Data-augmented machine learning approach for determination of over-excavation criteria in earth pressure balance...

Regulating the discharged muck volume is essential for preventing over-excavation in projects constructed by tunnel boring machines (TBMs). Over-excavation is typically identified when the over-excavation ratio (OER) exceeds a predefined criterion for over-excavation (C_OE). However, this criterion has traditionally been determined subjectively, and the site and operational conditions associated with anomalous over-excavation have not been systematically characterized. This study proposes a data-driven approach to objectively determine the optimal C_OE and to identify underlying anomalous conditions. Machine learning models, enhanced through data augmentation techniques, were developed to classify normal and over-excavation cases. An optimal C_OE of 1.15 was identified through an analysis of predictive performance and data patterns. The optimal model successfully identified 86.4% of over-excavation cases. The validity of the proposed C_OE was further confirmed by examining OER values under normal and abnormal over-excavation, including actual collapse events. Model interpretation revealed that elevated torque, particularly in deep, weathered ground with high water pressure, contributed to over-excavation. Beyond the specific C_OE identified in this study, the proposed framework provides a systematic and transferable approach for determining site-specific C_OE values in different tunnelling projects.