【作者】 聂建亮；

【导师】 杨元喜； 张勤；

【作者基本信息】 长安大学 ， 大地测量学与测量工程， 2007， 硕士

【摘要】 本文主要研究了神经网络在大地测量若干领域的应用,论文的主要内容概括如下:1.首先介绍了BP神经网络、Elman神经网络和Hopfield神经网络的基本原理,并分析三种神经网络的优缺点。2.阐述了目前动态数据处理中常用的几种滤波方法,并介绍了关于贝叶斯估计的有关理论。3.讨论了神经网络中最常用的学习算法——BP算法,分析了BP算法存在的缺点;并介绍了基于EKF、UKF滤波的神经网络学习算法,该方法能够提高前向神经网络收敛速度和泛化能力。4.在EKF、UKF训练神经网络连接权的基础上,提出了自适应EKF、自适应UKF训练神经网络的思路,进一步提高了神经网络的性能。并用高程拟合算例证明了基于自适应滤波训练连接权算法的有效性。5.对于Hopfield神经网络来说,选择合理的能量函数尤为重要,因此论述了ALR函数作为能量函数的Hopfield神经网络;另外,将抗差估计原理引入Hopfield神经网络,进一步提高了神经网络的抗差能力;分别采用高程拟合和水准网平差验证了这两种算法。6.研究了神经网络在导航中的应用。讨论了动态数据异常检验的三种方法,并用实测数据进行了验证;采用具有反馈层的Elman神经网络处理捷联惯导精对准问题,提高了精对准的精度;采用基于自适应Kalman滤波与BP算法组合的算法处理惯导数据;将抗差自适应滤波理论应用于UKF,进一步完善了UKF理论,并用模拟算例进行了验证;针对粒子滤波的粒子退化问题,提出了自适应UKF重点方法,进一步提高粒子滤波的精度。更多还原

【Abstract】 The applications of neural network in geodetic data processing is researched in this dissertation. The main works and contributions are summarized as follows:1. The theories of BP, Elman and Hopfield neural networks are introduced, the merits and shortcomings of the three kinds of neural networks are analyzed.2. Bayes estimation theory also is introduced. Several kinds of filters, which are widely applied in the dynamic data processing, are described.3. BP algorithm, which can overcome the training of the multi-layer feed forward network, is studied and analyzed. A new algorithm that uses EKF or UKF to train the weight of the neural network is proposed, which can improve the generalization ability and divergence of the BP algorithm.4. On the base of the new algorithms using EKF or UKF to train the weight, the adaptive EKF or UKF is introduced to compute the weight. The algorithms can further improve the generalization ability and divergence of neural network in the computation example of height fitting.5. Energy function is important in Hopfield neural network, the augmented Lagrangian with regularization(ALR) is chosen as an energy function. The robust estimation principle is applied to improve the robustness of Hopfield neural network. It proves that the two algorithms are efficient in examples of height fitting and adjustment of leveling network.6. The application of neural network in navigation is discussed. Three kinds of outlier detection outlier in Kalman filtering are introduced and tested in dynamic data; Elman neural network is used to improve the accuracy of refined initial alignment; the new algorithm combining the adaptive Kalman filtering and BP algorithm is applied in the data of processing INS; the theory of UKF combined with the robust adaptive filtering is applied in a simulated example; adaptive UKF is taken to improve the precision of sampling for particle filter has degeneracy.更多还原