Neural network augmented tightly coupled Kalman filter for low-cost reduced inertial navigation sensor and GPS integration

Current navigation systems rely upon an integrated Inertial Navigation System (INS) and a Global Positioning System (GPS) to determine vehicular position, velocity and attitude. Integrated INS/GPS modules have complementary characteristics that provide greater navigation accuracy over stand-alone systems. It has been common practice to utilize Kalman filtering (KF) as the integration technique to combine INS and GPS observations. The traditional KF architecture integrates INS and GPS in a loosely coupled fashion, which is easier to implement but have the major drawback of requiring at least four GPS satellites. Any less than four satellites becomes a GPS outage where the KF operates in prediction mode relying mostly on the INS error model. Further development has led to tightly coupled KF. In a tightly coupled system, the GPS raw measurements (i.e. pseudorange) are sent directly to a centralized KF. The advantage of a tightly coupled system is that GPS measurements are used to assist KF even if only one satellite is present.;The objective of this thesis is to explore a new hybrid AI/KF tightly coupled architecture to provide robust 2D positioning solution for low-cost sensors in challenging GPS environments. The method developed augments a tightly coupled KF with a Radial Basis Function Neural Network (RBFNN) to improve the integration of reduced inertial sensor system (RISS) measurements and GPS observations in order to realize the benefits of both techniques and improve the overall positioning accuracy. Several road test trajectories conducted in Ontario were utilised to examine the effectiveness of the proposed method. The performance of the hybrid KF/NN architecture proved to be significantly more effective in reducing the position errors in certain situations during 60 second partial GPS outages over that of a standalone KF. This thesis research advances the knowledge in the area of tightly coupled INS/GPS integration by involving robust hybrid fusion approaches relying on both KF and AI.;Keywords: Global Positioning System, Inertial Navigation System, Kalman Filter, Neural Network, Radial Basis Function, Tightly Coupled;KF has several shortcomings. These include the necessity of accurate stochastic models of the inertial sensor errors, a priori information of both INS and GPS used and the linearization of the INS dynamic errors. These shortcomings have motivated INS/GPS integration methods based on Artificial Intelligence (AI) techniques. Augmented AI/KF techniques have demonstrated an increase in navigation accuracy over an AI or KF standalone architecture on loosely coupled systems. However, as of this research, there have been no hybrid AI approaches towards a hybrid tightly coupled system, where the advantage of both KF and AI techniques can be brought together to provide one robust solution, processing the raw INS and GPS measurements.