基于级连相关神经网络的人工冻土本构模型1）

为解决BP (back propagation)神经网络收敛速度慢，网络结构需事先定义等缺点，采用了级连相关神经网络模型来建立人工冻土应力和应变之间的关系。基于该模型推导了冻土的一致刚度矩阵形式，利用人工冻土三轴试验数据对神经网络模型进行训练，并用其替换有限元计算中的传统本构模型，将计算结果与性质及含水率相同的冻土的试验结果进行了对比，发现该神经网络本构模型很好地反应了材料的非线性，能够改善数值计算结果，与实测结果吻合地很好，比具有相同隐含层神经元个数的BP模型更接近实测结果。

CONSTITUTIVE MODEL OF ARTIFICIAL FROZEN SOIL BASED ON CASCADE-CORRELATION NEURAL NETWORK 1)

In order to avoid the drawback of BP (back propagation) neural network of too slow velocity convergence, the network structure has to be defined in advance. A cascade-correlation artificial neural network model is adopted to create the relationship between stress and strain of the artificial frozen soil, the consistent stiffness matrix is derived based on this model for the frozen soil, the neural network model is trained by the triaxial test data to replace the traditional finite element constitutive model, and the calculated results of the properties and the moisture content of the frozen soil are compared with the experimental results. It is found that this neural network constitutive model can represent the nonlinear response of the material very well, can improve the numerical analysis results, with very good agreement with the measured results, and the results are closer to the measured results than the BP model with same number of neurons in the hidden layer.