Abstract:Speech recognition, which makes the voice readable and enables the computer to understand and respond to human language, is one of the key technologies for human-computer interaction in artificial intelligence. This study introduces the development of speech recognition, expounds the principles, concepts, and basic framework of speech recognition, and analyzes the research hotspots and difficulties in the related field. Finally, it summarizes the speech recognition technologies and presents an outlook on future research into this field.

Abstract:With the development of science and technology, the application of network-linked data in statistical learning, machine learning and other fields becomes increasingly common. In linear regression models, the current research on the variable selection of network-linked data mainly focuses on the homogeneous samples, namely that the individual effects of the samples are the same. In reality, however, the individual effects of most samples are heterogeneous. As a result, the neglect of the heterogeneity will lead to large deviations in the estimation and prediction of the models. Therefore, this paper proposes a new variable selection method SNC to cope with the situation when there is group heterogeneity in network-linked data. Using the network agglomeration effect, we carry out a joint penalty for the difference between the variable coefficient and the individual effect of the connected samples and solve the problem with ADMM algorithm, with the convergence of the algorithm proved. The results of numerical simulation and example analysis show that this method improves the accuracy of variable selection and reduces the prediction error.

Abstract:To solve the problem of low coherence in the interference superposition of the coherence map generated by the traditional single main image, this study proposes a weighted superposition method based on the differential interferometric synthetic aperture radar (D-InSAR) with multiple main images. This method uses fewer SAR images to generate interferograms, and relies on the weighted superposition method to avoid the involvement of coherence maps with low coherence in the superposition, so as to achieve high-accuracy and long-time deformation monitoring in a certain area. Firstly, multiple main images are selected to generate multiple interferogram sets. Secondly, the multiple sets of coherence maps are superposed according to the weight. Finally, the average velocity maps of surface deformation in different time periods are obtained. In this study, Sentinel-1A data of 9 scenarios from July 2019 to December 2019 were selected from San Diego City in the United States for experimental verification, and the experimental accuracy was verified by the comparison of the experimental results with groundwater monitoring data. The experimental results show that this method not only reduces the number of images required for the experiment, but also improves the quality of the interferograms involved in the superposition. It produces better monitoring results than the traditional D-InSAR in improving the monitoring accuracy.

Abstract:To address the problem that the public opinion data collection on food safety is not fast and accurate enough in the era of big data, this study proposes a public opinion monitoring probe on food safety based on the Bayesian network. Firstly, the MySQL database is used to establish a food safety keyword database. Secondly, the Bayesian network model is adopted to build a monitoring probe with the keyword database, and the public opinion monitoring system of the “Zhongyun Big Data” of PeopleYun is chosen for data collection. Thirdly, the monitoring probe is compared with traditional data collection technologies on public opinions and Web crawler technologies in three groups of comparative experiments (milk, wine, and tea) to verify its effectiveness. The results show that the data mining time of the three groups of experiments (milk: 3 s; alcohol: 2.5 s; tea: 2.4 s) is significantly reduced, and the data efficiency (milk: 83.6%, alcohol: 77%, tea: 77.9%) is considerably enhanced. Therefore, introducing a keyword database into the bayesian network model to form a monitoring probe can effectively improve the timeliness and accuracy of public opinion data collection on food safety.

Abstract:The sorting and restoring of source packets of satellite payload data is an important part of a ground data processing flow, and its correctness and completeness directly affect data-based scientific research. This study analyzes the abnormal data and challenges to source packet sorting in ground data processing brought by the satellite data organizations and satellite-ground data transmission mechanisms under the current international standard—the consultative committee for space data systems (CCSDS) standard. On this basis, it proposes a “multiple indexes and timestamp correction” packet sorting method, which not only solves the difficulties in duplicate data elimination and data registration of ground source packets caused by the frequent acquisition of data about space science satellites but also gives correct results in face of key index abnormalities such as abnormalities in the frame serial number, packet serial number, and timestamp. This method has effectively promoted the accuracy of source packet sorting and data processing efficiency of China’s space science missions.

Abstract:A brief case is a brief description of a case record made by a public security organ to improve the quality of information input in the Collaborative Case Handling System and ensure efficient information retrieval and joint investigation. A large amount of case information related to the victim and the perpetrator is between various entities. Therefore, in-depth excavation of brief case texts is an effective means to grasp the beginning and end of a case and to analyze the case. The dense distribution, inter-nesting, and abbreviation of entities in a brief case text bring great challenges to the accurate capture of the case entities. In response to the particularity and complexity of brief case texts, this study improves the method of character vector generation and proposes a Roberta-CNN-BiLSTM-CRF (RC-BiLSTM-CRF) network architecture. Compared with the mainstream Bert-BiLSTM-CRF architecture, this architecture can extract the character vector features, thereby solving the problem of a lengthy character vector brought by model pre-training. The model parameter number is reduced for a higher overall parameter convergence rate. In the comparative experiment, five mainstream architectures are selected and compared on the brief case dataset provided by the public security organs of Hunan Province. The method proposed in this study is proved to be the best in terms of accuracy, recall rate, and F1 value, and its F1 value reaches 88.02%.

