基于贝叶斯逻辑回归模型的边坡稳定性预测

Slope Stability Prediction by Bayesian Logistic Regression Model

该研究中提出了一种基于贝叶斯逻辑回归模型的边坡稳定性预测方法。该方法以边坡稳定性为预测对象,选取边坡的坡高、坡角以及岩土体的黏聚力、内摩擦角、重度、孔隙压力比6个指标作为特征参数,通过贝叶斯推断对逻辑回归模型中的自变量的回归系数和截距进行了估计。在收集大量边坡数据集的基础上,研究了数据预处理方法(标准化、归一化至[0,1]、归一化至[-1,1])及3种先验分布(正态分布、柯西分布、t分布)对模型精度的影响。结果表明:将数据进行归一化处理后得到的预测结果在准确性上与采用标准化处理后得到的结果较为接近;采用不同形式的先验分布,模型优化结果差别不大,但发现各先验分布的平均值和标准差会影响回归系数(截距)的后验结果;在采用五折交叉验证的情况下,当数据预处理方法为归一化至[-1,1]且先验分布为正态分布时,模型的预测准确率最高,其AUC值达到了0.860。

This paper presents a slope stability prediction method based on a Bayesian logistic regression model.The method focuses on predicting slope stability.Six predictive indicators of slope height,slope angle,cohesion of rock&soil,internal friction angle,weight and pore pressure ratio are selected to be characteristic parameters.Through Bayesian inference,the regression coefficients and intercept of the model's independent variables in logistic regression are estimated.Based on a substantial dataset of slope information,the impact of data preprocessing methods(data standardization,normalization to[0,1],normalization to[-1,1])and three different prior distributions(normal distribution,Cauchy distribution,t-distribution)on model accuracy are investigated in this study.The results indicate that the accuracy of prediction outcomes by normalized treatment is closely resembling those by standardization;There's not much differences in model optimization outcomes by different forms of prior distributions.But it is observed that the posterior results of regression coefficients(intercepts)will be impacted by the means and standard deviations of these prior distributions;Under a 5-fold cross-validation scenario,when data is normalized to the range of[-1,1]and a normal distribution prior is employed,the model attains the highest predictive accuracy,with an area under curve(AUC)value of 0.860.