A DualAttNet-Based Deep Adaptive Kalman Filter in Inertial Navigation

Inertial navigation systems (INSs) offer self-contained and highly autonomous navigation but suffer from unbounded error accumulation over time. The extended Kalman filter (EKF) is widely used to estimate and correct INS errors. However, EKF-INS encounters challenges in accuracy and adaptability caused by modeling assumption errors under complex motion modes, especially in pedestrian INS (PINS). To overcome these limitations, this article proposes a deep adaptive Kalman filter (DAKF) framework enhanced by a DualAttNet module. DualAttNet evaluates measurement confidence from raw inertial data in real time and converts it into a measurement noise covariance matrix. The system continuously performs measurement updates for error correction while adaptively adjusting correction weights to improve robustness. The Kalman filtering equations are embedded into the loss function, enabling end-to-end optimization that implicitly learns the mapping from deep features to noise covariance, thus mitigating the need for manually labeled ground truth. Experiments on a PINS dataset demonstrate that the proposed DAKF achieves real-time inference and reduces average position error to 0.48 m within a 100 m trajectory—50% lower than conventional methods.