Abstract:Curves are an important component of road traffic scenes. In road information reconstruction through visual information, it is difficult for the traditional world coordinate system built by the surveillance camera to represent the real road spatial information and vehicle position information in a curve scene. To solve this problem, this study proposes a mileage coordinate system based on road alignment. In this system, the horizontal direction represents distance information along the road section, and the vertical direction represents mileage information along the road alignment. For the construction of the mileage coordinate system, camera calibration and proposed result optimization are firstly carried out through the calibration algorithm based on a single vanishing point and the road prior information to obtain the real spatial positions of the lane lines or road edges. Then, with the real spatial information of lane lines or road edges in the world coordinate system, polynomial fitting is performed to obtain the fitted road alignment curve. Finally, the road identification points or vehicle track points are projected onto the fitted curve to acquire the distance information based on the road section direction and the distance information in the road direction. This scheme is tested in a simulated experimental curve scene and an actual highway curve scene. The experimental results show that the average position errors of the proposed mileage coordinate system are both less than 5% in the simulated curve scene and the actual scene. Compared with the traditional straight-line world coordinate system, this system, with favorable adaptability and high precision, can meet actual requirements.

Abstract:After a detailed analysis of the existing anti-intrusion system of gantry cranes, an anti-intrusion monitoring model based on machine vision and deep learning is built in view of the complex working environments and low intelligence levels of the system. In consideration of the advantages and disadvantages of each target detection algorithm and semantic segmentation algorithm, the semantic segmentation algorithm is adopted as the anti-intrusion model, and the ICNet is used as the main semantic segmentation network. Compared with other networks, ICNet displays the best accuracy, with a training accuracy of 99.37% and a training loss of 1.81%. The results prove the intelligence and feasibility of the anti-intrusion system based on semantic segmentation.

Abstract:To improve the abilities of meteorological archives management and social utilization, this paper proposes a method to construct the knowledge graph of meteorological records archives. With the meteorological records archives resources in Anhui Province, the ontology database is constructed by the design of conceptual layer and hierarchical relationship in a top-down way. A knowledge model is constructed based on the ontology database docking entity. Knowledge is extracted from archives in different storage formats with D2RQ, image recognition and custom wrapper for data layer construction. A synonym database is established to facilitate knowledge fusion. MySQL and Neo4j databases are employed for the storage of ontology database and entity database respectively, and an example demonstrates the knowledge graph construction of meteorological records archives. The knowledge graph construction in this research enriches the knowledge units and semantic relations contained in meteorological records archives, mainly including 28 core concepts, 280 general concepts, 6 ontological relationship models, and more than 10 000 volumes of archives entities. It can provide a reference for knowledge graph construction across industries of meteorology and archives.

Abstract:The steam injection thermal recovery technology has often been applied to the exploitation of heavy oil. In actual production, the heat dissipation loss is serious because the steam injection pipelines are very long and in the open air. This has greatly influenced the benefit of oilfield exploitation. Optimizing thermal insulation materials and thermal insulation structures can effectively reduce heat dissipation loss, but the calculation of related data is very complicated. To improve the efficiency of information collection and processing and to increase the speed and accuracy of the calculation, this study introduces information technology to the analysis project of the thermal insulation benefit of heavy-oil steam-injection pipelines. Through constructing an evaluation model of the thermal insulation benefit of heavy-oil steam-injection pipelines, we identify the main parameters affecting the evaluation of the thermal insulation benefit and build an evaluation system of the thermal insulation benefit of heavy-oil steam-injection pipelines. This system can automatically calculate heat loss and carry out visual analyses according to the actual operating environment and loading condition of a steam injection pipeline. At present, this system has been applied in the Xinjiang Oilfield. The calculation results of the system can benefit the accurate and fast optimization of thermal insulation materials and thermal insulation structures, thereby effectively improving the benefit of oilfield exploitation.

Abstract:In today’s society, Agile has gradually become the mainstream development mode in the industry, and more and more organizations have successfully realized agile transformation and made great achievements in improving research and development (R&D) efficiency and customer value. Agile has extended forward from the traditional pure R&D field to business agility and backward to the integration of DevOps development, operation, and maintenance. Therefore, our company also decided to develop agile transformation so that we could respond quickly and meet users’ needs, shorten the system construction cycle, avoid development waste caused by business changes, and ensure product delivery quality. Specifically, the Method 2.0 system and human resource 2.0 system are selected as pilot projects for agile iterative development. With users at the center of the initiative innovation and trial-and-error process, the systemic agile iteration is pushed forward with small quick runs, and the systemic construction mode is optimized with innovation. Two goals are expected. One is to achieve fast and high-quality project delivery through agile transformation, and the other is to summarize the transformation experience and sort out the system, standards, the process, and relevant supporting tools applicable to State Grid Agile projects. This paper introduces how to improve the software quality under the agile development mode from three aspects: working ideas, the practice process and methods, and the construction effect.

