<p>This study addresses the persistent challenge of localization drift in deep-sea mining vehicles (DSMV) caused by prolonged operational duration and inertial sensor bias. Unlike environmental disturbances such as underwater currents or electronic noise, these systematic errors gradually accumulate, resulting in significant deviation from true trajectories. To tackle this issue, we propose a novel algorithm that synergizes the Extended Kalman Filter (EKF) with Long Short-Term Memory (LSTM) neural networks, hereafter referred to as EKF-LSTM. In this architecture, the EKF suppresses short-term sensor noise and handles linear state estimation, while the LSTM captures long-term temporal dependencies in residual errors derived from the Inertial Navigation System (INS). The combined model offers robust drift correction by fusing physical motion models with learned error dynamics. Simulation experiments conducted on the Gazebo platform using a Clearpath Husky ground vehicle demonstrate that EKF-LSTM outperforms traditional EKF and standalone LSTM approaches, particularly under high-noise, nonlinear motion conditions. Compared to EKF alone, the proposed EKF-LSTM reduces Mean Squared Error (MSE), Mean Absolute Error (MAE), and Maximum Absolute Error (Max AE) by 33.8%, 15.1%, and 17.0%, respectively. These results highlight the method’s potential for accurate underwater localization even in the absence of acoustic positioning systems, offering a promising error correction framework for DSMV navigation.</p>