基于混合训练方法的RBF神经网络的曲面重构

Surface Reconstruction Based on Hybrid Training Method for RBFNN

根据径向基函数神经网络(RBFNN)具有很强的非线性逼近能力,以及强大的抗噪、修复能力等优点,讨论了目前神经网络训练方法,提出将径向基函数神经网络应用于带有噪声数据散乱数据点自由曲面的重构,并对该方法理论上的可行性和实践上的实用性进行了讨论和验证。结果表明:径向基函数网络用于曲面重构,不仅能够有效地逼近不完善的、带有噪声的曲面,而且拟合精度高、网络的训练速度快,说明了径向基函数神经网络应用于曲面重构问题的可行性,为解决反向工程的技术关键———自由曲面重构提供了一个新的途径。

Based on RBFNN＇s capabilities of approaching a no-linear function, powerful anti-noising, and repairing and so on, this paper describes the present training methods of RBFNN. The proposed method applies the RBFNN to the free surface reconstruction from an unorganized cloud of points in which always involve noise. Furthermore, it also discusses the proposed method＇s feasibility in theory and validates its practicability. The results show that the reconstruction method using RBFNN can not only approach the incomplete surface with noise effectively, but also have a high fitting precision and ahigh net-training speed. The above advantages indicate the feasibility of applying RBFNN to surface reconstruction. RBFNN provides a new method for RE＇s key technology：free surface recons.