Abstract:To address the problem of low accuracy and poor real-time performance of the facial attractiveness evaluation system, we propose a new facial attractiveness evaluation system based on deep learning. In this system, the HOG feature-based method and the FaceNet pre-training model are used for face detection and facial feature extraction respectively. Furthermore, a two-layer decision model based on the Softmax classification layer and ReLU regression layer is proposed, which is combined with the quantized values of local facial features to evaluate the facial attractiveness. Experimental results on the SCUT-FBP5500 dataset show that the accuracy of the system is 78.58%, and the average evaluation time of a single image is 2.98 seconds, which can meet the needs of practical applications.

Abstract:As the pace of digital oilfield construction continues to accelerate, the application of cloud-based intelligent video monitoring in oilfields has become a hotspot. The existing video monitoring system has the defects of difficult client deployment and poor system expandability, and the intelligent detection mainly relies on passive video monitoring to cope with the potential safety hazards at the operation sites. In addition, the monitoring log data cannot be effectively archived for management and comprehensive analysis. To meet the actual needs of oilfields, this study firstly designs the overall architecture of the cloud-based intelligent monitoring system, with the shortcomings of the existing monitoring architecture taken into consideration. Then, risk factors at the operation sites are detected and analyzed by intelligent video detection means. Finally, the hazard logs and other data generated by the system are managed and analyzed to provide decision support for site safety management. At present, the system has been deployed to an oilfield and has successfully passed the function and performance tests. It is of great significance for the improvement in effectiveness and real-time performance of monitoring, the enhancement of ability to deal with emergency events, and the safe operation and production in oilfields.

Abstract:At present, extensive offline management is still common in China’s textile equipment industry. Traditional mechanical automation equipment fails to adapt to the various changes and development of the market and consumes enormous labor costs. To tackle the problems in the textile equipment industry, this study develops a double twister monitoring system based on the Internet of things with some technologies such as the Internet of things and cloud computing, taking new industrialization as the goal. It also puts forward a remote data acquisition and transmission scheme with MQTT server and intelligent module under cloud at the core and realizes the storage, remote monitoring, and control of data from double twisters and other textile equipment with the help of a database and the Web server technology. The intelligent module under cloud adopts plug-and-play and self-learning station number modes to improve the convenience of sensor networking, avoiding plenty of routing difficulties. Practice test results show that the system can achieve the target functions and enhance the informatization level of workshops in textile enterprises. The cost of raw materials is reduced by about 2.1%, and the work efficiency of workers is increased by 25%.

Abstract:To intuitively analyze the variation of tritium diffusion data in the tritium safety containment system and realize the three-dimensional visualization of the calculated tritium diffusion data in space, this study developes a three-dimensional virtual simulation system for tritium safety containment based on a Unity3D engine. This system carries out a three-dimensional simulation of the tritium transport, leakage, and diffusion process with a particle system driven by calculated data. The tritium transport, leakage, and diffusion process in the tokamak exhaust processing (TEP) system is simulated, with the TEP system of the China Fusion Engineering Testing Reactor (CFETR) as an example. Visual analysis is carried out on key positions of infiltration and leakage, verifying the security of the tritium safety containment system. This study serves as a reference for future research in tritium safety containment systems and tritium infiltration and leakage.

Abstract:This study proposes a general evaluation scheme based on microservice for the analysis and evaluation of system log information such as system operation log and command-line text output. The framework of this scheme adopts a microservice architecture and a model of one single master node and multiple worker nodes. Each node has the same ability. The node role, namely, whether the node is a master or worker one, is configured on demand. The number of worker nodes can be flexibly expanded to support the big data analysis ability and the TB-level system log analysis and evaluation. Compared with the popular open-source technologies such as elasticsearch logstash kibana (ELK), this scheme not only has similar functions of log extraction and analysis and chart presentation but also designs and implements a general evaluation model of system logs based on the configuration driver. The model, adopting the principle of separating rules from the engine, is suitable for processing many kinds of system log information. Under the proposed evaluation scheme, a PON network health examination system is built with the Spring Boot/Spring Cloud microservice framework. The system has been applied in many PON network evaluation services around the world and has greatly improved service efficiency.

