Self-Learning Position Variance Using Crowdsourcing Data for Adaptive GNSS/SINS Integrated Navigation

A reasonable stochastic model, e.g., the pre-assigned a priori variance information determines the weight of measurements in the Kalman filter of GNSS/SINS and significantly affects the integrated navigation results. However, due to the nonlinear model, non-Gaussian noise, unmodeled error, and gross errors, GNSS may produce over-optimism estimation which cannot reflect the actual error and significantly affect the performance of GNSS/SINS solution. Hence, this work utilizes crowdsourcing datasets and data-driven learning approaches to resolve the inconsistency between estimated variance and actual errors in GNSS. A CNN-LSTM-Attention (CLA) model is constructed to enable long-time step propagation and resilient fusion of features. Furthermore, by leveraging the properties of IMU and the RTS algorithm, the label data can be self-generated, eliminating the reliance on the ground truth system. The CLA model is trained using GNSS features, enabling self-learning of position variance. Lastly, an adaptive fusion of GNSS/SINS is achieved through the semi-tightly coupled integration scheme, and crowdsourced data is employed for the continuous optimization of the CLA model. The test results showed that the consistency of GNSS position variance is improved and the CEP95 of position errors is improved by 64.33% compared to the conventional method.