BP神经网络模型在空间变异土坡可靠度分析中的预测精度  

Prediction accuracy of back propagation artificial neural network model in the reliability analysis of spatially variable slope

作　　者：李亮[1] 褚雪松[1] 

机构地区：[1]青岛理工大学土木工程学院,山东青岛266033

出　　处：《中国科技论文》2016年第13期1478-1482,共5页China Sciencepaper

基　　金：国家自然科学基金资助项目(51274126)

摘　　要：为了探讨强度折减法与BP神经网络预测模型在空间变异土坡可靠度分析中的预测精度,采用一维随机场模拟一边坡不排水强度的空间变异性,采用强度折减法计算训练样本的安全系数,将训练好的预测模型对不同相关距离下的随机场实现样本进行了预测,并与强度折减法计算结果进行了对比。结果表明:当土体材料的相关距离确定后,要合理选择随机场模拟的离散尺寸,离散尺寸越小,随机变量个数越多,训练后的神经网络模型预测精度越差;不同的训练样本对预测模型的预测精度也有较大影响,在应用BP神经网络预测模型时要尝试多组训练样本。In order to investigate the accuracy of the method of strength reduction procedure with Back Propogation Artificial Neu-ral Network （BPANN） model within the reliability analysis for spatially variable soil slope, the spatial variability of undrained shear strength, Su,is modeled by one dimensional random field. The factor of safety for each of training samples is calculated by strength reduction procedure and the trained BP ANN model is adopted to predict the results for randomly generated samples at different correlation lengths. The predicted results are compared with those calculated by strength reduction method. Results show that the proper choice of discretization size of random field is crucial for a given correlation length. The smaller the discretization size, the higher is the relative error between the predicted and calculated results. In addition,the training samples affect the performance of BP-ANN model. It is highly recommended that several sets of training samples are tried before a trained BP-ANN model is used.

关 键 词：边坡稳定 可靠度分析 BP神经网络 空间变异性 

分 类 号：TU43[建筑科学—岩土工程]