Abstract:Integrating context information into the traditional recommendation systems can effectively solve the problems such as data highly sparse, cold boot, and difficult to model user preference. Deep learning technology has become a research hotspot in the field of artificial intelligence in recent years, it will bring new opportunities and challenges to research in the field of recommendation while deep learning is applied into context-aware recommender systems. In this paper, some application models about the integration of deep learning technology into context-aware recommendation systems are mentioned, and deficiency of context-aware recommender systems based on deep learning and prospect in the future are elaborated at the same time, by introducing the related concepts of context-aware recommender systems and collating relevant research literatures worldwide.

Abstract:Pain is a personal experience that is generally divided into acute pain and chronic pain, which can result from injury, illness, surgery, or other health problems. If pain is not treated in time, it will cause great harm to the patient's physical and psychological health. Not all patients, e.g. dementia, are able to self-report pain. The continuity and objectivity of the assessment cannot be guaranteed when medical care personnel assessing patient's pain. Therefore, the demand for automatic pain recognition system is increasing. In the past decade, many researchers have made breakthroughs in this area. This paper reviews the automatic pain recognition system. On the one hand, it describes the structural composition of the automatic pain recognition system, including data acquisition, data preprocessing, feature extraction, and classification. On the other hand, it summarizes a great number of techniques from pain modal representation, i.e. behavior, speech, physiology, and multi-modal fusion. This paper also discusses the key technologies in automatic pain recognition system, and analyzes several challenges and directions in the field.

Abstract:In order to further enhance the prediction performance of oil futures price, this study proposes a novel CEEMDAN-PSO-ELM model for oil futures price forecasting based on CEEMDAN decomposition algorithm, extreme learning machine, and particle swarm optimization technology. Firstly, the original oil futures price series is decomposed by CEEMDAN algorithm into several intrinsic mode functions and a residual. Secondly, all the intrinsic mode functions and the residual are reconstructed based on Lempel-Ziv value. Then, the high, medium, and low frequency component are obtained respectively. Thirdly, the extreme learning machine optimized by particle swarm optimization algorithm is employed to predict each component and three component prediction results are obtained. Finally, integrate the prediction results of three components. The empirical research demonstrates that the CEEMDAN-PSO-ELM model proposed in this study has the best prediction performance compared with other 15 benchmark forecasting models. Moreover, the model confidence set and Diebold-Mariano test results further confirm the robustness of the proposed model.

Abstract:In 2009, Fujian Pingtan Comprehensive Experimental Zone was established as a window for cooperation between Fujian and Taiwan and the country's opening to the outside world. Its land use change is mainly affected by social and economic factors and natural geographical environment, and is also closely related to future land use planning. Landsat remote sensing image data of 1990, 2000, 2010, and 2017 is used to quantitatively analyze the impact of land use change on landscape pattern in the past 27 years. The results show that:(1) high accuracy of remote sensing land use classification can be obtained by using random forest method when selecting suitable training samples (the overall accuracy of the 4 remote sensing image classifications is above 87%, and the Kappa coefficient is above 0.84). (2) From 1990 to 2017, the water area decreased sharply by 31.04 km², and the lost water area is mainly converted into construction land and forest land; the construction land is increased by 40.98 km², and the annual average growth is 1.52 km². In the past ten years, it has shown a rapid growth trend with an average annual growth of 3.87 km². (3) At the plaque type level, the construction land is increasing year by year. The largest plaques accounted for the proportion of landscape area (LPI), degree of polymerization (AI), and edge density (ED), and the LPI was most affected by the increase of construction land. At the landscape type level, diversity (SHDI) and landscape shape (LSI) are declining.

Abstract:As one of the most influential networks in the field of deep learning, convolutional neural network is deeper and deeper, and proposes higher demand for computing capabilities. Various dedicated processors have emerged. In order to compare such processors fairly and help to optimize software and hardware, this study proposes macrobenchmarks and microbenchmarks for convolutional neural networks. The macrobenchmarks include mainstream convolutional neural networks for evaluating processors, the microbenchmarks include core layers in them for analyzing bottlenecks and guiding optimization. This study characterizes the behaviors of benchmarks from both system and microarchitecture aspects. The system metrics include I/O wait, cross-node communication and CPU utilization, the microarchitecture metrics include IPC, branch prediction, back-end resource competition and memory access. Based on the performance results, this study provides reliable advice for helping optimizing processors.

