This paper aims to address the challenge of seeking an optimal safe path for a UAV (unmanned aerial vehicle) from an initial position to a target position, while avoiding all obstacles in a three-dimensional environment. An improved APF (artificial potential field) method combined with the regular hexagon guidance method is proposed to solve unreachable and local minimum problems near obstacles as observed with traditional artificial potential field methods. First, we add a distance correction factor to the repulsive potential field function to solve problems associated with unreachable targets. Then, a regular hexagon-guided method is proposed to improve the local minimum problem. This method can judge the environment when the UAV is trapped in a local minimum point or trap area and select the appropriate planning method to guide the UAV to escape from the local minimum area. Then, 3D modeling and simulation were carried out via Matlab, taking into account a variety of scenes involving complex obstacles. The results show that this method has good feasibility and effectiveness in real-time path planning of UAVs. Lastly, we demonstrate the performance of the proposed method in a real environment, and the experimental results show that the proposed method can effectively avoid obstacles and find the optimal path.
Battlefield environments are combat spaces that contain geographic elements such as terrain and roads. Road modeling and simulation is an important part of battlefield simulation and plays a key role in complex combat decision-making. Traditional road modeling is unable to handle the complex terrain conditions present in the field; hence, this paper proposes road modeling and simulation method for field environments. In particular, in order to support road modeling and simulation of complex terrain environments, road construction designs oriented to typical battlefield environments are proposed. This method divides the road network into different sub-models according to their characteristics and models them separately, improving the demand for realism in battlefield simulation. Then, the proposed method uses OpenStreetMap geographic information data to drive road network construction. The model offers real-time, high accuracy road information content and complete classification that can meet the needs of military operations and modeling simulations for typical battlefield environments. Secondly, using terrain elevation data, road construction rules, and other auxiliary information, the road height is adjusted to adapt to the complex terrain conditions of the battlefield and possible multi-level road network structures. Lastly, the introduction of a $ {G}^{2} $ continuous Hermite interpolation spline can flexibly represent the center line of the road and improves the reusability of the road model through grid deformation. Experiments show that the proposed simulation method can more reliably restore the real details of a road network to effectively fit complex terrain and improve the reusability of road models. Finally, it provides a feasible analysis angle and modeling method for researching geographic elements in battlefield environments.
Adaptive learning is an educational method that uses computer algorithms to coordinate interaction with learners, and provides customized learning resources and learning activities to address the unique needs of each learner. With the impact of COVID-19, adaptive learning has become increasingly important. One of the challenges with adaptive learning is how to provide personalized learning resources for learners—i.e., how to generate personalized recommendation for learners from a large set of learning resources. Existing methodologies mainly generate recommendations based on a learner’s knowledge level; however, this approach has some limitations. Firstly, when assessing a learner’s knowledge level, learners’ forgetting phenomenon has to date not been well modeled. Secondly, recommendations are generated separately from knowledge tracing tasks, ignoring the interconnectedness between these aspects. In addition, learners’ preferences for the type of learning resources and learning strategies is normally ignored if the knowledge level alone is used. To solve the aforementioned problems, this paper proposes a knowledge and personality incorporated multi-task learning framework (KPM) to boost course recommendations (i.e., the above-mentioned learning resources); the proposed method regards an enhanced knowledge tracing task (EKTT) as an auxiliary task to assist the primary course recommendation task (CRT). Specifically, using EKTT, we design a personalized forgetting controller to enhance the deep knowledge tracing model for accurately assessing a learner’s knowledge level. With CRT, we combine the learner’s knowledge level and sequential behavior with their personality adapted to the specific context to obtain learner’s profile; this data is subsequently used to generate a course recommendation list. Experimental results on real-world educational datasets demonstrate the superiority of our proposed method in terms of hit ratio (HR), normalized discounted cumulative gain (NDCG), and precision, indicating that our method can generate more personalized recommendations.
With the rapid development of the Internet of Things, embedded hardware products face great challenges in data security. The AES (Advanced Encryption Standard) algorithm has the advantages of strong attack resistance, fast operation speed and flexible block length in the field of data encryption and decryption. The speed of this algorithm on microcontroller platforms is far inferior to general-purpose CPUs (Central Processing Units) which have an extended instruction set for AES encryption. To solve this problem, a speed optimized AES algorithm in CTR (Counter) mode based on the Cortex-M4 core instruction set is implemented using assembly language. The kernel’s unique barrel shifter and three-stage pipeline are used to optimize the round transformation of the algorithm, and the number of instruction cycles is reduced. Testing on an FRDM-K82F development board shows that the assembly optimization of the algorithm is substantially more efficient than the code implemented using the C language, and it offers more advantages in both cost and power consumption compared to hardware encryption based on the coprocessor.
