本期目录

2023年 30卷 4期 刊出日期：2023-08-31



Artificial Intelligence

Machine-learning based ocean atmospheric duct forecasting: a hybrid model-data-driven approach

Feng Yuting, Gong Haobing, Hao Xiaojing, Gao Hui, Guo Xiangming

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2011

摘要 ( 139 )

参考文献 | 相关文章 | 计量指标

The atmospheric duct is a vital radio wave environment. Conventional methods of forecasting the atmospheric duct mainly include statistical analysis based on sounding observation data and mesoscale numerical model-based prediction. The former can provide accurate duct information but is highly dependent on the acquisition of data sets. The latter is more practical but still lacks accuracy. This paper introduces machine learning to establish a novel meteorological parameter correction model for atmospheric duct prediction. In detail, using the weather research and forecasting (WRF) model data and spatiotemporal characteristics as input, sounding data as label and extreme gradient boosting (XGBoost) model for training, the meteorological parameter correction effect is the best, i. e. , the accuracy of forecast meteorological parameters is improved by about 65.4%. Combining the mapping relationship between meteorological parameters and corrected atmospheric refractive index ( CARI ), and the transition mechanism of CARI to duct parameters, a new duct forecasting mechanism is proposed. Due to the high efficiency of numerical model and the accuracy of sounding data, the new duct forecasting mechanism has excellent performance. By comparing the duct forecasting results, the forecasting accuracy of the new duct forecasting model is significantly higher than that of the mesoscale model.

Hardware-friendly Cycle-GAN and reconfigurable design

Xie Xiaoyan, Chai Miaomiao, Deng Junyong, Du Zhuolin, Yang Kun, Yin Shaorun

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2012

摘要 ( 110 )

参考文献 | 相关文章 | 计量指标

As a kind of generative adversarial network (GAN), Cycle-GAN shows an apparent superiority in image style translation. The more complicated architectures with large number of parameters and huge computational complexities, cause a big challenge in deployment on resource-constrained platform. To make full use of the parallelism of hardware under guaranteed image quality, this paper improves the generator network to a hardware-friendly Inception module. The optimized framework is named simplified Cycle-GAN (S-CycleGAN), with greatly reduced parameters of convolution, while avoiding the degradation of image quality from structural compression. Testing with the apple2organge and horse2zebra datasets, the experiment results show that the images generated by S-CycleGAN outperform the baseline and other models. The number of parameters reduces by 19.54%, memory usage cuts down by 9.11%, theoretical amount of multiply-adds (Madds) decreases by 17.96%, and floating-point operations per second (FLOPS) diminishes by 18.91%. Finally, the S-CycleGAN was mapped on the dynamic programmable reconfigurable array processor ( DPRAP ), which calculate the convolution and deconvolution in a unified architecture, and support flexible runtime switching. The prototype systems are implemented on Xilinx field programmable gate array (FPGA) XC6 VLX550 T-FF1759. The synthesized results show that, with 150 MHz, the hardware resource consumption is reduced by 52% compared to the recent FPGA scheme.

Medium and long-term traffic flow prediction based on ACNN-Trans model

Zhang Xijun, Zhang Guannan, Zhang Xianli, Zhang Hong, Zhang Minghu

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2013

摘要 ( 119 )

参考文献 | 相关文章 | 计量指标

Aiming at the failure of traditional medium and long-term traffic flow forecasting methods to effectively process long-time series data and extract the spatio-temporal characteristics between road nodes, a com-bined prediction model attention based on convolutional neural network (CNN) and transformer ( ACNN-Trans ) network is proposed. The spatial features between nodes on complex roads are extracted by CNN. The embedded attention mechanism is adaptively pay attention to the results of feature extraction. The spatial features of the node traffic flow are dug by the embedded attention mechanism. The traffic flow sequence is calculated by the transformer network. The temporal correlation of long-term series data is captured by the relative position weight between the data. Finally, the 30-min and 60-min traffic flow predictions are respectively modeled by the extracted temporal and spatial features. The results show that the prediction results of the ACNN-Trans combined model are better than other models on two different real datasets. Comparing to the baseline models, the root mean square error (RMSE) index of the prediction results is reduced by an average of 34.58%, which verifies the effectiveness of ACNN-Trans model.