Abstract:In the bridge maintenance industry, the transition of bridge disease detection from traditional mode to digital mode has become the development trend of the industry, and the 3D visualization of bridges is the basis of the realization of digital detection. General beam bridge parameterized modeling program has the disadvantage of "once written, it is not easy to modify". In order to avoid such defects and improve the flexibility and extensibility of parametric modeling, a method combining expert system with parametric modeling of girder bridges is adopted. By combining expert system theory knowledge, beam bridge structure knowledge and parametric modeling process of beam bridge components and whole bridge, this method designs a rapid modeling expert system for beam bridge. The main contents of the system include the classification design of knowledge library, design of database table for knowledge library, the selection of knowledge expression, the reasoning content of component modeling and full bridge modeling, and the algorithm of component, assembly and full bridge modeling in parametric modeling. Finally, through the analysis of the modeling results, it is found that the method can accurately complete the parametric modeling of bridge components and the whole bridge according to a few main parameters given by users. By modifying the knowledge library, users can change the structural calculation knowledge that the parametric modeling process depends on. This method increases the flexibility and extensibility of parametric modeling.

Abstract:The efficiency of warehousing and distribution in the turnover of architectural aluminum alloy formwork in China is low. In order to improve the turnover efficiency of architectural aluminum alloy formwork, an "octopus" sorting line is designed, which combines the main line shunting and grading sorting modes. Flexsim simulation software is used to make visual simulation of the operation of the sorting line. The operation bottleneck of the sorting line is obtained by analyzing the simulation results based on the data. The bottleneck problem was solved by optimizing the process flow and adjusting the distribution of sorters. It improves the sorting efficiency of architectural aluminum alloy formwork.

Abstract:With the advancement of science and technology, measuring instrument technology is gradually developing towards virtual instrument technology. Simple, universal and expandable development has become an important indicator for virtual instrument development. In order to solve the problem that the virtual instrument development software on the market is not easy to expand and the operation efficiency is low, a method for realizing universal virtual instrument front panel running platform is proposed. The main principle is that the platform instrument component exists in the form of jar package, which is convenient for expansion and platform loading. Different instrument scripts can generate different virtual instruments, and a one-to-many observer-based data exchange mechanism is proposed to facilitate communication between components and facilitate communication between platform and hardware system. An example of a signal generation display demonstrates the feasibility of the universal operating platform.

Abstract:In order to cope with the problem of insufficient function of intelligent fishing rod in domestic market, this study designs an intelligent fishing rod system, which has the functions of fishing region selection, intelligent alarm and automatic walking fish. The study mainly solves the attitude of fishing rodby using quaternion method, adopts a data fusion algorithm based on Kalman filter, and builds the uC/OS-II operating system with the core microprocessor STM32, combined with various environmental sensors and cascade PID method. The results indicate that this system can not only reduce the probability that fish escape from the short line, but also provide the chance of automatic fishing in harsh environment.

Abstract:The rapid development of cloud storage services also brings many security challenges. The existing fuzzy identity-based data integrity auditing scheme only focuses on static data, which is obviously not suitable for many practical applications. This study proposes a fuzzy identity-based dynamic data integrity auditing scheme, which combines the dynamic data structure of Merkle hash tree to realize the complete dynamic operations of cloud data. Compared with data integrity auditing schemes based on the public key infrastructure, the scheme avoids the processes of issuing, managing, and revoking public key certificates by using fuzzy identity-based cryptosystem, and reduces the communication cost. Furthermore, the proposed scheme supports batch verification and improves authentication efficiency. Finally, the new scheme is analyzed in terms of security and function, which resists forgery attack and preserves data privacy, and has certain advantages over other schemes in terms of function.

