Optimized ground settlement classification during TBM tunneling by combining machine learning with statistical analysis

Ground settlement management is crucial in tunnel boring machine (TBM) operations. Previous attempts to predict ground settlement have required substantial assumptions or information, complicating the explicit determination of their predictive criteria. This study developed an optimized system with simplicity and transparency for predicting ground settlements. By selecting three key features through correlation analysis and literature reviews, the optimized system was constructed to predict three settlement classes (heaving, normal, and large settlement) using a combination of machine learning and statistical analysis. The optimized system achieved an accuracy of 0.846, with recall values of 0.667 for heaving, 0.895 for normal, and 0.750 for large settlement. These results surpassed those of two comparison models that employed eight features and ensemble learning algorithms. Notably, the comparison models failed to correctly predict any instances of large settlement, highlighting the effectiveness of the optimized system in handling imbalanced datasets. Unlike conventional black-box models, the optimized system explicitly defined the predictive criteria. Moreover, among the four instances misclassified by the optimized system, three involved minor settlements within +-3 mm. The consistent decrease in accuracy when excluding each feature from the optimized system highlighted the importance of incorporating these features to accurately identify patterns in settlement predictions.