为进一步提高基于钻进参数的围岩智能分级模型判识准确率，引入围岩岩性作为已知条件，对不同岩性分别建立基于钻进参数的围岩智能分级模型。首先，依托宜兴高铁、郑万高铁隧道工程中凿岩台车采集到的钻进参数，将其根据围岩岩性划分为白云岩、花岗闪长岩、灰岩、砂岩、页岩5类子数据集； 然后，进行数据清洗、特征提取，对各子数据集及原始数据集按照4∶1的比例拆分为训练集和测试集，分别以提取的特征和原始特征均值2类特征集合作为输入；
最后，使用机器学习随机森林算法建立5个基于钻进参数和岩性的围岩智能分级模型以及基于钻进参数的围岩智能分级模型，评估每个模型的泛化性能。结果显示： 1）以提取特征集合为输入的白云岩、花岗闪长岩、灰岩、砂岩、页岩5种围岩智能分级模型在测试集上的准确率分别为85.48%、92.16%、88.62%、85.00%、89.47%，不考虑岩性的围岩智能分级模型的测试集准确率为84.91%。2）相比不考虑岩性的围岩智能分级模型，基于钻进参数和岩性的围岩智能分级模型测试集准确率提高0.09%~7.25%； 相比不进行钻进参数特征提取，围岩智能分级模型准确率提升1.61%~13.82%。由此表明：
进行钻进参数特征提取能够有效提升围岩智能分级模型的准确率； 相比不考虑岩性的围岩智能分级模型，考虑岩性的围岩智能分级模型稳定性提高，模型的泛化能力提升。
An intelligent model that
considers lithology in classifying surrounding rock is established based on
drilling parameters so as to further improve the accuracy of predicting the
surrounding rock. First, the drilling parameters collected by a drilling trolley
in the Yichang-Xingshan and Zhengzhou-Wanzhou high-speed railway tunnels are
divided into the five data subsets of dolomite, granodiorite, limestone,
sandstone, and shale. Next, data cleaning and feature extraction are conducted,
and each data subset and original data set are divided into a training set and
test set in a 4∶1 ratio. Furthermore, the extracted
features and the original features are used as inputs to establish five
intelligent classification models for surrounding rock based on drilling
parameters and lithology and an intelligent classification model for surrounding
rock based on drilling parameters using a machine learning random forest
algorithm. Finally, the generalization performance of each model is evaluated.
The results demonstrate the following. (1) The models that are based on
dolomite, granodiorite, limestone, sandstone, and shale achieve an accuracy on
the test set of 85.48%, 92.16%, 88.62%, 85.00%, and 89.47%, respectively, and
the accuracy of the model that does not consider lithology is 84.91%. (2)
Compared with the improvement in the model that considers lithology, the
improvement in the accuracy on the test set of the surrounding rock intelligent
classification model based on drilling parameters and lithology is 0.09%－7.25%, and the improvement in model accuracy is 1.61%－13.82% compared with no feature extraction. This shows that
extracting the features of the drilling parameters effectively improves the
accuracy of the model. The intelligent model that considers lithology in
surrounding rock classification is more stable and achieves better
generalization performance than the model that does not consider lithology.
