基于全局稀疏地图的AGV视觉定位技术

为了实现自动导引车（AGV）在复杂工业环境下的高精度定位,克服环境变化给定位带来的影响,提出了基于全局稀疏地图的视觉定位方法。首先,设计了大容量二维编码点,作为人工路标铺设在工业环境的地面;然后,基于一种四边形识别算法,在复杂工业环境中准确分割和识别二维编码点;最后,利用二维编码点提供的编码信息,鲁棒匹配图像中的特征点,并以此为基础,使用一种分参数块优化的三维重建策略,实现了工业环境的大规模地图构建,为AGV视觉定位提供了一种稀疏电子地图。AGV视觉的定位通过匹配车载视觉传感器图像中的特征点和稀疏电子地图实现。停车重复定位精度小于0.5 mm,角度偏差小于0.5°,轨迹平均位移误差小于0.1%。实际应用结果表明,该方法能在复杂工业环境中实现AGV视觉的定位,定位的速度和精度方面都满足工业应用的要求,为AGV的视觉定位提供了新的思路。      

分布式容错配电系统设计与仿真

阐述了分布式容错配电系统的功能需求及构型设计,完成了分布式容错配电仿真验证系统的研制,并进行了试验验证。结果表明:系统运行可靠,控制逻辑正确,系统设计满足功能要求。     

共聚焦三维数据表面重建的一种反走样方法

用边界体素的集合表示物体表面,将边界体素作为点投影到屏幕上形成光照图像时,可能出现空洞和图像走样.改进了体素表面算法:用剖分立方体算法得到物体的边界体素集,然后用边界体素的中心点构造投影表面点,用脚印函数计算表面点对显示图像平面上像素的影响范围,以此值作为该像素的权值,形成最后的图像.这样,边界体素不是只投影到一个点上,而是有一个投影区域,从而避免空洞现象,并减轻走样现象.实验证明,该方法具有较快的表面重建速度,消除了空洞,改善了显示效果.     

涡扇发动机气路传感器故障诊断

为了实现对某涡扇发动机传感器故障的在线诊断,提出并设计了1种基于在线贯序极端学习机的故障诊断算法。其核心思想是在定位某传感器故障后,在线建立针对该故障传感器"预学习"的信号重构算法,解决多故障混叠问题。在线信号重构算法以泛化能力指标为判定条件,利用选择策略对算法网络权值进行选择性更新,提高了故障诊断系统的实时性。以某型涡扇发动机为对象开展了传感器故障诊断与重构仿真,结果表明:该算法能够对发动机单、双传感器故障进行准确地诊断与信号重构,且具有良好的实时性。     

发动机孔探图像三维测量与立体重建的实现

作为一种新兴的可视探测技术，孔探成像分析在现代发动机故障诊断中一直发挥着重要的作用。研制和开发基于孔探图像的三维测量与立体重建系统对提高故障诊断水平和预测准确度，降低工作中的人力物力投入，节约维护成本都有着重要的现实意义。本文以整体系统的构架为线索，阐述了系统的图像采集，摄像机标定，图像预处理，立体匹配，3D计算以及深度重建等功能模块的研制思路和基本算法，并给出部分研制结果。本系统的开发为基于也探图像的发动机故障监测与诊断系统的奠定了重要的基础。     

基于相空间重构的网络流量RBF神经网络预测

应用混沌理论,分析了网络流量,用单变量的网络流量时闯序列重构与网络动力系统等距同构的相空间,进而计算了实际网络的关维数和Lyapunov指数,并证实了网络流量存在混沌特性;据此建立了基于径向基函数（RBF）神经网络的模型,并对实际网络数据流进行了预测。仿真结果表明,相对于其他前馈神经网络预测,基于混沌理论的RBF神经网络预测方法学习速度快,预测精度高。     

基于流的遥感图像超分辨率重建

为了提高遥感图像超分辨率重建的质量,提出了一种基于流的遥感图像重建算法。首先,在Glow模型的基础上引入改进后的RRDB架构用于低分辨率图像特征提取,通过构建更多层和连接以提升训练的稳定性。然后,以一种纯数据驱动的流模型来训练分布的参数,通过最大化负的对数似然的方法进行优化,得到该算法的损失函数。实验证明该模型在网络训练过程中能够快速达到稳定收敛的状态,且具有很强的泛化能力。用重建出的图像质量对比SRCNN、SRGAN、ESRGAN,经过测试后发现,提出的算法远远优于SRCNN算法,与其他算法相比也有明显优势。重建出的图像不仅在指标上有所提升,例如与SRCNN相比,PSRN和SSIM分别提升了15%和40%,且人眼观察时有更高的清晰度,高频细节更为丰富。     

Substitution of satellite-based land surface temperature defective data using GSP method

Land surface temperature (LST) as an important environmental variable provides valuable information for earth environmental system modelling. Currently, LST is obtained through satellite thermal sensors at various spatial and temporal resolutions. Although spatially continuous satellite-based LST measurements are intended to overcome the shortcomings of sparse ground-based LST measurements, LST images often contain anomalous values due to the existence of clouds or sensor malfunctioning. The problem becomes more serious where the users deal with high spatial resolution characterized by low temporal resolution. This study examines the capability of a newly developed graph signal processing (GSP) method using two-dimensional single-date thermal data. For this purpose, four Landsat/TIRS datasets are analyzed. The data of five elliptical regions on thermal images are eliminated and then reconstructed through the GSP method and using the LST values of the enclosing rectangles containing the ellipsoids. The results indicate that the temperature variation determined by the GSP method generally conforms to the original image LST values. According to a correlation test conducted on the original image LST and those obtained through the GSP method, the values vary from 58% to 95%, which is an above-the-average rate (RMSE from 0.69 to 2.27). The statistical analysis of the original image LST in both the elliptical regions and the enclosing rectangles containing the ellipsoids indicates that an increase in the variance of LST data causes an increased error in the calculation of temperature by the GSP method, and vice versa. The results of the analysis of variance (ANOVA) and Duncan test indicated that an increase in the number of the non-zero spectral bins would result in increased RMSE values for all the dates and the regions. Moreover, the model errors were significant at the 0.05 level across all the image date and five elliptical study regions. Based on the results, the use of this method is recommended for the reconstruction of LST missing values, where dissimilarity of atmospheric conditions limits the use of other methods that depend on the time series data of various dates and a great deal of data calculation.     

View-dependent progressive transmission and rendering for lunar model based on bicubic subdivision-surface wavelet

Given the fact that the back of the digital lunar model is invisible from any viewpoint, a view-dependent progressive transmission and rendering method for lunar model is proposed and a client-server browsing system is implemented. On the server side, the semi-regular lunar surface is partitioned into dozens of segments, and each segment is compressed independently using bicubic subdivision-surface wavelet and SPIHT. For any viewpoint, only the data of the visible segments are transmitted to the client side progressively, then decoded and locally reconstructed there. This approach not only improves the efficiency of transmission, but also reduces the reconstruction time greatly. We can browse the lunar model on the client almost in realtime.