Augmented Kalman filter/artificial intelligence for inertial sensors/GPS data fusion

For land vehicle applications, in order to obtain an accurate positioning solution in denied Global Positioning System (GPS) environments, low cost Inertial Measurement Units (IMUs) can be integrated with GPS. Kalman Filtering (KF) is usually used to fuse the position and velocity information obtained from both systems. It benefits from the availability of the dynamic mathematical model of the Inertial Navigation System (INS) position, velocity and attitude errors. However, KF requires stochastic models of the inertial sensor errors and a priori information about the covariances of both INS and GPS data. With low cost INS, it is usually difficult to come up with accurate enough stochastic models for each inertial sensor error, which may lead to large position errors. Moreover, nonlinear and non-stationary INS errors become very serious when low cost inertial sensors are utilized.;Alternative INS/GPS integration methods, such as Neural Networks (NN), have received more attention. The main advantage of NN over KF is that it can solve non-linear problems that map input to output data without relying on a priori information. However, NN-based methods do not benefit from the knowledge of the dynamic model of INS errors and are in general computationally expensive.;This thesis augments the KF with NN in order to realize the benefits of both techniques and to improve the overall positioning accuracy. The KF benefits from the INS error model and is capable of removing part of the INS errors. In addition, Radial Basis Function Neural Network (RBFNN) is used as an NN module to model and predict the residual stochastic and nonlinear parts of the INS errors. The ultimate objective of this thesis is to obtain a consistent level of accuracy over relatively long GPS outages (60 seconds) using a low-cost MEMS-based INS integrated with GPS.;Two different architectures of the augmented solution were built; the Position Update Architecture (PUA) and the Velocity Update Architecture (VUA). Both architectures use a non-overlapping window large enough to cover the longest GPS outage. Packets are introduced to counteract the disadvantages of using large window sizes.;In order to validate the effectiveness of the proposed method, several road tests were conducted in Ontario in a land vehicle. The performance of the KF/NN solution was tested with the data collected and it proved to be significantly more effective in reducing the position errors than a standalone KF, especially for relatively long GPS outages that may be experienced in urban canyons and downtown areas.;Keywords: Global Positioning System, Inertial Navigation System, Kalman Filter, Artificial Intelligence, Neural Network, Radial Basis Function Neural Networks...