基于神经网络对有限元强度折减法分析

Analysis on Theory of Strength Reduction FEM Based on Artificial Neural Networks

基于改进BP神经网络方法对目前较为流行的强度折减法的理论进行分析。人工神经网络可以模拟人脑的思维,可以在完全不知道数据分布形式和变量之间确切关系的的情况下处理边坡各参数之间的复杂非线性映射,利用这一优势预测分析不同的理论在实际应用中的准确性。将已经通过工程手段计算出结果的数据以输入单元、隐含层和输出单元的形式代入系统进行神经网络训练,不同的屈服准则对应训练出不同的网络系统。用训练好的网络对检验数据进行预测分析,还使用这一方法预测了不同剪胀角对边坡破坏的影响程度及趋势,结果显示,对于平面应变问题,在有限元强度折减法中使用DP4和DP5准则所得到的效果较好,DP1准则的误差最大。

Application of theory on strength reduction FEM was analyzed based on improved BP neural networks. BP neural networks can simulate human brain and deal with complicated nonlinear relationship in different slope parameters under the condition of unknown relation between data distribution form and variables. Just taking the advantage the authors forecast the accuracy of every strength criterion in evaluating the slop stability. Take known data to the system to train the networks. According to different strength criterions, the different networks were trained. After that, the networks with new data were checked up. In the same way, the authors also analyze different slope damages considering character of dilatancy. The results show that using DP4 and DP5 criterions may obtain a favorable effect;meanwhile, the error is the largest in DP1 criterion. The conclusions would offer useful information for further application of strength reduction FEM.