An accurate nonlinear stochastic model for MEMS-based inertial sensor error with wavelet networks
An accurate nonlinear stochastic model for MEMS-based inertial sensor error with wavelet networks
复制标题
DOI：
10.1515/jag.2007.022
复制
发表时间：
2007-12-01
影响因子：
1.4
通讯作者：
Pagiatakis, Spiros
中科院分区：
文献类型：
--
作者：
El-Diasty, Mohammed;El-Rabbany, Ahmed;Pagiatakis, Spiros
全球定位系统（GPS）与惯性导航系统（INS）的组合已经广泛地用于定位和定向目的的许多应用中。传统上，随机游走（RW），高斯-马尔可夫（GM）和自回归（AR）过程已被用来开发经典卡尔曼滤波器中的随机模型。经典卡尔曼滤波的主要缺点是非线性动态系统的潜在不稳定线性化。因此，由于预期的线性化误差，非线性随机模型在基于导数的滤波器中不是最优的。采用无导数滤波器（Unscented Kalman Filter或Divided Difference Filter），可以对复杂的高度非线性动态系统进行无线性化误差的滤波处理。提出了一种基于小波网络的惯性传感器误差非线性随机模型。小波网络是一种高度非线性的模型，近年来已被引入作为建模和预测的有力工具。静态和动态数据集收集使用基于MEMS的IMU（DQI-100）开发的随机模型在静态模式，然后在运动模式下实施。使用GM、AR和提出的基于小波变换的过程的基于无导数的滤波方法来验证新模型。结果表明，一阶小波基的非线性随机模型给出了上级定位结果的一阶GM和AR模型与30%的整体改善时，30和60秒的GPS中断。
The integration of Global Positioning System (GPS) with Inertial Navigation System (INS) has been widely used in many applications for positioning and orientation purposes. Traditionally, random walk (RW), Gauss-Markov (GM), and auto-regressive (AR) processes have been used to develop the stochastic model in classical Kalman filters. The main disadvantage of classical Kalman filter is the potentially unstable linearization of the nonlinear dynamic system. Consequently, a nonlinear stochastic model is not optimal in derivative-based filters due to the expected linearization error. With a derivativeless-based filter such as the unscented Kalman filter or the divided difference filter, the filtering process of a complicated highly nonlinear dynamic system is possible without linearization error. This paper develops a novel nonlinear stochastic model for inertial sensor error using a wavelet network (WN). A wavelet network is a highly nonlinear model, which has recently been introduced as a powerful tool for modelling and prediction. Static and kinematic data sets are collected using a MEMS-based IMU (DQI-100) to develop the stochastic model in the static mode and then implement it in the kinematic mode. The derivativeless-based filtering method using GM, AR, and the proposed WN-based processes are used to validate the new model. It is shown that the first-order WN-based nonlinear stochastic model gives superior positioning results to the first-order GM and AR models with an overall improvement of 30% when 30 and 60 seconds GPS outages are introduced.