Transfer learning for collapse warning in TBM tunneling using databases in China

Tunnel boring machine (TBM) has been widely used for tunnelling in harsh environments with deeply embedded tunnels and hard rocks. The collapse of surrounding rocks seriously jeopardizes the safety of the machine and the driver. Prediction of TBM tunneling response is vitally important to alert the hazard in advance, which is still a challenge topic, especially for a new-constructing tunnel where scarce data are available. This study proposes a transfer learning model that is capable to predict TBM tunneling responses for a new project using the boring data obtained from the other existing tunnels. The key problem of transfer learning model, i.e., what to transfer, is first studied based on the big data obtained from Yin-song project and Yin-chao project in China. Then, when to transfer is investigated by comparing with the machine learning model that developed using only the data in the new tunnel. Finally, possible collapses in the new Yin-chao project are analyzed using the TBM responses that obtained from the transfer learning model. Results show that the types of transferred TBM parameters significantly affect the prediction accuracy. A filter procedure is proposed to determine the number and types of parameters to be transferred. The predicted accuracy for the torque and thrust reach 93 % and 95 %, respectively, for the first 100 steps from Ch. 66060 m to Ch. 65366 m in the new Yin-chao project when the most important 10 parameters are used in the transfer learning model. The predicted accuracy of the transfer learning model is better than that of the machine learning model in the early stage of the new tunnelling, especially in the first 2000 steps. The proposed model shows superior capability in prediction of the TBM responses and in collapse warning for a new tunnelling project, which paves the way for an automatic construction of TBM.