合成孔径雷达图像目标识别问题研究

首先对合成孔径雷达 (SAR)图像目标识别的研究背景和国内外发展现状作了简单介绍 ,然后分析了SAR图像目标识别问题的复杂性 ,讨论了SAR目标识别系统的一般方案与方法 ,提出了基于多源信息融合和人机协作的识别框架 ,进而展望了SAR图像目标识别技术今后的发展方向。     

基于"SOM脸"的选择性单训练样本人脸识别

基于“SOM脸”计算模型提出一种新的人脸局部区域重要程度度量方法,并用于进行选择性单训练样本人脸识别。从机器人脸识别的角度,并未预先人为设定任何重要区域,而是通过学习来自动发现这些对机器而言相对重要的人脸区域,即包含类信息相对丰富的区域,并将其进行可视化。实验结果表明,在利用了人脸局部区域重要程度信息后,识别算法的性能和效率均得到提高;特别是仅选择人脸图像中若干部分重要的区域用于识别时．在提高识别效率的同时．识别性能来见明显下降。     

Analysis of ionospheric TEC response to solar and geomagnetic activities at different solar activity stages

Studying the relationship of total electron content (TEC) to solar or geomagnetic activities at different solar activity stages can provide a reference for ionospheric modeling and prediction. On the basis of solar activity indices, geomagnetic activity parameters, and ionospheric TEC data at different solar activity stages, this study analyzes the overall variation relationships of solar and geomagnetic activities with ionospheric TEC, the characteristics of the quasi-27-day periodic oscillations of the three variables at different stages, and the delayed TEC response of solar activity by conducting correlation analysis, Butterworth band-pass filtering, Fourier transform, and time lag analysis. The following results are obtained. (1) TEC exhibits a significant linear relationship with solar activity at different solar activity stages. The correlation coefficients |R| are arranged as follows: |R|_EUV > |R|_F10.7 > |R|sunspot number. No significant linear relationship exists between TEC and geomagnetic activity parameters (|R| < 0.35). (2) TEC, solar activity indices, and geomagnetic activity parameters have a period of 10.5 years. The maximum amplitudes of the Fourier spectrum for TEC and solar activity indices are nearly 27 days and those of geomagnetic activity parameters are nearly 27 and 13.5 days. (3) The deviations of the quasi-27-day significant periodic oscillation of TEC and solar activity indices are consistent. (4) No evident relationship exists between the quasi-27-day periodic oscillation of TEC and geomagnetic activity parameters. (5) The delay time of TEC for the 10.7 cm solar radio flux and extreme ultraviolet is always consistent, whereas that for sunspot number varies at each stage.     

无人车基于激光雷达/视觉的目标体积自动测量方法

近些年，基于激光雷达和视觉的目标感知在无人系统中得到了广泛应用。目标的体积测量在很多应用场景可以发挥极其重要的作用，然而对识别感知目标的体积测量，目前尚无大量研究。首次提出了一种基于激光雷达/视觉的无人车目标体积自动测量方法，实现了无人车与目标体积测量功能的结合。通过在LeGO-LOAM算法中加入点云畸变补偿，相较于原始LeGO-LOAM算法，无人车在高速情况下的构图精度得到提升；通过将激光雷达与视觉进行深度融合，实现了目标的自动识别与全局定位；通过基于平面拟合的地面分割与欧式聚类，实现了目标点云轮廓的实时获取；通过设计一种基于切片法的不规则物体体积测量方法，实现了无人车在运动情况下对目标体积的自动估计。最终，分别通过Gazebo仿真和实际试验验证了算法的有效性。试验结果表明，所提算法在无人车运动的情况下对静态目标物的实时体积测量精度优于3%，具有较好的工程应用价值。     

基于深度学习的航空器异常飞行状态识别

飞行设备快速存取记录仪(Quick Access Recorder,以下简称QAR)保留了原始航班各类重要飞行参数在内的航行信息,使研究分析航空器实时状况和保障飞行质量成为可能。针对QAR数据高维大样本的特点,在如今大数据背景下,除了传统机理建模分析航空器飞行状态外,采用深度学习的方式建立基于数据驱动的航空器飞行状态识别模型,理论与实用意义兼具。通过对真实QAR飞行数据的研究,开发了基于深度稀疏受限玻尔兹曼机的异常飞行状态识别程序。首先利用小波降噪技术对原始飞行数据进行预处理清洗,在一系列典型飞行参数上提取经典时域特征以及小波奇异熵等信息熵特征构成特征集。在此基础上,分别利用经典的线性主元分析技术和深度稀疏玻尔兹曼机对特征集进行有效降维,最后采用四折交叉验证方式,通过高斯过程分类器实现对飞行状态的辨识。实验结果显示,基于深度受限玻尔兹曼机-高斯过程分类的飞行状态识别具有较高分类准确性。     

