为提高水平定向钻勘察中岩性识别的效率，基于机器学习算法采用钻进参数识别围岩岩性。以新疆某隧道工程为例，通过对水平定向钻的工作原理进行分析，采用钻进速度、校正后的钻孔底部压强、泥浆压力和进浆流量作为输入特征预测围岩岩性。对KNN（k-nearest neighbor）算法和随机森林算法各设置48个超参数，测试集的平均准确率分别为83.28%和93.04%，模型不存在欠拟合和过拟合问题。将五分类问题转化为5个二分类问题，2种算法的准确率、精确率、召回率、F1值基本均在90.00%以上，受试者工作特征（receiver operating characteristic,ROC)曲线中曲线下面积（area under curve，AUC）也接近于1。〖ＪＰ２〗使用Smote过采样后的KNN算法和随机森林算法都具有良好的鲁棒性和泛化能力，但综合各项评价指标可知，使用随机森林模型预测围岩岩性的效果更佳。
In horizontal directional drilling surveys, the
lithology of the surrounding rock is generally determined by coring or drilling
into rock chips. However, owing to the limitation of coring efficiency during
drilling, determining the lithology of the surrounding rock throughout the
entire borehole line based on rock chips is a tedious process. Therefore, to
improve the efficiency of lithology identification, the authors use machine
learning algorithms to identify the surrounding rock lithology using drilling
parameters. A case study is conducted on a tunnel in Xinjiang, China, and the
drilling speed, corrected bottom pressure of the borehole, mud pressure, and
feed flow rate are input features to estimate the surrounding rock lithology
based on the analysis of the working principle of horizontal directional
drilling. Fortyeight hyperparameters are set for each knearest
neighbor(KNN) and the random forest algorithms and the average accuracies of
the test set are 83.28% and 93.04%, respectively, with no underfitting or
overfitting concerns in the models. Furthermore, the five classification issues
are transformed into five binary classification problems; the accuracy,
precision, recall, and 〖WT５《TNR#I》〗F〖WT５《TNR》〗1 values of both algorithms are above 90.00%. The area undercurve values in the receiver operating characteristic curves are
close to 1. The results show that the KNN and random forest algorithms have
good robustness and generalization ability after using Smote oversampling;
however, the random forest model is more effective in predicting the rock
properties of the surrounding rocks when the evaluation metrics are combined.