The concept of knowledge tracking involves tracking changes in a student’s knowledge level based on historical question records and other auxiliary information, and predicting the result of a student’s subsequent answer to a question. Since the performance of existing neural network knowledge tracking models needs to be improved, this paper proposes a deep residual network based on a stacked gated recurrent unit (GRU) network named the stacked-gated recurrent unit-residual (S-GRU-R) network. The proposed solution aims to address over-fitting caused by too many parameters in a long short-term memory (LSTM) network; hence, the solution uses a GRU instead of LSTM to learn information on the sequence of questions. The use of stacked GRU can expand sequence learning capacity, and the use of residual connections can reduce the difficulty of model training. Experiments on the Statics2011 data set were completed using S-GRU-R, and AUC (area under the curve) and F1-score were used as evaluation functions. The results showed that S-GRU-R surpassed other similar recurrent neural network models in these two indicators.
Neural architecture search algorithms aim to find more efficient neural network structures in a huge neural network structure space using computer heuristic search instead of manual search. Previous studies have addressed the problem of inefficient and time-consuming search for early neural network structures by introducing various constraints on the search space. While constraints on the search space can improve and stabilize the performance of the model, they ignore potentially efficient model structures. Hence, in this study, we constructed a recursive model search space that focuses more on the macroscopic structure of neural networks. We proposed a neural architecture search algorithm that explores this search space through a step-by-step incremental search approach. Experiments showed that the algorithm can efficiently perform neural architecture search tasks in complex search spaces, but still fell slightly short of the latest constrained search space-based neural architecture search algorithms.
With the development and iteration of financial technology(FinTech) software programs, the size of test suites will gradually increase, which may introduce inherent redundancy. In order to effectively quantify test redundancy, a test redundancy evaluation metric called MVI (Most Valuable Item) is proposed in this study. To verify the validity of the MVI metric, the MVIR (Most Valuable Item Reduction) test case reduction algorithm is proposed. Experimental results show that the MVIR can achieve a test case reduction ratio of more than 89.88% assuming the test performance loss is less than 9.20%, this demonstrates that the MVI metric is valid.
The identification and classification of garment patterns are important technologies for intelligent clothing production and management. This paper proposes a method to convert garment patterns into graphic data and subsequently proposes a lightweight graph neural network GPC-GCN (Garment Pattern Classification Graph Convolutional Network) that can process this graphic data. The proposed graph data modeling method can not only maintain information on the shape of each component in the garment pattern but also deal with the arbitrariness of the position of components in garment patterns. Experiments show that the proposed graph neural network GPC-GCN achieves a better result for the classification of garment patterns compared to convolutional neural networks and graph convolutional networks.
Traditional virtual terrain modeling commonly uses a procedural generation method based on manual design, which cannot be used for competent simulation modeling tasks that need to restore real environments, such as in military applications. In this paper, we proposed a landscape simulation modeling method based on remote sensing images. The core of our proposed method is a landscape blended texture generation network (LBTG-Net); this method uses a blended texture generator (BTG) to generate landscape blended textures with the supervision of a style discriminator (SD) and multi-stage classification loss. Then, we procedurally build the complete virtual environment based on the blended texture generated by LBTG-Net. Our method has two main features: (1) accurate land-cover classification ability of remote sensing image inputs; and (2) high quality landscape blended texture outputs to guarantee virtual landscape modeling quality. We used multispectral image data from the Sentinel-2 satellite as the experimental dataset. The experimental results showed that our method offered high performance under mainstream land-cover classification evaluating indicators and can accurately reproduce the environmental distribution of input remote sensing images while completing high-quality virtual terrain simulation modeling.
With the development of the Internet of Things, a large number of sensor devices can be connected to a network. Anomaly detection of data generated by these devices is related to the stability of system services. A time series database is a database system optimized for time series data. As an important component of a monitoring system, time series databases are responsible for storing and querying continuous streams of time series data. The current time series database, however, cannot fully utilize system computing resources and cannot meet the latency requirements when coping with queries from multiple data sources. To address these drawbacks, we redesigned the query execution model of a time series database based on the well-known InfluxDB, and we proposed InfluxDB-PP (parallel processing) as a method to address the aforementioned problems. The experimental results show that InfluxDB-PP reduces query latency by about 85.7% compared to InfluxDB for real-time anomaly data query scenarios.
