Interpretable Hybrid Context-Sensitive Learning Framework for Tunnel Squeezing Classification

Abstract

The early-stage classification of tunnel squeezing is essential for the mitigation of geotechnical risks and the safety of the infrastructure. This study introduces a hybrid context-sensitive machine learning framework, where hybrid context refers to the combined use of numerical inputs (diameter, depth, rock quality index, and support stiffness) and categorical site-specific text description inputs (tunnel name, lithology, and location) that together define the engineering and geotechnical context of a tunnel. The three key elements proposed are (1) utilizing hybrid context data set of 274 real-world tunnel cases; (2) a two-tier class imbalance handling strategy combining six preprocessing techniques and algorithm-level class weighting; and (3) comprehensive empirical benchmarking and interpretability analysis across nine supervised algorithms. Trained on real-world tunnel cases, the model achieves 81.63% accuracy and up to 90% on benchmark validations under class-weighted hybrid context input. These results exceed the performance of prior state-of-the-art non-meta-ensembled models. Shapely additive explanation (SHAP) and permutation importance plots validate the feature relevance, enhancing interpretability and domain confidence. The approach demonstrates the Technology Readiness Level 5 from the perspective of deployment for practical geotechnical engineering.