Abstract:Innovative applications of cloud platform and related mechanism research in recent years are hot research topics in supply chain and the field of "smart". From the perspective of industry interconnected, cloud platform oriented supply chain information coordination problems are studied, and the industrial resources collaboration efficiency path is explored. This paper presents a cloud platform solution for the smart supply chain information collaboration. The scheme takes the distributed collaborative theory into account to deal with huge technological advantage of supply chain business data, builds and shares data under the distributed real-time dynamic updating algorithm combining the resource pool supply chain enterprise data connectivity cooperation model. Finally, the home furnishing industry supply chain is analyzed and demonstrated to verify the feasibility and effectiveness of cloud platform model construction.

Abstract:The layout of tower cranes directly affects the construction efficiency and construction cost in super high-rise buildings. This study combines the BIM technology with the expertise of tower crane planning, constructs the Revit parameterization family of the tower crane, defines the structural parameters and performance parameters of the tower crane family, and develops the computer aided tower crane planning system using Revit's API. Finally, the function of tower crane model management, tower crane quantity estimation, tower crane layout and auxiliary positioning has been realized. The system provides three-dimensional auxiliary tools for super high-rise construction tower crane planning, which frees designers from abstract drawing, improves efficiency, and reduces planning errors.

Abstract:Based on the requirement of intelligent driving, an APP of remote vehicle control is designed and developed. When designing and developing this APP, we use the XMPP communication protocol for mobile and vehicle communication, and use icons to present vehicle tire pressure, tire temperature, temperature inside the car, endurance mileage, and other conditions. If the car falls in fault, the alarms arise with bright red on the page. Amap API is also embedded to provide the function of the remote vehicle location and navigating to the car position. It also has the feature of remote vehicle calling, automatic car parking, and remote automatic door opening. The APP has been tested by binding with real cars, and the results show that the APP is stable and available.

Abstract:Space environmental surveillance is of great significance to space activities. A visual experiment system for the space environment surveillance is developed based on the Matlab GUI platform and then released in the website by converting into an executable program. The visual experiment system is mainly used for analyzing the space environment including the magnetic storm, solar wind and IMF conditions, ion upflow in the polar cusp, and so on. One can examine or monitor the space environment parameters or even to calculate and forecast in this system. The system is convenient to use and can improve the efficiency of monitoring or analyzing space environment.

Abstract:With the frequent loss of pets, a pet-tracking system based on Message Queuing Telemetry Transport (MQTT) protocol is designed. The function realization of the system mainly relies on MQTT server, Web server and application program based on Android platform. The embedded device is made from a WZ-203CS demo board which integrates the MTK2503 chipset. With the help of the embedded IoT card, the device can transfer information about its location to Android mobile device under the MQTT protocol. Due to integrating the Amap navigation system, the application provides voice navigation, path planning, and other functions by processing the received information about location. Thanks to the basic structure, the system's function is not merely pet-tracking, but also has reserved interfaces for future uses, making the system more practical and has strong extensible at the same time.

Abstract:The optimization of compilation options provides a feasible and effective solution to reduce the energy consumption of embedded software. GA-FP algorithm applies frequent pattern mining into the evolutionary process and has achieved better results than other algorithms. However, GA-FP still has the disadvantages, such as the large size of transaction table, incomplete and obsolete heuristic information provided by frequent option patterns and inefficient single point mutation, which potentially affects the solution quality and convergence rate. Aiming to these problems, this study proposes an evolutionary optimization algorithm for embedded software energy consumption at GCC compile time, called GA-MFPM. GA-MFPM replaces the reference point and transaction table mechanism by generation to reduce size of the transaction table. Further, a frequent compilation option mining algorithm is designed to acquire more heuristic information. It adopts a generation-by-generation mining strategy to help maintain the timeliness of frequent option patterns. Based on the frequent option patterns, a maximum frequent pattern matching algorithm is designed to perform multi-point mutation to improve optimization quality and convergence rate. The comparative experiments are done on 8 typical cases in 5 different fields between GA-MFPM and GA-FP. The experimental results indicate that the GA-MFPM can not only reduce the energy consumption of software more effectively (the average and maximal reduction ratios are 2.4% and 16.1% respectively), but also converge faster (the average of 57.6% faster and up to 97.5% faster) than GA-FP in this study.