Abstract:With the advancement of technology, the systems of various parts of the picking robots have been increasingly improved. The design of the visual positioning system largely affects the work efficiency of a picking robot, especially its target detection speed, fruit picking accuracy, and target picking environment adaptation. In this study, we propose to use a binocular stereo vision system to acquire images of camellia oleifera fruit targets and then collect and calculate depth information to build our own VOC dataset of Camellia oleifera fruits. The you only look once v3 (YOLOv3) target detection algorithm is adopted to achieve Camellia oleifera fruit recognition in complex environments. The function of locating Camellia oleifera fruit targets is intuitively demonstrated by a newly designed upper computer interface. Experimental results show that compared with other methods, the proposed method has a higher recognition rate and a faster recognition speed, which demonstrates the superiority of its algorithm in complex environments.

Abstract:On-orbit anomaly diagnosis of satellites and payloads is an essential support for the efficient and safe operations of satellites. Intelligent and efficient methods of satellite anomaly detection are one of the focuses of research in satellite ground systems. Under the background of the satellite missions of China’s Strategic Priority Program on Space Science, this study proposes an intelligent anomaly detection method of satellite engineering parameters based on the data characteristics and data anomaly forms of the space-science satellites and the gradient boosting decision tree (GBDT). The engineering data of the Quantum Science Experimental Satellite “Micius” are employed for application verification and analysis. Compared with the original “threshold + regular expression” anomaly detection method, the proposed method has an average accuracy of over 98%, with an increase of about two percentage points. False negatives and false positives can be effectively reduced, and the detection speed is increased by about six times.

Abstract:In recent years, deep reinforcement learning has achieved great success in many sequential decision-making problems, which makes it possible to provide effective and optimized decision-making strategies for complex and high-dimensional multi-agent systems. However, in complex multi-agent scenarios, the existing multi-agent deep reinforcement learning algorithm has a low continuous convergence speed, and the stability of the algorithm cannot be guaranteed. Herein, we propose a new multi-agent deep reinforcement learning algorithm, which is called multi-agent distributed distributional deep deterministic policy gradient (MA-D4PG). We adapt the idea of value distribution to multi-agent scenarios and retain the complete distribution information of expected return, so that agents can obtain a more stable and effective learning signal. We also introduce a multi-step return to improve the stability of the algorithm. In addition, we use a distributed data generation framework to decouple empirical data generation and network update for the purpose of taking full advantage of computing resources to speed up the convergence. Experiments show that the proposed method has better stability and a higher convergence speed in multiple continuous/discrete controlled multi-agent scenarios and the decision-making ability of agents has also been significantly enhanced.

Abstract:The prediction and accurate warning of CDN bandwidth outliers have always been the focus and difficulty of network operation. For this reason, the study proposes and implements a new algorithm framework, the serial LSTM (long short-term memory) network with locally weighted regression, based on the LSTM network with time series. The framework uses the time-series interpolation sampling method to construct the data set, and the local weighting algorithm is integrated into the fitting model based on least square regression for initial prediction. The prediction result is serialized with the LSTM time series model for the final bandwidth outlier prediction. The 4sigma method is used to determine whether the bandwidth is abnormal at a certain moment, and an abnormal alarm is issued according to the grade standard. The experimental results show that the model is effective for the prediction and alarm of bandwidth outliers.

Abstract:Airliner stowage is a key step in aircraft service. With the industrial Internet development and the higher requirements of the aviation industry for its own service, the efficiency of manual stowage operations has decreased, and the airport operation businesses will inevitably develop towards automation and intelligence, so will the digitalization of airliner stowage. Some of the existing algorithms cannot adapt to the multi-constraint and variable environment of airliner stowage and rely on the manual adjustments by ground staff. Therefore, the structural analysis of the constraints involved in the stowage environment and the stowage process is carried out in this study, and the constraint compensation dynamic scoring solution is proposed based on the genetic algorithm. The test results show that different stowage results can be obtained according to different constraints, so as to optimize the center of gravity shifts and real-time stowage calculation. On the basis of this, the scalability of the system structure is ensured, so the system can adapt to various new models and new requirements.

Abstract:Unmanned aerial vehicle (UAV) based on visual positioning is prone to drift in a weak texture environment. To solve the problem, this study proposes a method integrating the improved Snake algorithm and ORB (oriented FAST and rotated BRIEF) feature optical flow to estimate the operating parameters of UAV. Firstly, Gaussian filtering is used to denoise the collected video frames. Then, the gray level co-occurrence matrix of video frames is calculated, and whether they are all in the weak texture area is judged. If they are all weak texture parts, ORB optical flow algorithm is used to estimate the parameters. If there is a non-weak texture area, the center of gravity of the area is calculated by the improved Snake model, by which the drift is estimated. With the homography matrix, the rotation component is obtained by decomposition. Laboratory experiments verify that the accuracy of parameter estimation is higher than 95% and the average processing time is as short as 0.025 s in both weak and non-weak texture areas.

