Modeling of INS Position and Velocity Errors Using Radial Basis Function Neural Networks for INS/GPS Integration

L. Semeniuk and A. Noureldin

Abstract:	As the global demand for a highly affordable, reliable and accurate positioning system increases, techniques capable of providing a reliable back-up to the commonly used Global Positioning System (GPS) are continually being developed. Presently, Kalman filter (KF) is widely used to combine GPS information with Inertial Navigation System (INS) data, overcoming each of their individual shortcomings and providing a reliable navigation solution. Unfortunately, tailoring the Kalman filter for use with different systems and for different applications can be very challenging as precise stochastic modelling of the sensor errors and the fine-tuning of many parameters are required to achieve the ideal system performance. The goal of this research is to develop an algorithm that is a suitable replacement for the KF for INS/GPS integration, thus providing metre-level positioning accuracy and being applicable for use with both navigation and tactical grade sensors. As such, a position and velocity update architecture utilizing artificial neural networks, called the Artificial- Intelligence-based Segmented Forward Predictor or ASFP, is proposed. This technique is simple to apply to new sensors, requiring little or no prior knowledge of the sensor characteristics. The use of Radial Basis Function Neural Networks (RBFNN) was chosen since this type of neural network has few tuneable parameters and can therefore be implemented in a simple and standard manner. Along each of the North, East and Vertical directions, two RBFNNs are used for updating both the INS velocity and position. During the update procedure, the RBFNN velocity and position networks are trained using the INS and GPS data. These neural networks are then employed during GPS outages to process current INS information to estimate the upcoming INS velocity and position errors. By correcting the INS data, a more accurate solution during GPS outages is obtained. For the purpose of real-time implementation, this approach has been applied using each of the non-overlapping and sliding window techniques. The performance of the proposed INS/GPS integration method was tested using real navigation data acquired from an Ashtech Z12 GPS receiver and a Honeywell LRF-III navigation-grade INS mounted inside a land vehicle. The algorithm was tuned for reliable use with these sensors and proficiently corrected for multiple outages simulated at varying locations along the trajectory. The results indicate that a positioning accuracy similar to that found using KF is achieved while using ASFP applied with a non-overlapping window.
Published in:	Proceedings of the 2006 National Technical Meeting of The Institute of Navigation January 18 - 20, 2006 Hyatt Regency Hotel Monterey, CA
Pages:	479 - 489
Cite this article:	Semeniuk, L., Noureldin, A., "Modeling of INS Position and Velocity Errors Using Radial Basis Function Neural Networks for INS/GPS Integration," Proceedings of the 2006 National Technical Meeting of The Institute of Navigation, Monterey, CA, January 2006, pp. 479-489.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In