基于Kalman／UKF组合训练神经网络的初始对准方法

针对基于Unscented卡尔曼滤波(UKF)的神经网络训练学习方法存在的计算量大,实时性差的问题,提出了一种基于Kalman/UKF组合滤波原理的神经网络学习方法,该方法综合了Kalman滤波对线性系统和UKF对非线性系统的最优估计的优势,在保证神经网络权值估计精度的同时,有效降低了神经网络权值学习的计算量,提高了神经网络训练的实时性。最后将该利用方法训练的神经网络应用于惯性导航系统的非线性初始对准过程中,并进行了仿真研究。仿真结果表明利用提出的算法训练的神经网络与基于UKF训练的神经网络具有相同的对准精度和实时性,而提出的算法的有效降低了神经网络训练的计算量,提高了训练的运行效率,是解决惯性导航系统初始对准的一种有效和实用的方法。

Initial Alignment Method Based on Kalman/UKF Integrated Training Neural Network

The Unscented Kalman Filter (UKF) was studied as a state estimation method for the nonlinear system and was used to train multilayered neural network by augmenting the state with unknown connecting weights.Whereas the computer time of UKF depended on the dimension of the inertial navigation system model state vector.Any more number of states would take leave of real time.A learning algorithm for multiplayer neural network based on the Kalman/UKF integrated filter was studied.The theoretical procedure of the algorithm was described in detail.Then,it was used to the nonlinear initial alignment of inertial navigation system.Simulation results prove the availability of the neural network algorithm for nonlinear initial alignment of inertial navigation system.Not only can surely alignment accuracy and alignment time be obtained,which is similar to that of the UKF,but also the computation time is reduced considerably.Consequently,an available algorithm of neural network for the nonlinear initial alignment of inertial navigation system is discovered.