Moving Magnetic Target Localization Using Automatically Initialized KalmanNet Driven by Euler Deconvolution

The task of moving magnetic target localization involves recovering hidden states from noisy, nonlinear magnetic field observations. While nonlinear filters like the extended Kalman filter (EKF) and the unscented Kalman filter (UKF) are effective, they depend heavily on knowledge of underlying noise statistics. Besides, a lack of prior knowledge can lead to poor accuracy and robustness in state estimation. To address these two limitations, this letter proposes a novel filtering algorithm that automatically estimates the initial state value. An initial value estimator is created using three-axis magnetic scalar gradient data and the Euler deconvolution to quickly determine the target’s initial state. This value is then used as prior knowledge in a new neural network-aided Kalman filter called KalmanNet for state prediction and updating. Unlike traditional methods that transform nonlinear equations into linear forms, the proposed method does not require specific assumptions, preserving target characterization generality. Furthermore, this method provides a reliable initial state value while not relying on statistical noise knowledge, enhancing estimation accuracy. Simulation and field experiments show that the estimation results closely align with true values. In the field experiment, the root mean square error (RMSE) values for position, velocity, and magnetic moment vectors were reduced by 69.21%, 86.30%, and 60.55% compared to the EKF, and by 75.49%, 83.74%, and 80.45% compared to the UKF.