Abstract: The transformation of GNSS elevation and leveling normal elevation is a key project in the modern construction of benchmark.The randomness in the process of selecting the center value of radial basis function often cause the unsatisfactory fitting results of the fitting model of RBF network.Based on the research of regularization of RBF neural network,a center optimized model is established in this paper by applying robust estimation to the determination of the value of network hidden layer.Through the fitting test of engineering data in a small range of a county,the results show the advantages of the anti-error optimized model that the external coincidence accuracy of the model can reach centimeter level,which can meet the requirements of general engineering applications.

Key words: the regularization of RBF neural network, robust estimation, elevation fitting

摘要： GNSS高程和水准测量的正常高程转换是现代化测绘基准体系基准建设中的重点项目。RBF网络拟合模型中径向基函数中心值设置过程的随机性往往导致其拟合结果不太理想。文中在研究正则化RBF神经网络拟合模型的基础上,将选权迭代的稳健估计应用于网络隐层中心值的确定方法中,构建了一种中心优化的正则化RBF神经网络高程拟合模型。通过对某县区小范围内的工程数据进行拟合测试,结果表明经抗差优化的RBF拟合模型具有一定的优越性,其外符合精度可达厘米级,可以满足一般工程应用所需的精度。

关键词:  , 正则化RBF网络, 稳健估计, 高程拟合