A Novel Robust Deep Learning Aided Kalman Filter for State Estimation

Given the performance degradation of the Kalman filter due to unknown and time-varying process and measurement noise, this paper proposes a novel robust state estimation method based on system state-space models and deep learning. The objective is to enhance the accuracy and robustness of traditional Kalman filtering in complex noise environments. Specifically, we employ two deep neural networks to learn the process and measurement noise covariance matrices from data, capturing their time-varying characteristics. These learned matrices are then embedded into the Kalman filter framework to improve the precision and robustness of state estimation while preserving the interpretability of the system. A series of simulations and dataset experiments validate the proposed method's superior performance in terms of accuracy, convergence, and robustness to noise variations.