A hybrid deep learning approach for dynamic attitude and position prediction in tunnel construction considering spatio-temporal patterns

This study proposes a hybrid deep learning approach for dynamic attitude and position prediction of the tunnel boring machine (TBM) with high accuracy. By utilizing the key operational parameters as well as the historical value of TBM's positions, the proposed deep learning model with graph convolutional network (GCN) and long short-term memory (LSTM), named GCN-LSTM, is constructed and trained to predict the vertical and horizontal deviations at the articulation and tail of TBM. Shapley Additive exPlanations (SHAP) analysis is then performed to improve the model's interpretability and determine the key contributing factors. Data obtained from a realistic tunnel project in Singapore's Thomson-East Coast line is utilized as a case study. The results indicate that: (1) The proposed GCN-LSTM approach provides accurate prediction with an average MAE of 1.009 mm, RMSE of 1.445 mm and R2 of 0.941. (2) The historical values of the deviation and adjustment are the major contributions to the current deviations, while the present adjustment could only influence the deviation in the future. (3) The pro-posed GCN-LSTM model outperforms the state-of-the-art methods in most metrics for the four outputs and thus is the most suitable method for the prediction. The proposed approach provides a reliable estimation of TBM's position which assists in improving the overall project quality and reduces the risk of tunnel misalignments.