基于自适应粒子滤波算法的捷联惯导初始对准方法研究

发布时间：2018-04-01 19:35

本文选题：捷联惯导　切入点：初始对准　出处：《哈尔滨工程大学》2014年硕士论文

【摘要】：粒子滤波(Particle Filter,简称PF)是一种性能优越的非线性滤波算法。它对系统的状态方程和量测方程以及噪声统计特性均未加任何限制,因而相对于传统的非线性滤波具有更宽广的应用范围。而且它突破了卡尔曼滤波体系下非线性滤波(EKF、UKF、CKF)的架构束缚,摒弃了对状态变量均值和方差估计的思想,转而通过从后验概率密度中抽取采样粒子,并对其进行迭代的预测和更新,来不断逼近真实的后验概率密度分布,从而更加的贴近最优估计的本质。粒子滤波具有滤波精度高,收敛速度快等优点,使得它已经成为了在处理非线性、非高斯系统下状态滤波和参数估计的主流滤波算法.本文围绕粒子滤波在捷联惯导大方位失准角背景下展开了如下的研究工作:首先,根据欧拉平台误差角推导了大方位失准角条件下捷联惯导系统初始对准的误差方程。在研究贝叶斯信号处理和蒙特卡洛(MonteCarlo)积分的的基础上,引出序贯重要性重采样(Sequential Importance Sample)粒子滤波算法,并给出目标跟踪模型下的PF和UKF仿真对比。其次,针对标准粒子滤波算法中的缺点和不足,提出以下改进方法:一是标准粒子滤波当中直接选取先验概率密度函数作为重要性密度函数进行采样粒子,导致最新时刻的量测信息丢失,使得采样的粒子过分依赖于状态模型,当似然概率密度呈现尖峰状态或是位于先验概率密度函数尾部的时候很容易造成粒子退化。提出根据最新时刻的量测信息,给予重要性密度函数有目的的调整、修正,从而使得重要性密度函数能够最大程度上的向后验概率密度分布偏移。本文设计的算法是使用基于EKF、UKF、CKF滤波估计之后的均值和方差来产生了新的采样粒子集群。给出分段非线性模型下的PF、EKPF、UPF、CPF的对比仿真分析。二是高维状态估计中所需粒子数成级倍数增长,因而计算延迟,实时性不够理想的情况,提出动态调节粒子数,减少计算量。具体做法是将自适应技术引入到粒子滤波的重采样之前,根据上一时刻对信号的估计精度来确定下一时刻估计所需要的粒子数。设计在目标跟踪模型下的仿真来对比APF和PF的滤波性能,验证自适应的有效性。三是将CPF和APF结合形成ACPF算法,新算法既通过CKF设计重要性密度函数,提高了采样效率;又根据上一时刻的估计状态预测下一时刻所需要的粒子数,实现了粒子动态调节,减少了计算量。最后在飞行器机动飞行模型中,验证ACPF的有效性。最后仿真在捷联惯导初始对准非线性误差模型下进行,给出ACPF和PF的仿真对比分析,证明算法改进的正确性和有效性。
[Abstract]:Particle filter Particle filter (PFR) is a nonlinear filtering algorithm with excellent performance. It has no restrictions on the state equation, measurement equation and noise statistical characteristics of the system. Therefore, compared with the traditional nonlinear filtering, it has a wider range of applications. Moreover, it breaks through the structural shackles of nonlinear filtering in Kalman filtering system and abandons the idea of estimating the mean and variance of state variables. Instead, sampling particles are extracted from the posterior probability density and iterated to predict and update them to continuously approach the true posterior probability density distribution, which is closer to the essence of the optimal estimation. Particle filter has high filtering accuracy. The advantages of fast convergence have made it more and more effective in dealing with nonlinearity. The main filtering algorithms of state filtering and parameter estimation in non-#china_person0# system. This paper focuses on particle filtering in the background of large azimuth misalignment angle of sins. According to the error angle of Euler platform, the error equation of sins initial alignment under the condition of large azimuth misalignment is derived. Based on the study of Bayesian signal processing and Monte Carlo integral, The sequential importance resampling Importance sampling (Sequential Importance sample) particle filter algorithm is introduced, and the simulation comparison between PF and UKF in the target tracking model is given. Secondly, the shortcomings and shortcomings of the standard particle filter algorithm are discussed. The following improved methods are proposed: first, the priori probability density function is directly selected as the importance density function to sample the particles in the standard particle filter, which results in the loss of measurement information at the latest time. When the likelihood probability density is in a peak state or at the end of a priori probability density function, it is easy to cause particle degradation. Give importance density function purposeful adjustment, correction, Therefore, the importance density function can offset the posterior probability density distribution to the maximum extent. The algorithm designed in this paper uses the mean value and variance after estimation based on EKFU UKFU CKF filter to generate a new sample particle cluster. The comparison and simulation analysis of the PFEK PFU CPF under the piecewise nonlinear model are given. The second is the increase of the number of particles required in the high dimensional state estimation. Therefore, when computing delay and real time are not ideal, a dynamic adjustment of particle number is proposed to reduce the computational load. The specific method is to introduce adaptive technology to the resampling of particle filter. The number of particles needed to estimate the signal at the next time is determined according to the estimation accuracy of the signal at the previous time. Simulation based on the target tracking model is designed to compare the filtering performance of APF and PF. The third is to combine CPF and APF to form ACPF algorithm. The new algorithm not only improves the sampling efficiency by designing the importance density function of CKF, but also predicts the number of particles needed at the next moment according to the estimated state of the previous time. Finally, the effectiveness of ACPF is verified in the flight model of aircraft maneuvering. Finally, the simulation is carried out under the nonlinear error model of initial alignment of sins, and the comparison between ACPF and PF is given. The correctness and validity of the improved algorithm are proved.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN96;TN713

【参考文献】