Abstract:To address the problem that the sampling strategies of existing open-source distributed tracing systems or frameworks collect redundant traces of normal executions that are less helpful to tasks such as fault analysis, we propose a dynamic sampling strategy. With two data structures of sampling strategy tree and execution trace graph, it realizes the automatic adjustment of trace sampling rate and finds the way to quickly and accurately determine services that need to adjust the trace sampling rate. The collaboration between the above data structures enhances the trace proportion of anomalous executions. The experimental results show that the proposed method effectively improves the trace proportion of anomalous executions in any time period.

Abstract:With the innovation and progress of Internet technologies, the application of three-dimensional (3D) simulation technologies in games and animations has become increasingly important. In response to the lack of detail and the distortion of the mountains generated by the random midpoint displacement method, this paper proposes a method of generating simulated 3D terrain environments based on the OpenGL technology and the Perlin noise algorithm. The elevation maps generated by the traditional Perlin noise algorithm are relatively flat. Therefore, they are optimized with the fractal and turbulence algorithms, which produce obvious conflicts that are more in line with the characteristics of the mountains. For the transitional fault phenomena in the process of terrain image mapping, a layered sampling strategy is adopted to make the image more natural. In the case of added light, the plane normal vector is obtained through the multiple binary averaging algorithm, which delivers a smoother mountain light representation. The experimental results show that the method not only realizes the 3D grid construction of 3D terrain but also shows a favorable simulation effect.

Abstract:In the process of implementing recommendations, the browsing order of users is important information for the recommendation algorithm. The same user’s different preferences for items at different times also affect the recommendation results. Under the framework of the neural collaborative filtering model, this study proposes to integrate long short-term memory networks with generalized matrix factorization and capture both the user’s long-term and short-term preferences. The new model utilizes the strong fitting ability of long short-term memory networks to time series data to learn the user’s short-term preference and capture the long dependence relationship of the sequence. The user’s long-term preference is learned through generalized matrix factorization. The recommendation algorithm is thereby optimized, and the recommendation performance is improved. Experiments are carried out on the Movielens-1M dataset and the results show that the new model has a higher convergence rate and better recommendation performance.

Abstract:Kubernetes is a popular open-source container orchestration engine. Its default scheduling algorithm only considers CPU and memory and uses unified weight to calculate the score of candidate nodes, which cannot meet the requirements of different Pod applications. In view of this, the paper expands the Kubernetes performance indexes, with bandwidth, disk capacity, and IO rate added. The subjective weight is calculated by the analytic hierarchy process (AHP) and the objective weight of resource indexes is calculated by the entropy weight (EW) method in real time according to the resource utilization rate of performance indexes of nodes in the Pod application deployment process. We combine the two weights and apply them to a multi-attribute decision algorithm based on the improved technique for order preference by similarity to an ideal solution (TOPSIS) to select appropriate candidate nodes. The experiment results show that with the increase in the deployed Pod number, the standard deviation of the integrated load increases by 18% compared with that of the Kubernetes default scheduling algorithm under the condition of a large cluster load.

Abstract:The deep learning-based algorithms of action recognition are often difficult to achieve fast performance and high accuracy due to the complexity of neural networks. In view of this, we modularize the existing temporal shift and split attention module as an end-to-end trainable block which can be easily plugged into the classical two-stream action recognition pipeline. In the RGB and optical flow branch network, we adopt a random sampling strategy with sparse temporal grouping to realize long-term modeling. Furthermore, we use the Temporal Shift module to replace some channels in the time dimension so as to enhance the sequential characterization ability with information of adjacent frames. In addition, the Split Attention module integrating multi-paths and feature map attention mechanism improves the recognition performance of the network. Experiments show that our method achieves appealing performance on two public benchmark datasets including UCF101 (recognition accuracy of 95.00%) and HMDB51 (recognition accuracy of 72.55%), demonstrating its effectiveness.

Abstract:The graph convolutional network (GCN) is a very important method of processing graph-structured data. The latest research shows that it is highly vulnerable to adversarial attacks, that is, modifying a small amount of data can significantly affect its result. Among all the adversarial attacks on a GCN, there is a special attack method—the universal adversarial attack. This attack can produce disturbances to all samples and cause an erroneous GCN result. This study mainly studies targeted universal adversarial attacks and proposes a GTUA by adding gradient selection to the existing algorithm TUA. The experimental results of three popular datasets show that only in a few classes, the method proposed in this study has the same results as those of the existing methods. In most classes, the method proposed in this study is superior to the existing ones. The average attack success rate (ASR) is improved by 1.7%.

