基于自适应UKF的BP神经网络及其在高程拟合中的应用

The application of BP neural network based on UKF height fitting

针对BP神经网络收敛速度慢、容易陷入局部极小值点和泛化能力差等问题,基于自适应Kalman滤波理论,提出一种自适应非线性滤波(UKF)训练BP神经网络的方法。该方法采用Kalman滤波框架,引入自适应因子,对神经网络的连接权进行训练,提高了神经网络的学习质量。高程异常拟合算例表明,基于自适应UKF的BP神经网络比标准BP神经网络收敛速度快,泛化能力强,从而证明了该方法是一种有效的连接权训练方法。

BP neural network based on adaptive UKF is introduced in this paper for the standard BP neural network has slow converges,local minimum value and weak generalization ability.It can train weight of BP and improve the efficiency of BP without linearization by using the frame of Kalman filters and adaptive factor to adjust the variance of dynamic model.In the height fitting,BP based on UKF has quicker converges and stronger generalization ability than the standard BP.So it shows that BP neural network based on adaptive UKF is a kind of efficient neural network.