Application of Machine Learning and Linear Regression for Estimating Cutter Life Index (CLI)

Accurately predicting cutter wear is critical when budgeting construction schedules and costs for tunneling projects involving tunnel boring machines (TBMs). To address this need, the cutter life index (CLI), developed by the Norwegian University of Science and Technology (NTNU), is usually calculated using Sievers’ J-value (SJ) and the Abrasion Value Steel (AVS). However, measuring the CLI is challenging owing to the need for specialized test equipment, as well as the time-consuming and complex nature of the experimental process. This study aims to estimate CLI values based on common mechanical rock properties and the Cerchar Abrasiveness Index (CAI). The considered mechanical rock properties include the uniaxial compressive strength (UCS), Brazilian tensile strength (BTS), equivalent quartz content (EQC), and brittleness indices (B1–B5), as well as the cohesion derived as a function of uniaxial compressive and Brazilian tensile strengths. To develop an estimation model, we used linear regression analysis and various machine learning techniques, including extreme gradient boosting (XGBoost), extra trees (ET), gradient boosting machines (GBM), random forests (RF), decision trees (DT), and adaptive boosting (AdaBoost). For the linear regression, CAI, EQC, BTS, and the rock type were defined as the input variables, taking multicollinearity into account. For the machine learning models, appropriate input variables were selected. A literature survey was conducted to compile a suitable dataset, which was then divided into 80% training data and 20% test data. The accuracy of the predictive model was evaluated using the Root Mean Squared Error (RMSE) and coefficient of determination (R2) of five k-fold cross-validation values and considering the skewness of the data. The results show that the gradient boosting machine model provides the best predictive performance, with an R2 of 0.93, 0.70, and 0.89 for the training, test, and entire datasets, respectively.