Feature-based GNSS positioning error consistency optimization for GNSS/INS integrated system

The estimated GNSS positioning error, which is commonly represented by measures such as variance or standard deviation, will determine the weight of measurements in the Kalman filter of GNSS/INS integration and significantly affect the integrated navigation results. However, there is a substantial inconsistency between the estimated error provided by GNSS solutions and the actual GNSS positioning error, especially in harsh environments like urban canyons. Previous research has primarily focused on detecting and processing GNSS gross errors to reduce the impact on GNSS/INS integrated systems, while the consistency of the estimated GNSS error has not received much attention. Hence, this work focuses on optimizing GNSS error estimation based on machine learning to improve the consistency with the actual positioning error and the reliability of the GNSS/INS integrated system. An integrated classification and regression tree and bootstrap aggregating (CART-Bagging) algorithm was applied to construct the classification model, and the observation-based features were employed for consistency optimization. Field datasets covering typical urban scenes (e.g., open-sky environment and complex urban environment) over 24 h were collected to assess the accuracy of the GNSS quality control method. The test results showed that the classification accuracy of the estimated GNSS positioning error is more than 90%, and the consistency of GNSS positioning error and the accuracy of integrated positioning are improved by approximately 70% and 30%, respectively.