Neural Network Aided Information Filtering for Model Uncertainty

The extended information filter (EIF) effectively addresses state estimation problems in the information form by utilizing information matrices. Unlike the extended Kalman filter (EKF), the EIF provides improved robustness to initialization. However, its performance can deteriorate significantly in the presence of a mismatched system model or unknown noise statistics. To address this issue, this paper proposes a neural network aided information filter (NNAIF) to reduce the reliance of EIF on accurate model information and noise statistics. Specifically, two specially designed lightweight neural networks are embedded into the prediction and update steps in the EIF framework to replace the components that are most sensitive to modeling errors and statistical uncertainties. This design not only preserves the interpretability and structural advantages of EIF but also introduces data-driven learning capabilities to effectively compensate for model inaccuracies and unknown noise characteristics. In addition, by incorporating a width-preserving feature transformation, NNAIF reduces parameter size and enhances tolerance to learning rate settings without compromising its modeling capacity. Simulations including state estimation of lithium batteries validate the effectiveness of the proposed NNAIF under both model-matched and model-mismatched conditions, and confirm its robustness to initialization.