基于UKF的神经网络自适应全局信息融合方法

Study on Nonlinear Optimal Estimation for Neural Networks Data Fusion

为了解决神经网络(NN)在数据融合过程中权值实时更新问题,依据神经元激活函数的非线性特点,提出了一种利用Unscented卡尔曼滤波(UKF)实现神经网络权系数自适应调整的模型及方法,从而使全局融合信息最优.并分别以仿真数据及DGPS/GPS/RLC/罗经等设备组成的舰船导航系统实测数据为例,首先对各局部滤波器进行UKF滤波,然后分别利用神经网络卡尔曼滤波(NNKF)及神经网络非线性卡尔曼滤波(NNUKF)进行数据融合,仿真和试验结果表明,所提方案对提高整个系统的精度和运算速度是行之有效的.

A new model and algorithm to realize adaptive adjustment of the weights of NN and to make global fusion information optimal were presented. The method utilizes Unscented Kalman filter （UKF） for nonlinear optimal estimation to solve the problem that weights of neural networks are not be on-line trained in data fusion. Applies the above project to a multi-sensors vessel integrated navigation system, obtains actual data from the integrated navigation system of DGPS/GPS/RLC/compass. First, using UKF methods estimates and filters the location information, then, NNKF and NNUKF are used to fuse them. The results of experiment and simulation show that the proposed approach is very useful for improving the accuracy and calculation speed of the system.