An LSTM-PPO-Based Adaptive Filter Parameter Tuning Method for Tightly Coupled GNSS/INS EKF in Urban Environments

Abstract To improve the robustness of conventional tightly coupled GNSS/INS extended Kalman filtering (EKF) in complex urban environments, this paper proposes an LSTM-PPO-based adaptive filter parameter tuning method (LPAFPTM). In conventional tightly coupled EKF, key parameters are usually fixed or empirically tuned, limiting adaptability under rapidly varying observation conditions. To address this issue, the proposed method reformulates filter parameter tuning as a sequential decision-making problem while preserving the standard INS prediction-GNSS update framework. A compact decision state is constructed from internal filter statistics to jointly characterize GNSS measurement quality, satellite geometry, residual variation, and the consistency between INS mechanization and filter estimation. Based on this state, a discrete action space is designed to dynamically adjust the measurement noise covariance matrix through adaptive scaling of the measurement error model, while also regulating the prior uncertainty of the INS attitude states. Furthermore, an LSTM network is incorporated into the PPO framework to capture temporal dependence and error accumulation in navigation states. A two-stage offline training-online deployment strategy is adopted to support real-time implementation. Experiments on two vehicular datasets in complex urban environments compare EKF, REKF, RAEKF, PAFPTM, and LPAFPTM over 10 matched evaluation realizations, with PAFPTM and LPAFPTM examined at training episode lengths of 50, 100, and 150 steps. At 50, 100, and 150 steps, the three-direction average RMSE values of LPAFPTM are 0.0240, 0.0237, and 0.0239 m on CUMT and 0.4826, 0.4787, and 0.4873 m on Tokyo, respectively. When averaged over the three training settings, LPAFPTM gives the lowest three-direction average RMSE among the five methods on both datasets. Compared with the PPO-based method, LPAFPTM further improves overall performance, demonstrating that reinforcement learning-based dynamic covariance adaptation with temporal modeling is effective for tightly coupled GNSS/INS navigation in challenging urban scenarios.