Multi-level Machine Learning-Driven Tunnel Squeezing Prediction: Review and New Insights: Fid-Bau Portal

Multi-level Machine Learning-Driven Tunnel Squeezing Prediction: Review and New Insights

Fathipour-Azar, Hadi

Abstract Tunnel squeezing is a time-dependent phenomenon that often happens in weak or over-stressed rock masses, considerably impacting tunnel construction costs and timelines. This paper addresses the severity of potential squeezing problems based on the predicted squeezing levels. To answer this problem, a multi-level data mining decision making assessment methodology is developed based on four commonly available parameters at the early design stage, that is, diameter (D), buried depth (H), support stiffness (K), and rock tunneling quality index (Q) to predict squeezing conditions in rock tunnels. Six different types of machine learning classifier groups viz., decision tree (19 algorithms), rule (11 algorithms), miscellaneous (4 algorithms), function (3 algorithms), Bayesian (2 algorithms), and lazy (2 algorithms) classifiers are comparatively included in the proposed method to assist users in determining which method should be utilized to achieve the required degree of accuracy. The proposed models are trained and tested on a dataset collected from published literature comprising 117 tunnel squeezing inventories. The model is validated using tenfold cross-validation on the training set (containing 80% of data) and also using a 20% test dataset of case histories (24 cases) that had not been originally utilized to train the model. The results show that the proposed multi-level decision-making models can effectively be utilized to analyze tunnel squeezing problems and yield the promised results. The prediction accuracy of developed multi-level models can reach above 90%, which is a very encouraging result for squeezing risk mitigation and prevention. Moreover, based on the well-established confusion matrix, predictive capabilities of the taken tunnel squeezing models is calculated in term of precision, recall, and F-measure. These robust data-centric methods can be extended to various geoengineering classification challenges. By measuring the variable importance based on the best-performed classifiers, i.e., RandomForest, CSForest, and SimpleCart algorithms, in general, it can be concluded that K is the most sensitive predictor to squeezing severity development, followed by Q, H, and D parameters, each with a different degree of significance. However, different methods demonstrate different degrees of importance for these input variables. The results of this study provide beneficial insights for squeezing severity strategies and minimizing potential risks and costs via early-stage four supplied attributes. Moreover, this study leverages various state-of-the-art machine learning classification algorithms, which highlight their application as a tool of data mining technology in geoengineering.