Predicting floor heave risk in road tunnels with machine learning

Floor heave is a common defect in mountainous tunnels. It is critical but challenging to predict the risk of floor heave, as traditional methods often fail to characterize this phenomenon effectively. This study proposes a data-driven approach utilizing a support vector machine (SVM) optimized by the sparrow search algorithm (SSA) to address the issue. The model was developed and validated using a dataset collected from 100 tunnels. Shapley value analysis was conducted to identify the key features influencing floor heave defects. Moreover, a committee-based uncertainty quantification method is presented to evaluate the reliability of each prediction. The results show that: (1) Data feature engineering and SSA play pivotal roles in expediting the convergence of the SVM model. (2) Groundwater and high in situ stress are key factors contributing to tunnel floor heave. (3) In comparison to backpropagation (BP) neural networks, the SSA-SVM demonstrates superior robustness in handling imperfect and limited data. (4) The committee-based uncertainty quantification method is proven effective to evaluate the trustworthiness of each prediction. This data-driven surrogate model offers an effective strategy for understanding the factors that impact tunnel floor defects and accurately predicting tunnel floor heave deformation.