Abstract:

A predictive model integrating weighted random forest(RF) and variational mode decomposition(VMD) is constructed to accurately predict the penetration rate(PR) of tunnel boring machines (TBMs). Data from the KS tunnel and the water conveyance tunnel in the Lanzhou water source area, encompassing various lithologies, are collected and decomposed four times using VMD to eliminate highfrequency noise while preserving data characteristics. The traditional RF model′s feature contributions are assessed using Shapley additive explanations(SHAP), facilitating the application of weighted RF. Feature selection and hyperparameter optimization for the weighted RF are achieved through recursive feature elimination with crossvalidation and grid search. The model′s performance is validated in practical engineering settings, and it is explained both globally and locally based on SHAP theory. The results indicate: (1) High prediction accuracy with mean square error(MSE), mean absolute error(MAE), and determination coefficient(〖WT５《TNR#I》〗R〖WT５《TNR》〗2) of 0.064 9 (m/h)2, 0.187 5 m/h, and 0.925 4, respectively. (2) In practical engineering applications, the MSE, MAE, and 〖WT５《TNR#I》〗R〖WT５《TNR》〗2 are 0.050 3 (m/h)2, 0.161 3 m/h, and 0.950 5, respectively, demonstrating superior accuracy and performance compared to commonlyused deep neural networks, support vector regression, and unweighted RF. (3) VMD processing enhances PR prediction accuracy, showing improvements in MSE, MAE, and 〖WT５《TNR#I》〗R〖WT５《TNR》〗2 by 82.50%, 59.00%, and 33.25%, respectively. (4) The uniaxial compressive strength of rock is crucial for accurate PR prediction, and the interaction of geological parameters significantly outperforms that of tunneling parameters. For comprehensive analysis in critical tunnel sections, both global and local perspectives should be considered.

Key words: tunnel boring machine (TBM) tunnel, TBM tunneling performance, net penetration rate prediction, variational mode decomposition, random forest