A multi-channel decoupled deep neural network for tunnel boring machine torque and thrust prediction

Accurate prediction of thrust and torque plays a crucial role in the control parameters optimization and intelligent tunneling of tunnel boring machines (TBMs). Currently, researchers seldom utilize the correlation and difference between thrust and torque to improve the performance of the model when building a data-driven thrust and torque prediction model. In this study, a novel multi-channel decoupled deep neural network (MDDNN) model is designed to predict the thrust and torque, which is capable of taking the correlation and difference between the thrust and torque into consideration. The inputs of the model are tunneling speed, cutterhead speed, and geological type, and the outputs of the model are thrust and torque. On-site data collected from a Singapore tunnel project was utilized to verify the effectiveness and superiority of the proposed decoupled modeling strategy and the built MD-DNN model. The results demonstrate that the MD-DNN model outperforms the support vector regression, k-nearest neighbors, random forest, AdaBoost, XGBoost, independent deep neural network (DNN), and conventional dual-output DNN models in both thrust and torque predictions. Compared with the independent DNN models, the proposed model can take the correlation between thrust and torque into consideration. Compared with the conventional dual-output DNN model, the proposed model can take the difference between the thrust and torque into consideration. Therefore, the established MD-DNN model is effective and superior, and it is of great significance for improving the intelligence level of TBM. Moreover, the proposed decoupled modeling strategy has a certain guiding value for other scenarios where multiple correlated outputs are predicted with the same input.

Automated summary

In this study, a novel multi-channel decoupled deep neural network (MDDNN) model is designed to accurately predict the thrust and torque of tunnel boring machines. The model takes the correlation and difference between thrust and torque into consideration, and utilizes tunneling speed, cutterhead speed, and geological type as input parameters. On-site data from a Singapore tunnel project is used to verify the effectiveness and superiority of the proposed model. The results show that the MD-DNN model outperforms other models in predicting both thrust and torque, and its consideration of correlation and difference makes it effective and superior.