基于数值样本和随机森林分类器的岩爆风险快速预测代理模型

A surrogate model for the rapid prediction of rockburst risk based on numerical samples and random forest classifier

该文依托滇中引水工程香炉山隧洞,开展深埋隧洞掘进过程中岩爆风险预测研究,提出基于数值样本和随机森林(random forest,RF)的岩爆风险快速预测代理模型。以地应力和围岩本构参数作为模型输入特征,以围岩最大弹性应变能密度作为输出特征,开展使用隧道掘进机(tunnel boring machine,TBM)掘进深埋隧洞过程中的多工况数值仿真,基于正交试验设计构建了611组数值样本,并通过围岩能量演化分析和相关性分析,验证输入、输出参数的合理性。进一步以RF为基分类器,利用10折交叉验证优化超参数,建立了岩爆风险快速预测代理模型,并对比多种机器学习算法,验证所提代理模型的预测准确性和适用性。结果表明:所提方法具有良好的预测性能和泛化能力,测试集样本的预测准确率达82.02%,优于用于比较的其余4种机器学习模型,为快速预测深埋隧洞施工期岩爆风险提供了一种研究路径。

[Objective]This study aims to address the significant challenge of predicting rockburst risks during the excavation of deep tunnels using tunnel boring machine(TBM)tunnel boring machine and develop a rapid prediction model to provide the basis for rockburst prevention and control,enhancing the safety and efficiency of deep tunnel construction.The proposed model leverages numerical samples and random forest(RF)algorithms to overcome the limitations of existing methods,which often do not achieve real-time and rapid prediction or consider the underlying mechanisms and factors influencing rockbursts.[Methods]Considering the Xianglushan Tunnel within the Dianzhong Water Diversion Project,we introduced a model that utilizes geostress and rock constitutive parameters as inputs and the elastic strain energy density of the surrounding rock as output.Numerical simulations of tunneling using the TBM under various working conditions arewere conducted,and 611 numerical samples were crafted through an orthogonal experimental design.We employed RF as the underlying classifier,with hyperparameters optimized through 10-fold cross-validation to create an efficient prediction model.The accuracy and applicability of the model were confirmed by comparing several machine learning algorithms.[Results]We conducted a series of numerical simulations of excavation using the TBM,employing an elastoviscoplastic constitutive model with internal variables.These simulations disclosed the energy evolution within the rock mass throughout the excavation process.Energy concentration occurred during transient unloading and the time-dependent deformation of the surrounding rock,leading to two distinct peaks in strain energy density.The second peak indicative the final energy storage during the creep phase of the surrounding rock postexcavation and unloading.Notably,a higher value at the tunnel wall—under identical conditions—correlated with an elevated risk of strainburst.We verified the rationality of the input and output parameters by analyzing energy evolution and correlation.The predictive accuracy and computational efficiency of the model were enhanced following the optimization of the hyperparameters using a 10-fold cross-validation.The input parameters partially mirrored the factors influencing rockburst,while the output parameters measured the energy storage status of the surrounding rock before potential rockburst failure.The RF-based rockburst risk prediction proxy model exhibited commendable performance on the training and testing sets,achieving accuracies of 99.75%and 82.02%,respectively.The performance of the RF-based rockburst risk prediction proxy model was superior to that of four other machine learning models—decision tree,K-nearest neighbors,support vector machine,and logistic regression—achieving prediction accuracies of 82.02%,76.40%,79.77%,75.28%,and 76.40%for all samples,respectively.This result indicateds the robust predictive capability and generalization performance of the RF-based rockburst risk prediction proxy model in assessing rockburst risk levels.[Conclusions]We offer a novel approach and framework for the rapid prediction of rockburst risks during the excavation phase of deep tunnels.The RF-based rockburst risk prediction proxy model is reportedly an effective tool for rockburst risk prediction,marking a significant advancement in rockburst risk management.We provide a research path and framework for the rapid prediction of rockburst risk during the excavation period of deep tunnels.