基于AI+神经网络的数字化故障电弧识别方案研究

线路故障电弧的特征随着电力电子技术的提高更趋复杂，传统的故障电弧识别方式已无法应对新型用电系统。对比传统故障电弧识别方案，研究故障电弧产生类型和新型检测技术，提出基于谐波因素、总谐波畸变率、电流零休时间、电流变化速率、电流周期性等典型时域、频域特征的AI+神经网络的数字化故障电弧识别方案。该方案运用回归算法向量机对特征图进行分类处理，通过正向传输运算、反向传输运算和迭代回归运算三个步骤执行卷积神经网络，频繁迭代回归以获得最佳特征识别图。本文以调光器的故障电弧识别为例，对故障识别方案进行验证，结果表明采用此方案可实现更高效、更精准、更稳定地识别故障电弧。     

载人月球探测月面活动发展设想

载人月球探测是浩大的系统工程，组成复杂，关键技术多，研制周期长，经费投入高，任务风险大。月面活动是载人月球探测任务体现能力、保障安全、提升效益的重要组成部分，必须回答“为什么”“去哪里”“怎么干”等问题。本文从国内外月球探测历程出发，结合当今世界月球探测领域趋势与局势，从月球战略布局、月球科学内涵开展分析。针对月面驻留与活动工程实施，对比Apollo与Artemis计划方案，提出兼顾可行性和工程效益的月面实验室方案。结合未来发展方向，提出载人月球科研站建设发展思路和月球熔洞探测设想，探索具有中国特色可持续发展的载人月面活动发展道路。     

猪IFN-α在巴斯德毕赤酵母中的分泌表达

为了获得高效抗病毒活性的猪α-干扰素蛋白,采用PCR法扩增获得猪α-干扰素基因完整的开放阅读框(ORF),构建重组表达质粒pPICZαC-IFN,电击转化毕赤酵母感受态细胞X33.挑阳性菌落诱导表达,离心取上清,SDS-PAGE和 Western-blotting分析,可见1条相对分子质量约19 400的清晰蛋白条带,与理论值相符,表明表达产物为猪α-干扰素.在Vero细胞上检测该干扰素抗VSV的活性为4.04×10~6 IU/L.     

A model of foreground emission in UV using GALEX deep observations

Foreground emission, mainly airglow and zodiacal light, is a significant contributor in an ultraviolet observation especially from low earth orbit. Its careful estimation and removal are tedious yet unavoidable processes in the study of diffuse UV radiation and by extension interstellar dust studies. Our analysis of deep GALEX observations show that airglow is not only a function of Sun angle but also a strong function of Solar activity at the time of observation. We present here an empirical model of airglow emission, derived from GALEX deep observations, as a function of 10.7 cm Solar flux and Sun angle. We obtained the model by training machine learning models on the data using a variant of the regression algorithm that is both resilient toward outlier data and sensitive to the complexities of the provided data. Our model predictions across various observations show no loss in generalization as well as good agreement with the observed values. We find that the total airglow in an observation is the sum of a baseline part (AG_c) that depends on the Solar flux and Sun angle, and a variable part (AG_v) that depends on the Sun angle and the time of observation with respect to local midnight. We also find that the total airglow can vary between 85 – 390 photon units in FUV and 80 – 465 photon units in NUV.     

基于深度学习的雷达干扰识别方法研究

针对复杂电磁环境下雷达干扰信号识别问题，从优化卷积神经网络结构的角度出发，本文提出了一种对卷积神经网络结构LeNet-5增加批量归一化层和改变激活函数的方法。该方法能够加速网络收敛，提升网络的学习效率。本文首先建立舰船目标模型，分析了噪声调幅干扰、噪声调频干扰、梳状谱干扰和无干扰的真实目标回波信号在时频域的差异，然后利用四种信号对舰船目标模型生成数据集，最后通过本文所提方法实现雷达干扰的自动识别。仿真结果表明：在全信噪比条件下，本文所提方法对四种信号的识别准确率达到98.1%，表明所提方法有着较好的稳定性和鲁棒性。