Tunnel boring machine performance prediction using knowledge-driven transfer learning

Abstract

Machine learning (ML) emerges as a powerful tool in tunnel boring machine (TBM) performance prediction. A reliable ML model requires sufficient training data which, however, is usually absent for a new tunnelling project. Data-driven transfer learning offers a potential solution to this issue but usually experiences reduced performance due to the inconsistent data distributions among TBM projects. In this paper, we propose a knowledge-driven deep transfer learning-based approach for TBM performance prediction, aiming at improving TBM performance for new and ongoing tunnelling projects with no or limited boring data. The proposed method first trains source models using the knowledge-driven transformed data. The data are transformed by a proposed TBM invariant transformation method, which is developed based on TBM mechanical and empirical relationships. Subsequently, deep transfer learning is applied to fine-tune the source models for the target project using available small data. Three TBM tunnelling projects in China (i.e. the Yinsong project, the Yinchao project and the YE project) are taken as case studies to investigate the feasibility of the proposed method. The proposed knowledge-driven transfer learning method outperforms data-driven transfer learning in all tested scenarios and achieves satisfactory prediction performance in both data-limited and data-rich cases. Significant improvements over the conventional deep learning method can also be observed in the most data-limited condition (i.e. 100 training boring cycles): R squares are increased by 0.17 and 0.31 for torque and total thrust prediction, respectively, corresponding to mean absolute percentage error (MAPE) decreases of 3.65% and 5.82%. The optimal frozen strategy for TBM transfer learning is also investigated. By empowering knowledge sharing among different TBM tunnelling projects, the proposed method reveals a smart and promising way to address the TBM data scarcity problem and improve TBM performance prediction for new and ongoing projects.