Abstract:The major characteristics of vehicles for hazardous chemicals transportation are the danger sign on the roof and the dangerous goods sign beside the license plate, which are difficult to detect for most object detection algorithms. To improve the detection accuracy and enhance the detection speed, this study proposes a novel detection algorithm for these vehicles based on the residual network (ResNet) and bidirectional feature pyramid network. A data set of vehicles for hazardous chemicals transportation is first made by the interception of the highway surveillance video, and then feature extraction is performed with the ResNet. In this novel model, the recurrent residual module is used to replace the middle convolution layer of the residual block. Then the bidirectional feature pyramid network is employed for feature fusion. Finally, the prediction results are obtained with the prediction network. Performance verification is carried out on the test set, and the results show that the indicators of the proposed model are superior to those of other networks overall. It has the detection accuracy up to 0.961 and the frames per second (FPS) of 43.5, showing a good industrial application prospect.

Abstract:To tackle the problem of poor recognition accuracy caused by large changes of crowd target feature in a high-density scenario, this study proposes two kinds of multi-scale feature fusion structures: attention-weighted fusion module (AWF) and bottom-up fusion module (BUF). The AWF module uses the attention branch to learn the weights of feature maps, and the weighted multi-scale features are superposed finally. The BUF module uses dilated convolution to obtain more scale information during feature processing, and the shallow feature maps are merged by stitching. The processed feature map has stronger expressive ability, and the predicted density map is more accurate. Taking ResNet50 as the backbone network for feature extraction, the algorithm presented in this study uses AWF and BUF modules for feature fusion respectively, and experiments are conducted on public datasets. The results show that the crowd counting algorithm with the AWF module can reduce the mean absolute error (MAE) to 45.54 (part A) and 7.6 (part B) and the mean square error (MSE) to 100.28 (part A) and 11.4 (part B) on the Shanghai Tech dataset. On the UCF_CC_50 dataset, the MAE and MSE are decreased to 212.42 and 323.06, respectively. Regarding the algorithm with the BUF module, the MAE is reduced to 51.6 (part A) and 8.0 (part B), and the MSE is decreased to 102 (part A) and 12.8 (part B) on the Shanghai Tech dataset. On the UCF_CC_50 dataset, the MAE and MSE are decreased to 242.6 and 359.5, respectively. Experiments indicate that the AWF module and BUF module can both effectively integrate deep and shallow feature information, thus able to optimize feature maps and improve counting accuracy.

Abstract:The existing regional intrusion detection methods for oilfield operation are exposed to problems of poor detection efficiency and failure in real-time detection due to the long distance and small targets. Therefore, this study proposes an offshore intrusion detection model to improve the safety supervision efficiency and ensure the accuracy and timeliness of intrusion detection in the offshore perimeter of oilfield operation sites. This model is based on the combination of SOLOv2 and CenterNet. First, the model uses SOLOv2 to segment the perimeter of the offshore area and identifies the danger area from these segmentation results. Then, CenterNet is adopted to detect and locate the operator and calculate the location of the operator and the danger area to realize the intrusion detection in the perimeter of the offshore area. The experimental results prove that this method can effectively solve the intrusion detection problems in offshore perimeter areas including complex backgrounds, high detection accuracy requirements, and small targets, and the accuracy of the model can reach 94.7%. This model has been successfully implemented in the oilfield operation sites with good performance.

Abstract:To alleviate the class imbalance problem of software defect prediction and avoid the influence of overfitting on the accuracy of the defect prediction model, this study proposes an oversampling method for software defect prediction based on heterogeneous distance ranking (HDR). First, a minority of instances are distinguished by three classes to remove noise instances and reduce overfitting caused by noise data. Then, instances are ranked based on heterogeneous distances and paired with highly similar ones to generate new instances for the improvement of new instance diversity. Valuable minority instances that were deleted are restored afterward. The experiment compares the HDR algorithm with the SMOTE and the Borderline-SMOTE algorithms, and the RF classifier is used on the eight actual project data sets of NASA. The results show that there are 7.7% and 10.6% performance improvements on the F1-measure and G-Mean indicators respectively. Experimental results show that the HDR algorithm is significantly better than other algorithms in processing software defect prediction data sets with large data volumes and high imbalance rates.

Abstract:In the construction and training stage of the machine translation model, the maximum likelihood estimation principle used in the end-to-end machine translation framework training can lead to the low quality of the translation model. To alleviate the problem, this study uses the adversarial learning strategy to train generative adversarial networks and improves the translation quality of the generator with the assistance of discriminators. Through experiments, the machine translation framework, Transformer, is chosen for its better performance with generators, and the convolution neural network for its better performance with discriminators. The experimental results verify that adversarial training can improve the naturalness, fluency, and accuracy of the translation. In the model optimization stage, the Mongolian-Chinese machine translation quality is still unsatisfactory due to the lack of Mongolian-Chinese parallel data sets. For improvement, the dual-generative adversarial networks (Dual-GAN) algorithm is introduced to the Mongolian-Chinese machine translation. Through the effective use of a large number of Mongolian-Chinese monolingual data, the dual learning strategy is adopted to further improve the quality of the Mongolian-Chinese machine translation model based on adversarial learning.