Distant supervision relation extraction captures noisy instances while reducing the burden of manual annotation, which hinders the training and testing process. To alleviate this problem, we proposed a de-noising method based on the influence function. The influence function measures the influence of each training point; the influence of one training point is defined as the change in test loss after removing the training point. We observed that this property could be used to determine whether a training instance involves noisy data. First, we designed a scoring function based on the influence function. Then, we integrated the scoring function into a bootstrapping framework to obtain the final denoising dataset from a small clean set. Using this preprocessing method, every distantly supervised dataset could be denoised by our method. Experimental results showed that the proposed denoised dataset can achieve good performance on a public dataset.
The popularity of positioning devices has generated a large volume of vehicle driving data, making it possible to use historical data to predict the driving time of vehicles. Vehicle driving data consists of two parts: the sequence of road segments that the vehicle travels through, the departure time, the total length of the path, and other external information. The questions of how to extract sequence features in road segments and how to effectively fuse sequence features with external features become the key issues in predicting the travel time. To solve the aforementioned problems, a transformer-based travel time prediction model is proposed, which consists of two parts: a road segment sequence processing module and a feature fusion module. First, the road segment sequence processing module uses the self-attention mechanism to process the road segment sequence and extract the road segment sequence features. The model can not only fully consider the spatiotemporal correlation of road speeds between each road segment and other road segments, but also ensures the parallel input of data into the model, avoiding the low efficiency problem caused by sequential input of data when using recurrent neural networks. The feature fusion module fuses the road segment sequence features with external information, such as departure time, and obtains the predicted travel time. On this basis, the number of road segments connected by the intersection is determined by the upstream and downstream intersection features of the road segment, and the input model is combined with the road segment characteristics to further improve the prediction accuracy of the driving time. Comparative experiments with mainstream prediction methods on real data sets show that the model improves prediction accuracy and training speed, reflecting the effectiveness of the proposed method.
In the era of big data and with the continuous expansion of data, there are significant challenges with efficient access to data. Hence, designing an efficient index structure is of great significance. ALEX (updatable adaptive learned index) is a learned index that uses a machine learning model to replace the traditional B-tree index structure. Although it offers good time and space performance, it suffers from frequent page faults. In order to solve this problem and further improve the performance of ALEX, a memory pre-allocation strategy based on huge pages is proposed, on the basis of ALEX, that can help reduce the rate of memory page faults and improve the overall performance of ALEX. In the memory allocation phase, the pre-allocation strategy is adopted, and the memory free phase adopts a delayed release strategy. Experiments on the Longitudes dataset show that this strategy offers good performance.
The issue of how to unfold an irregular cylindrical mural from the top surface of a cave corridor into a panorama is a challenge for researchers involved with ancient mural protection and secondary development. This paper presents a method of dividing cylindrical murals into many overlapping small areas for sampling firstly, and then stitching these sampled images into a panorama. The constituents of this method include the following key elements: ① Reconstructing the 3D model with the sampled image set; ② Mapping the image texture to the 3D model; ③ Fitting the reconstructed irregular 3D cylindrical surface to the ideal cylindrical surface which is closest to the original form; and ④ Projecting the mural of the ideal cylindrical surface to a panorama. The method proposed in this paper was verified on an actual cave image set. The experimental results showed that the proposed method can generate the panorama in full; moreover, there was no evidence of stitching traces or texture deformation on the panorama. The proposed method offers practical value for mural protection.
In real-world scenarios, various events in the news are not only too nuanced and complex to distinguish, but also involve multiple entities. To address these problems, previous event-centric methods are designed to detect events first and then extract arguments, relying on imperfect performance for event trigger detection; this process, however, is unfit to deal with the sheer volume of news in the real world. Given that the performance of named entity recognition (NER) is satisfactory, we shift our perspective from an event-centric to a target-centric view. This paper proposes a new task: target-dependent event detection (TDED), which aims to extract target entities and detect their corresponding events. We also propose a semantic and syntactic aware approach to support thousands of target entity extractions first and subsequently the detection of dozens of event types; this approach can be applied to data from massive corporations. Experimental results on a real-world Chinese financial dataset demonstrated that our model outperformed previous methods, particularly in complex scenarios.
