摘要
针对系统误差的不确定性可能会引起滤波精度降低或发散的问题,提出一种新的基于模型预测滤波的前向神经网络算法。首先,采用模型预测滤波估计网络在正向传递过程中的模型误差,并对其进行修正,以弥补模型误差对隐含层权值更新的影响;然后,利用模型预测滤波为神经网络提供精确的训练样本,学习待估计系统的非线性关系。将提出的算法应用于SINS/CNS/BDS组合导航系统,并与扩展卡尔曼滤波进行比较,仿真结果表明,提出的算法得到的姿态误差、速度误差和位置误差分别在[-0.25′,+0.25′]、[-0.05 m/s,+0.05 m/s]和[-5 m,+5 m]以内,滤波性能明显优于扩展卡尔曼滤波算法,表明该算法能提高组合导航定位的解算精度。
In view that the uncertainty of system errors may reduce the filtering precision or cause the problem of filter divergence, this paper proposes a novel feed-forward neural network algorithm based on model predictive filtering. First, a model predictive filtering (MPF) is used to estimate the neural network model errors in the forward transfer process and revise it to resist the impact on weights updating. Then, the MPF is used to provide accurate learning samples to neural network for approximating the nonlinear relationship. Finally, the proposed algorithm is applied to the SINS/CNS/BDS (SINS/celestial navigation system/BeiDou navigation satellite system) integrated navigation system and compared with extended Kalman filter (EKF). Simulation results demonstrate that the attitude angle error, velocity error and position error obtained by the improved neural network algorithm are within [-0.25', +0.25'], [-0.05 m/s, +0.05 m/s] and [-5 m, +5 m], respectively, and the filtering performance is significantly superior to that of the EKF. The results show that the proposed algorithms effectively improve the positioning precision of the navigation system.
出处
《中国惯性技术学报》
EI
CSCD
北大核心
2014年第2期221-226,共6页
Journal of Chinese Inertial Technology
基金
国家自然科学基金(61174193)
中国航天科技集团公司卫星应用研究院开放基金资助