Towards a more generalised ROP prediction of TBMs through streamlined parameters: application of feature importance and machine learning

This study explores machine learning approaches for predicting the rate of penetration of tunnel-boring machines using different combinations of input parameters. This research centres on geology-related parameters to build models for the rate of penetration prediction. Feature importance was conducted using decision trees, random forests, XGBoost, correlation analysis, and permutation importance, leading to three different frameworks based on streamlined influential parameters. In the second phase, different advanced ML algorithms were used to construct predictive models. The models were tuned to achieve better generalisation, reducing the likelihood of overfitting and leading to more reliable and accurate predictions by integrating hyperparameter tuning with cross-validation, while a voting regressor combined the previous algorithms. Of the 21 generated models, the top three for each framework achieved R2 values of 0.96, 0.97, and 0.74, respectively. Model inference was then conducted on out-of-sample and unseen datasets, demonstrating that the workflow effectively predicts ROP and significantly improves accuracy and generalisation. Overall, the employed workflow demonstrates how more robust and generalisable ROP predictions than prior models and empirical formulas can be obtained by expanding the considered feature set and streamlining inputs via feature-importance analysis, leveraging modern ML with extensive hyperparameter search and tuning, introducing a VR-based estimation.