Abstract:In order to realize high-precision synchronization of OFDM (Orthogonal Frequency Division Multiplexing) under arbitrary carrier frequency offset, this study proposes a system timing offset estimation method based on auxiliary data. Firstly, under the Gaussian white noise channel, based on the auxiliary data without special structure, the optimal synchronization algorithm is derived under the maximum likelihood criterion. Then, under the consideration that the maximum likelihood method is too complex, a sub-optimal method, which has reduced computational complexity, is proposed. Finally, the performance of the proposed timing method is evaluated by Monte Carlo simulations under the frequency selective Rayleigh fading channel. The simulation results show that the timing performance of the proposed method is significantly better than the traditional algorithms.

Abstract:In the traditional segmented data stream clustering algorithm, the inaccuracy of micro-cluster threshold radius T in the online part as well as the oversimplifying of the dealing process with the micro-cluster by the offline part leads to a low clustering quality. In order to break through such limitation, a data stream clustering algorithm on the basis of artificial bee colony optimization for offline part processing is proposed based on the existing dynamic sliding window model. This algorithm consists of two parts:(1) The online part dynamically adjusts the size of the window and improves the value of the micro-cluster threshold radius T according to the length of time that the data stays in the window so as to get micro clustering step by step. (2) The offline part uses the improved bee colony algorithm to continuously adjust dynamically to find the optimal clustering result. The experimental results show that this algorithm not only bears a high clustering quality, but also has fairly good ductility and stability.

Abstract:In this study, a method of fast estimation of stability margin based on clustering is proposed integrating the traditional causal analysis method and data core thinking method. Firstly, from the massive historical quantitative analysis or simulation calculation results, the transient stable mode is extracted according to faults, and the stable mode of all faults is taken as the key characteristic quantity. Secondly, each fault is clustered according to characteristic quantity to generate safe operation knowledge base. Finally, based on the knowledge base, each fault in the current mode is automatically matched and the stability margin is quickly estimated. This method improves the speed of analysis and calculation, provides a basis for the rapid decision-making of power network security and stability, and provides a new idea for the analysis and evaluation of power system transient stability.

Abstract:Face recognition is an important field of visual recognition. Because of the large scale of variations in face recognition, namely drastic changes in illumination and pose, occlusion problems, and complex image background, it is difficult to recognize the face under such unrestricted conditions. In order to solve these problems, a multi-Inception model based on Tensorflow platform is proposed in this study. By combining multiple Inception knots, a multi-Inception-V3 model based on Tensorflow platform is proposed. The structure is connected in series, which realizes the convolution and re-aggregation of multiple dimensions at the same time, and improves the accuracy of face recognition. The experimental results show that the proposed method can extract more discriminant face features with fewer parameters. Compared with the classification loss method and the fusion of other metric learning methods, it improves the accuracy of face recognition under unconstrained conditions.

Abstract:As a supervised learning model, traditional extreme learning machine assigns input weights and bias of nodes of hidden layer arbitrarily, and completes learning process by calculating output weights of nodes of hidden layer. Aiming at the problem that traditional extreme learning machine does not work well in prediction research, an improved extreme learning machine model based on simulated annealing algorithm was proposed. Firstly, traditional extreme learning machine method was used to learn the training set, and output weight of hidden layer is obtained. The evaluation standard of prediction result was selected to assess prediction result. Then, using the simulated annealing algorithm, input weights and bias of hidden layer of traditional extreme learning machine were regarded as the initial solution, and the evaluation standard was regarded as the objective function. The optimal solution was found in cooling process that was input weights and bias of hidden layer of extreme learning machine with the smallest prediction error. Iris classification data and Boston house price forecast data were used to conduct experiments. The experiment finds that compared with traditional extreme learning machine, extreme learning machine based on simulated annealing is better on classification and regression.