With the continuous development of China’s electric power system, the security and reliability of power supply directly affects regional production output and people’s economic life. As an important part of the power dispatch system, traditional fault locations rely on the cumulative experience and manual judgment of dispatchers. Faced with increasing demands, fault locations that rely solely on the traditional method are likely to result in an increase in misjudgment rates and pose a threat to the stable operation of the power system. This paper proposes a Boolean equation based on Kirchhoff’s law and the grid fault location algorithm to address this challenge. The fault location issue can effectively be converted to Boolean linear mixed programming problems and combined with simulated annealing algorithms. When these genetic algorithms are applied to the idea of a network and realized in the grid for fast positioning of small faults, the scheduling error rate can be reduced and the time difference from fault occurrence to fault isolation and fault processing can be shortened; in turn, this saves human resources and improves scheduling efficiency.
Acute kidney injury is a clinical disease with a high morbidity rate, and early identification of potential patients can facilitate medical interventions to reduce morbidity and mortality. In recent years, electronic health records have been widely used to predict an individual’s potential risk. Most of the existing acute kidney injury prediction models tackle the issue of sparsity and irregularity in the physiological variables data by aggregating data or imputing the missing value, but ignore the patient’s health status implied by the missing information. Moreover, they do not consider the characteristics of and correlation between the various modalities. To solve the above issues, we present a multi-modal disease prediction model for acute kidney injury. The proposed model considers a variety of modal data, including physiological variables, disease, and demographic data. A new mask and time span based long short term memory (LSTM) network is designed to learn the time span and missing information of individual Physiological variables, and furthermore, to capture their numerical changes and frequency changes. The multi-head self-attention mechanism is introduced to promote interaction learning of each modality representation. Experiments on the real-world application of acute kidney injury risk prediction and mortality risk prediction demonstrate the effectiveness and rationality of the proposed model.
With the continuous development of industrial intelligent inspection technology, the equipment element state recognition system based on digital image processing is widely used. In order to improve the accuracy of power distribution cabinet(PDC) equipment element state recognition in a distribution room, a ResNet(residual networks)-based equipment element state recognition method is proposed. Firstly, the data acquisition system is set up and the data set is constructed. Then, for the PDC image, the preset device component target area is cropped to generate the device component image. For device component images, a ResNet-based component state recognition model was constructed and trained, and the trained model was used to identify component states. Taking the data set for power distribution cabinet equipment element in substation distribution rooms as the research object, a network of single prediction heads is adopted as the component with complex features, and the network of multiple prediction heads is adopted as the component with simple features. Then, the compact and pruning model compression method is used to reduce the number of parameters and the calculation amount under the condition of less accuracy loss. Finally, the architecture design of the inspection system is introduced. A JetSon Nano edge terminal is used as the running hardware of the algorithm module to reduce the communication cost.
In this paper, we propose a method for fast splicing of three-dimensional point clouds based on the lock pin model on a container terminal using high overlapping views. This experiment first uses an Azure Kinect depth camera to collect scene point clouds, and subsequently preprocesses the point cloud. The target point cloud is thus obtained. For lock pins with slightly different views, the sample consensus initial algorithm (SAC-IA) is used on the basis of the classic iterative closest point (ICP) algorithm to determine the overlapping position relationship of the two point clouds. In the overall splicing process, the relative size of the bounding box area projected by the lock pin in the z-direction of the camera is adopted to estimate the general shape of the lock pin; the relative size of the bounding box area is also used to select an appropriate number of point cloud views with high overlap in order to ensure the accuracy of registration and reduce processing time by comparing the difference between the area of adjacent views. The experimental results show that the proposed method has a lower relative registration error for the lock pin, and can quickly establish a workpiece model suitable for type matching.
Chinese sentiment analysis is one of important researches in natural language processing, which aims to discover the sentimental tendencies in the Chinese text. In recent years, research on Chinese text sentiment analysis has made great progress in efficiencies, but few studies have explored the characteristics of the language and downstream task requirements. Therefore, in view of the particularity of Chinese text and the requirements of sentiment analysis, using the Chinese text sentiment analysis method that integrates multi-granularity semantic features, such as characters, words, radicals, and part-of-speech is proposed. This introduces radical features and emotional part-of-speech features based on character and word features. Additionally, this integration uses bidirectional the long short-term memory network (BLSTM), attention mechanism and recurrent convolutional neural network (RCNN). The softmax function is used to predict the sentimental tendencies by integrating multi-granularity semantic features. The comparative experiment results on the NLPECC (natural language processing and Chinese computing) dataset showed that the F1 score of the proposed method was 84.80%, which improved the performance of the existing methods to some extent and completed the Chinese text sentiment analysis task.
中国综合性科技类核心期刊(北大核心)