Research on statistical characteristics of engine backward RCS based on K-KDE algorithm

Fu Li, Cui Zhe, Deng Hongwei

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2014

摘要 ( 84 )

参考文献 | 相关文章 | 计量指标

In order to solve the problem of low accuracy of traditional fixed window width kernel density estimation (KDE) in radar cross section (RCS) statistical characteristics analysis, an improved Epanechnikov KDE (K-KDE) algorithm was proposed to analyze the statistical characteristics of the engine's backward RCS. Firstly, the K-nearest neighbor method was used to calculate the dynamic window width of the K-KDE, and the Euclidean distance of each adjacent sample was used to judge the local density of the sample, and then the window width of the kernel function was adjusted by the distance between the sample point and the nearest neighbor to complete the KDE. Secondly, based on the K-KDE and the traditional KDE algorithm, the cumulative probability density function (CPDF) of four RCS random distribution sample points subject to fixed parameters was calculated. The results showed that the root mean square error of the K-KDE was reduced by 31.2%, 38.8%, 38.1% and 31.9% respectively compared with the KDE. Finally, the K-KDE combined with the second generation statistical analysis models were used to analyze the statistical characteristics of the engine backward RCS.

Surveys on the application of neural networks to event extraction

Duan Li, Chen Lin, Luo Bing

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2015

摘要 ( 96 )

参考文献 | 相关文章 | 计量指标

Event extraction (EE) is a significant part of natural language information extraction, and it is widely adopted in other natural language processing (NLP) tasks such as question answering and machine reading comprehension. With the development of the NLP field, numerous datasets and approaches for EE are promoted, raising the need for a comprehensive review. In this paper, the resources for EE are reviewed, and then the numerous neural network models currently employed in EE tasks are classified into three types: Word sequence-based methods, graph-based neural network methods, and external knowledge-based approaches. And then the methods are compared and contrasted in detail, and their flaws and difficulties are analyzed with existing research in this survey. Finally, the future research tendency is discussed for EE.

Workflow optimization algorithm based on virtualization and nonlinear production quality under time constraints

Luo Zhiyong, Tan Shanxin, Xu Haifeng, Liu Xintong

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2016

摘要 ( 72 )

参考文献 | 相关文章 | 计量指标

Balancing time, cost, and quality is crucial in intelligent manufacturing. However, finding the optimal value of production parameters is a challengingnon-deterministic polynomial (NP)-hard problem. In the actual production process, the production process has the characteristics of multi-stage parallel. Therefore, aiming at the difficult problem of multi-stage nonlinear production process optimization, this paper proposes a workflow optimization algorithm based on virtualization and nonlinear production quality under time constraints (T-OVQT). The algorithm proposed in this paper first abstracts the actual production process into a virtual workflow model, which is divided into three layers: The bottom production process collection layer, the middle layer of service node partial order composition layer, and the high level of virtual node collection layer. Then, the virtual technology is used to reconstruct the node set and divide the task interval. The optimal solution is obtained through inverse iterative normalization and forward scheduling, and the global optimal solution is obtained by algorithm integration. Experimental results demonstrate that this algorithm better meets actual production requirements than the traditional minimum critical path (MCP) algorithm.

Wireless

Design and implementation of unmanned aerial vehicle localization and light tracking system

Xu Ming, Dong Chen, Li Hanlu, Luo Qiming, Wang Lizi, Zhang Wanru, Liu Huixin

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2017

摘要 ( 99 )

参考文献 | 相关文章 | 计量指标

Unmanned aerial vehicles (UAV) are applied widely and profoundly in various fields. Moreover, high-precision positioning and tracking in multiple scenarios are the core requirements for UAV usage. To ensure stable communication of UAVs in denial environments with substantial electromagnetic interference, a systematic solution is proposed based on a deep learning algorithm for target detection and visible light for UAV tracking. Considering the cost and computational power limitations on the hardware, the you only look once (YOLO) v4-Tiny model is used for static target detection of the UAV model. For UAV tracking, and a light tracker that can adjust the angle of emitted light and focus it on the target is used for dynamic tracking processing. Thus, achieving the primary conditions of UAV optical communication with good secrecy is also suitable for dynamic situations. The UAV tracker positions the UAV model by returning the coordinates and calculating the time delay, and then controls the spotlight to target the UAV. In order to facilitate the deployment of deep learning models on hardware devices, the lighter and more efficient model is selected after comparison. The trained model can achieve 99.25% accuracy on the test set. The dynamic target detection can reach 20 frames per second (FPS) on a computer with an MX520 graphics processing unit (GPU) and 6 GB of random access memory (RAM). Dynamic target detection on a Jetson Nano can reach 5.4 FPS.