Abstract:Aiming at the low accuracy of road data information obtained by airborne LiDAR, an algorithm for extracting and segmenting road information by dynamic fitting based on low-altitude scanning three-dimensional point cloud data of UAV (Unmanned Aerial Vehicle) is proposed. Firstly, the principal component analysis algorithm is used to obtain the normal vector of road point data. Then, combining elevation information with normal vector information, the range of road elevation and normal vector is obtained by clustering algorithm, and the road point cloud data is extracted afterwards by range. Secondly, polynomial fitting is used to model the road data. Then the dynamic polynomial fitting is used to extract the data of the whole section of road surface, assets on the road, pedestrian and vehicle data. Finally, the region growth algorithm is used to segment the assets and pedestrian vehicle data on the road surface. The experiment shows that the proposed algorithm has a strong anti-interference ability to block objects on the road. It can extract the road surface and segment the data on the road surface. The proposed algorithm in this study is more sensitive to the road surface data by comparing with the region growth algorithm.

Abstract:In order to improve the security and credibility of wireless networks, a novel algorithm based on node trust value is proposed based on the random parallel cluster head selection algorithm. With the distributed strategy, the algorithm can select, identify, and delete cluster head nodes fairly and evenly. Simulation results show that the improved algorithm is more effective than the traditional algorithm in preventing malicious nodes from participating in data communication, and can make wireless network communication more secure and reliable.

Abstract:With the research boom of deep learning, vehicle target detection has gradually changed from machine learning to deep learning in recent years. At present, most of the deep learning methods have different degrees of error detection and omission in vehicle target detection. The vehicle detection method based on incremental learning dataset is proposed to solve the problems of small targets error detection and the omission of truncated and overlapping targets. This method is combined with faster R-CNN algorithm to detect and classify vehicle targets. At the end of the experiment, the influence of with/without incremental learning method on the experimental results was compared from two aspects of subjective judgment and objective test data. The result show that the vehicle detection method based on incremental learning and faster R-CNN has significantly improved the performance in subjective judgment of the missed targets. Objective data also show that the VGG16 network mAP value is increased by 4% and the ResNet101 network mAP value is increased by 6% compared with the incremental learning method.

Abstract:Based on the rapid development of intelligent transportation, this work studies the lane detection and vehicle tracking technology of high-speed sections. For multi-lane detection, the road surface is segmented by using the feature that the gray level difference between the road surface and the dividing line is rather large. Then, the line equation and the Catmull-Rom Spline interpolation algorithm are used to fit the lane dividing line. For single-lane detection, the single lane is first effectively segmented based on the HSV color space and Sobel edge extraction method, and then the lane separation coordinate points are extracted in the perspective transformation space and the segmentation line is fitted with a quadratic polynomial. Aiming at the vehicle detection, the HOG+Gentle-Adaboost classification algorithm is firstly used to detect the vehicle in front of the unmanned vehicle, and then the shadow of the vehicle is detected based on the characteristics of the shadow at the bottom to verify the authenticity of the vehicle area detected by the learning algorithm. For vehicle tracking, the dynamic second-order autoregressive model method is used to predict the state of the vehicle. For the inherent particle degradation problem of particle filtering, this study innovatively introduces the Thompson-Taylor algorithm to improve the defects of particle degradation and low diversity. The lane detection and vehicle tracking algorithms in this study can be easily transplanted on the embedded platform with high reliability and accuracy, and further to realize the lane departure warning and forward collision avoidance system.

Abstract:Traffic sign recognition equipment has low power consumption and hardware performance, while the existing convolutional neural network model has high memory footprint, slow training speed, and high computational overhead, which cannot be applied to the recognition equipment. To solve this problem, in order to reduce model storage and improve training speed, deep separation convolution and mixed wash grouping convolution are introduced and combined with the ultimate learning machine. Two lightweight convolutional neural network models are proposed:DSC-ELM model and SGC-ELM model. The proposed models use the lightweight convolutional neural network to extract the features, and then send the features to the extreme learning machine for classification, which solve the problem of slow parameter training in the full connection layer of the convolutional neural network. The new models combine the advantages of lightweight convolutional neural network model with low memory footprint, good feature extraction quality, good generalization of ELM, and fast training and classification. Experimental results show that compared with other models, the hybrid model can accomplish traffic sign recognition tasks more quickly and accurately.

