Improving Adaptive Kalman Filter in GPS/SDINS Integration with Neural Network

Kalman filter (KF) can provide optimal solutions if the system dynamic and measurement models are correctly defined, and the noise statistics for the measurement and system are completely known. The conventional way of determining the covariance matrices of process noise and observation errors relies on analysis of empirical data from each sensor in a system, which is called KF tuning. In practice, however, the process noise and observation errors vary with time and environment, which causes uncertainty in the covariance matrices of process noise and observation errors and results in system performance degradation. Adaptive KF (AKF) has been intensively investigated, which can tune a filter continuously so as to eliminate empirical data analysis and aims to improve filtering performance. The covariance matching technique in AKF uses innovation-based estimation that attempts to make the filter residual covariances consistent with their theoretical covariances estimated with samples. This paper presents a neural network aided AKF based on covariance matching technique, for integrated GPS/INS system. Instead of using a limited window for estimation as conventional AKF, all the previous samples are counted in according to their character using neural network (NN). The covariance matching is conducted then its relation with the corresponding character is mapped with the NN. The adjustment of the AKF is based on both the NN training result and the updated covariance matching result. The purpose of doing so is to eliminate estimation noise, and to keep the selected samples ergodic. The objective of this research is to develop a system that is self-adaptive to the change of operation environment or hardware components, such as the type of INS and system configuration etc. with the help of AKF. The principle of this hybrid method and the NN design are presented. Field test data are processed to evaluate the performance of the proposed method. Different types of INS are tested to demonstrate the effectiveness of the proposed adaptive algorithm.