为探究TBM施工中掘进参数与围岩等级的关联，梳理某高海拔铁路TBM隧道600余万组掘进数据，采用数理统计方法研究TBM掘进参数间、掘进参数与围岩等级的相关性，利用四分法分析TBM掘进不同等级围岩时的数据分布特征和规律，提出针对不同围岩等级的掘进参数优化建议； 按照环号提取TBM掘进数据的24项分布特征参数，利用SMOTE算法对数据不均衡问题进行处理； 采用PSO-SVM、GA-SVM、SVM、Transformer等机器学习算法开展围岩等级反演，并对比4种模型的预测精度。研究表明： 1）TBM掘进过程中刀盘转矩与掘进速度相关性最大，相关系数为0.92，其余掘进参数间相关性不明显，单一掘进参数均值与围岩等级相关性较小； 2）Ⅱ、Ⅲ级围岩TBM各掘进参数相对于Ⅳ、Ⅴ级围岩，数值分布更加集中，波动范围更小，随着围岩等级的增加，掘进速度总体增大、刀盘转矩减小、刀盘转速和总推力先减小后增加； 3）基于PSO-SVM模型的围岩等级预测精度高于SVM、GA-SVM、Transformer模型，且混淆矩阵分布合理，总体准确率达到88.6%，Ⅱ、Ⅲ、Ⅳ、Ⅴ级围岩的预测精准率分别为85.0%、93.1%、82.8%、71.4%。

More than 6 million sets of relevant data are collected from a high-altitude railway tunnel to examine the correlations among tunnel boring machine (TBM) tunneling parameters as well as between tunneling parameters and surrounding rock grades. The distribution characteristics and patterns of tunneling parameters for different rock grades are examined via quartile analysis, and a table for optimizing these parameters is proposed. Furthermore, 24 ring-based parameters are extracted from the tunneling data and analyzed using the synthetic minority over-sampling technique to address data imbalance issues. Four machine learning algorithms—particle swarm optimization (PSO)-support vector machine (SVM), genetic algorithm (GA)-SVM, SVM, and Transformer—are employed for surrounding rock grade inversion, and their prediction accuracies are compared. The results indicate that: (1) during TBM tunneling, cutterhead torque exhibits the highest correlation with tunneling speed, with a correlation coefficient of 0.92, and no significant interrelations with other tunneling parameters; the mean value of individual parameters shows limited correlation with rock grades; (2) TBM tunneling parameters for surrounding rocks of Grades Ⅱ and Ⅲ are more concentrated and fluctuate less than those for rocks of Grades Ⅳ and Ⅴ; with increasing rock grade, tunneling speed generally increases, cutterhead torque decreases, and cutterhead rotation speed and total thrust initially decrease and then increase; and (3) the PSO-SVM model is more accurate in predicting the rock grade compared to the SVM, GA-SVM, and Transformer models, exhibiting a reasonably distributed confusion matrix. Its overall accuracy reaches 88.6%, whereas individual prediction accuracies for surrounding rocks of Grades Ⅱ,Ⅲ,Ⅳ, and Ⅴ are 85.0%, 93.1%, 82.8%, and 71.4%, respectively.