Abstract: Aiming at the defects in the training process of BP neural networks, a novel learning algorithm of neural networks is designed based on the improved fading unscented Kalman filtering in this paper. In this algorithm, the filter gain matrix in the algorithm of UKFNN is adjusted by introducing the adaptive factor, which is calculated by an improved calculation method. Therefore, the influence of sample noise on the weights updating is limited and thus the training accuracy is improved. Meanwhile, the calculation burden of fading factor could be simplified by the improved calculation method. Finally, the proposed algorithm is applied to INS/GPS integrated navigation system for establishing an error estimation model. The experiment results demonstrate that the fitting precision of the prediction model could be advanced by the proposed algorithm, and the adaptive ability for noise samples could be improved effectively, as well as has better application prospects.

Key words: Neural networks, unscented Kalman filter, fading factor, integrated navigation error estimation model

摘要： 针对BP神经网络在训练过程中存在易陷入局部极小值、对样本噪声缺乏自适应性等问题，设计一种改进的渐消无迹卡尔曼滤波神经网络(Improved fading unscented Kalman filter neural networks, IF-UKFNN)训练算法。该算法以UKF为基础，在滤波过程中引入渐消因子，实时调整滤波增益，限制样本噪声对权值更新的影响，进而提高网络训练的精度；同时，采用一种改进的自适应因子计算方法，使计算过程简化。将提出的算法应用于组合导航系统建立误差估计模型，仿真结果表明：提出的算法不仅可以提高模型的估计精度，而且增强了网络模型对噪声样本的适应性和鲁棒性，具有更好的应用效果。

关键词: 神经网络, 无迹卡尔曼滤波, 渐消因子, 组合导航误差估计模型