基于支持向量机的边坡稳定状态预测及参数优化

Slope stability prediction and parameter optimization using support vector machine

以边坡稳定状态为预测对象,在搜集了171个边坡数据的基础上,选取容重、黏聚力、内摩擦角、孔隙压力比、坡角、坡高作为预测指标,建立了基于支持向量机的边坡稳定状态预测模型。在此基础上,研究了模型训练集抽取比例(50%、60%、70%、80%)、数据预处理方法(标准化、归一化至[0,1]、归一化至[-1,1])及模型最优参数确定方法对预测结果的影响。提出了两种模型最优参数搜索方案。方案一是“N个搜索网格,逐个网格优化搜索”,方案二是“单个搜索网格,逐个网格点N次采样平均”。结果表明:预测模型在各种情况下的平均准确率均在80%左右;随训练集抽取比例增加,模型的预测准确性显著提升;当训练集抽取比例为80%,将数据归一化至[-1,1],取SVM模型的惩罚参数c和核函数参数g分别为2^(5)、2^(0)时,模型的平均准确率最高,达到了84.39%;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后的结果较为接近;最优参数搜索方案二在准确率、计算效率上要优于方案一。最后,使用建立的模型预测了22个稳定边坡和17个不稳定边坡的稳定状态,总体准确率达92.31%。

In this study,to predict slope stability,a support vector machine(SVM)-based model was established.A dataset of 171 slope cases with six indicators(e.g.,unit weight,cohesion strength and friction angle of slope mass,pore pressure ratio,slope angle,and slope height)was used to train and validate the SVM model.To optimize model performance,the effects of training dataset percentage,data preprocessing method,and optimized parameter searching method were investigated.Four different percentages(50%,60%,70%,and 80%)of the total dataset in combination with three preprocessing methods of input data(standardization,normalization of[0,1],and normalization of[-1,1])were used in this study.Two searching methods were developed to optimize the model parameters.The first method is the so-called“N grids with each grid optimization”.This method creates N searching grids and each grid was optimized using one randomly sampled dataset.The second method is the so-called“One grid with N sampling average of each grid cell”.This method creates only one searching grid and each grid cell of the grid was optimized using the average value of N sampling datasets.The results show that:(1)the SVM model has a good prediction performance with an average accuracy greater than 80%for all conditions;(2)prediction accuracy increases with increasing the training data percentage;(3)the highest accuracy is obtained when the training data percentage is 80%in which the input data is normalized into[-1,1]and the c,g parameter of SVM are 2^(5)、2^(0),respectively;(4)No obvious differences are observed between models with input data preprocessed using normalization and standardization;(5)“One grid with N sampling average of each grid cell”method is recommended to search the best SVM model parameters.The developed model is validated using a dataset of 22 stable slopes and 17 unstable slopes and the optimized SVM has an overall accuracy of 92.31%.