混合估计自适应平滑算法研究

【作者】 贾宇岗；

【导师】 潘泉； 张洪才；

【作者基本信息】 西北工业大学 ， 模式识别与智能系统， 2003， 硕士

【摘要】 本文对自适应估计领域中的混合估计算法进行了较为深入，系统的研究，主要工作如下： 1．对混合估计的应用背景和发展现状做了简要概述，着重介绍了混合估计的主流算法—交互式多模型算法(IMM)。并采用Monte Carlo方法对算法采用正态分布来近似描述每个滤波模型的似然函数作初步的验证。仿真结果表明，IMM中各子滤波器滤波残差基本服从正态分布，但其均值随滤波模型与系统实际运动模式匹配程度变化而变化，即二者匹配时，滤波结果无偏；不匹配时，滤波结果有偏。 2．比较了目前已有的几个混合估计平滑算法；并提出了一种新的基于系统扩维的任意步固定滞后平滑算法。该算法将IMM算法应用于系统状态和模型概率同时扩维的系统，能够实时计算模型概率平滑值，为实时判断系统模式切换提供依据，并弥补了Chen算法的任意步固定滞后平滑算法的理论缺陷。 3．在两种典型的仿真环境下比较了IMM算法和固定滞后平滑算法的精度；以及本文提出的固定滞后平滑算法和Chen的固定滞后平滑算法的精度。仿真结果表明本文所提出的新算法的精度优于Chen算法，其一步滞后平滑精度与Chen算法两步滞后平滑精度相当。 4．为了满足有些实时性要求很强的系统的要求，提出了一种多步迭代的实时IMM算法，仿真表明利用该方法是可行的。 5．提出了一种基于Unscented Kalmal Filter(UKF)的交互式多模型算法，该算法结合了IMM算法对系统结构突变的跟踪能力和UKF算法对系统状态或量测方程非线性的处理能力，并在二维非线性机动目标跟踪中显示了其比基于EKF的IMM算法的优越性。 6．分析了目前广泛应用于非线性非高斯系统状态估计的“粒子”滤波(Particle Filter)算法的基本思想，指出其存在的问题和可能的研究方向。更多还原

【Abstract】 This thesis is focused on several fundamental problems of hybrid estimation techniques, which becomes a most important field of adaptive filtering in the recent two decades. The main contributions are as follows:1 . Interacting Multiple Model Algorithm (IMM) has been demonstrated to be the most cost-effective algorithm in hybrid estimation. The transient process of IMM is analyzed via Monte Carlo simulation. Simulation results shows that the residual of each filter in IMM is still Gaussian distribution, but its mean is not zero if the dominating filter cannot match the real target model.2. After reviewing several smoothing algorithms for hybrid estimation ,we presented a sub optimal approach to the d step fixed-lag smoothing problem for Markovian switching system by applying the basic IMM structure to the system with augmented system state and mode probability. The new fixed-lag smoothing. algorithm can provide us with the smoothed model probability that can be used to judge the system jump or abruptly change.3. The smoothing algorithms are compared in two typical target tracking simulation examples. Simulation results show that a significant improvement on the filtering algorithm is achieved by smoothing algorithms at the cost of a small time delay. The new smoothing algorithm proposed in this thesis is superior to other smoothing algorithms.4. A new enhanced IMM algorithm based on IMM smoothing algorithm is presented. Simulation results show that this new algorithm is superior to standard IMM algorithm without time delay.5. The mechanism of Unscented Kalman filter (UKF) is analyzed and a new multiple model estimator based on UKF and IMM is proposed for maneuver target tracking with non-linear measurements. Simulation results show the superiority of this new algorithm in dealing with such systems.6. We discuss a new popular approach, which is called particle filter, to state estimation problem for non-linear, non-Gaussian system. In addition, some potential problems of the particle filter have been presented.更多还原