文献阅读“Precise positioning of smartphones using a robust adaptive Kalman filter” ,2023,ION,Anurag Raghuvanshi
一、基本KF滤波器模型
二、自适应滤波模型
GNSS信号的变化本质上对应的是噪声模型的变化,包括观测噪声Rk和状态噪声Qk。
因此,考虑自适应调整上面两个噪声项,通过滤波增益Kk调整滤波器模型。
三、模型参数设计
In the current study, standardized measurement residual is used to scale measurement covariance and ratio of variance is used for prediction covariance scaling. 即使用标准化之后的观测量残差(公式13),和方差比例(公式14)分别对Rk和Qk进行自适应调整。
这里的ri是观测量残差,Qri是预测的状态量方差。
xk是估计值和预测值的差,γ是所有卫星的观测量残差向量,n是观测量数目,Qxk是残余状态向量的方差矩阵,Qγ是观测量残余的方差矩阵,m是状态量数目。
author注1:xk、γ、Qxk和Qγ的物理含义到底是什么?
author注2:理论上归一化之后的ri~N(0,1)?
The learning statistics are used to form an adaptive factor and different types of adaptive factors have been explained in Yang (2010); Yang & Gao (2005); Vana & Bisnath (2023). In the current research, a 3- segment adaptive factor is used for measurement scaling as the measurements are divided into three regions, good measurements, measurements that need be weighted and outliers that need to be removed. The predicted covariance is scaled using an exponential adaptive factor, as the learning statistic is used to decide whether dynamic model adaptiveness is required or not. The two adaptive factors are defined by Equations (15) and (16).
观测噪声权重β采用的是三段模式,分为:可用,降权使用,不可用。
状态噪声权重仅在残差较大,即观测量降权使用时打开,注意,这里的ri和c0与β中的两个参数含义相同。
author注3:状态噪声越大,α越小,理论上α是可以大于1的?大于1意味着把现有的状态噪声减小,合理吗?
author注4:根据实际经验,Q和R对权重K的影响是协同的,同时调整两个参数是否存在作用抵消的可能,以及,是否有引起滤波器震荡的风险,需要评估。