A machine learning based attitude control model for open-type TBMs on the spiral ramp of Beishan underground research laboratory
Qixin Xu, Qiuming Gong*, Liu Huang, Ju Wang, Hongsu Ma, Mengxue Qi, Bei Han
Deep Resources Engineering. 2026, 3(3): 100250. 
doi.org10.1016j.deepre.2026.100250.pdf
Abstract: Precise attitude control of Tunnel Boring Machines (TBM) is critically important for ensuring tunneling quality and safety, especially in tunnels with steep slopes, long curves, or small turning radii. However, the tunneling trajectory is susceptible to deviations from the designed tunnel axis (DTA) due to the effect of a variety of factors including rock mass conditions, TBM tunneling parameters, operator’s expertise, and the designed tunnel axis complexities. This study addresses this vital need through a machine learning based approach for attitude adjustment. The theoretical analysis of the steering mechanism of the open-type TBM shows the horizontal attitude control is primarily governed by the stroke difference of lateral gripper cylinders, while vertical control is governed by the mean stroke of the torque cylinders. Depending on the spiral ramp in the Beishan Underground Research Laboratory project in Gansu Province, China, TBM tunneling data and guidance data were collected. Through a series of preprocessing steps, four key feature categories were fused, and a comprehensive attitude database was set up. Three machine learning models were developed for the real-time attitude adjustment. Among three models, the LightGBM algorithm delivered the best performance, achieving a mean absolute error (MAE) of 1.764 mm and root mean square error (RMSE) of 2.082 mm in horizontal control, and an MAE of 1.370 mm with RMSE of 1.564 mm in vertical control. The model validation using data from other sections of the spiral ramp, including both straight and curved sections, demonstrated that the performance was consistent with test results. It confirmed the model’s robustness and applicability in practical engineering condition.
Highlights:
• The steering mechanism of the “Beishan NO.1” Open-Type TBM was theoretically derived.
• Four feature categories were fused into a comprehensive attitude database via tunneling cycle extraction, preprocessing, and feature engineering.
• The optimal LightGBM model was proposed to control both horizontal and vertical attitudes based on TBM attitude control mechanism.
• Engineering applicability evaluation confirmed the proposed model’s effectiveness in both straight and curved tunnel sections.
Keywords: Open-type TBM; Spiral ramp; Attitude control; Machine learning method
Cite: Xu, Q.X.; Gong, Q.M.; Huang, L.; Wang, J.; Ma, H.S.; Qi, M.X.; Han, B., A machine learning based attitude control model for open-type TBMs on the spiral ramp of Beishan underground research laboratory, Deep Resources Engineering 2026, 3 (3), 100250. https://doi.org/10.1016/j.deepre.2026.100250
Citations: ……