Abstract:Traditional facial expression recognition (FER) methods have focused on the six basic facial expressions. However, compound facial expressions are also used by humans in the real world. Compound facial expression means that it is a combination of the basic facial expressions. However, the traditional methods of recognizing basic facial expressions are unable to handle compound facial expressions. Moreover, the compound facial expression datasets have insufficient training data. To address the difficulties in the compound FER, this study proposed a graph convolutional network in multi-label learning for compound facial expression recognition (GCN-ML-CFER). The global features of facial expression and the local features of the regions of interest were extracted by the feature extraction network. According to the prior knowledge of basic and compound facial expressions, a relationship graph of facial expression categories was constructed by a data-driven method. The expression category classifiers learn the graph via a graph convolutional network (GCN). Finally, compound FER was carried out by the classifiers. Experiments were conducted on the RAF-DB and EmotioNet datasets. The results show that this method achieves a 4%–5% increase in the compound FER accuracy compared with those of the VGG19 and ResNet50 methods.

Abstract:In the data transmission process of the Internet of Things, it is necessary to authenticate the identities of the communicating parties and encrypt the transmitted data. At present, a large number of authentication and key agreement schemes have been designed, but they are vulnerable to various attacks, such as smart card stolen attack and denial of service attack. In view of the problems of existing schemes, this study proposes a lightweight one-to-many authentication and key agreement scheme, using elliptic curve cryptography and XOR operation on the user and the sensor sides respectively to achieve mutual authentication and the method of a pre-shared key to expand the sensor side to multiple ones. Finally, through the comparison of functions, computation cost, and communication cost, it is shown that the proposed scheme is better than other schemes and is more suitable for multi-sensor scenarios.

Abstract:With the continuous increase in software scale, software security faces increasingly severe challenges. As an effective means of detecting software system security, formal proof aims to use mathematical methods to complete rigorous verification of software attributes. Commonly used formal proof methods prove theorems with pattern matching, which, however, suffer from defects such as incomplete strategy generation. This study proposes a command prediction framework based on the attention mechanism. It combines long short-term memory (LSTM) with Coq to predict the strategies and parameters during theorem proving. The experimental results show that the model proposed in this study is superior to existing ones in the accuracy of command generation (the accuracy of command prediction is 28.31% in this paper).

Abstract:To address the crowd evacuation problem at Jinji Lake city square in Suzhou under emergencies, this study builds a real-time dynamic planning model of evacuation network paths and analyzes the influence of complex environments in a large public area on crowd evacuation efficiency. Meanwhile, the Dijkstra algorithm is improved with the time of the crowd escaping from the dangerous area as the weight and the number of people evacuating from each exit is rationally set by the feedback compensation mechanism to realize the dynamic adjustment and path planning of crowd evacuation. The improved algorithm is verified by Pathfinder simulation, which shows that planning the crowd to escape from specific exits in advance is conducive to achieving better crowd evacuation and thus ensuring the safety of life and property under emergencies.

Abstract:Cryptographic technology is the foundation of cloud computing security. The high-performance cryptographic cards supporting SR-IOV virtualization technology are suitable for cloud cipher machines, which can realize the encryption protection of virtualization data for cloud computing environments and meet the security requirements. However, these cryptographic cards have unsatisfactory compatibility, limited expansibility, poor migration, and low cost performance when applied in cloud cipher machines. Thus, this study proposes a software virtualization method of cryptographic cards based on an I/O front-end and back-end model. With shared memory or virtio as the communication mode, it completes the efficient communication between multiple virtual machines and the host by designing the front-end and back-end driver or service program of cryptographic cards and realizes that common cryptographic cards can be shared by multiple virtual machines. This method can effectively lower the hardware threshold of cloud cipher machines and makes cryptographic cards possess good compatibility and expansibility and high performance, thus showing broad application prospects in information technology applications and innovation.

Abstract:In the bottom layer of Agent2D, single defensive and offensive methods often fail to complete the offensive and defensive tasks required by a game. In response to this problem, this study proposes a multiplayer defensive strategy and a triangle offensive strategy. The multiplayer defensive strategy is when the opponent is on the offense, we send two or more teammates to force the opponent’s team player with the ball. The triangle offensive strategy is realized through the running cooperation of three teammates in suitable positions. A core offensive player and two auxiliary offensive players form a triangle offensive team. The two strategies are applied to teams and experimented on the RoboCup2D simulation platform. The experimental results verify the effectiveness of the two strategies, with the teams using the two strategies all having improved winning rates and defensive and offensive efficiencies.

