ABSTRACT
Ensuring the safety of employees and adjacent buildings during tunnel construction is vital, and considerable effort is undertaken to reduce the risk represented by the ground settlement. In the past decades, there have been many solutions for modeling the ground settlement risk using machine learning or deep learning algorithms. Despite their satisfactory performance, the existing research needs to pay more attention to the sequence characteristics of tunnel construction and being uninterpretable. To this end, we aim to address the following question: How can we accurately model the ground settlement risk using complex sequence features and improve its explainability? We proposed a hybrid two-stage data-driven approach to improve the accuracy and explainability of the ground settlement's prediction. Our proposed approach includes (1) linear trend prediction using seasonal autoregressive integrated moving average model (SARIMA), (2) nonlinear residuals prediction using the deep neural network (DNN), and (3) the posthoc explain technique using Shapley additive explanation (SHAP). Our proposed approach (ARIMA-DNN) is validated in a real-life tunnel construction in Wuhan, China. The prediction results with R2 of 0.895, RMSE of 1.085, MAE of 0.841, and MAPE of 9.07% show the superiority and applicability of the ARIMA-DNN method.
