Neural network-based GPS GDOP approximation and classification

• Main Authors: Chien-Cheng Lai, 賴建成
• Other Authors: Tai-Sheng Lee
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/23544377371515923704

碩士 === 國立海洋大學 === 導航與通訊系碩士班 === 91 === When using the Global Positioning System (GPS) for navigation and positioning, we adopt all the signals from the satellites in view, except the satellites at low mask angles. This not only increased the amount of usable data for measurement but also boosted the accuracy, yet for this same reason, the time required for calculation also increased immensely. Sometimes a receiver may have only a limited number of channels, or it may be integrated with some other device, so we have to obviate some satellites or to choose only a certain group of satellites with acceptable Geometry Dilution of Precision (GDOP). Therefore we use the neural network as the tool for choosing satellites. Errors that occur in GPS positioning come from two different sources: the error in measurements and the error induced by the geometry satellite constellation. After Clinton, President of the United States, ordered to discontinue the SA in May 1st 2000, errors in measurement are reduced and the accuracy improved greatly, but it does not really help to solve the geometric problems. GDOP is an indicator of how well a satellite constellation is arranged geometrically. It could be viewed as a multiplicative factor that magnifies ranging error. Therefore the greater the GDOP value is, the less accurate the navigation solution is. In recent years, neural network has been used extensively in various fields. In this research, we use the four most poplar Back Propagation Neural Network (BPNN), the Optimal Interpolative Net (OI Net), the Probabilistic Neural Network (PNN), and the General Regression Neural Network (GRNN) to process the GDOP approximation and classification of GPS satellites so that we could save the processing time required for the countless matrix inversions when calculating. In this thesis, it can be a good performance which uses four neural networks for classification and approximation problem. In the accuracy of the classification, four kinds of networks can reach the correct classification rate of 93% ~ 100%. In the accuracy of the function approximation, RMSE of the BPNN and GRNN can reach lower than 0.1. Especially the GRNN uses the sample enough to train the RMSE can be lower than 0.02.