Real-time classification of ground conditions ahead of a TBM using supervised machine learning algorithms

• Published in: Modeling earth systems and environment 2024-10, Vol.10 (5), p.6173-6186
• Main Authors: Sebbeh-Newton, Sylvanus, Seidu, Jamel, Ankah, Mawuko Luke Yaw, Ewusi-Wilson, Rodney, Zabidi, Hareyani, Amakye, Louis
• Format: Article
• Language: English
• Subjects: Algorithms; Boring machines; Chemistry and Earth Sciences; Classification; Computer Science; Decision trees; Drilling & boring machinery; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Ecosystems; Environment; Geological hazards; Learning algorithms; Machine learning; Math. Appl. in Environmental Science; Mathematical Applications in the Physical Sciences; Original Article; Parameters; Physics; Pitching; Raw water; Real time operation; Rock; Rock mass rating; Rocks; Statistics for Engineering; Supervised learning; Support vector machines; Thrust; Torque; Tunnel construction; Tunnels
• ISSN: 2363-6203; 2363-6211
• DOI: 10.1007/s40808-024-02093-1

Accurately predicting the ground conditions ahead of a tunnel boring machine (TBM) in real-time is crucial for preventing geological hazards as well as for the adaptive adjustment of TBMs. The subjectivity in ground characterization is a major challenge in rock engineering. There is therefore the need for data-driven approaches. In this study, four machine learning classification models, namely support vector machine (SVM), k-nearest neighborhood (KNN), random forest (RF), and extremely randomized trees (ERT) were used to develop real-time rock mass classification models based on TBM operational parameters from Pahang-Selangor Raw Water Tunnel (PSRWT), Malaysia. Nine TBM operational parameters were used as input parameters. These include boring energy, cutterhead torque, cutterhead thrust force, revolution per minute (RPM), rate of penetration, stroke speed, gripper cylinder pressure, pitching, and motor current amps. An aggregated dataset of TBM operation data and rock mass data were created by adjoining the rock mass record for a particular chainage interval to all the TBM records in that interval. A balanced training set was obtained by the synthetic minority oversampling technique (SMOTE) for unbiased learning. The hyper-parameters of each classifier are optimized using the grid search method. The prediction results indicate that the ERT classifier has a better performance than other classifiers, and it shows a more powerful learning and generalization ability. The results suggest that ERT has the potential to correctly predict rock masses conditions ahead of a TBM in real-time by utilizing TBM operation parameters.