Abstract
Accurate identification of the tunnel-surrounding rock conditions is crucial for TBM excavation to prevent disasters such as excessive deformation and TBM jamming, which could significantly affect the project schedule. To mitigate premature convergence and insufficient local refinement of the caterpillar fungus optimizer (CFO) during machine-learning hyperparameter search, a Gaussian elite mutation-enhanced CFO (GCFO) is developed; it is applied to optimize hyperparameters of multiple machine learning models—including random forest, support vector machine, extreme gradient boosting, gated recurrent unit, and convolutional neural network—to compare and select an accurate predictive model for TBM tunnel surrounding rock classification. First, based on conditions such as penetration depth, torque, total thrust, cutterhead speed, and jacking shoe clamping force all being non-zero, stable data from the excavation cycle are extracted to calculate their mean values and standard deviations, forming data variables that characterize the rock-machine interaction process. A dataset of 6719 samples is constructed. However, because Class IV surrounding rock has relatively few samples, the SMOTE technique was applied to expand the dataset to 10,059 samples. Then, the GCFO and CFO algorithms are used to build optimized machine learning models, while the SSA algorithm is introduced to compare the optimization performance against the previous two algorithms. After comparisons, it is found that SMOTE can obviously boost the model prediction accuracy of the IV-class rock mass samples and GCFO-RF is finally developed for tunnel rock mass prediction, achieving high performance (training set: ACC = 0.99652, F1-score = 0.99696; validation set: ACC = 0.89741, F1-score = 0.90112; testing set: ACC = 0.82292, F1-score = 0.74402). Finally, a feature importance analysis using the SHAP (Shapley Additive Explanations) framework identifies the mean jacking shoe clamping force as the model’s most influential variable.