• Three-stage ML framework predicts TBM performance using minimal data. • Ring-averaged data enables reliable prediction in complex geology without extensive laboratory testing. • CatBoost achieves R 2 = 0.81 for PR prediction; 83.5% accuracy for ground classification. • Ratio-based approach estimates total operation time with 4.1% cumulative error. • Parameter optimization yields 17–76% PR improvements across six GC classes. Accurate prediction of tunnel boring machine (TBM) performance and ground condition (GC) identification are critical for efficient tunnelling, yet most existing approaches rely on extensive laboratory testing, detailed geological surveys, or dense instrumentation rarely available in practice. This study presents a three-stage machine learning framework that integrates performance prediction, ground classification, and parameter optimization using minimal universally monitored TBM parameters—face pressure, thrust force, cutterhead torque, and rotation speed—aggregated as ring-weighted averages. Six tree-based ensemble algorithms were evaluated using data from a 3,000 m slurry shield TBM traversing complex geology in Singapore. Stage 1 achieved penetration rate ( PR ) prediction with R 2 = 0.81 and developed a ratio-based approach for operation time estimation (4.1% cumulative error). Stage 2 demonstrated GC classification accuracy of 83.5% and fault detection AUC of 0.941. Stage 3 employed Bayesian Optimization and Genetic Algorithms, achieving 17–76% PR improvements across six GC classes. By relying exclusively on minimal, universally available machine monitoring data combined with basic geological information from sparse borehole data, the framework eliminates the dependency on extensive laboratory testing and comprehensive geological surveys, enabling practical deployment in projects with limited site investigation budgets.
Sharghi et al. (Sat,) studied this question.