The Training Method of General Regression Neural Network for GDOP Approximation

Abstract

Aimed at the problem of real-time and accurate Geometry Dilution of Precision (GDOP) approximation, a new method using general regression neural network (GRNN) was proposed firstly, and the training samples selection and normalization method was studied by using spectrum analysis. The computation results show that the symmetrical constellation needs 24 hours continuous samples while the hybrid one needs 72 hours to train the neural network sufficiently. Finally, the performance analysis shows that this new method has excellent performance on temporal and spatial generalization approximation accuracy, when trained GRNN are used, the GDOP computational error is less than 0.25 within 30 days, and the error is less than 0.3 within 10 degrees latitude/longitude area.