RBF神经网络模型在砂土液化判别中的应用研究

以时松孝次收集的砂土液化数据为研究对象,选取黏粒含量ρc、相对密实度Dr、临界深度ds、竖向有效应力σ’、地下水位dw、地震震级M、最大地面水平加速度αmax和标准贯入次数SPT-N等8个砂土液化的主要影响因素作为RBF神经网络的输入参数,利用MATLAB7.0中的神经网络工具箱,对部分样本数据进行训练和测试。并利用建立的RBF神经网络模型分析了各因素对砂土液化的影响规律。结果表明:砂土液化判别指标随αmax的增加而增大,随SPT-N和dw的增加而减小。研究成果表明,建立的RBF网络模型完全满足砂土液化判别的精度要求,能够精确模拟输入和输出之间复杂的非线性映射关系,具有较高的预测精度,具有重要的工程应用价值。

Application of RBF Neural Network Model to Evaluating Sand Liquefaction

The neural network toolbox of MATLAB7.0 was used to train and test some sample data of sand liquefaction collected by Tokimatsu Kohji. Eight eigenvectors clay content (ρ_c),relative compaction(D_r),critical depth of soil layer(d_s),vertical effective stress(σ′),groundwater level(d_w),magnitude of earthquake(M),maximum horizontal ground acceleration(α_max) and standard penetration number(SPT-N) were selected as input parameters of the RBF neural network. Furthermore,the established RBF neural network model was used to analyze the effect of each factor on the sand liquefaction. Results of the relative contribution of each factor showed that α_max was the biggest influencing factor on the evaluation index of sand liquefaction,followed by SPT-N and d_w. The evaluation index increased with the rise of α_max,while reduced with the increase of SPT-N and d_w. The evaluation index shows a logarithmic relation with α_max,cubic polynomial relation with SPT-N,and a negative linear relation with d_w . It’s revealed that the established RBF network model fully meets the requirement of evaluation accuracy for sand liquefaction. It can simulate the complex nonlinear mapping relation between the input and output data and also gives high prediction precision.