三维地应力场BP反分析的改进

从回归分析出发,对地应力场的神经网络反分析进行了改进:采用线弹性有限元计算进行线性回归分析,获得优化参数的大致范围;应用均匀设计来确定计算参数不同水平的组合,进行弹塑性有限元计算获得训练的样本;采用Levenberg-Marquardt算法来训练BP神经网络,以提高效率;使用及早停止和正规化方法来避免神经网络的过拟合问题;得到离散化的应力值后,用神经网络拟合出了以坐标为参量的全场应力函数。

MODIFIED BP BACK ANALYSIS OF 3D IN-SITU STRESSES

Based on regression analysis,an improved method is provided for back propagation (BP) back analysis of in-situ stresses. Linear elastic FEM calculation and linear regression analysis are implemented to determine the general bounds of optimized parameters. Uniform design method is carried out to settle different combinations of factor levels. Training samples are gained by elastoplastic FEM analysis. Levenberg-Marquardt algorithm is applied to achieve better performance during the training process of BP neural networks. Early stopping and regularization are employed to avoid over-fitting problem. When discrete stress values are obtained,neural networks are also used to construct a global stress function with coordinates as parameters.