岩质边坡稳定性 SVM 预测模型的新型核函数应用

导致岩体破坏的影响因素复杂多样，一般分为几何因素和物理因素。目前针对均质体的岩质边坡物理参数取值方面的研究较多，而关于改进的二折线岩质边坡模型的影响因素分析还处于起步阶段，且寻找能够解决各因素之间非线性关系的分析方法是十分必要的。基于二折线边坡计算模型，引入能够解决高维非线性问题的支持向量机方法，通过重要参数的敏感性分析，提出一种新型的综合核函数，论证该方法在岩质边坡稳定性预测分析中的可行性。通过各因素的敏感性分析可知，关于边坡的几何因素，采用RBF核函数所建预测模型精度较高，Sigmoid核函数适用性较差；关于边坡的物理力学因素，采用Linear核函数所建预测模型精度较高，Polynomial核函数和Sigmoid核函数适用性较差。经核函数矩阵组合得到一种新型综合核函数，并与4种常规核函数进行预测效果对比，结果表明，采用新型综合核函数所得岩质边坡稳定性预测精度最高，绝对误差不超过0.010 3，相对误差不超过2.83%。研究结论可为岩质边坡稳定性分析提供一种新思路。

Application of New Kernel Function of SVM Prediction Model for Rock Slope Stability

The influencing factors leading to rock mass failure are complex and diverse, which are geometric factors and physical factors. At present, there are many studies on the physical parameter choice of homogeneous rock slope, but the analysis on the influencing factors of the improved binary rock slope model is still in its infancy. It is necessary to find an analysis method that can solve the nonlinear relationship among the factors. Based on the second line of slope calculation model, the support vector machine (SVM) method is introduced to solve the high-dimensional nonlinear problem. A new synthetic kernel function is proposed by sensitivity analysis of important parameters to prove the feasibility of this method in the predictive analysis on the stability of rock slope. By analyzing the sensitivity of each factor, it can be known that the prediction model established by RBF kernel function for the geometric factors of slope has higher accuracy, but the applicability of Sigmoid kernel function is poor. And for the physical mechanical factors of slope, the prediction model established by Linear kernel function has higher accuracy, but the applicability of Polynomial kernel function and Sigmoid kernel function is poor. A new synthetic kernel function is obtained by the combination of kernel function matrix, and the prediction effect is compared with 4 conventional kernel functions. The result shows that prediction accuracy of the rock slope stability given by the new synthetic kernel function is the highest, the absolute error does not exceed 0.010 3, and the relative error is less than 2.83%. The research conclusion can provide a new idea for the stability analysis of rock slope.