基于BP神经网络的土体细观力学参数反演分析

利用离散元方法对颗粒材料的细观力学特性研究, 目前确定数值计算模型的细观力学参数大多数通过反复调试获取, 效率低、可重复性差。本文采用开源的颗粒离散元程序LMGC开展了土体双轴压缩数值试验, 通过25组土体细观力学参数计算得到相应的宏观力学参数, 建立了BP人工神经网络反演系统。利用土体物理试验得到的土体宏观力学参数, 输入BP神经网络, 反演得到土体的细观力学参数。将所得细观力学特性参数输入所建立的土体数值计算模型, 得到土体破坏过程中的应力-应变关系曲线, 以及土体颗粒的力链图和旋转变形云图。所建立的土体数值试验模型能够较好地模拟土体变形破坏过程, 利用BP神经网络反演细观力学参数以及数值模型计算得到的土体宏观力学参数与物理试验吻合较好, 误差在10%左右, 土颗粒间力链云图以及旋转变形云图较好地揭示了土体变形破坏的机理。

BACK-ANALYSIS ON MICROMECHANICAL PARAMETERS OF SOIL MASS USING BP NEURAL NETWORK

The micromechanical parameters of numerical model are obtained by repeated tests. Researches on particle material based on DEM are done. This article builds a numerical model of soil using LMGC, an open source software. And the 25 groups of macromechanical and micromechanical parameters calculated by the numerical model build a BP neural network. The BP neural network can back-calculates the micromechanical parameters if the macromechanical parameters of real physical tests are inputted into the neutral network. The numerical model can get the stress-strain line, stress train and particle rotation graph if the micromechanical parameters are set. The results show that the numerical model can well simulate the soil destruction process. The stress-strain line of numerical model matches well with the result of physical test. And error of macromechanical parameters calculated by BP neural network is about 10%.Besides, the stress train and particle rotation graph reveal the mechanism of soil destruction process.