DGPS-aided INS land vehicle navigation systems

The need of a Global Positioning/Inertial Navigation (GPS/INS) augmentation system stems from the shortcomings of each individual technology. The primary limitation of GPS is its requirement for line-of-sight visibility of GPS satellites. On the other hand, the stand-alone INS position errors are unbounded. The motivation for GPS/INS integration is to develop a navigation system that overcomes the shortcomings of each system.; The goal of this research is to develop a method for integrating a speed-based Inertial Navigation System and a Differential Global Positioning System (DGPS) for land vehicle navigation that overcomes the KF method disadvantage by being more stable, having less computation load, having better immunity to noise effects, and being more robust to false statistics. This method of integration relies on the use of an artificial linear neuron. The neuron adaptively estimates the scale factor and the bias INS error source values during the availability of the DGPS, and then uses these estimated values to aid the INS during DGPS outages or unsuitable DGPS solutions. The linear neurons send these error source estimations with the corresponding statistics to correct the INS position solution. A statistical propagator propagates these statistics to the next DGPS epoch and reflects them onto the INS position solution. Then a statistical combiner optimally combines the DGPS and the INS position solution. The overall INS/DGPS accuracy is improved over the stand-alone DGPS accuracy by the continuous correction to the INS followed by the INS/DGPS optimal combination. An optimality criterion is obtained by deriving a new optimal gradient descent training method, as well as the statistical development of the entire algorithm.; The experimental results demonstrate the advantages of the new approach over a typical loosely coupled Kalman filter (KF) and neural network (NN) methods in terms of performance and computation efficiency. The way the linear neurons are set up, their decoupled training, and the finer search of the gradient descent method near the optimal solution, makes the linear neuron a better estimator of the INS error sources than the KF method. The system can fill a 30 second gap online with about 0.27m accuracy in each axis. The NN demonstrates poor performance in limiting the bias and scale factor effects on the INS solution. The instantaneous accuracy for the NN is very low, while its long-term accuracy is quite acceptable.