Improving GPS-denied navigation through temperature error compensation in atomic gyroscopes and accelerometers using a hybrid adaptive Kalman Filtering–Machine learning approach

This paper investigates the impact of temperature variations on inertial navigation performance in GPS-denied environments by focusing on atomic gyroscopes and accelerometers. Temperature fluctuations introduce bias drift and degrade system stability, which significantly affects long-term navigation accuracy. To address this challenge, a hybrid framework combining Adaptive Kalman Filtering (AKF) and Machine Learning (ML) is proposed, in which a physics-informed temperature-dependent error model is developed using laboratory thermal characterization data. In the proposed approach, a hybrid dataset consisting of field-collected INS/GNSS measurements and laboratory-based thermal calibration data is utilized to construct a physics-informed emulated atomic inertial sensor model. The field data provide real navigation trajectories under practical operating conditions, while the laboratory measurements are used to identify temperature-dependent sensor error characteristics. These components are integrated to emulate the behavior of atomic gyroscopes and accelerometers under varying thermal conditions. First, slowly varying bias components are estimated and mitigated using an Adaptive Kalman Filter (AKF) to reduce accumulated inertial errors. Then, a machine learning module is employed to capture nonlinear and time-varying temperature-induced effects that are not fully addressed by classical filtering techniques. By learning from the combined dataset, the model establishes a mapping between temperature variations and sensor error dynamics. The proposed framework is evaluated using real navigation measurements from the INS/GNSS system together with the emulated atomic sensor model. Results demonstrate significant improvements in navigation performance under varying thermal conditions. The hybrid AKF–ML approach effectively reduces temperature-induced drift and enhances overall estimation accuracy across velocity, acceleration, angular rate, and attitude states.