UTILIZATION OF THE RMR SYSTEM AND MACHINE PARAMETERS FOR PERFORMANCE ESTIMATION OF HARD-ROCK TBMS: A STATISTICAL AND MACHINE LEARNING APPROACH

Abstract

Accurate estimation of TBM performance is crucial for project planners, as it helps predict the expected excavation time and overall project costs. The rock mass rating (RMR) system has been used widely for this purpose because it is easy to apply and accounts for the effect of rock mass discontinuities. However, TBM penetration models based on the RMR system are limited, with many existing ones developed using relatively small datasets and relying solely on RMR as the input variable, often overlooking the critical influence of machine parameters. This thesis aims to develop models for estimation of TBM penetration per revolution (Prev) using RMR and machine parameters as inputs. To achieve the study aim, a database containing 908 data points was compiled from seven hard rock tunnels in Italy and Iran. This database was further developed into 137 datasets, each containing RMR, TBM operational parameters, and performance metrics, categorized based on tunnel lithologies. After establishing the datasets, simple linear regression (SLR) was initially performed to examine the relationships between individual input and the Prev. The SLR results showed that single input variable could not provide an accurate estimation of Prev. However, all the input parameters, except the number of cutters (Nc), were statistically significant at p-values less than 0.005, leading to the exclusion of Nc from the input variables used for developing the models. After selecting the input parameters, linear and non-linear multivariable regressions were conducted using SPSS (V.28) to establish the empirical models for estimating Prev. Five non-linear models using RMR, thrust, cutterhead rotational speed, and cutterhead diameter as input variables, were introduced. The introduced non-linear models achieved an average determination coefficient (R²) of 0.75 and 0.72 on training and testing datasets, respectively. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms were employed to estimate Prev using the same datasets and input variables employed for the non-linear models. The results showed that the RF models achieved an average R² of 0.91 and 0.89, respectively on the training and testing datasets, while the XGBoost models achieved an average R² of 0.93 and 0.89, respectively on the training and testing datasets. It is necessary to understand the impact and feature importance of individual input in the machine learning-based models. Therefore, a sensitivity analysis was performed using SHAP. The analysis showed that cutterhead diameter has the highest impact on Prev, followed by RPM, thrust, and RMR. Conclusion from the study is that introduced non-linear models may be used for estimating Prev under conditions similar to those observed in this study.