Enhancing Rock Mass Classification in TBM Tunnelling Using Ensemble Machine Learning Algorithms

It is common practice to use empirical rock mass classification systems for rock mass characterization in TBM tunnelling. However, complex geological conditions and limited access to the tunnel face significantly impact rock mass classification. The aim of this study is to address existing limitations by applying ensemble machine learning algorithms such as bagging, boosting, and random forest to characterize rock mass conditions using machine operational and geological parameters of double shield TBM tunnel excavated in Nepal Himalaya. To do so, the TBM database of 2,859 segmental rings were preprocessed and transferred into a usable format. The database then split into training and testing sets and the selected algorithms were trained with a training set with optimal hypermeters. The performance of trained models was validated with test database. Based on a comparison of evaluation metrics, the random forest model demonstrated superior performance for rock mass classification, achieving the highest accuracy of 94%. Notably, the net penetration rate (PRnet), cutter head penetration rate (PRchd), and cutter head speed (CRS) exhibited a negative impact on rock mass quality, while thrust and torque showed a positive influence. These findings provide critical insights into the relationship between TBM operational parameters and rock mass quality.