XGBoost 머신러닝 기반 쉴드 TBM 지반침하 예측

초록

This study developed an XGBoost (eXtreme Gradient Boosting) machine learning model to predict ground settlement during urban shield TBM (tunnel boring machine) tunnel construction and evaluated its performance. While previous studies have primarily focused on predicting ground settlement behind the tunnel, this study used real-time shield TBM construction data to predict both rear and forward settlement. For this purpose, field data from a TBM tunnel construction project in Hong Kong, provided by a local construction company, was analyzed. This data included information on ground conditions, TBM advancement, and tunnel geometry. A total of 17 input variables were selected for the machine learning model, which were grouped into three prediction ranges: the forward range (25 segment rings ahead of the tunnel face, CASE 1), the central section (upper part of the TBM body, CASE 2), and the rear range (25 segment rings behind the tunnel face, CASE 3). The relationships between the input variables and settlement were analyzed for each of these ranges. For each case (forward, central, and rear positions), an XGBoost model was developed to predict ground settlement, with hyperparameters optimized through Bayesian optimization and 5-fold cross-validation. The results showed that the rear settlement model performed the best, achieving a coefficient of determination (R2) of 0.82, while the forward settlement model had a lower performance, with an R2 of 0.52. These results suggest that rear settlement predictions are more accurate than forward predictions, with the latter being more affected by ground variability and excavation factors, resulting in lower accuracy. Overall, the findings highlight that machine learning models can be effective tools for predicting ground settlement during TBM tunnel construction, especially for rear settlement. However, further research is needed to enhance the accuracy of forward settlement predictions.

키워드