Estimation of Uniaxial Compressive Strength of Rock Using Shield TBM Operation Data and Supervised Machine Learning Methods

Abstract

Uniaxial compressive strength (UCS) is a frequently used parameter in predicting TBM excavation performance and disc cutter wear. However, due to financial and time constraints in tunnel projects, typically only UCS information obtained from limited geotechnical investigations is available. In this study, we predicted UCS using operational data collected during TBM excavation and machine learning algorithms based on supervised learning. The models, including KNN, RF, XGBoost, LightGBM, and CatBoost, were evaluated, with CatBoost demonstrating superior performance in terms of the lowest average Root Mean Square Error (RMSE) and the highest average R-squared (R²) for the test dataset. SHAP analysis identified the key variables influencing UCS prediction, with FPI, feed pressure, and cutterhead rotation speed ranking as the most significant factors.