Diagnosing tunnel collapse sections based on TBM tunneling big data and deep learning: A case study on the Yinsong Project, China

Abstract

The Yinsong Water Diversion Project in China’s northeast region contains a 20 km long tunnel section, which was drilled by a tunnel boring machine (TBM) and monitored in real time to generate continuously measured data. During the tunnel construction, 18 tunnel wall collapse failures were documented. This study reviews the boring performance of the TBM during these failures based on the field-collected data for the penetration rate, cutterhead rotation speed, torque, and thrust force, which were obtained at a 1-second interval. The main task in this study includes the development of a time series forecasting (TSF) approach combined with a deep belief network (DBN) that predicts the torque associated with a given penetration rate and rotation speed through a parameter called the drilling efficiency index, TPI, in a neural network. This study begins with a pilot case of a tunnel collapse that eventually caused the TBM construction to be abandoned. In this case, the TSF&DBN algorithm clearly identifies the deviations of the observed TPI values from those given by the neural network, indicating the successful prediction of an unfavorable geological condition. In addition, the cases of 13 other collapse sections are reviewed; of these, 11 clearly exhibit similar performance to the pilot case, whereas the remaining 2 provide no sufficient indications to exclude possible unstable roofs. This preliminary study shows that a systematic and high-quality TBM performance database can be useful in diagnosing adverse geological conditions in conjunction with the proper use of big data and machine learning techniques.