Multiple Interactive Model for MEMS IMU in GPS/INS Integrated Navigation System

摘要

Enhancing the accuracy of navigation systems that includes MEMS based IMU by GPS aiding through Kalman filter (KF) estimator is a well-known accepted technique. The performance of overall system is hardly dependent on the IMU errors during GPS outage periods. The system accuracy is dramatically degraded, during GPS outages, in case of using MEMS-based IMU if there is no accurate modeling used to estimate the IMU errors in denied GPS environment. Kalman filter has become a durable approach for estimating the linear sensor errors while the inherent, non-stationary and nonlinear errors of MEMS-based inertial sensors are still a serious problem. Modeling of higher order inertial sensor errors has been a challenging task to mitigate the navigation system errors. Many researchers suggested using KF augmented by a modeling algorithm to virtually aid the system during GPS outage. Some researchers model the overall system error during the GPS availability using the inertial sensor readings to estimate this error during GPS absence. Others kept the KF as linear errors estimator and proposed different modeling techniques such as fast orthogonal search (FOS), parallel cascade identification (PCI), and AI-based modeling techniques to model the nonlinear error part. FOS is one of the most competitive algorithms since it is a general purpose numerical method and gives challenging results. It is able to efficiently sense the system nonlinearity with a tolerance of arbitrary stochastic system noise. FOS has been used in different approaches to model the nonlinearity of the inertial sensors where the difference key in these approaches is the candidates used in the modeling process. One approach models the nonlinear azimuth error by using the azimuth linear error delivered from KF as modeling candidates. Some researchers used the whole KF state vector which includes position, velocity and attitude linear errors as combined candidates to model the nonlinearity in azimuth errors. Another suggestion was using vehicle state with current sensor readings as candidates to model the inertial system errors in position, velocity, and attitude. After building the model, most of the proposed KF augmentation approaches omit that the prediction process depends on the KF output. This KF output is already suspicious due to its dependence on GPS information during its update process and this dependency may affect the system performance. In this paper, we propose a combined solution of two parts to isolate the KF output effect on the prediction result. The first part employs the autoregressive (AR) concept using FOS for azimuth nonlinear error modeling. During GPS availability, the resultant nonlinear azimuth error (difference between the sensor reading and linear error estimated by KF) is delivered to the FOS algorithm to be modeled as an AR model. In this model, the nonlinear azimuth error, at certain instant, is calculated in terms of previous nonlinear error terms. The obtained model is used during GPS outage to predict the current azimuth nonlinear error. The second part uses the support vector machine (SVM) approach to model the GPS outputs (position and velocity) during GPS availability using the outputs of the inertial mechanized as candidates. This model is used in GPS denied conditions to predict the GPS position and velocity information which are delivered to the KF to virtually imitate the GPS availability. Using this combined approach enables to segregate between predicting the linear and nonlinear azimuth errors during GPS outages. In addition, the prediction process is independent on the GPS reading, which eliminates the suspicion in the KF output during GPS outage. The proposed solution gives promising results, where it reduces the position and azimuth errors better than the usage of KF only. Moreover, it gives better results in the modeling process, where the obtained models are more accurate when compared to reference data.