Abstract:In this study, air conditioning control software is designed with neural network technology, and the traditional manual control mode and neural network controller are compared. First, Energy Plus is used to build a real high-speed railway station building and its multi-connected air conditioning system, with 424 working conditions of the air conditioning system set up to complete the operation simulation for a whole year. Then the neural network controller is trained with data having excellent predicted mean vote (PMV)-based thermal comfort and energy consumption which are extracted from millions of simulation data. Finally, the prototype system of air conditioning control software for the high-speed railway station is developed with Java Enterprise Edition (JavaEE), and the dynamic control of air conditioners is realized by using Energy Plus simulation data and simulation with a machine learning prediction model. The simulation results based on this prototype software system show that the intelligent controller can reduce energy consumption in comparison with manual control based on fixed settings under typical working conditions in winter and summer.

Abstract:Effective tool life prediction holds important research value in that it can improve the machining efficiency and ensure the machining accuracy of a workpiece. However, accurate tool life prediction is difficult to achieve as it is influenced by many factors such as tool material, cutting parameters, and machining material. So we propose a method of tool life prediction based on a radial basis function (RBF) neural network optimized by the particle swarm optimization (PSO) algorithm. Firstly, the main parameters of the RBF neural network, namely the center value c, width σ, and connection weight w, are optimized by the PSO algorithm. Then, tool life prediction is carried out, with the factors affecting tool life as input neurons of the PSO-RBF neural network model and tool life as the output neuron. The experimental results show that the proposed method of tool life prediction based on the PSO-RBF neural network is feasible, with an average relative error reduced by 17.14% from that of the standard RBF neural network to 6.16%.

Abstract:How to accurately and efficiently forecast sales data is an important issue for enterprises. Although the traditional time series prediction method is dominant in research and practice, it has some limitations. With the development of big data, e-commerce enterprises can obtain unprecedented data volume and data characteristics, and it is difficult to accurately predict sales only by using past behaviors and trends. This study proposes a risk aversion-biased combination forecasting model based on the random forest, GBDT, and XGBoost algorithm and used the cost data of each commodity to realize the accurate sample weighting and to output the forecasting results. The experimental results show that the combination forecasting model can predict sales more accurately, which is of great significance for e-commerce enterprises to reduce the cost of commodity management.

Abstract:An affective speech generation technology based on a conditional generative adversarial network (GAN) is proposed in this study. After the introduction of affective conditions and the learning of affective information from the phonetic database, a brand new affective speech with specified emotions can be generated independently. GAN is composed of a discrimination network and a generator. With TensorFlow as the learning framework, the conditional GAN model is employed to train plenty of affective speech, and the speech generation network G and generation network D are used to form a dynamic “game process” for better learning and observation of the conditional distribution of speech emotion data. The generated sample is close to the natural speech signal of the original learning content, which has diversity and can approximate the speech data consistent with the real emotion. The proposed solution is evaluated on the interactive emotional dyadic motion capture (IEMOCAP) corpus and the self-built emotional corpus. It generates more accurate results than the existing affective speech generation algorithms.

Abstract:With the popularization and application of formal methods, there are increasingly more cases in which the theorem prover HOL4 cannot automatically complete the termination proof in the process of formal modeling. Manual termination proof still lacks a general idea. In response, a standardized manual termination proof method is proposed. Starting from the nature of the problem, the method guarantees that the target has the necessary conditions for solving the termination problem. Then, the proof target is simplified by equivalent substitution. Finally, on the basis of the original theorem library, the lacking lemma in the proof process is found to advance the proof. The example shows that this method has a clear logic and can solve the manual termination proof problem of the HOL4 in most cases.

Abstract:The proficiency of students’ knowledge points is an important basis for teachers to make learning plans. To tackle the problem that the students’ proficiency for knowledge points cannot be described in a probabilistic way in cognitive diagnosis, this study proposes a prediction method of embedding knowledge points as features. This method establishes knowledge point vectors for students and test questions respectively and constructs a convolutional neural network for supervised learning to adjust students’ proficiency for knowledge points according to their answering records. Compared with existing related methods, the proposed method greatly improves the accuracy.

Abstract:Cloud storage provides users with outsourcing file storage. However, as the number of data outsourcing surges, deduplication has become critical. At present, a quite effective and commonly used method for data compression is deduplication, that is, to identify redundant data blocks in the data and store only one copy of them. The previous scheme can satisfy different users to encrypt the same file into the same ciphertext, which exposes the consistency of the file. Later, another scheme is proposed in which a centralized server is taken as a deduplication aid. However, with the increase in the number of users, the deduplication efficiency also reduces. In view of the current security and low deduplication efficiency of cloud storage, this study proposes a cross-domain deduplication scheme. The proposed scheme generates random tags and fixed-length random ciphertext for each data to ensure data confidentiality in the case of cross-domain deduplication, resist violent attacks and protect the information equality from being disclosed. In addition, the implementation of the proposed scheme and its function are analyzed. The safety analysis shows that the scheme realizes the privacy protection regarding data content, equality information and data integrity while resisting brute force attacks. The performance analysis demonstrates that it outperforms the existing scheme in terms of repeated search calculation cost and time complexity, which enables lightweight characteristics.