Abstract:The existing image super-resolution reconstruction method based on deep learning is easy to generate pseudo texture, and the rich local feature layer information in the original low-resolution image is not fully utilized. In order to improve image quality, a super-resolution reconstruction method based on attentive generative adversarial is proposed. The generator part of the method is constructed by attention recursive network, and a dense residual block structure is also introduced in the network. First, the generator extracts the local feature layer information of the image by using the self-encoding structure to improve the resolution. Then, the image is corrected by the discriminator. Finally, the image is reconstructed into a high-resolution image. In a variety of networks for peak signal-to-noise ratio super-resolution evaluation methods, the experimental results show that the designed network exhibits stable training performance, improves the visual quality of the image, and has strong robustness.

Abstract:The current stock price forecast is a hot issue in research. People are paying more and more attention to the establishment of stock price forecasting model, and improving the accuracy of stock price forecast has practical application value for stock investors. At present, the forecasting methods of stock prices are endless, among which the typical ones are traditional technical analysis and ARMA models. In order to improve the accuracy of prediction and consider the nonlinearity of stock market, this study proposes an improved stock price forecasting model of echo state neural network. The improved particle is applied to the characteristics of Echo State Neural Network (ESN). The group algorithm (GTPSO) searches the output connection weight of the ESN, and finally obtains the optimal solution, i.e., the optimal output connection weight of the ESN. The GTPSO algorithm is generally in the traditional Particle Swarm Optimization (PSO) algorithm. Based on the idea of taboo in the Tabu Search algorithm (TS) and the idea of mutation in the Genetic Algorithm (GA), the PSO is reduced to a local minimum during the learning process, and the ability of the PSO to search globally is improved. The forecasting model is used in the daily closing price forecast of individual stocks, and the closing price of the 11th day is predicted using the closing price of every 10 days. The correctness of the model is verified by experiments, and it is proved that the model has a good prediction effect.

Abstract:The light guide plate is an important part of the backlight module of liquid crystal display screen. In the production process, defects such as bright spots, scratches, crushing, and shadows are inevitable, which directly affects the display effect. Aiming at this problem, according to the arrangement characteristics of the dots and the imaging effect of the defects, this study proposes a method based on machine vision for detecting defects of light guide plates. Firstly, according to the degree of density of the dots, the image is divided into a sparse zone and a dense zone. Secondly, because the dot in sparse zone will cause great interference to the defect detection, the idea of separating the dots is proposed, and the filter is used to separate the dots and the background. Suspected defect extraction is performed separately. The Gaussian derivative filter is designed to filter the dense area. On the basis of this, the gray area morphology and image calculation methods are used to extract the suspected defects in the dense area. Further, for different areas of the light guide plate, according to the quality inspection requirements of the production line, the characteristics of the defect area, and the screening rules are used to determine the defects. Finally, a large number of on-site tests are performed on the proposed detection method on the self-developed light guide plate defect detection system. The test results show that the detection precision for bright white spots, crushing and scratching defects has a detection accuracy of over 99%, which can basically meet the requirements of industrial testing.

Abstract:For the problem of deblurred confocal image restoration under Poisson noise, a regularization method based on Hessian matrix norm was proposed to solve the stairs effect existing in traditional methods. Based on the Poisson probability model, the method introduced the Hessian matrix norm as a regular condition, and applied the alternating direction multiplier method and gradient projection method to solve the optimization model. The quality of the reconstructed image obtained in the confocal laser scanning microscopy experiment was better than that of the traditional method. This result proves that the method can effectively restore the deblurred confocal image under Poisson noise.

