Abstract:Kalman filter (KF) is the main algorithm in the field of optimal estimation. KF can realize the optimal estimation only under the premise that both the function model and the random model are known. In most practical applications, random models are often unknown, and unknown or biased random models will make filtering suboptimal. Besides, the gross errors will cause the innovation to appear abnormal, making the innovation statistics abnormal. The abnormal innovation statistics will lead to the deviation of the estimated noise covariance information. Therefore, when there are gross errors, how to ensure the random model's correct estimation is a major problem faced by the current random model estimation method. To solve this problem, we propose a robust autocovariance least-squares noise covariance estimation algorithm (RALS). In the new method, we first introduce the robust Kalman filtering method (RKF) based on the chi-square test to resist the influence of abnormal innovation on subsequent epoch state estimation; Then, based on the basic idea of the correlation robustness method, we constructed the correlation robust innovation function sequence model, taking into account the correlation between the innovation sequences, to repair the abnormal innovation statistics, so as to obtain accurate post-test innovation statistics; Subsequently, we use the autocovariance least square method to eliminate the coupling effect between the two types of noise covariance information, to correctly estimate the two types of noise covariance information; Finally, we adopt an iterative strategy to eliminate the influence of a priori random model deviation, to realize the separation of the gross error coupled in the abnormal innovation and the prior random model deviation. Two experimental results show that: compared with the KF, RKF, and ALS methods, the new method has higher noise covariance estimation accuracy and filtering accuracy.

Recurrent Neural Network Based Kalman Filtering: A Suboptimal Approach

RKFNet: A Novel Neural Network Aided Robust Kalman Filter

KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics

Nonlinear Kalman Filter Using Cramer Rao Bound

RNN-UKF: Enhancing Hyperparameter Auto-Tuning in Unscented Kalman Filters Through Recurrent Neural Networks

A Novel Kalman Filter for Nonlinear Non-Gaussian Systems and Its Applications in Self-localization

Optimal Sequential Kalman Filtering with Cross-Correlated Measurement Noises

Performance Evaluation of Deep Learning-Based State Estimation: A Comparative Study of KalmanNet

Structural Dynamic Response Reconstruction Based on Recurrent Neural Network–Aided Kalman Filter

Adaptive Kalman Filter with LSTM Network Assistance for Abnormal Measurements

Extended Kalman Filter Using A Kernel Recursive Least Squares Observer

Kalman Filtering Based on Dynamic Perception of Measurement Noise

Neural network-aided variational Bayesian adaptive cubature Kalman filtering for nonlinear state estimation

A redundant adaptive robust filtering algorithm based on cubature Kalman fliter

Switching Recurrent Kalman Networks

Optimal robust non-fragile Kalman-type recursive filtering with finite-step autocorrelated noises and multiple packet dropouts

EKFNet: Learning System Noise Statistics from Measurement Data

Iterative Learning Kalman Filter for Repetitive Processes

Robust autocovariance least-squares noise covariance estimation algorithm

Adaptive KalmanNet: Data-Driven Kalman Filter with Fast Adaptation

An Efficient Estimation Method for Dynamic Systems in the Presence of Inaccurate Noise Statistics