Serviceability evaluation of highway tunnels based on data mining and machine learning: A case study of continental United States

• Published in: Tunnelling and underground space technology Vol. 142; p. 105418
• Main Authors: Xue, Ya-Dong, Zhang, Wei, Wang, Yi-Lin, Luo, Wei, Jia, Fei, Li, Sheng-Teng, Pang, Hao-Jun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2023
• Subjects: Infrastructure Operation; SHAP; SMOTE; TOPSIS-CRITIC-Kmeans; TPE-LightGBM; Tunnel Service Performance Index (TSPI); TOPSIS-CRITIC-Kmeans; SHAP; SMOTE; Infrastructure Operation; Tunnel Service Performance Index (TSPI); TPE-LightGBM
• ISSN: 0886-7798; 1878-4364
• DOI: 10.1016/j.tust.2023.105418

•A comprehensive evaluation method for operational tunnels based on full data flow is proposed.•An unsupervised clustering-based service classification method for operational tunnels is proposed.•A TPE-LightGBM classifier that utilizes the information features of tunnels to predict their service performance is developed.•Factors that contribute to the degradation of operational tunnels at different stages are investigated. The appraisal of a tunnel service condition usually takes into account the lining deformations and defects of the lining structure, with the assessment technique heavily rely on the expertise of the evaluation. In this study, based on the US National Tunnel Inventory (NTI) database, a TOPSIS-CRITIC evaluation system was established to comprehensively evaluate the service performance of tunnels during the operation period from three categories of structural, traffic & civil, and non-structural components, resulting in the Tunnel Service Performance Index (TSPI). Then, a K-means clustering method was used to unevenly divide the TSPI of tunnels into service levels I ∼ IV. After consulting with experts, 23 important features were selected from 88 basic characteristics of tunnels, and further refined this list to 19 key factors using Spearman correlation analysis. These features were grouped into five categories: traffic, geological, space & time, planning & construction, and operational. Due to the imbalance of categories, SMOTE oversampling method was employed for data augmentation. Further, a TPE Bayesian optimized LightGBM classifier was proposed to predict the service level of tunnels and compared with nine common machine learning models (LR, KNN, SVM, DT, RF, GBDT, XGB, ET and MLP). After comprehensive experiments and numerical testing, the TPE-LGBM classifier achieved the highest overall performance among the ten models. Lastly, the SHapley Additive exPlanations (SHAP) interpretability model in conjunction with GIS spatial geographic information was employed to conduct an in-depth analysis of how full life cycle features impact tunnel service performance during operation.