmented foreground pedestrians, a method of semi-automatic labeling of pedestrians is proposed. The experimental results show that the proposed method can provide a large number of datasets for the pedestrian detection system, and the efficiency is obviously superior to the traditional manual annotation method.

Abstract:With the emergence of large image sets and the rapid development of computer hardware especially GPU, Convolutional Neural Network (CNN) has become a successful algorithm in the region of artificial intelligence and exhibit remarkable performance in various machine learning tasks. But the computation complexity of CNN is much higher than traditional algorithms, however, the restrict of limited resources on embedded devices become a challenging issue for making efficient embedded computing. In this study, we propose a efficient convolutional neural networks based on embedded devices for electrical equipment inspection, this efficient neural network is evaluated in term of processing speed. The results show that the proposed algorithm can meet the requirement of real-time video processing on embedded devices.

Abstract:One may often need to conduct the inquiry and comparison of information related to weapons and equipment during the teaching and training in the academy and field troops. Traditionally, this can be realized through relation database technology. However, on the one hand, there is no common "data model" available due to the big difference between equipment types. On the other hand, it is very difficult for the relation database to deal with the "connections" between different types of equipment. This study proposes an idea of graph database-based "association" inquiry of the equipment information, aiming to solve the problem of the traditional system.

Abstract:The recognition of rust target on power equipment has very high application value in power security, nevertheless, the rust has the characteristics of irregular size and shape, thus the detection efficiency and accuracy of traditional machine learning algorithm are not high. Aiming at this problem, the characteristics of rust stain are studied and analyzed, and a rust detection and recognition method for power equipment based on Mask R-CNN is proposed. Faster R-CNN is used to complete the function of target detection, FCN accurately completes the function of semantics segmentation, realizes the classification and recognition at the pixel level, and better solves the problem of irregular rust detection. The experimental results show that the accuracy of rust detection of power equipment based on Mask R-CNN is high.