Designing the main drive motor capacity of Earth Pressure Balanced Tunnel Boring Machines (EPB TBMs) is a crucial task for every EPB tunnelling project. The machine needs to be equipped with sufficient power to master the geotechnical conditions of the respective project. On the other hand, overpowering the machine should be avoided for economic and sustainability reasons. Main drive torque estimation for EPB TBMs is challenging due to a multitude of impact factors and reciprocal mechanisms between the geotechnical conditions and the tunnelling process. In EPB TBM tunnelling active tunnel face support is achieved in soft and mixed ground or weak and unstable rock by generating a pressurized earth paste in the tool gap and excavation chamber of the machine. Complexity arises due to tribological and rheological effects of the active tunnel face support. These elements of uncertainty, the expected main drive torque is frequently overestimated to prevent a jamming of the machine in the ground. Mean main drive torque values often lie below 50 % of the installed nominal main drive torque capacity. In scope of this research machine learning algorithms, such as regressions, decision trees, tree ensembles, support vector machines and gaussian process regressions, have been used to predict the main drive torque. Models have been trained and tested on data collected from 9 different reference projects and validated on the data of 3 additional reference projects to test the transferability of the model. TBM diameters of the reference projects vary between 6,5 and 15,9 m and TBMs have been operating in a wide range of geotechnical boundary conditions. Different feature selection algorithms have been used and prediction results have been compared to models trained on manually selected features. Models using tree ensembles and manually selected features showed best prediction results and model performance. The machine learning approach returned a smaller and more accurate torque estimation range than traditional estimation approaches and prediction accuracy has been improved. Transparent and robust tree ensembles proofed to be suitable tools for TBM torque estimation.

A significant part of the energy consumed during the tunnelling process of Earth Pressure Balanced (EPB) Tunnel Boring Machines (TBMs) is related to the main drive, consisting of a set of motors driving the rotation of the cutting wheel. An energy efficient EPB design requires the optimization of the main drive to avoid over- or under powering of the machine. Key aspect is therefore a precise and reliable estimation of the expected cutting wheel torque. In this paper state-of-the-art torque estimation models are compared to supervised machine learning (ML) approaches, including classification and regression trees (CART), support vector machines (SVM), Gaussian process regression (GPR) and decision tree ensembles (DTE). Feature selection algorithms are compared to models using manually selected input features. ML models are evaluated using accuracy metrics, residual analyses, and model validation. Torque prediction for a real-world validation project shows that utilization rates can be increased distinctively due to the application of ML techniques.