基于反向传播神经网络改进的增益修改卡尔曼滤波算法

Improved modified gain extended Kalman filter algorithm based on back propagation neural network

增益修改的卡尔曼滤波(MGEKF)算法在实际应用时,一般使用带有误差的测量值代替真实值进行增益修正计算,导致修正结果也被误差污染。针对这一问题,提出一种基于反向传播神经网络(BPNN)改进的MGEKF算法,该算法使用训练后的神经网络代替MGEKF的增益修正函数。该算法在网络训练阶段,以实际测量值作为神经网络的输入,真实值修正后的结果作为训练目标;在实际应用中,使用网络的输出修正卡尔曼增益。针对移动单站只测向目标定位问题进行了实验,实验结果表明:该算法与扩展卡尔曼滤波(EKF)、MGEKF、平滑增益修改的卡尔曼滤波(sMGEKF)算法相比:定位精度至少提升10%,并且有更强的稳定性。

In practical application,Modified Gain Extended Kalman Filter（ MGEKF） algorithm generally uses erroneous measured values instead of the real values for calculation,so the modified results also contain errors. To solve this problem,an improved MGEKF algorithm based on Back Propagation Neural Network（ BPNN）,termed BPNN-MGEKF algorithm,was proposed in this paper. At BPNN training time,measured values were used as the input,and modified results by true values as the output. BPNN-MGEKF was applied to single moving station bearing-only position experiment. The experimental results shows that,BPNN-MGEKF improves the positioning accuracy of more than 10% compared to extended Kalman filter,MGEKF and smoothing modified gain extended Kalman filter algorithm,and it is more stable.