基于贝叶斯正则化 BP 神经网络的 GPS 高程转换

为了改善BP神经网络在GPS高程转换过程中过拟合的现象,提出了用贝叶斯正则化算法的BP神经网络转换GPS高程的新方法,并利用区域GPS/水准数据,将新方法和未采用正则化算法的BP神经网络进行GPS高程转换的比较.结果表明:在较大区域和高程异常呈不规则的情况下,新方法不仅可以有效提高GPS高程转换的精度,而且通过贝叶斯正则化算法可以改善网络结构,抑制过拟合现象.在约10 km的GPS基线尺度上,新方法可以得到精度达0.050 m的正常高.

Conversion of GPS Height Based on Bayesian Regularization BP Neural Network

In order to improve the over-fitting in GPS(global positioning system) height conversion using BP(back propagation) neural network,a new method of GPS height conversion based on the Bayesian regularization BP neural network was proposed.Using the GPS/leveling data in a certain area,this new method was compared with the BP neural network without using the regularization algorithm for GPS height conversion.The research results show that the new method can not only improve the precision of GPS height conversion but also restrain the over-fitting through using the Bayesian regularization algorithm to improve the structure of neural networks in cases with a big area and anomalous height anomaly.The precision of GPS height conversion can achieve 0.050 m to an about 10 km baseline with the new method.