Advance prediction of collapse for TBM tunneling using deep learning method

Collapse hazards pose a serious threat to the tunnel construction, especially when a tunnel boring machine (TBM) is used. The current methods are mainly focus on the prediction of adverse geology, which cannot provide the specific area that collapse may happen. Recent development of TBM is capable to offer immense amounts of monitored data by the sensors in the machine. A three-stage method was proposed in this study to predict the collapse area in a real-time manner by combining the TBM-generated data with a deep learning algorithm. The method starts with constructing a long short-term memory model to predict torque and thrust based on the data in non-collapse area. This model can then be used to predict the torque and thrust in a real-time manner for the following tunneling. The predicted parameters will be consistent with the measured ones when the tunneling are stable. On the contrary, the prediction accuracy will be decreased when a collapse is prone to happen. A three stage method is then proposed based on this principle, which is further examined using the big data from Yin song tunneling Project in China. There were 18 collapses along this 20 km tunnel, 12 of which have been successfully predicted in the second stage and the remaining 6 have been predicted in the third stage. Essentially, the proposed method here is capable to accurately predict tunnel collapse and provide early warning in advance for tunneling, paving the way to a self-driving TBM in harsh geological conditions.