Channel estimation based on multi-armed approach for maritime OFDM wireless communications

Zhang Qianqian, Xu Yanli

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2018

摘要 ( 97 )

参考文献 | 相关文章 | 计量指标

With the development of maritime informatization and the increased generation of marine data, the demands of efficient and reliable maritime communication surge. However, harsh and dynamic marine communication environmentcan distort transmission signal, which significantly weaken the communication performance. Therefore, for maritime wireless communication system, the channel estimation is often required to detect the channel suffered from the impacts of changing factors. Since there is no universal maritime communication channel model and channel varies dynamically, channel estimation method needs to make decision dynamically without pre-knowledge of channel distribution. This paper studies the radio channel estimation problem of wireless communications over the sea surface. To improve the estimation accuracy, this paper utilizes multi-armed bandit (MAB) problem to deal with the uncertainty of channel state information (CSI), then proposes a dynamic channel estimation algorithm to explore the global changing channel information, and asymptotically minimize the estimation error. By the aid of MAB, the estimation is not only dynamic according to channel variation, but also does not need to know the channel distribution. Simulation results show that the proposed algorithm can achieve higher estimation accuracy compared to matching pursuit (MP)-based and fractional Fourier transform (FrFT)-based methods.

Others

Optimal decomposition level selection approach in wavelet threshold denoising algorithm for ECG signal

Yao Yindi, Yi Jun, Zeng Zhibin, Li Xiong, Wang Chen, Li Yuhang

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2019

摘要 ( 74 )

参考文献 | 相关文章 | 计量指标

The effect of electrocardiogram (ECG) signal wavelet denoising depends on the optimal configuration of its control parameters and the selection of the optimal decomposition level. Nevertheless, the existing optimal decomposition level selection scheme has some problems, such as lack of reliable theoretical guidance and insufficient accuracy, which need to be solved urgently. To solve this problem, this paper proposes an optimal decomposition level selection method based on multi-index fusion, which is used to select the optimal decomposition level for wavelet threshold denoising of ECG signal. In the stage of index selection, in order to overcome the limitation of a single evaluation index, the optimal multi-evaluation index is selected through the joint analysis of the geometric and physical significance of traditional evaluation indexes. In the stage of index fusion, based on the method of weighting the selected multiple indexes by the information entropy weight method and the coefficient of variation method, an optimal decomposition level selection method based on the evaluation index Z is proposed to improve the accuracy of the optimal decomposition level selection. Finally, extensive experiments are carried out on the real ECG signal from the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) arrhythmia database and simulated signal to test the performance of the proposed method. The experimental results show that the accuracy of this method is superior to other related methods, and it can achieve better denoising effect of ECG signal.

Research on security situation awareness algorithm of Markov differential game block-chain model

Luo Zhiyong, Wang Shuyi, Song Weiwei, Liu Jiahui, Wang Jianming

中国邮电高校学报(英文), 2023, 30 (4). doi： 10.19682/j.cnki.1005-8885.2023.2020

摘要 ( 70 )

参考文献 | 相关文章 | 计量指标

To analyze the impact of solar eclipse attacks on the block-chain network and comprehensively and accurately evaluate the security of the current system, a method based on the Markov differential is proposed to solve the problem that it is difficult to meet the multi-stage continuous real-time randomness in the current block-chain network attack and defense process. Block-chain network security situation awareness method based on game model. This method analyzes the security data generated by the solar eclipse attack, establishes the corresponding attack graph, and classifies the offensive and defensive strengths of the offensive and defensive parties. Through a multi-stage offensive and defensive game, the security level of each node of the blockchain system is combined with the final objective function value comprehensively evaluates the real-time security status of the system. The results of simulation experiments show that the proposed model and algorithm can not only effectively evaluate the overall security of the blockchain network, but also have feasibility in predicting the future security status.