Extrapolative Model of DGPS Corrections using a Multilayered Neural Network Based on the Extended Kalman Filter

Abstract

This paper presents an accurate DGPS land vehicle navigation system using a multilayered neural network (NN) based on the extended Kalman filter (EKF). The network setup is developed based on a mathematical model to avoid excessive training. The proposed method uses an EKF training rule, which achieves the optimal training criterion. The NN predicts the DGPS corrections for accurate positioning. The proposed method is suitable for DGPS systems sampled at different rates. The experimental results on collected real data demonstrate the suitability of this method in developing an accurate DGPS land vehicle navigation method. The experiments show that the prediction total RMS error is less than 1.65m and 0.67m, before and after SA, respectively. Also, tests with real data demonstrate that the prediction accuracy is better than 1.1m for 10 second prediction and 1.9m for 30 second prediction, respectively, which can maintain the vehicle navigation in the required accuracy for a period of 30 seconds