基于深度Q网络的无迹卡尔曼滤波噪声参数自适应估计方法
An Adaptive Estimation Method for Noise Parameter Matrices of Unscented Kalman Filter Based on Deep Q-Network
倪代天 1孙璐悦 2呼羽 1杜涛1
作者信息
- 1. 北方工业大学信息学院,北京 100144
- 2. 北方工业大学信息学院,北京 100144;中国特种设备检测研究院,北京 100029
- 折叠
摘要
针对一类含有噪声参数不确定性的非线性系统,如何自适应调节噪声参数,实现高精度抗干扰信息融合是亟待解决的问题.信息融合中系统准确的噪声参数是滤波器设计中首先考虑的问题,当实际系统噪声参数与先验信息不一致时,基于卡尔曼滤波无法实现准确的状态估计,导致信息融合性能下降.针对一类噪声参数未知的非线性系统的信息融合问题,借助深度强化学习方法搭建并训练深度Q网络,实现自适应调节系统噪声矩阵大小,进而用于设计扩展卡尔曼滤波器和无迹卡尔曼滤波器,实现抗干扰信息融合.为验证方法的有效性,开展了数值仿真实验.结果表明:基于深度Q网络自适应调节噪声参数使无迹卡尔曼滤波器均方误差保持在0.02 以内,扩展卡尔曼滤波器均方误差保持在 0.6 左右,有效提升了传统卡尔曼滤波器精度,实现了未知噪声参数下的抗干扰信息融合.
Abstract
For a class of nonlinear systems with uncertainty in noise parameter,how to adaptively adjust the noise parameters to achieve high-precision anti-disturbance information fusion is an urgent problem to be solved.In information fusion,the accurate estimation of system noise parameters is the primary consideration in filter design.When the system noise parameters are inconsistent with the prior information,the Kalman filters cannot achieve accurate state estimation,leading to a decline in the performance of information fusion.To address the information fusion problem of a class of nonlin-ear systems with unknown noise parameters,a deep Q-network is constructed and trained with the help of deep reinforce-ment learning methods to adaptively adjust the size of the system noise matrix,which is then used to design extended Kal-man filters and unscented Kalman filters to realize the interference-resistant information fusion.To verify the effectiveness of the method,numerical simulation experiments are conducted.The results show that with the adaptive estimation of the noise parameters based on deep Q-network,the mean square error of the unscented Kalman filter is kept within 0.02 and the mean square error of the extended Kalman filter is kept around 0.6,which effectively improves the accuracy of the tradi-tional Kalman filter and realize the anti-disturbance information fusion under unknown noise parameters.
关键词
干扰估计/信息融合/深度强化学习/卡尔曼滤波Key words
disturbance estimation/information fusion/deep reinforcement learning/Kalman filtering引用本文复制引用
出版年
2024
NETL
NSTL
万方数据