Abstract:Existing systems are generally based on traditional single-application architectures. With the continuous improvement of user requirements and the continuous improvement of website functions, there are shortcomings such as monotonous application function module obfuscation, low deployment efficiency, difficulty in extension functions, and high cost of technology and iterative methods. Therefore, the microservices technology gets attentions and adopted in project development. The microservices architecture has the advantages of easy development and maintenance, partial modification and deployment, easy extension functions, and unlimited technical options. In this work, the advantages of microservices application was studied, and setup an B/S system for performance testing. An experimental scheme for thread response time, throughput, and deployment time was designed and tested using Jmeter testing tools. The test data of 20 and 50 concurrent users were analyzed. Experimental results show that microservices has greater performance advantages and higher efficiency.

Abstract:The speed and position state equation of the indoor robot positioning based on the traditional MSCKF algorithm needs to integrate the measurement data of the accelerometer in the IMU which causes the drift and cumulative errors, and its accelerometer is always interfered by gravity. Aiming at this problem, this study proposes an improved MSCKF algorithm. Under the premise of not using the accelerometer sensors, the improved MSCKF utilizes the advantages of wheel odometer sensors which measure the amount of translation more accurately, fuses the data of the wheeled odometer with the data of the gyroscope in the IMU, and improves the state equation of Extended Kalman Filter (EKF) for MSCKF algorithm. First, the improve posture equation of the EKF is obtained by using the angular velocity data of the gyro sensor. Then, after combining the translation data of the wheel odometer sensor with the rotation information of the posture equation, the improve velocity and position equation of the EKF are obtained. Finally, the MSCKF and its improved algorithm are implemented on the Robot Operating System (ROS), and verified in an indoor scene with the Turtlebot2 robot. The experimental results show that the improved MSCKF algorithms' motion trajectory is closer to the real trajectory, and its positioning accuracy is also improved. Compared to the average closed-loop error which is 0.429 m, its average closed-loop error is 0.348 m.

Abstract:In order to reduce the negative impact of excessive grinding temperature on the thermal damage of parts, and to improve the yield and quality of parts, this study establishes a surface grinding temperature prediction model based on convolutional neural network. Firstly, the temperature data is obtained through finite element simulation, and pre-processing is performed. Then, the convolutional neural network program is written by Google's open-end learning tool TensorFlow, and finally the prediction result is obtained and compared with the simulation value. The results show that the grinding temperature prediction model based on convolutional neural network has strong learning ability and nonlinear fitting ability, which greatly improves the prediction accuracy of grinding temperature.

Abstract:It is an important communication way for webcast video watchers to produce and consume time-sync comments, which can be beneficial to understand the webcast video users. Based on data related to time-sync comment collected from 3 hot live streaming platforms (Douyu, Panda and Zhanqi), a hypothesis testing based method is proposed to analyze webcast video watchers from user attribute and user behavior, a user activity model is constructed based on user behavior feature time series analysis. Research results show that, the number of live streaming platform online users has obvious characteristics of periodic changes, source of live streaming platform online users tends to be distributed in inshore developed cities, and the proposed user activity model can effectively predict activity of users in live streaming platforms.

Abstract:Unmanned Aerial Vehicle (UAV) inspection has become an important way of transmission line inspection. However, most UAV inspections are based on post-processing and cannot achieve real-time detection, which is mainly limited by the performance of the front-end equipment. A real-time detection method is proposed to detect construction vehicles under the transmission lines by UAV based on Android using neural network. Real-time object detection of the construction vehicle is realized by collecting the data of construction vehicles obtained by UAV, using data enhancement methods, and integrating SSD-MobileNet algorithm model into the Android platform.

Abstract:Corrosion detection of power equipment is a very important part of power system malfunction detection and needs to be quickly and accurately identified. This study proposes an algorithm of power equipment corrosion object detection based on attention model, which can effectively detect the rust area of power equipment. The proposed algorithm model uses the depthwise separable convolution instead of the standard convolution to compress the model greatly. Based on this, an upsampling feature fusion strategy based on the attention model is proposed to compensate for the loss of precision caused by the reduced model structure. Compared with the standard SSD on the RustDetection dataset, the proposed algorithm can improve the accuracy of 10.47% and the average accuracy of 5.99% when the parameter quantity is reduced by 63.6% and the speed is increased by 46.7%.