基于RBF混合神经网络的自由曲面重构

Freeform Surface Reconstruction Based on RBF Mixture Neural Network

为解决大样本集的简化建模和快速训练问题,以层次划分为思想基础,提出了基于径向基函数神经网络的混合网络(RBFMNN)模型,对自由曲面进行重构.用减聚类方法划分样本空间,对各子样本空间用正交最小二乘法进行RBF子网络训练.最后,利用最大似然法来校正RBF子网输出层的参数,以进一步提高混合网络输出精确度.试验结果表明,该网络模型对已知理想的曲面拟合误差为106星级.

In order to solve the question of simplified modeling and the fast training for big sample collection, on basis of the thought of the level division, RBF mixture neural network model was proposed by this article to restructure the freeform surface. It divided sample space by the way of reduce-gathers first, then trained to each subsample space by using OLS on the RBF sub-network, finally adjusted the parameter of the RBF subnet output level with the maximum likelihood method to enhance the output precision of mixture neural network. The test result indicated the fitting precision of freeform surface was enhanced distinctly by this network model.